CN118010848A - Intelligent anchorage device ponding detection method and system - Google Patents
Intelligent anchorage device ponding detection method and system Download PDFInfo
- Publication number
- CN118010848A CN118010848A CN202410413990.5A CN202410413990A CN118010848A CN 118010848 A CN118010848 A CN 118010848A CN 202410413990 A CN202410413990 A CN 202410413990A CN 118010848 A CN118010848 A CN 118010848A
- Authority
- CN
- China
- Prior art keywords
- anchor
- vector
- characterization
- ultrasonic signal
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 67
- 239000013598 vector Substances 0.000 claims abstract description 802
- 238000012512 characterization method Methods 0.000 claims abstract description 403
- 239000011159 matrix material Substances 0.000 claims abstract description 329
- 238000009826 distribution Methods 0.000 claims abstract description 263
- 238000000605 extraction Methods 0.000 claims abstract description 126
- 238000011156 evaluation Methods 0.000 claims abstract description 73
- 230000002787 reinforcement Effects 0.000 claims abstract description 27
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 128
- 238000000034 method Methods 0.000 claims description 109
- 238000009825 accumulation Methods 0.000 claims description 94
- 238000005728 strengthening Methods 0.000 claims description 85
- 238000002604 ultrasonography Methods 0.000 claims description 83
- 230000008569 process Effects 0.000 claims description 56
- 238000012545 processing Methods 0.000 claims description 35
- 238000005457 optimization Methods 0.000 claims description 20
- 238000001914 filtration Methods 0.000 claims description 19
- 230000004927 fusion Effects 0.000 claims description 14
- 230000009467 reduction Effects 0.000 claims description 14
- 239000000203 mixture Substances 0.000 claims description 10
- 238000005192 partition Methods 0.000 claims description 7
- 238000011282 treatment Methods 0.000 claims description 6
- 230000003044 adaptive effect Effects 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 5
- 238000004422 calculation algorithm Methods 0.000 description 28
- 238000004458 analytical method Methods 0.000 description 25
- 238000013527 convolutional neural network Methods 0.000 description 21
- 230000006870 function Effects 0.000 description 16
- 238000013528 artificial neural network Methods 0.000 description 15
- 238000000513 principal component analysis Methods 0.000 description 12
- 238000010801 machine learning Methods 0.000 description 11
- 239000000284 extract Substances 0.000 description 8
- 238000003062 neural network model Methods 0.000 description 8
- 230000009466 transformation Effects 0.000 description 8
- 239000000470 constituent Substances 0.000 description 7
- 238000012423 maintenance Methods 0.000 description 7
- 238000003860 storage Methods 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 6
- 238000013136 deep learning model Methods 0.000 description 6
- 238000011176 pooling Methods 0.000 description 6
- 238000012706 support-vector machine Methods 0.000 description 6
- 238000000844 transformation Methods 0.000 description 6
- 238000012880 independent component analysis Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 230000003416 augmentation Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000007667 floating Methods 0.000 description 3
- 238000011478 gradient descent method Methods 0.000 description 3
- 230000036541 health Effects 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 230000000306 recurrent effect Effects 0.000 description 3
- 238000012216 screening Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000003066 decision tree Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 239000003550 marker Substances 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 241000521257 Hydrops Species 0.000 description 1
- 206010030113 Oedema Diseases 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 238000012067 mathematical method Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 230000006798 recombination Effects 0.000 description 1
- 238000005215 recombination Methods 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000026676 system process Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000011179 visual inspection Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
The application provides an intelligent anchor ponding detection method and system, which are characterized in that after characterization vector extraction is carried out on each anchor ultrasonic signal matrix in an anchor ultrasonic signal distribution tensor comprising a bridge anchor, distribution information reinforcement is carried out on the basis of distribution conditions respectively corresponding to each anchor ultrasonic signal matrix on the basis of a distribution information reinforcement assembly after debugging, whether ponding exists is determined on the basis of a commonality evaluation coefficient between the reinforced characterization vector and a contrast characterization vector, and finally anchor ponding identification information of the anchor ultrasonic signal distribution tensor is generated. Because each contrast characterization vector is extracted from the ponding-free ultrasonic signal data of the same bridge anchor, and compared with the ponding-free characterization vector by the commonality evaluation coefficient, whether each anchor ultrasonic signal matrix has ponding can be accurately and efficiently determined. Meanwhile, because the sample without ponding is easy to obtain, the generalization of ponding detection is improved, and the application difficulty is reduced.
Description
Technical Field
The application relates to the technical fields of data processing and artificial intelligence, in particular to an intelligent anchorage device ponding detection method and system.
Background
With the continuous development of bridge construction and the extension of service time, the safety and stability of bridge anchors serving as key components in bridge structures are increasingly concerned. Water that may occur inside the anchor is one of the important factors affecting its performance, and therefore regular detection and maintenance of the anchor is of paramount importance. Conventional detection methods often rely on manual inspection and visual inspection, but these methods are not only inefficient, but also difficult to find some hidden water accumulation. In recent years, the ultrasonic detection technology is widely applied to bridge anchorage detection by virtue of the advantages of non-destructiveness, high precision, real-time performance and the like. The internal structure and state information of the anchor can be obtained through ultrasonic signals, and powerful support is provided for subsequent fault diagnosis and preventive maintenance. However, due to the complexity of the bridge anchor structure and the nature of the ultrasonic signal itself, extracting useful information directly from the original ultrasonic signal and performing accurate water logging identification remains a challenging task. Firstly, the ultrasonic signals generated by the anchors with different structures at different positions are difficult to converge when water is accumulated, and secondly, most of the anchors are provided with water accumulation preventing devices at present, so that the size of an anchor sample actually generating the water accumulation is not large, but more learning samples are needed for debugging a neural network model with strong generalization, and great application barriers are brought to intelligent detection.
Disclosure of Invention
Accordingly, the embodiment of the application at least provides an intelligent anchorage device ponding detection method and system. The technical scheme of the embodiment of the application is realized as follows:
in one aspect, an embodiment of the present application provides a method for detecting accumulated water of an intelligent anchor, which is applied to a computer system, and the method includes: acquiring an anchor ultrasonic signal distribution tensor aiming at a bridge anchor, and extracting characterization vectors of each anchor ultrasonic signal matrix contained in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector; based on the distribution information strengthening component after debugging, respectively carrying out distribution information strengthening operation on the corresponding matrix characterization vectors according to the distribution conditions of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor to obtain corresponding strengthening characterization vectors containing the distribution conditions; respectively determining water accumulation condition prediction coefficients corresponding to the anchor ultrasonic signal matrixes according to the commonality evaluation coefficients between the reinforced characterization vectors and the contrast characterization vectors deployed in advance; each contrast characterization vector is obtained according to the extraction of ponding-free ultrasonic signal data, and each contrast characterization vector comprises the distribution condition of a corresponding anchor ultrasonic signal matrix, wherein the ponding-free ultrasonic signal data corresponds to a ponding-free bridge anchor; and determining the anchor ponding identification information of the anchor ultrasonic signal distribution tensor according to the obtained ponding condition prediction coefficients. The method for determining the water accumulation condition prediction coefficients respectively corresponding to the anchor ultrasonic signal matrixes according to the commonality evaluation coefficients between each strengthening characterization vector and the contrast characterization vector deployed in advance comprises the following steps: for each strengthening characterization vector, the following processing is respectively carried out: for one enhancement token vector, determining a collation token vector having a greatest common assessment coefficient with the one enhancement token vector; determining a ponding condition prediction coefficient of an anchor ultrasonic signal matrix corresponding to the one enhanced characterization vector according to the error between the contrast characterization vector and the one enhanced characterization vector; or determining the water accumulation condition prediction coefficients corresponding to the anchor ultrasonic signal matrixes respectively according to the commonality evaluation coefficients between the reinforced characterization vectors and the contrast characterization vectors deployed in advance, wherein the water accumulation condition prediction coefficients comprise: for each strengthening characterization vector, the following processing is respectively carried out: for one enhancement token vector, determining a collation token vector having a greatest common assessment coefficient with the one enhancement token vector; determining a plurality of contrast token vectors with the commonality evaluation coefficients not larger than a preset commonality evaluation coefficient threshold value between the contrast token vectors with the largest commonality evaluation coefficients in the contrast token vectors deployed in advance; and determining the ponding condition prediction coefficient of the anchor ultrasonic signal matrix corresponding to the one strengthening characterization vector according to the comparison characterization vector with the maximum commonality evaluation coefficient and errors between the plurality of comparison characterization vectors and the one strengthening characterization vector.
In some embodiments, before the debugging-completed distribution information strengthening component performs a distribution information strengthening operation on the corresponding matrix characterization vectors according to the distribution condition of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor, so as to obtain corresponding strengthening characterization vectors including the distribution condition, the method further includes: repeatedly debugging the distributed information enhancement component to be debugged according to a plurality of anchor ultrasonic signal distribution tensor samples and the prior marks of the samples corresponding to the anchor ultrasonic signal distribution tensor samples respectively until the debugging cut-off requirement is met, and obtaining the distributed information enhancement component after the debugging is completed; wherein, each sample prior mark represents the confidence coefficient of each anchor ultrasonic signal matrix in the corresponding anchor ultrasonic signal distribution tensor sample; wherein, in each round of debugging, the following operations are included: for one anchor ultrasonic signal distribution tensor sample, extracting characterization vectors of each anchor ultrasonic signal matrix of the anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors; the distribution information strengthening component is used for carrying out distribution information strengthening operation on the corresponding sample characterization vectors according to the distribution condition of each anchor ultrasonic signal matrix in one anchor ultrasonic signal distribution tensor sample so as to obtain corresponding strengthening characterization vectors containing the distribution condition; according to the corresponding strengthening characterization vectors of each anchor ultrasonic signal matrix of the anchor ultrasonic signal distribution tensor sample, anchor ponding identification information of the anchor ultrasonic signal distribution tensor sample is obtained; and optimizing the parameter of the distributed information strengthening component for the round of debugging according to the error between the obtained accumulated water identification information of the plurality of anchors and the corresponding sample prior marks.
In some embodiments, before determining the water accumulation condition prediction coefficients respectively corresponding to the anchor ultrasonic signal matrixes respectively according to the commonality evaluation coefficients between the reinforced characterization vectors and the contrast characterization vectors deployed in advance, the method further comprises: based on a second characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix contained in each ponding-free ultrasonic signal data respectively to obtain a basic characterization matrix, wherein the basic characterization matrix comprises basic characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data; annotating a plurality of anchor ultrasonic signal distribution tensor samples respectively according to the basic representation matrix to obtain corresponding sample prior marks, and repeatedly debugging a ponding detection network according to the anchor ultrasonic signal distribution tensor samples and the corresponding sample prior marks respectively until the debugging cut-off requirement is met; the ponding detection network comprises a first characterization vector extraction component and the distribution information strengthening component, wherein the first characterization vector extraction component and the second characterization vector extraction component are twin components, and the same parameter is used; according to the first characterization vector extraction component in the ponding detection network after debugging, respectively extracting characterization vectors of each ponding-free ultrasonic signal data, and carrying out distribution information strengthening operation on matrix characterization vectors obtained by extraction on the basis of the distribution information strengthening component in the ponding detection network after debugging, so as to obtain a comparison characterization vector set, wherein the comparison characterization vector set comprises the previously deployed comparison characterization vectors.
In some embodiments, annotating the plurality of anchor ultrasonic signal distribution tensor samples according to the basic characterization matrix to obtain corresponding sample prior marks includes: for the plurality of anchor ultrasonic signal distribution tensor samples, respectively carrying out the following processing: for one anchor ultrasonic signal distribution tensor sample, based on the second characterization vector extraction component, respectively extracting characterization vectors of each anchor ultrasonic signal matrix in the one anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors; determining a composition priori mark of the corresponding anchor ultrasonic signal matrix according to the obtained commonality evaluation coefficient between each sample characterization vector and the basic characterization matrix, wherein each composition priori mark characterizes the confidence degree of the corresponding anchor ultrasonic signal matrix as a ponding anchor ultrasonic signal matrix; obtaining a sample prior mark of the anchor ultrasonic signal distribution tensor sample according to the obtained each composition prior mark; the first characterization vector extraction component in the ponding detection network according to the completion of the debugging performs characterization vector extraction on each ponding-free ultrasonic signal data, and performs distribution information reinforcement operation on matrix characterization vectors obtained by extraction based on the distribution information reinforcement component in the ponding detection network after the completion of the debugging, so as to obtain a comparison characterization vector set, and the method comprises the following steps: according to the first characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data respectively, and performing distribution information reinforcement operation on the extracted matrix characterization vector based on the distribution information reinforcement component in the ponding detection network after debugging is completed to obtain an iteration characterization vector set, wherein the iteration characterization vector set comprises iteration characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data; and determining partial iterative characterization vectors in the iterative characterization vector set based on the commonality evaluation coefficients between every two iterative characterization vectors in the iterative characterization vector set, so as to obtain the comparison characterization vector set.
In some embodiments, the determining a portion of the iterative token vectors in the iterative token vector set based on the commonality evaluation coefficients between every two iterative token vectors in the iterative token vector set, to obtain the collation token vector set, includes: determining one or more iterative characterization vectors from the iterative characterization vector set, and constructing a basic set for the contrast characterization vector set; iteratively optimizing the set of control token vectors; wherein, in one round of optimization, the following processes are performed: for each iteration characterization vector in the iteration characterization vector set, respectively determining a corresponding characterization vector doublet, wherein each characterization vector doublet comprises one iteration characterization vector and a comparison characterization vector with the largest common evaluation coefficient with the one iteration characterization vector in the comparison characterization vector set; for each contrast token vector contained in each token vector doublet, adding the iterative token vector with the smallest commonality evaluation coefficient with each contrast token vector to the contrast token vector set.
In some embodiments, based on the distribution information strengthening component after the debugging is completed, according to the distribution condition of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor, respectively performing a distribution information strengthening operation on the corresponding matrix characterization vector to obtain a corresponding strengthening characterization vector including the distribution condition, including: based on the distributed information strengthening assembly after debugging, the following treatments are respectively carried out on each anchor ultrasonic signal matrix: for an anchor ultrasonic signal matrix, obtaining a corresponding sequence number representation vector according to the sequence number distribution condition of the anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor; and obtaining a corresponding strengthening characterization vector according to the matrix characterization vector and the serial number characterization vector of the anchor ultrasonic signal matrix.
In some embodiments, extracting the characterization vector of each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector includes: carrying out multi-round characterization vector extraction on the anchor ultrasonic signal distribution tensor; wherein each round performs the following processes: according to a preset filtering matrix, filtering characterization vector extraction is carried out on the input of the current wheel characterization vector extraction, and a filtering characterization vector is obtained; wherein, when the first round of characterization vector extraction is performed, the input is the distribution tensor of the anchor ultrasonic signal, and when the other rounds of characterization vector extraction is performed, the input is the output of the previous round of characterization vector extraction; performing cross-layer identity connection according to the filtering characterization vector and the input to obtain a fusion characterization vector; if the current wheel representation vector extraction is the last wheel representation vector extraction, taking the fusion representation vector as the output of the current wheel representation vector extraction; and if the current wheel representation vector extraction is not the last wheel representation vector extraction, performing feature adaptive dimension reduction on the fusion representation vector, and taking the obtained dimension reduction representation vector as the output of the current wheel representation vector extraction.
In some embodiments, extracting the characterization vector of each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector includes: obtaining corresponding target reference bridge anchor data from a plurality of reference bridge anchor data which are deployed in advance according to the bridge tag of the bridge anchor; according to preset anchor distribution information in the target reference bridge anchor data, carrying out data alignment on the anchor ultrasonic signal distribution tensor to obtain an alignment matrix; according to a preset matrix area in the target reference bridge anchor data, obtaining a matrix area partition where the bridge anchor is located in the alignment matrix; and extracting characterization vectors of each anchor ultrasonic signal matrix contained in the matrix region blocks to obtain corresponding matrix characterization vectors.
In another aspect, the application provides a computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, the processor implementing the steps of the method described above when the program is executed.
The beneficial effects of the application at least comprise: after the characterization vector extraction is carried out on each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor comprising the bridge anchor, the distribution information reinforcement component is based on the distribution conditions respectively corresponding to each anchor ultrasonic signal matrix after debugging, the distribution information reinforcement is carried out on the corresponding matrix characterization vector to obtain the corresponding reinforced characterization vector, further, whether each anchor ultrasonic signal matrix has water accumulation conditions or not is determined based on the commonality evaluation coefficient between the reinforced characterization vector and the contrast characterization vector, and finally anchor water accumulation identification information of the anchor ultrasonic signal distribution tensor is generated. Because each contrast characterization vector is extracted from the ponding-free ultrasonic signal data of the same bridge anchor, belongs to the ponding-free characterization vector, and based on the comparison of the commonality evaluation coefficients with the ponding-free characterization vector, whether the anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor has ponding can be accurately and efficiently determined. Meanwhile, because the samples without water accumulation are easy to obtain, a sufficient amount of samples without water accumulation can be obtained to extract contrast characterization vectors, thereby helping to increase generalization of water accumulation detection and reducing application difficulty.
Further, after matrix characterization vectors of the anchor ultrasonic signal matrix are extracted, distribution information reinforcement is carried out according to the distribution condition of the anchor ultrasonic signal matrix to obtain reinforced characterization vectors, each comparison characterization vector is a characterization vector which contains distribution information after distribution information reinforcement operation is completed, when the comparison is carried out with the comparison characterization vector, the specific anchor ultrasonic signal matrix is compared according to the distribution condition, the data processing range is reduced, and the identification precision and speed are improved.
Drawings
Fig. 1 is a schematic implementation flow chart of an intelligent anchorage device ponding detection method provided by an embodiment of the application.
Fig. 2 is a schematic diagram of a hardware entity of a computer system according to an embodiment of the present application.
Detailed Description
The technical solution of the present application will be further elaborated with reference to the accompanying drawings and examples, which should not be construed as limiting the application, but all other embodiments which can be obtained by one skilled in the art without making inventive efforts are within the scope of protection of the present application. The term "first/second/third" is merely to distinguish similar objects and does not represent a particular ordering of objects, it being understood that the "first/second/third" may be interchanged with a particular order or precedence, as allowed, to enable embodiments of the application described herein to be implemented in other than those illustrated or described herein.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing the application only and is not intended to be limiting of the application.
The embodiment of the application provides an intelligent anchorage device ponding detection method which can be executed by a processor of a computer system. The computer system may refer to a server, a notebook computer, a tablet computer, a desktop computer, or other devices with data processing capability.
Fig. 1 is a schematic implementation flow chart of an intelligent anchorage device ponding detection method provided by an embodiment of the application, as shown in fig. 1, the method includes:
Step S10: and obtaining an anchor ultrasonic signal distribution tensor aiming at the bridge anchor, and extracting characterization vectors of each anchor ultrasonic signal matrix contained in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector.
In bridge maintenance, the anchor is a critical component, and whether it is intact is directly related to the safety of the bridge. In order to detect whether water is present inside the anchor, an ultrasonic detection technique may be employed. The number of anchors on a bridge is often thousands, and through detection and analysis one by one, the detection efficiency is obviously not realistic, so the application proposes the concept of anchor ultrasonic signal distribution tensor. That is, the computer system is responsible for acquiring these ultrasound signal data in this step and storing and processing it in a specific data structure, tensor. In particular, the anchor ultrasound signal distribution tensor is formed by stacking a plurality of matrixes. Each matrix corresponds to a particular anchor on the bridge. This means that if there are N anchors on the bridge, the tensor will consist of N such matrices. Each matrix contains data of multiple signal acquisitions from different points in time or times. The data collected each time is stored in the matrix in the form of a row, and each row contains a plurality of floating point numbers. These floating point numbers represent various characteristic values of the ultrasound signal, such as amplitude, frequency, time domain data, skewness, energy, etc. These feature values are key information of the ultrasound signal and together describe the shape, intensity and variation characteristics of the signal.
Taking amplitude data as an example, it may represent the variation of the amplitude of the ultrasound signal during propagation. The frequency data may then reflect the content of the different frequency components in the signal. Time domain data generally describes the time-varying condition of a signal, while skewness and energy, etc., may reveal statistical properties and energy distribution of the signal.
For example, assume that there are s anchors on a bridge, and d ultrasonic signal acquisitions are performed. The anchor ultrasound signal distribution tensor may then be represented as a three-dimensional array of s x d x M, where s represents the number of anchors, n represents the number of acquisitions, and M represents the number of ultrasound signal features per acquisition.
In practice, the anchor ultrasonic signal distribution tensor is formed by stacking s d×m matrices, each matrix corresponding to one anchor. Here, M may be any positive integer depending on the accuracy and setting of the ultrasound signal acquisition device.
Taking m=5 as an example, assume that 5 features of amplitude, frequency, time domain data, skewness, and energy are collected. One possible anchor ultrasound signal distribution tensor is then as follows (only values are given here for simplicity, these values will in fact be floating point numbers, and the range will depend on the actual situation):
Tensor T:
[
ultrasonic signal matrix of [ (# anchor 1)
[1.0, 2.0, 3.0, 4.0, 5.0], # First acquired eigenvalue
[1.1, 2.1, 3.1, 4.1, 5.1] # Feature values acquired a second time
],
Ultrasonic signal matrix of [ (# anchor 2)
[1.2, 2.2, 3.2, 4.2, 5.2], # First acquired eigenvalue
[1.3, 2.3, 3.3, 4.3, 5.3] # Feature values acquired a second time
],
Ultrasonic signal matrix of [ (# anchor 3)
[1.4, 2.4, 3.4, 4.4, 5.4], # First acquired eigenvalue
[1.5, 2.5, 3.5, 4.5, 5.5] # Feature values acquired a second time
]
]
In this example, the tensor T contains ultrasound signal data for 3 anchors. The data for each anchor is a 2x5 matrix, where 2 represents the number of acquisitions and 5 represents the number of features. For example, the amplitude eigenvalue of the first acquisition of anchor 1 is 1.0, the frequency eigenvalue is 2.0, and so on.
From the tensor example given above, a particular anchor ultrasound signal matrix can be extracted. Taking the anchor 2 as an example, the corresponding ultrasonic signal matrix is as follows:
Matrix m_anchor 2:
[
[1.2, 2.2, 3.2, 4.2, 5.2], # first acquired eigenvalue
[1.3, 2.3, 3.3, 4.3, 5.3] # Feature values acquired a second time
]
This matrix is a2 x 5 two-dimensional array in which each row represents an acquired ultrasound signal characteristic value. In this example it can be seen that the characteristic values of the anchor 2 in both acquisitions differ, which may be due to changes in environmental conditions, equipment errors or changes in internal states of the anchor. These eigenvalues will be used for subsequent prediction and identification of the water accumulation situation.
In the anchor ultrasound signal distribution tensor, the arrangement order of each matrix corresponds to the spatial distribution order of each matrix on the bridge. This means that by looking at the structure of the anchor ultrasound signal distribution tensor it is possible to intuitively know which matrix corresponds to which anchor, and the relative position of these anchors on the bridge. Once the anchor acoustic signal distribution tensor is obtained, the computer system further processes it to extract a characterization vector from the ultrasonic signal matrix for each anchor. The characterization vector is obtained by a feature extraction technology, contains key information in an original matrix, and is simpler in form and easier to process. Feature extraction may be implemented using various algorithms or models, such as Principal Component Analysis (PCA), self-encoder neural networks, convolutional neural networks, and the like. These algorithms or models can extract the most representative features from the raw data, thereby forming a low-dimensional but informative token vector.
For example, if feature extraction is performed using principal component analysis, the computer system first calculates a covariance matrix for each matrix, and then finds feature values and feature vectors for this covariance matrix. The feature vectors corresponding to the largest feature values form the characterization vector. This token vector contains the most important component of the original matrix that is the most important direction of change and information.
Through the processing of step S10, a set of token vectors containing rich information is obtained. These characterization vectors not only reflect the ultrasound signal characteristics of the individual anchors, but also imply the spatial relationship between them. This provides powerful data support for subsequent prediction and identification of water logging conditions.
Step S20: and based on the distribution information strengthening component after debugging, respectively carrying out distribution information strengthening operation on the corresponding matrix characterization vectors according to the distribution condition of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor to obtain corresponding strengthening characterization vectors containing the distribution condition.
In step S20, the computer system performs an enhancement operation on the matrix characterization vector extracted from the anchor ultrasonic signal distribution tensor by using the pre-debugged distribution information enhancement component (or position enhancement component). The purpose of this strengthening operation is to integrate the spatial distribution information of each anchor on the bridge into the characterization vector, thereby enhancing the expressive power of the data and the accuracy of subsequent analysis.
Specifically, the distributed information strengthening component is a neural network layer, and can purposefully strengthen corresponding matrix characterization vectors according to the distribution condition of an anchor ultrasonic signal matrix in tensors through debugging and optimization. The distribution condition refers to the position information of each anchor ultrasonic signal matrix in tensors, and the position information reflects the actual distribution sequence and spatial relationship of the anchors on the bridge.
By way of example, assume that there is a tensor comprising three matrices of anchor ultrasound signals, each matrix corresponding to a particular anchor on the bridge. The three matrices are arranged in left to right order of the anchors on the bridge. After the characterization vector of each matrix is extracted from the tensor, the distribution information strengthening component strengthens the corresponding anchors on the bridge according to the positions (i.e. left, middle and right) of the anchors.
The augmentation operation may include adjusting the weights of the token vectors, adding position-coding information, or capturing and augmenting the spatial distribution features with specific neural network structures (e.g., convolutional neural network CNN, recurrent neural network RNN, etc.). For example, the location information for each anchor may be encoded into a vector using a location embedding (Positional Encoding) technique, and this vector is then added to the corresponding matrix characterization vector to obtain an enhanced characterization vector that contains the location information.
Through the processing of the step S20, a group of reinforced characterization vectors are obtained, wherein the vectors not only contain the characteristic information of the original ultrasonic signals, but also integrate the spatial distribution information of the anchors on the bridge. The enhanced characterization vector can provide richer and more accurate data support for subsequent water accumulation condition prediction.
Step S30: respectively determining water accumulation condition prediction coefficients corresponding to each anchor ultrasonic signal matrix according to the commonality evaluation coefficients between each enhanced characterization vector and a contrast characterization vector deployed in advance; each contrast characterization vector is obtained according to the extraction of ponding-free ultrasonic signal data, and each contrast characterization vector comprises the distribution condition of a corresponding anchor ultrasonic signal matrix, wherein the ponding-free ultrasonic signal data corresponds to a ponding-free bridge anchor.
In step S30, the computer system compares the enhanced characterization vector obtained in the previous step with the contrast characterization vector deployed in advance to evaluate the commonality between them, and determines the ponding condition prediction coefficient corresponding to each anchor ultrasonic signal matrix according to the commonality.
Specifically, the control characterization vector is obtained based on extraction of ultrasound signal data acquired without water accumulation. These data correspond to an anchor without water accumulation on the bridge and thus can be used as a reference standard for determining whether the anchor is water accumulation. Similar to the enhanced characterization vector, the contrast characterization vector also contains distribution information of the corresponding anchor ultrasonic signal matrix to ensure the accuracy and effectiveness of the comparison.
In performing the commonality assessment, the computer system may use a similarity measure (e.g., cosine similarity, euclidean distance, etc.) to calculate the similarity between the enhanced token vector and the reference token vector. This similarity value reflects the proximity of the two vectors in the feature space, i.e. the similarity of the anchor ultrasound signals they represent in the absence of water and in the possible presence of water.
And according to the calculated similarity value, the computer system further determines a water accumulation condition prediction coefficient corresponding to each anchor ultrasonic signal matrix. This coefficient may be a scalar value representing the magnitude of the likelihood of water accumulation for the anchor; or a vector containing information in multiple dimensions for more fully describing the water accumulation of the anchor. By way of example, assume that there are three strengthening characterization vectors A, B and C, corresponding to three anchors on a bridge, respectively. Meanwhile, three contrast characterization vectors X, Y and Z which are deployed in advance represent the ultrasonic signal characteristics of the three anchors under the condition of no ponding respectively. The computer system calculates the similarity between A and X, B and Y, C and Z, resulting in three similarity values. And then, according to the magnitude of the similarity values, determining the ponding condition prediction coefficient corresponding to each anchor ultrasonic signal matrix. For example, if the similarity between a and X is high, then it is considered that the corresponding anchor is likely to have no water accumulation, and therefore its water accumulation condition prediction coefficient may be low; conversely, if the similarity is low, the prediction coefficient may be high.
And (3) obtaining a group of water accumulation condition prediction coefficients corresponding to each anchor ultrasonic signal matrix through the processing of the step S30. The coefficients can be used for subsequent water accumulation condition judgment and analysis, and an important basis is provided for bridge maintenance.
Step S40: and determining anchor ponding identification information of the anchor ultrasonic signal distribution tensor according to the obtained prediction coefficients of each ponding condition.
In step S40, the computer system comprehensively analyzes and determines the anchor water accumulation identification information corresponding to the anchor ultrasonic signal distribution tensor according to the water accumulation condition prediction coefficients obtained in the previous step. Specifically, the water accumulation condition prediction coefficient is obtained by comparing the strengthening characterization vector with the comparison characterization vector in step S30, and reflects the potential of the water accumulation of the anchorage represented by each anchorage ultrasonic signal matrix. These coefficients may be scalar values representing the probability or extent of water accumulation; or may be a vector containing more dimensional information for more fully describing the water accumulation situation.
In step S40, the computer system takes these water accumulation prediction coefficients as input, and performs a comprehensive process by using a certain algorithm or model. The algorithm or model may be a simple threshold decision, or may be a complex machine learning model such as a Support Vector Machine (SVM), decision tree, neural network, etc. The specific choice of which algorithm or model depends on the requirements of the actual application and the nature of the data.
Taking a simple threshold judgment as an example, assuming that a water accumulation probability threshold is set, if the water accumulation condition prediction coefficient of a certain anchor exceeds the threshold, considering that the anchor has water accumulation; otherwise, no water accumulation is considered. The computer system traverses all the ponding condition prediction coefficients, judges the ponding state of each anchorage device according to the threshold value, and generates corresponding anchorage device ponding identification information. If a more complex machine learning model, such as a neural network, is used, the computer system may use the trained neural network model to process the water accumulation condition prediction coefficients. The neural network model learns and predicts the ponding state of each anchor according to the characteristics and modes of the coefficients. The method has the advantages of being capable of processing more complex nonlinear relations and improving the accuracy of accumulated water identification.
Whichever method is employed, the final output of step S40 is the anchor water identification information of the anchor ultrasonic signal distribution tensor. This information may be a simple classification result (water accumulation/no water accumulation) or a more detailed description such as the degree of water accumulation, location, etc. The information has important guiding significance for maintenance and management of the bridge, can help workers to find and treat the water accumulation problem of the anchor in time, and ensures the safety and stability of the bridge.
In some embodiments, in step S20, based on the distribution information strengthening component after the debugging is completed, according to the distribution condition of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor, respectively performing a distribution information strengthening operation on the corresponding matrix characterization vector, and before obtaining the corresponding strengthening characterization vector including the distribution condition, the method may further include a debugging process of the distribution information strengthening component, and specifically may include:
Step S110: repeatedly debugging the distributed information enhancement assembly to be debugged according to the plurality of anchor ultrasonic signal distribution tensor samples and the sample prior marks corresponding to the anchor ultrasonic signal distribution tensor samples respectively until the debugging cut-off requirement is met, and obtaining the distributed information enhancement assembly after debugging is completed; the priori marks of each sample represent the confidence that each anchor ultrasonic signal matrix in the corresponding anchor ultrasonic signal distribution tensor sample is the ponding anchor ultrasonic signal matrix.
In step S110, the computer system uses a plurality of known anchor ultrasonic signal distribution tensor samples and corresponding sample prior marks to repeatedly debug the distributed information enhancement component to be debugged until meeting the debugging stop requirement, thereby obtaining the debugged distributed information enhancement component. The process is a supervised learning process in machine learning, and aims to enable a distribution information strengthening component to accurately strengthen corresponding matrix characterization vectors according to the distribution condition of an anchor ultrasonic signal matrix in tensors.
Specifically, each anchor ultrasound signal distribution tensor sample is constructed based on actual acquired ultrasound signal data, and comprises a plurality of anchor ultrasound signal matrices, each matrix corresponding to a particular anchor. The sample prior mark is the confidence degree allocated to each anchor ultrasonic signal matrix according to the known anchor ponding condition, and the confidence degree represents the possibility that the corresponding anchor of the matrix is a ponding anchor.
Wherein, in each round of debugging, the following operations are included:
step S111: for one anchor ultrasonic signal distribution tensor sample (i.e. a sample), extracting characterization vectors of each anchor ultrasonic signal matrix of one anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors;
Step S112: the distribution information strengthening component is used for carrying out distribution information strengthening operation on the corresponding sample characterization vectors according to the distribution condition of each anchor ultrasonic signal matrix in one anchor ultrasonic signal distribution tensor sample, so as to obtain corresponding strengthening characterization vectors containing the distribution condition;
step S113: according to the reinforced characterization vectors respectively corresponding to the anchor ultrasonic signal matrixes of one anchor ultrasonic signal distribution tensor sample, anchor ponding identification information of one anchor ultrasonic signal distribution tensor sample is obtained;
step S114: and optimizing the parameter of the distributed information strengthening component for the round of debugging according to the error between the obtained accumulated water identification information of the plurality of anchors and the corresponding sample prior marks.
In step S111, the computer system performs a token vector extraction on each anchor ultrasonic signal matrix in each anchor ultrasonic signal distribution tensor sample, so as to obtain a corresponding sample token vector. This token vector extraction process may be implemented based on a specific feature extraction algorithm or neural network model in order to convert the raw ultrasound signal data into a vector form that can be processed by the computer.
In particular, token vector extraction is the process of converting raw ultrasound signal data into a vector form that can be processed and understood by a computer. In the process, the computer system utilizes a specific algorithm or model to carry out deep analysis on the anchor ultrasonic signal matrix, and extracts key information capable of reflecting the essential characteristics of the anchor ultrasonic signal matrix. Such information may include physical characteristics of the signal such as frequency, amplitude, phase, etc., or may be a higher level of characteristics obtained by complex mathematical transformations. In practical operation, the extraction of the characterization vector may be implemented by various methods, such as classical dimension reduction algorithms, such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA), or deep learning models, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). These methods are all capable of extracting useful feature information from the raw data and encoding it into a vector form.
Taking convolutional neural network as an example, the computer system can take an anchor ultrasonic signal matrix as input, and gradually extract spatial features and abstract features in the matrix through the processing of a plurality of convolutional layers, a pooling layer and a full-connection layer. Finally, after the fully connected layers, the computer system obtains a fixed length vector, which is a characterization vector characterizing the anchor ultrasonic signal matrix.
Next, in step S112, the computer system uses the distribution information enhancement component of the current wheel debug to perform a distribution information enhancement operation on the corresponding sample characterization vector according to the distribution condition of each anchor ultrasonic signal matrix in the sample. This augmentation operation may be to adjust the weights of the vectors, add position-coding information, or capture and augment spatially distributed features through a specific neural network structure. The enhanced characterization vector contains more distribution information, which is helpful for subsequent ponding identification.
In particular, the distributed information enhancement component may be a complex machine learning model, such as a neural network, whose internal structure and parameters are continually optimized during the debugging process to better capture and enhance the distributed characteristics of the anchor ultrasound signal matrix. The model receives as input the sample characterization vector and performs a series of transformations and processes on the vector based on the matrix distribution in the tensor.
For example, if the distributed information enhancement component is a neural network model, it may comprise a plurality of hidden layers, each hidden layer being composed of a plurality of neurons. These neurons perform nonlinear transformations on the input data through specific activation functions and connection weights to capture higher level features. In the strengthening operation process, the neural network can adjust the connection weight of each neuron and the parameters of the activation function according to the distribution condition of the anchor ultrasonic signal matrix, so that the output vector can more accurately reflect the distribution characteristics of the matrix.
Specific implementations of the reinforcement operations may include operations such as weighting, stitching, pooling of vectors, or more complex transfer and transformation of information between neural network layers. These operations aim to incorporate the distribution information of the anchor ultrasound signal matrix into the characterization vector, enabling it to more fully describe the characteristics of the matrix. After the reinforcement operation, the computer system obtains a series of reinforcement characterization vectors including distribution conditions. The vectors not only contain the information of the original characterization vectors, but also integrate the distribution characteristics of the matrix in tensors, and provide richer data support for subsequent ponding identification.
In step S113, the computer system obtains anchor water identification information of an anchor ultrasonic signal distribution tensor sample according to the reinforced characterization vector. This identification information may be generated based on classification algorithms (e.g., support vector machine, decision tree, etc.) or neural network models for predicting the water accumulation state of each anchor.
Specifically, the strengthening characterization vector is obtained after the strengthening operation of the distribution information in step S112, and includes the distribution condition of each anchor ultrasonic signal matrix in the distribution tensor and the corresponding characteristic information thereof. These enhanced characterization vectors are points in high dimensional space, each representing a particular anchor ultrasound signal matrix and containing the position and features of the matrix in space. In step S113, the computer system processes and analyzes the enhanced token vectors using a suitable algorithm or model, such as a Support Vector Machine (SVM), random Forest (Random Forest), deep learning model, and the like. For example, if a deep learning model is used, a neural network may be designed that includes a plurality of hidden layers that receive the enhanced token vector as input and generate corresponding anchor water identification information by learning the relationship between the vectors and their mapping to anchor water conditions. The anchor water accumulation identification information can be a label (such as 'water accumulation' or 'no water accumulation') or a probability value (which indicates the possibility of water accumulation). The information is the comprehensive judgment result of each matrix in the anchor ultrasonic signal distribution tensor sample, and is obtained based on the spatial distribution and characteristic information of the whole tensor.
By way of example, assume that an anchor ultrasound signal distribution tensor sample contains the enhanced characterization vectors of three anchor ultrasound signal matrices A, B and C. By using a trained neural network model, the computer system may take as input the three enhanced token vectors and output a label or probability value representing the sample water identification result. This result integrates the position information, interrelationships, and the respective characteristic information of matrices A, B and C in the tensor.
Finally, in step S114, the computer system calculates the error between the anchor water accumulation identification information obtained in the previous step and the corresponding sample prior mark according to the comparison. Based on this error, the computer then adjusts and optimizes the parameters of the distributed information enhancement component. Specifically, the prior marking of the samples is the result of marking the anchor ultrasonic signal distribution tensor samples in advance, and represents the actual ponding condition of the samples. The anchor ponding identification information is a result obtained after being processed by the distribution information strengthening component, and represents the judgment of the computer on the ponding conditions of the samples. The error is the difference between the two. In calculating the error, various metrics may be used, such as mean square error, cross entropy loss, etc., with the specific choice of which depends on the nature and requirements of the problem. For example, in classification problems, cross entropy loss is often used to measure the difference between the probability distribution of model predictions and the true probability distribution.
Once the error is calculated, the computer uses an optimization algorithm to adjust the parameters of the distributed information enhancement component to minimize the error. The optimization algorithm may be a gradient descent method, a random gradient descent method, adam, or the like. These algorithms update the values of the parameters based on the gradient information of the error so that the model can make more accurate predictions at the next iteration.
Taking the gradient descent method as an example, the computer calculates the gradient of the error to the parameter, i.e. the direction and rate of change of the error with the parameter. The computer then updates the values of the parameters along the opposite direction of the gradient to reduce the error. This process is iterated until the error converges to a small value or a preset number of iterations is reached. Through the optimization process of step S114, the parameters of the distribution information strengthening component are gradually adjusted to an optimal state, so that the model can better capture the distribution characteristics of the anchor ultrasonic signal matrix and accurately perform water accumulation identification.
As an implementation manner, in step S30, before determining, respectively, the water accumulation condition prediction coefficients corresponding to each anchor ultrasonic signal matrix according to the commonality evaluation coefficients between each enhanced characterization vector and the contrast characterization vector deployed in advance, the method further includes a process for generating the contrast characterization vector set, which specifically may include:
Step S1: and based on the second characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix contained in each ponding-free ultrasonic signal data respectively to obtain a basic characterization matrix, wherein the basic characterization matrix comprises basic characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data.
In step S1, the computer system uses the second token vector extraction component to extract token vectors from each anchor ultrasonic signal matrix included in the ponding-free ultrasonic signal data. This process converts the raw ultrasound signal data into a vector form that can be understood and processed by the computer. The second token vector extraction component may be a model built based on a deep learning or other machine learning algorithm that is trained to extract key feature information from the anchor ultrasound signal matrix and encode this information into a vector form. These vectors not only contain the main features of the original signal, but can also be more efficiently utilized in subsequent calculations and analyses.
For example, if the second token vector extraction component is a Convolutional Neural Network (CNN), it performs a series of convolution, pooling, and activation function operations on the input anchor ultrasound signal matrix, ultimately outputting a fixed length vector, i.e., a base token vector. This vector is a highly abstract and compressed representation of the original matrix, which contains the principal feature information in the matrix.
By doing so for each anchor ultrasonic signal matrix in the water-free ultrasonic signal data, the computer system will obtain a base characterization matrix. The matrix contains basic characterization vectors of all ponding-free ultrasonic signal data, and is a basis for carrying out ponding condition prediction and ponding detection network debugging in the subsequent steps.
It should be noted that the second token vector extraction component mentioned in step S1 and the first token vector extraction component in the subsequent step are twinning components, which use the same parameters. This means that the two components are identical in structure and function, but differ in the data processed. The design can ensure that the data without water accumulation and with water accumulation are processed consistently and fairly in the follow-up water accumulation condition prediction and water accumulation detection network debugging.
Step S2: annotating a plurality of anchor ultrasonic signal distribution tensor samples respectively according to the basic representation matrix to obtain corresponding sample prior marks, and repeatedly debugging a ponding detection network according to the anchor ultrasonic signal distribution tensor samples and the corresponding sample prior marks respectively until the debugging cut-off requirement is met; the ponding detection network comprises a first characterization vector extraction component and a distribution information strengthening component, wherein the first characterization vector extraction component and the second characterization vector extraction component are twin components, and the same parameter is used.
And S2, annotating a plurality of anchor ultrasonic signal distribution tensor samples by using a basic characterization matrix, further generating corresponding sample prior marks, and debugging a ponding detection network through the marks.
First, the computer system annotates a plurality of anchor ultrasound signal distribution tensor samples according to the basic characterization matrix generated in step S1. The annotation process herein is understood to be the characterization of the anchor ultrasound signal matrix in each sample to determine if it represents a water free condition. Because the basic characterization matrix is generated by the ponding-free ultrasonic signal data, the labels can be used as prior labels of the sample, namely real labels known in advance. Next, the computer system uses these a priori labeled anchor ultrasound signal distribution tensor samples to debug the water logging detection network.
In step S2, as an optional implementation manner, annotating a plurality of anchor ultrasonic signal distribution tensor samples according to the basic characterization matrix to obtain corresponding sample prior marks, which may specifically include:
for a plurality of anchor ultrasonic signal distribution tensor samples, respectively carrying out the following treatments:
step S21: and for one anchor ultrasonic signal distribution tensor sample, based on the second characterization vector extraction component, respectively extracting characterization vectors of each anchor ultrasonic signal matrix in one anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors.
Specifically, the computer system performs feature extraction and vectorized representation of the anchor ultrasound signal matrix in each sample using the second token vector extraction component. In performing step S21, the computer system first receives as input an anchor ultrasound signal distribution tensor sample. This example typically contains a plurality of matrices of anchor ultrasound signals, each of which records the distribution of ultrasound signals over time and space. Such information is critical to determining if there is water accumulation within the anchor.
Next, a second token vector extraction component is invoked, which is a pre-trained machine learning model whose task is to extract key feature information from the input anchor ultrasound signal matrix. Such characteristic information may include the intensity, frequency, propagation velocity, etc. of the ultrasonic signal, which may reflect the structure and state of the interior of the anchor. The second eigenvector extraction component will process each input anchor ultrasonic signal matrix independently. Through a series of mathematical operations and transformations, it converts each matrix into a fixed length vector, which is the sample characterization vector. The sample characterization vector is a highly abstract and compact representation of the original matrix information, which contains critical information for subsequent analysis and judgment.
For example, if the input anchor ultrasound signal matrix is a two-dimensional array in which each element represents the ultrasound signal intensity at a particular location, the second token vector extraction component may employ the structure of a Convolutional Neural Network (CNN) to extract features. The CNN converts the input two-dimensional array into a one-dimensional sample characterization vector through a series of convolution, pooling, and full-join layers. Each element in this vector is an abstract representation of a certain region or feature in the original matrix.
Through the processing of step S21, the computer system can obtain a sample characterization vector corresponding to each matrix in each anchor ultrasonic signal distribution tensor sample. These vectors will be used as inputs for subsequent steps (e.g., step S22 and step S23) for further analysis and determination.
Step S22: and determining the composition priori marks of the corresponding anchor ultrasonic signal matrixes according to the obtained commonality evaluation coefficients between the sample characterization vectors and the basic characterization matrixes, wherein each composition priori mark characterizes the confidence degree of the corresponding anchor ultrasonic signal matrix as a ponding anchor ultrasonic signal matrix.
Specifically, the computer system first calculates the similarity between each sample characterization vector and the base characterization matrix. This similarity can be measured by various algorithms, such as cosine similarity, euclidean distance, etc. The basic characterization matrix is generated from ultrasound signal data in a water-free state, so that sample characterization vectors that are highly similar to the basic characterization matrix are more likely to represent the water-free state. After calculating the similarity, these similarity values are used as a common evaluation coefficient. These commonality assessment coefficients reflect how close the sample characterization vector is to the underlying characterization matrix in feature space. The calculation of the commonality assessment coefficients may involve normalization, weight adjustment, etc. steps to ensure that the contribution of features in different dimensions to the final assessment is reasonable.
Next, a component a priori signature for each anchor ultrasonic signal matrix is determined based on the commonality assessment coefficients. These markers are typically a number or set of numbers that indicate the confidence that the corresponding matrix is a water anchor ultrasonic signal matrix. The confidence level may be calculated based on a preset threshold or a series of rules. For example, if the commonality assessment coefficient is above a certain threshold, the corresponding matrix may be considered more likely to represent a water-free state, so its constituent a priori signature would reflect a higher confidence level.
For example, assuming that there is a commonality evaluation coefficient of 0.95 for a sample characterization vector and the underlying characterization matrix, this high value indicates that the sample is likely to represent a water-free state. Thus, the component a priori flag of the corresponding anchor ultrasound signal matrix may be set to a value close to 1 (e.g., 0.9 or 0.95), indicating a high confidence level. Conversely, if the commonality assessment coefficient is low (e.g., 0.3), then the constituent a priori flag may be set to a low value (e.g., 0.1 or 0.2) indicating low confidence.
Through the process of step S22, the computer system can assign a component prior signature to each anchor ultrasonic signal matrix, which will be used in subsequent steps to comprehensively determine the water accumulation condition of the entire sample.
Step S23: and obtaining a sample prior mark of an anchor ultrasonic signal distribution tensor sample according to the obtained each composition prior mark.
Step S23 combines the individual constituent prior markers obtained in step S22 to form a complete sample prior marker. This sample prior signature will contain water confidence information corresponding to each anchor ultrasonic signal matrix in the sample. In performing step S23, the computer system collects all of the constituent prior markers that respectively represent the confidence that the corresponding anchor ultrasound signal matrix is in the water-bearing state. Each of the component prior marks is an independent evaluation result and reflects the association degree between the corresponding matrix and the ponding state.
In order to combine these constituent prior labels into one sample prior label, the computer may employ different strategies. A simple method is to splice or combine all the constituent prior markers in some way to form a vector or array containing all the confidence information. This vector or array constitutes the sample prior marker. For example, if each constituent prior label is a number indicating the confidence that the corresponding matrix is in a water-bearing state, the sample prior label may be an array containing all of these numbers. Each element of the array corresponds to a component prior mark, and the integrity of the original confidence information is maintained.
The objective of step S23 is to integrate confidence information dispersed in each anchor ultrasonic signal matrix to form a comprehensive and unified sample prior mark. This flag will be one of the important criteria for determining whether the whole sample is in a water accumulation state in the subsequent step. Through the processing of step S23, the computer system can generate a sample prior mark containing rich information, which provides powerful support for subsequent ponding detection work.
The ponding detection network comprises a first characterization vector extraction component and a distribution information strengthening component. The first token vector extraction component and the second token vector extraction component used in step S1 are twinning components that have the same structure and parameters ensuring consistency in processing different types of ultrasound signal data.
In the debugging process, the ponding detection network receives a sample with a priori mark as input, extracts matrix characterization vectors in the sample through the first characterization vector extraction component, and then performs strengthening operation on the vectors through the distribution information strengthening component. And then, the network predicts the ponding condition according to the reinforced characterization vector and compares the ponding condition with the prior mark of the sample to calculate a prediction error.
Based on the prediction error, the computer can adjust the parameter of the ponding detection network by adopting an optimization algorithm so as to reduce the prediction error and improve the accuracy of the network. This process is repeated until a preset debug cutoff requirement is reached, such as the prediction error being below a certain threshold or the number of debugs reaching an upper limit.
Through the debugging process of the step S2, the ponding detection network can gradually learn the ability of accurately identifying ponding situations from the anchor ultrasonic signal matrix. The method lays a solid foundation for carrying out ponding detection on actual ultrasonic signal data by using the network in the follow-up step.
Step S3: according to a first characterization vector extraction component in the water accumulation detection network after debugging, respectively extracting characterization vectors of each piece of water accumulation-free ultrasonic signal data, and carrying out distribution information strengthening operation on matrix characterization vectors obtained by extraction on the basis of a distribution information strengthening component in the water accumulation detection network after debugging, so as to obtain a comparison characterization vector set, wherein the comparison characterization vector set comprises comparison characterization vectors deployed in advance.
Step S3, processing and strengthening the ponding-free ultrasonic signal data by utilizing the optimized network component, and generating a comparison characterization vector set. This step provides the necessary reference standard for subsequent water accumulation detection.
In specific execution, the computer system first invokes the first token vector extraction component in the debugged water logging detection network. This component is structurally identical to the second token vector extraction component used in step S1 and their parameters have reached a preferred state after the debugging of step S2, thus enabling efficient extraction of key features from the incoming water-free ultrasound signal data. These features are then encoded into matrix characterization vectors, each of which contains information about a certain anchor ultrasound signal matrix in the raw data.
For example, if the input water-free ultrasound signal data is a three-dimensional tensor comprising a plurality of two-dimensional anchor ultrasound signal matrices, the first token vector extraction component may employ a Convolutional Neural Network (CNN) structure to extract the features of these matrices layer by layer. The final output matrix characterization vector may be a high-dimensional array with each element corresponding to an abstract representation of a particular region or feature in the input tensor.
These matrix characterization vectors are then fed into a distribution information enhancement component in the commissioned water logging network. The role of this component is to further enhance the expressive power of these vectors, in particular their relevance in terms of spatial distribution. These vectors are enhanced into more discriminative and robust contrast token vectors through a series of nonlinear transformations and feature recombination.
Finally, these enhanced collation token vectors constitute a collation token vector set. Each vector in this set is a highly abstract and informative representation of the corresponding water-free ultrasound signal data that will be used as a reference standard in subsequent water detection operations. By comparing and matching the characterization vectors generated by the data to be detected, the comparison characterization vectors can help the computer system to more accurately identify the water accumulation condition.
As an implementation manner, step S3, according to a first token vector extraction component in the water accumulation detection network after the debugging is completed, performs token vector extraction on each water accumulation-free ultrasonic signal data, and performs a distribution information reinforcement operation on a matrix token vector obtained by extraction based on a distribution information reinforcement component in the water accumulation detection network after the debugging is completed, so as to obtain a comparison token vector set, which specifically may include:
Step S31: according to the first characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data respectively, and performing distribution information reinforcement operation on the matrix characterization vectors obtained by extraction based on the distribution information reinforcement component in the ponding detection network after debugging is completed, so as to obtain an iteration characterization vector set, wherein the iteration characterization vector set comprises iteration characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data.
In performing step S31, the computer system first processes the water-free ultrasound signal data using the first token vector extraction component. Such data typically contains a plurality of anchor ultrasound signal matrices, each of which records ultrasound signal information at a particular time and location. The first token vector extraction component is a pre-trained machine learning model that functions to extract key feature information from these matrices and convert this information into token vectors. These token vectors are abstract representations of the original matrix information that capture the dominant features and patterns in the matrix. For example, the first token vector extraction component may be a Convolutional Neural Network (CNN) that processes the input anchor ultrasound signal matrix through a series of convolutional, pooling, and fully-connected layers. In this process, the CNN learns the way to extract valid features from the original matrix and encodes those features into a token vector. These token vectors are then used as inputs for subsequent operations.
The computer system then uses the distributed information enhancement component to further process the extracted matrix characterization vector. The goal of the distribution information enhancement component is to enhance the spatial and temporal correlation between the characterization vectors in order to better capture the ultrasound signal distribution characteristics in the water-free state. This may be achieved by applying a series of mathematical operations and transformations on the token vector, such as matrix multiplication, convolution operations, or Recurrent Neural Networks (RNNs), etc. Through the distributed information augmentation operation, the computer system is able to obtain a set of enhanced iterative token vectors. Each iterative characterization vector in this set corresponds to an anchor ultrasound signal matrix in the water-free ultrasound signal data, and they have been optimized to better reflect the characteristics of the matrix in the water-free state.
In general, step S31 achieves in-depth processing and analysis of the water-free ultrasound signal data by combining the functions of the first token vector extraction component and the distribution information enhancement component. This provides more accurate and reliable underlying data support for subsequent water logging detection.
Step S32: and determining part of the iterative token vectors in the iterative token vector set based on the commonality evaluation coefficients between every two iterative token vectors in the iterative token vector set to obtain a comparison token vector set.
Step S32 is responsible for screening out a part of representative vectors from the iterative token vector set to form a collation token vector set. In performing step S32, the computer system first calculates a commonality assessment coefficient between each two iterative token vectors in the set of iterative token vectors. These coefficients are used to measure similarity and correlation between vectors, and may be calculated using various algorithms, such as cosine similarity, euclidean distance, etc. By calculating these coefficients, the computer system is able to learn the relative position and relationship between the different vectors in the feature space.
Next, it is determined which iteration characterization vectors are more representative based on the computed commonality assessment coefficients. This typically involves setting a threshold or applying some sort of ordering algorithm to select partial vectors from the set. For example, a vector whose common evaluation coefficient is higher than a certain threshold value may be selected, or the first N vectors may be selected in order of coefficients from high to low. Through this step, the computer system can screen out those iterative token vectors that have significant features in the no-water state and have a high similarity to each other. These selected vectors constitute a set of reference characterization vectors that will play an important role in the subsequent water accumulation detection process.
For example, assume that the set of iterative characterization vectors includes 10 vectors that each represent a characteristic of a different anchor ultrasound signal matrix. The computer system calculates the commonality evaluation coefficients between the 10 vectors to obtain a 10×10 coefficient matrix. It may then select a portion of the vectors with higher coefficients from this matrix according to a pre-set threshold or ordering rule. For example, if the threshold is set to 0.8, only those vector pairs with coefficients higher than 0.8 will be selected. Ultimately, these selected vectors constitute a set of collation token vectors. It should be noted that in practical applications, the calculation method of the commonality evaluation coefficient and the selection of the threshold/ordering rule need to be adjusted and optimized according to the specific data set and the detection requirement.
As an implementation manner, step S32, based on the commonality evaluation coefficient between every two iterative token vectors in the iterative token vector set, determines a part of iterative token vectors in the iterative token vector set, to obtain a comparison token vector set, which specifically may include:
Step S321: and determining one or more iterative characterization vectors in the iterative characterization vector set, and constructing a basic set for the contrast characterization vector set.
Step S321 is an initialization step when building a set of collation token vectors. The main goal of this step is to set a starting point for the subsequent iterative optimization process, ensuring that the most representative vector is effectively screened out of the iterative token vector set. Upon execution of step S321, the computer system first processes the iterative token vector set. This set contains the characterization vectors extracted and enhanced from the anchor ultrasound signal matrix for each of the water-free ultrasound signal data. Each vector is a high-dimensional representation of the original matrix data, capturing key features therein. Next, one or more iterative token vectors are determined from the set of iterative token vectors for initializing the set of collation token vectors. The "determining" herein may be a random selection, a selection based on some preset rule, or a selection of the most representative vector using some algorithm (e.g., a clustering algorithm). The initialization process is intended to provide a basis for subsequent optimizations, so this step does not require immediate screening of all vectors that will ultimately be included in the set of collation token vectors.
Taking a random selection as an example, the computer may randomly extract a certain number of vectors from the iterative token vector set as initial members of the contrast token vector set. These selected vectors will serve as the starting point for the subsequent optimization process. Once the initialization is complete, a basis for further optimization is provided against the set of token vectors. In a subsequent step (e.g., step S322), the computer will utilize different algorithms and techniques to continuously optimize the set to ensure that it contains the most representative token vector, thereby improving the accuracy and efficiency of the water accumulation detection. It should be noted that the specific implementation of the initialization step may vary depending on the application scenario, the characteristics of the data set, and the detection requirements. In practical application, the most appropriate initialization strategy should be selected according to the specific situation.
Step S322: iterative optimization (against a set of token vectors).
Wherein, in one round of optimization, the following processes are performed:
Step S3221: for each iteration token vector in the iteration token vector set, a corresponding token vector tuple is respectively determined, and each token vector tuple comprises an iteration token vector and a comparison token vector with the largest common evaluation coefficient with the iteration token vector in the comparison token vector set.
In step S322, the computer system will perform a number of iterative optimizations on the initially constructed set of control token vectors. The goal of the optimization is to update the vectors in the set step by step, ensuring that they can more accurately reflect the ultrasound signal characteristics in the water-free state.
In particular, the optimization process may involve a variety of strategies and techniques including, but not limited to:
vector reordering: the vectors in the set are rearranged based on some evaluation criterion (e.g., a commonality evaluation factor) so that more representative vectors are ranked in front.
Vector screening: vectors in the set that are less similar or redundant are removed to reduce computational complexity and improve overall quality of the set.
New vector introduction: a new, more representative vector is selected from the set of iterative token vectors and added to the set of reference token vectors.
After each iteration, the computer system evaluates the performance of the current set of control token vectors and adjusts the optimization strategy based on the evaluation to further refine the set in the next iteration.
For example, assume that the initial set of cross-reference token vectors includes 5 vectors. In the first iteration, the computer system may find that two of the vectors are very similar, and thus decide to remove one to reduce redundancy. At the same time, it may select a new vector from the iteration characterization vector set, which is greatly different from the vector in the current set, to be added. Thus, after the first iteration, the set of reference characterization vectors is optimized to a certain extent.
In particular embodiments, the computer system will process each vector in the set of iteratively characterized vectors. For each iterative token vector in the set, the computer system calculates the commonality evaluation coefficients (e.g., cosine similarity) between it and all vectors in the set of comparison token vectors, and finds the comparison token vector having the greatest commonality evaluation coefficient with the current iterative token vector. These two vectors (an iterative token vector and a cross token vector) then form a token vector doublet.
In this way, the computer system can find one of the most similar objects in the set of collation token vectors for each vector in the set of iteration token vectors. These tuples provide a basis for the subsequent optimization step (step S3222), helping the computer system to more accurately understand which vectors are more representative in the water-free state, and accordingly adjust the composition of the set of reference characterization vectors.
Step S3222: for each contrast token vector contained in each token vector doublet, adding the iterative token vector with the smallest commonality evaluation coefficient with each contrast token vector to the contrast token vector set.
Step S3222 is responsible for updating the set according to the commonality evaluation coefficients in the token vector tuples. The purpose of this step is to ensure that the reference token vector set can contain more diverse and representative vectors, thereby improving the accuracy of the water accumulation detection. In performing step S3222, the computer system traverses all formed token vector tuples. For the contrast token vector in each dyadic, find the iterative token vector with the smallest commonality evaluation coefficient. This minimum co-evaluation coefficient means that in the feature space this iterative token vector differs most from the current reference token vector, and thus it may contain some important information or features that are currently lacking in the reference token vector set.
After finding the iterative token vector with the smallest common evaluation coefficient, adding the iterative token vector to the comparison token vector set. Thus, a new vector which is greatly different from the existing vector is added to the set, so that the diversity and the representativeness of the set are improved. It is noted that the size limitation of the set may need to be considered when adding a new iterative token vector. If there is a size limitation against the set of token vectors, then it may be necessary to remove some of the more highly similar or less important vectors in the set before adding a new vector to ensure that the size of the set is within a controllable range.
For example, assume that there is one token vector doublet, where the commonality evaluation coefficients against token vector A and iterative token vectors B and C are 0.8 and 0.5, respectively. Since 0.5 is less than 0.8, the computer system will choose to add the iterative token vector C to the set of collation token vectors, rather than vector B. The reason for this is that vector C differs from vector a more in feature space, so it is more likely to bring new, valuable information to the collection.
By repeatedly executing step S3222 (and other optimization steps), the computer system is able to build a more complete, more representative set of control token vectors, providing a solid basis for subsequent water accumulation detection.
In step S20, according to the distribution condition of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor, the distribution information strengthening operation is performed on the corresponding matrix characterization vector to obtain a corresponding strengthening characterization vector including the distribution condition, which specifically may include:
based on the distributed information strengthening assembly after debugging, the following treatments are respectively carried out on each anchorage ultrasonic signal matrix:
Step S21: and for an anchor ultrasonic signal matrix, obtaining a corresponding sequence number representation vector according to the sequence number distribution condition of the anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor.
In executing step S21, the computer system selects an anchor ultrasonic signal matrix as a processing object. The matrix is extracted from the ultrasonic signal data of the anchor to be detected and contains rich signal characteristic information. Then, the computer generates a corresponding serial number representation vector according to serial number distribution conditions of the matrix in the anchor ultrasonic signal distribution tensor. The sequence number characterization vector is a vector that describes the location information of the matrix in the distribution tensor. Its generation can be understood as encoding the position of the matrix in the distribution tensor. Specifically, the computer system traverses the anchor ultrasound signal distribution tensor to find the position (i.e., sequence number) corresponding to the currently processed matrix, and converts the position information into a vector form. This vector, called the sequence number token vector, reflects the distribution of the matrix in the distribution tensor.
This procedure is illustrated by way of example: it is assumed that there is an anchor ultrasound signal distribution tensor T that contains a plurality of anchor ultrasound signal matrices. One of the matrices M is now processed, and its position in the tensor T can be represented by a three-dimensional coordinate (i, j, k). Then, this position information can be converted into a sequence number characterization vector v= [ i, j, k ]. This vector V describes the position information of the matrix M in the tensor T. It should be noted that the sequence number characterization vector is only a simple example, and in practical application, a more complex coding manner may be designed according to specific requirements and data characteristics. However, the core function of the sequence number characterization vector is to describe the position information of the matrix in the distribution tensor, so as to provide a basis for the subsequent strengthening operation.
After the generation of the sequence number token vector is completed, the computer system may proceed to step S22, where the enhanced token vector is generated according to the original token vector and the sequence number token vector of the matrix. The process combines the characteristic information and the position information of the matrix to provide more comprehensive and more accurate input data for subsequent ponding detection.
Step S22: and obtaining a corresponding strengthening characterization vector according to the matrix characterization vector and the serial number characterization vector of one anchor ultrasonic signal matrix.
In step S22, the computer system will generate a new enhanced token vector containing more information using the previously generated matrix token vector and sequence number token vector.
Specifically, the computer system obtains a matrix characterization vector of an anchor ultrasound signal matrix. The matrix characterization vector is extracted from the original ultrasonic signal data by a specific feature extraction method (such as principal component analysis, deep learning model and the like), and contains main features and key information in the matrix. The computer then obtains a sequence number characterization vector corresponding to this matrix. A sequence number characterization vector is generated in step S21, which describes the position information of the matrix in the anchor ultrasound signal distribution tensor.
Next, the computer system performs a stitching operation on the two vectors. The stitching operation is a method of combining multiple vectors into a longer vector that can fuse together information from different sources to form a more comprehensive representation. In this example, by concatenating the matrix characterizing vector and the sequence characterizing vector, an enhanced characterizing vector may be obtained that contains both matrix characterizing information and position information.
This procedure is illustrated by way of example: assuming a matrix characterization vector v1= [0.1, 0.2, 0.3], it describes the characteristic information of a certain anchor ultrasonic signal matrix; there is also a sequence number characterizing vector v2= [1, 2, 3], which describes the position information of the matrix in the distribution tensor. Then, by the stitching operation, an enhancement characterization vector v= [ V1, V2] = [0.1, 0.2, 0.3, 1, 2, 3] can be obtained. The enhanced characterization vector V contains both the characteristic information of the matrix (embodied by V1) and the position information (embodied by V2), so that the enhanced characterization vector V can provide more comprehensive and accurate input data for subsequent ponding detection.
It should be noted that the splicing operation is only a simple implementation, and more complex reinforcement methods may be designed according to specific requirements and data characteristics in practical applications. However, the core function of the enhanced characterization vector is to fuse information from different sources together to form a more comprehensive and representative characterization, thereby providing a more accurate and reliable basis for subsequent ponding detection.
In step S10, as an implementation manner, extracting a characterization vector from each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector, which may specifically include:
Carrying out multi-round characterization vector extraction on the distribution tensor of the ultrasonic signals of the anchorage device; wherein each round performs the following processes:
Step S11: according to a preset filtering matrix, filtering characterization vector extraction is carried out on the input of the current wheel characterization vector extraction, and a filtering characterization vector is obtained; the input is an anchor ultrasonic signal distribution tensor when the first-round characterization vector is extracted, and is the output of the previous-round characterization vector extraction when the other-round characterization vectors are extracted.
Specifically, when performing the first round of token vector extraction, the computer system receives as input an anchor ultrasound signal distribution tensor. The distribution tensor is a multi-dimensional array containing various characteristic information extracted from the anchor ultrasound signal. The computer then convolves this input tensor with a preset filter matrix. The filter matrix is usually a matrix with specific size and weight parameters, which can perform sliding window type processing on input data according to preset rules, so as to extract characteristic information matched with the filter matrix.
For example, assume that the input is a three-dimensional tensor of the shape (10, 10, 10), representing a 10x10 feature distribution of the anchor ultrasound signal image at different depths. The preset filter matrix is a 3x3x3 matrix for extracting local features in the image. The computer system applies the filter matrix to the input tensor, and calculates the characteristic response value at each position through convolution operation, so as to obtain a new three-dimensional array as a filter characterization vector. This new array may have the same spatial dimensions as the input tensor, but the value at each location represents the degree to which the feature near that location matches the filter matrix.
In subsequent rounds, the process is similar, but the input data becomes the output of the token vector extraction of the previous round. This means that each round of extraction is further refined and abstracted based on the features of the previous round of extraction. In this way, the computer system is able to progressively extract more complex and abstract feature information, providing an efficient representation of the features for subsequent analysis and recognition tasks.
It should be noted that the selection and design of the filter matrix may be chosen to suit the particular requirements and data characteristics.
Step S12: and performing cross-layer identity connection according to the filtering characterization vector and the input to obtain a fusion characterization vector.
Step S12 fuses the filtered token vector with the input by means of cross-layer identity connection (also referred to as residual connection), resulting in a richer and more comprehensive fused token vector. This step aims at preserving important information of the original input and combining the features after the filtering process to jointly construct a more powerful feature representation. Specifically, when the computer system executes to step S12, two inputs are received simultaneously: one is a filtering characterization vector, namely, the characteristic information extracted after the filtering matrix processing; the other is the original input, i.e. the input data of the current round token vector extraction. These two inputs may differ in dimension and shape, so appropriate adjustments need to be made to ensure that they can be fused effectively.
Cross-layer identity connection is a common feature fusion method, and is particularly common in deep learning. The core idea is to directly add or splice together the feature information of different levels in order to use these features simultaneously in the subsequent network layers. The method is helpful to relieve the problem of gradient disappearance or gradient explosion of the deep learning model in the training process, so as to accelerate the convergence of the model and improve the performance.
In step S12, the computer system fuses the filtered token vector with the original input using a cross-layer identity connection. In particular, it may stitch the filtered token vector with the original input in the channel dimension to form a new tensor as the fused token vector. The new tensor contains detailed information in the original input and incorporates the characteristic information after filtering processing, so that the subsequent network layer can use the information of the two aspects to perform more accurate prediction or classification.
In deep learning, cross-layer identity connections, also known as jump connections or short-circuit connections, are used to help deep networks learn identity mapping, thereby alleviating gradient vanishing or representing bottleneck problems. In step S12, the filtered token vector is added to the input vector (the original data before the multi-layer convolution or processing or the output of the previous layer), e.g., there is an input vector input_vector, which is the output of the previous token vector extraction (or the original distribution tensor of the anchor ultrasound signal in the first round).
A filtered vector is also obtained by the filtering process, which contains the specific features extracted from the input vector. The residual connection is achieved by adding the filtered_vector to the input_vector (in the case where they have the same dimensions) or by matching their dimensions by a suitable transformation. This addition procedure represents a residual connection. If the dimensions of the filtered_vector and the input_vector are different, their dimensions may be adjusted by 1x1 convolution, upsampling, downsampling, or other methods so that they may be added.
The result of the addition is a fused vector that fuses the original information and the filter characteristics, and that is passed on to the next round (if any) as the output of the current round or directly for the subsequent task.
For example: let input_vector be a tensor in the shape (batch_size), where batch_size is the batch size and features is the feature quantity. After filtering, a filtered_vector with the same shape is obtained. The residual connection adds the two tensors:
residual_connection = input_vector + filtered_vector
fused_vector = residual_connection
In this example, the integrated_vector now contains the original information of the input_vector and the filter characteristics in the filtered_vector. If this is the last round of token vector extraction, then the summed_vector will be the final output.
In this way, step S13 ensures that even after multiple rounds of token vector extraction, important information of the original signal is not lost and is effectively combined with the extracted features.
Step S13: if the current round token vector extraction is the last round token vector extraction, the fusion token vector is taken as the output of the current round token vector extraction.
Specifically, when the computer system proceeds to step S13, it is first checked whether the current round is the last round of the feature vector extraction preset. This preset number of rounds is set according to the specific application requirements and data processing complexity. If the current round is indeed the last round, the computer system will take the fusion token vector generated in the previous step as the output of the current round. The fusion characterization vector is obtained through cross-layer identity connection in the step S13, integrates information of the filtering characterization vector and the input vector, and has richer feature representation capability. In the last round of token vector extraction, the fusion token vector is taken as output, which means that the whole extraction process has completed layer-by-layer extraction and abstraction of input data, and a vector representation which contains original information and is subjected to feature extraction and optimization is obtained.
This output vector may be used for various subsequent tasks such as classification, regression, clustering, etc. For example, in the analysis of the anchor ultrasound signal, this output vector may be used to determine the quality of the anchor, identify potential defects, or predict the life of the anchor, etc.
It should be noted that the output of step S14 is not the end point of the entire process flow. In practical applications, further processing and analysis of the output vector is required according to specific requirements to obtain a final result or decision. For example, the output vector can be input into a classifier to automatically classify the anchor; or fusing the data with information of other data sources to perform more complex decision analysis.
Step S14: if the current-wheel representation vector extraction is not the last-wheel representation vector extraction, performing feature adaptive dimension reduction on the fusion representation vector, and taking the obtained dimension reduction representation vector as the output of the current-wheel representation vector extraction.
When the computer system judges that the current round is not the last round of feature vector extraction, feature adaptability dimension reduction is carried out on the fusion feature vector, and then the feature vector after dimension reduction is used as the output of the current round for the next round of extraction. Feature adaptive dimension reduction is a process of reducing feature dimensions to simplify model complexity according to factors such as importance, relevance or information quantity of features. In step S15, the computer system implements this dimension reduction process using a specific algorithm or neural network model. For example, it may use classical dimension reduction algorithms such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), and may also learn and extract key features using deep learning models such as self-encoders and Convolutional Neural Networks (CNNs).
Taking principal component analysis as an example, the computer system performs PCA processing on the fusion token vector, and finds a new feature space by calculating the principal components of the data, where the new space can maximally preserve variance information in the original data. The computer system then projects the fused token vector onto this new feature space to obtain a reduced-dimension token vector. The dimension-reduced characterization vector is lower in dimension, redundant and noise information in the original data is removed, and the subsequent rounds of characterization vector extraction is facilitated.
It should be noted that the method for dimension reduction and the parameter setting should be determined according to the specific application scenario and the data characteristics. In the processing of anchor ultrasound signals, proper dimension reduction processing is critical to improving computational efficiency and model performance due to the high dimension and complexity of the signals.
Through the feature adaptive dimension reduction processing of step S15, the computer system can reduce the dimension and complexity of the data while maintaining the key information, and provide better input for the subsequent rounds of token vector extraction. The method is favorable for the smooth progress of the whole processing flow, and finally improves the accuracy and reliability of the ultrasonic signal analysis of the anchorage device.
In one embodiment, step S30, according to the commonality evaluation coefficients between each enhanced characterization vector and the contrast characterization vector deployed in advance, determines the ponding condition prediction coefficients corresponding to each anchor ultrasonic signal matrix respectively, which specifically may include:
For each enhancement characterization vector, the following treatments were performed:
Step S31: for one enhancement token vector, a collation token vector having the greatest common evaluation coefficient with the one enhancement token vector is determined.
Step S32: and determining a ponding condition prediction coefficient of an anchor ultrasonic signal matrix corresponding to the strengthening characterization vector according to the error between the comparison characterization vector and the strengthening characterization vector.
In the embodiment of step S30, the computer system determines the water accumulation condition prediction coefficients corresponding to the anchor ultrasonic signal matrices according to the commonality evaluation coefficients between each enhanced characterization vector and the previously deployed contrast characterization vector. Specifically, step S31 requires the computer system to find, for each enhanced token vector, a collation token vector with which the coefficient of commonality is the greatest. The commonality evaluation coefficient is an index for measuring similarity or commonality between two characterization vectors, and can be obtained by calculating cosine similarity, euclidean distance and the like between the vectors. In this step, the computer system traverses all of the collation token vectors, calculates the commonality assessment coefficients between each collation token vector and the currently processed reinforcement token vector, and selects the collation token vector having the largest coefficient as the matching result. For example, assuming an enhanced token vector A, the computer system calculates a commonality assessment coefficient between A and each of the reference token vectors (e.g., B1, B2, B3, etc.). If the calculation shows that the commonality assessment coefficient of A and B2 is the largest, then B2 is determined as the comparison characterization vector having the largest commonality with A.
This process is performed once for each enhancement token vector to ensure that each enhancement token vector finds the most similar reference token vector. The output of this step is a set of matching pairs, each pair containing an enhanced token vector and a reference token vector with the greatest coefficient of commonality.
Next, in step S32, the computer system determines a water accumulation prediction coefficient of the anchor ultrasonic signal matrix according to the error between the contrast characterization vector and the reinforcement characterization vector. This error can be obtained by calculating a difference measure (e.g., euclidean distance, manhattan distance, etc.) between the two vectors. The prediction factor may be a function of this error, e.g. the smaller the error, the larger the prediction factor, which indicates that the anchor is more watery. In this way, the computer system is able to generate a water accumulation condition prediction coefficient for each anchor ultrasonic signal matrix for subsequent health assessment and prediction.
In step S32, the computer system calculates the difference or error between the contrast token vector and the reinforcement token vector. Such errors can be measured by a variety of mathematical methods, such as euclidean distance, cosine similarity, or loss functions in more complex machine learning algorithms, etc. The magnitude of the error reflects the proximity of the two vectors in the feature space, i.e. their similarity or difference.
Taking Euclidean distance as an example, assume that the enhanced token vector is a point A in a multidimensional space, and the reference token vector is another point B. The euclidean distance between point a and point B can be obtained by calculating the sum of squares of their differences in each dimension. The smaller this distance value, the closer the points a and B are, i.e., the more similar the reinforcement characterization vector is to the control characterization vector; conversely, a larger distance value indicates a larger difference between them. After the error is determined, the computer system calculates a water accumulation condition prediction coefficient based on the error value. This coefficient may be a function directly based on the error value, such as the inverse of the error or some normalized value. The specific calculation mode of the prediction coefficient may be different according to the application scene, but the prediction coefficient is generally designed to be larger when the error is smaller, so that the possibility of water accumulation in the anchorage device is reflected to be higher.
For example, if the error value is calculated as the Euclidean distance and this distance is small, the prediction factor may be set to a value close to 1, indicating that there is a high likelihood of water accumulation inside the anchor. Conversely, if the euclidean distance is large, the prediction factor may be set to a value close to 0, which indicates that there is a low likelihood of water accumulation inside the anchor. Through such a process flow, step S32 can generate a water accumulation condition prediction coefficient corresponding to each reinforcement characterization vector. These prediction coefficients may then be used for further analysis and decision support, such as input features for machine learning models to predict the health of the anchor or to formulate maintenance strategies, etc.
As another embodiment, step S30, according to the commonality evaluation coefficients between each enhanced characterization vector and the contrast characterization vector deployed in advance, determines the ponding situation prediction coefficients corresponding to each anchor ultrasonic signal matrix respectively, which specifically may include:
For each enhancement characterization vector, the following treatments were performed:
Step S301: for one enhancement token vector, a collation token vector having the greatest common evaluation coefficient with the one enhancement token vector is determined.
In this alternative embodiment of step S30, the computer system, when processing the enhanced token vector to determine the hydrops condition prediction coefficients corresponding to the anchor ultrasound signal matrix, focuses on the reference token vector having the greatest common evaluation coefficient with the enhanced token vector, as well as other reference token vectors in close proximity thereto.
Step S301 requires the computer system to find, for each enhanced token vector, a collation token vector with which the coefficient of commonality is the greatest. The commonality evaluation coefficient is an index for measuring similarity between two vectors, and can be obtained by calculating cosine similarity between the vectors, pearson correlation coefficient and the like. In this step, the computer traverses all previously deployed collation token vectors, calculates the commonality assessment coefficients between them and the currently processed reinforcement token vector, and selects the collation token vector with the largest coefficient.
For example, assume that there is an enhanced characterization vector V_enhanced, which represents the key features of a certain anchor ultrasound signal matrix. The computer system calculates V _ enhanced and each of the comparison token vectors (e.g. V _ control1, V control2, where, V controlN) are used. If the calculation result shows that the commonality evaluation coefficient of V_enhanced and V_CONTROL3 is the largest, then V_CONTROL3 is determined as the most similar comparison characterization vector with V_enhanced.
The output of this step is a reference token vector that is most similar to each of the enhanced token vectors. These control characterization vectors will be used in subsequent steps to further analyze the characteristics of the enhancement characterization vector and determine the water accumulation condition prediction coefficients.
Step S302: among the collation token vectors deployed in advance, a plurality of collation token vectors having a commonality evaluation coefficient not larger than a preset commonality evaluation coefficient threshold value are determined with the collation token vector having the largest commonality evaluation coefficient.
In step S302, the computer system focuses on not only the reference token vector having the greatest common evaluation coefficient with the enhanced token vector, but also further considers other reference token vectors that are close to the greatest common evaluation coefficient. Specifically, the computer system first determines a predetermined commonality assessment factor threshold. This threshold is set according to the actual application scenario and requirements, which determines which of the collation token vectors will be considered sufficiently similar to the collation token vector with the largest commonality assessment coefficient. The setting of the commonality evaluation coefficient threshold may be determined based on experience, experimental data, statistical analysis, or the like. Once the commonality assessment coefficient threshold is determined, the computer system traverses all of the collation token vectors previously deployed. For each of the collation token vectors, the computer system calculates a commonality evaluation coefficient between it and the collation token vector having the largest commonality evaluation coefficient. If this calculated commonality assessment coefficient is not greater than a preset threshold value, then the collation token vector is considered to be a similar part to the collation token vector having the largest commonality assessment coefficient.
This procedure is illustrated by way of example: assume that an enhanced token vector V _ enhanced is present and that a comparison token vector V _ max _ similar has been determined that has the greatest coefficient of commonality with it. Now, a commonality evaluation coefficient threshold T is set. The computer system calculates V _ max _ similar and all other reference token vectors (e.g. V _ control1, V control2, where, V controlN) are used. If the commonality assessment coefficient between a certain contrast token vector V_ controlX and V_max_similar is less than or equal to T, then V_ controlX is selected as one of the contrast token vectors that is similar to V_max_similar. This process continues until the computer system finds all of the cross-reference token vectors that satisfy the condition. These selected control token vectors will be used in subsequent water logging prediction coefficient determination steps along with the control token vector having the greatest common assessment coefficient. Step S302 helps to improve the accuracy and reliability of the prediction coefficients by considering a plurality of similar reference token vectors.
Step S303: and determining the ponding condition prediction coefficient of the anchor ultrasonic signal matrix corresponding to the strengthening characterization vector according to the comparison characterization vector with the maximum commonality evaluation coefficient and errors between the plurality of comparison characterization vectors and the strengthening characterization vector.
In step S303, the computer system comprehensively considers the information of the contrast token vector most similar to the reinforcement token vector (i.e. the contrast token vector having the largest commonality evaluation coefficient) and the plurality of contrast token vectors close thereto (i.e. other contrast token vectors having commonality evaluation coefficients not greater than the preset threshold value) to determine the final water accumulation condition prediction coefficient. Specifically, the computer system first calculates the error between the enhanced token vector and the reference token vector having the greatest common evaluation coefficient. This error may be calculated by euclidean distance between vectors, cosine similarity differences, or other suitable metrics. The magnitude of the error reflects how close the enhanced token vector is to the most similar reference token vector in feature space.
Next, the computer system calculates errors between the enhanced token vector and a plurality of reference token vectors that are deemed similar. These error values are also calculated based on the distance or similarity differences between the vectors. By considering a plurality of similar reference characterization vectors, the computer system can obtain more comprehensive information to more accurately evaluate the ponding condition of the anchor ultrasonic signal matrix represented by the enhanced characterization vectors.
After these error values are obtained, the computer system uses a predetermined algorithm or model to determine the water accumulation condition prediction coefficients. The algorithm or model may be a simple mathematical function (e.g., a weighted average of errors) or may be a more complex machine learning model (e.g., support vector machine, neural network, etc.). The specific choice of which algorithm or model depends on the complexity and data characteristics of the application scenario. For example, if a simple weighted average method is used, the computer system may assign a higher weight to the control token vector with the greatest commonality assessment coefficient and a lower weight to other similar control token vectors. A weighted average error is then calculated from these weights and the corresponding error values. Finally, the weighted average error is mapped to the water accumulation condition prediction coefficient through a certain conversion function (such as an inverse proportion function, an exponential function and the like).
It should be noted that the conversion function and the weight distribution method are designed according to practical application requirements and experience. Their objective is to make the water accumulation condition prediction coefficient larger when the error between the enhanced token vector and the reference token vector is smaller (i.e., commonality is larger); conversely, when the error is larger (i.e., the commonality is smaller), the water accumulation condition prediction coefficient is smaller.
Through such a process flow, information of a plurality of similar reference characterization vectors is comprehensively considered, and a more accurate and reliable water accumulation condition prediction coefficient is determined based on errors between the information and the enhanced characterization vectors. This prediction factor can then be used for further analysis and decision support, such as assessing the health of the anchor, formulating maintenance strategies, etc.
In step S10, as an implementation manner, extracting a characterization vector from each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector, which may specifically include:
step S101: and obtaining corresponding target reference bridge anchor data from a plurality of reference bridge anchor data which are deployed in advance according to the bridge tag of the bridge anchor.
In step S101, data related to the target bridge anchor, that is, target reference bridge anchor data, is accurately selected from a large number of reference bridge anchor data stored in advance. Specifically, the computer system first identifies the bridge tag carried by the bridge anchor to be processed. This tag typically contains a unique identifier of the bridge, such as a number, name or specific location information. Once the bridge tag is identified, the computer system accesses its internal storage or external database, which stores data for a plurality of reference bridge anchors. These data were collected historically for analysis and comparison of ultrasonic signal characteristics for different bridge anchors.
When accessing the data, the computer system screens and matches the data according to the bridge tag as a search condition. This process is similar to using the book name or author name in a library to find a particular book. By matching the tags, the computer system is able to quickly locate reference data associated with the target bridge anchor, which data is referred to as target reference bridge anchor data.
For example, assume that ultrasound signal data for an anchor labeled "bridge A" is being processed. The computer system will look up all historical anchor ultrasound signal data associated with the "bridge a" tag in its reference database. The data may include ultrasonic measurements of bridge a at various points in time, positional information of the anchor, and any history of water accumulation or damage. In this way, step S101 provides an accurate and targeted reference data basis for subsequent data alignment, region segmentation and token vector extraction. The importance of this step is that it ensures the accuracy and relevance of the subsequent analysis. By using reference data directly related to the target bridge anchor, the ultrasonic signal characteristics of the current anchor can be more accurately understood, thereby more effectively predicting and identifying potential water accumulation or damage conditions.
Step S102: and carrying out data alignment on the anchor ultrasonic signal distribution tensor according to preset anchor distribution information in the target reference bridge anchor data to obtain an alignment matrix.
The purpose of step S102 is to ensure that the anchor ultrasound signal distribution tensor to be analyzed is structurally consistent with the information in the target reference bridge anchor data, thereby facilitating subsequent comparison and analysis.
Specifically, the computer system first accesses target reference bridge anchor data that contains detailed information about the bridge anchor, most importantly preset anchor distribution information. This information describes the position of the anchor on the bridge. Next, the computer system uses these preset anchor distribution information as a reference to perform a data alignment on the current anchor ultrasonic signal distribution tensor to adjust the data of different data sources or different time points to a process under a unified coordinate system to ensure comparability between them. In the embodiment of the application, the data in the anchor ultrasonic signal distribution tensor is adjusted to be matched with the preset anchor distribution information in the target reference bridge anchor data. For example aligned by the serial number of the anchor.
After the data alignment is completed, the computer system will get an alignment matrix. The alignment matrix is an adjusted anchor ultrasonic signal distribution tensor, wherein data is consistent with preset anchor distribution information in target reference bridge anchor data. This means that the subsequent analysis steps can be performed on a uniform and accurate data basis.
Step S103: and obtaining matrix area blocks where the bridge anchor is positioned in the alignment matrix according to the preset matrix area in the target reference bridge anchor data.
In performing this step, the computer system first accesses the target reference bridge anchor data that has been stored. These data contain not only the basic information of the bridge anchor, but more importantly they also the preset matrix area information. These predetermined matrix areas are typically determined based on historical data, expert knowledge, or algorithmic analysis, and represent areas of particular interest or significance in the detection of ultrasound signals, such as those historically prone to water accumulation, damage, or other anomalies. The computer system then further processes the alignment matrix generated in the previous step according to the preset matrix area information. The alignment matrix is a two-dimensional or multi-dimensional array containing bridge anchor ultrasound signal data whose data structure has been aligned with the target reference data by a previous alignment step. On the basis of the above, the computer system divides the alignment matrix, extracts the data part corresponding to the preset matrix area, and forms independent matrix area blocks. These matrix area partitions are key objects for subsequent analysis. Because they directly correspond to areas of bridge anchors where problems may exist or particular attention is required, detailed analysis and comparison of these segments helps to more accurately identify and locate water accumulation, damage or other anomalies in the anchors.
By way of specific example, assume that the anchor ultrasound signal distribution tensor of a bridge is aligned to form a 10×10 matrix, and that the target reference bridge anchor data is provided with a region of interest that corresponds to a 4×4 sub-matrix of (3, 3) through (6, 6) in the aligned matrix. Then the operation of step S103 is to extract the 4×4 sub-matrix from the 10×10 alignment matrix to form an independent matrix area block for further analysis and processing.
Step S104: and extracting characterization vectors of each anchor ultrasonic signal matrix contained in the matrix region blocks to obtain corresponding matrix characterization vectors.
The core task of step S104 is to extract the token vector from the matrix area partitions obtained in the previous step. A token vector is a mathematical representation that can represent key features of raw data, commonly used in machine learning and data analysis. In performing step S104, the computer system operates for each matrix area block. First, it will analyze the data structures and patterns in these partitions. Such data may include characteristics of signal strength, frequency, waveform, etc., which reflect internal structural and status information of the anchor.
Next, the computer system utilizes a particular algorithm or model to extract the token vector from the partitions. These algorithms may be statistical-based methods such as Principal Component Analysis (PCA) or Independent Component Analysis (ICA) that are capable of extracting the dominant direction of change or independent component in the data as a characterization vector. Another common approach is to use neural networks, in particular Convolutional Neural Networks (CNNs), which are particularly suitable for processing data structures of images or similar images and from which hierarchical feature representations can be learned.
For example, if a convolutional neural network is used to process matrix region tiles, the computer system would first input the tiles into the network. The network processes and converts the data through a series of convolution layers, pooling layers and full connection layers, and finally outputs a vector with a fixed length as a representation vector of the block. This vector captures key information in the original data such that similar tiles have similar representations in vector space, while different tiles have larger differences. Finally, through the process of step S104, the computer system obtains a set of token vectors corresponding to the respective matrix area partitions. These vectors not only reduce the dimensionality of the data, but also retain sufficient information for subsequent analysis and identification.
It should be noted that, in the embodiment of the present application, if the above-mentioned intelligent anchor water accumulation detection method is implemented in the form of a software function module, and is sold or used as an independent product, the method may also be stored in a computer readable storage medium. Based on such understanding, the technical solution of the embodiments of the present application may be essentially or some of contributing to the related art may be embodied in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read Only Memory (ROM), a magnetic disk, an optical disk, or other various media capable of storing program codes. Thus, embodiments of the application are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.
The embodiment of the application provides a computer system, which comprises a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor executes the program to realize part or all of the steps of the method.
Fig. 2 is a schematic diagram of a hardware entity of a computer system according to an embodiment of the present application, as shown in fig. 2, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program executable on the processor 1001, the processor 1001 implementing the steps in the method of any of the embodiments described above when the program is executed.
The memory 1002 stores computer programs executable on the processor, the memory 1002 being configured to store instructions and applications executable by the processor 1001, and may also cache data (e.g., image data, audio data, voice communication data, and video communication data) to be processed or already processed by the respective modules in the processor 1001 and the computer system 1000, which may be implemented by a FLASH memory (FLASH) or a random access memory (Random Access Memory, RAM).
The steps of the intelligent anchor water detection method of any one of the above are implemented when the processor 1001 executes a program. The processor 1001 generally controls the overall operation of the computer system 1000.
Embodiments of the present application provide a computer storage medium storing one or more programs executable by one or more processors to implement the steps of the intelligent anchor water detection method of any of the embodiments above.
It should be noted here that: the description of the storage medium and apparatus embodiments above is similar to that of the method embodiments described above, with similar benefits as the method embodiments. For technical details not disclosed in the embodiments of the storage medium and the apparatus of the present application, please refer to the description of the method embodiments of the present application.
The foregoing is merely an embodiment of the present application, but the scope of the present application is not limited thereto, and any person skilled in the art can easily think about changes or substitutions within the technical scope of the present application, and the changes and substitutions are intended to be covered by the scope of the present application.
Claims (10)
1. An intelligent anchorage device ponding detection method, which is characterized by being applied to a computer system, comprising the following steps:
acquiring an anchor ultrasonic signal distribution tensor aiming at a bridge anchor, and extracting characterization vectors of each anchor ultrasonic signal matrix contained in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector;
Based on the distribution information strengthening component after debugging, respectively carrying out distribution information strengthening operation on the corresponding matrix characterization vectors according to the distribution conditions of each anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor to obtain corresponding strengthening characterization vectors containing the distribution conditions;
Respectively determining water accumulation condition prediction coefficients corresponding to the anchor ultrasonic signal matrixes according to the commonality evaluation coefficients between the reinforced characterization vectors and the contrast characterization vectors deployed in advance; each contrast characterization vector is obtained according to the extraction of ponding-free ultrasonic signal data, and each contrast characterization vector comprises the distribution condition of a corresponding anchor ultrasonic signal matrix, wherein the ponding-free ultrasonic signal data corresponds to a ponding-free bridge anchor;
determining anchor ponding identification information of the anchor ultrasonic signal distribution tensor according to the obtained ponding condition prediction coefficients;
the method for determining the water accumulation condition prediction coefficients respectively corresponding to the anchor ultrasonic signal matrixes according to the commonality evaluation coefficients between each strengthening characterization vector and the contrast characterization vector deployed in advance comprises the following steps:
For each strengthening characterization vector, the following processing is respectively carried out:
For one enhancement token vector, determining a collation token vector having a greatest common assessment coefficient with the one enhancement token vector;
Determining a ponding condition prediction coefficient of an anchor ultrasonic signal matrix corresponding to the one enhanced characterization vector according to the error between the contrast characterization vector and the one enhanced characterization vector;
Or determining the water accumulation condition prediction coefficients corresponding to the anchor ultrasonic signal matrixes respectively according to the commonality evaluation coefficients between the reinforced characterization vectors and the contrast characterization vectors deployed in advance, wherein the water accumulation condition prediction coefficients comprise:
For each strengthening characterization vector, the following processing is respectively carried out:
For one enhancement token vector, determining a collation token vector having a greatest common assessment coefficient with the one enhancement token vector;
determining a plurality of contrast token vectors with the commonality evaluation coefficients not larger than a preset commonality evaluation coefficient threshold value between the contrast token vectors with the largest commonality evaluation coefficients in the contrast token vectors deployed in advance;
And determining the ponding condition prediction coefficient of the anchor ultrasonic signal matrix corresponding to the one strengthening characterization vector according to the comparison characterization vector with the maximum commonality evaluation coefficient and errors between the plurality of comparison characterization vectors and the one strengthening characterization vector.
2. The method of claim 1, wherein before the debug completion based distribution information enhancement component performs a distribution information enhancement operation on the corresponding matrix characterization vector according to the distribution condition of each of the anchor ultrasonic signal matrices in the anchor ultrasonic signal distribution tensor, respectively, to obtain a corresponding enhanced characterization vector including the distribution condition, the method further comprises:
Repeatedly debugging the distributed information enhancement component to be debugged according to a plurality of anchor ultrasonic signal distribution tensor samples and the prior marks of the samples corresponding to the anchor ultrasonic signal distribution tensor samples respectively until the debugging cut-off requirement is met, and obtaining the distributed information enhancement component after the debugging is completed; wherein, each sample prior mark represents the confidence coefficient of each anchor ultrasonic signal matrix in the corresponding anchor ultrasonic signal distribution tensor sample;
wherein, in each round of debugging, the following operations are included:
For one anchor ultrasonic signal distribution tensor sample, extracting characterization vectors of each anchor ultrasonic signal matrix of the anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors;
The distribution information strengthening component is used for carrying out distribution information strengthening operation on the corresponding sample characterization vectors according to the distribution condition of each anchor ultrasonic signal matrix in one anchor ultrasonic signal distribution tensor sample so as to obtain corresponding strengthening characterization vectors containing the distribution condition;
According to the corresponding strengthening characterization vectors of each anchor ultrasonic signal matrix of the anchor ultrasonic signal distribution tensor sample, anchor ponding identification information of the anchor ultrasonic signal distribution tensor sample is obtained;
And optimizing the parameter of the distributed information strengthening component for the round of debugging according to the error between the obtained accumulated water identification information of the plurality of anchors and the corresponding sample prior marks.
3. The method of claim 1, wherein prior to separately determining respective corresponding water accumulation condition prediction coefficients for each of the anchor ultrasound signal matrices based on the respective enhanced characterization vector and the respective previously deployed contrast characterization vector, the method further comprises:
Based on a second characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix contained in each ponding-free ultrasonic signal data respectively to obtain a basic characterization matrix, wherein the basic characterization matrix comprises basic characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data;
Annotating a plurality of anchor ultrasonic signal distribution tensor samples respectively according to the basic representation matrix to obtain corresponding sample prior marks, and repeatedly debugging a ponding detection network according to the anchor ultrasonic signal distribution tensor samples and the corresponding sample prior marks respectively until the debugging cut-off requirement is met; the ponding detection network comprises a first characterization vector extraction component and the distribution information strengthening component, wherein the first characterization vector extraction component and the second characterization vector extraction component are twin components, and the same parameter is used;
According to the first characterization vector extraction component in the ponding detection network after debugging, respectively extracting characterization vectors of each ponding-free ultrasonic signal data, and carrying out distribution information strengthening operation on matrix characterization vectors obtained by extraction on the basis of the distribution information strengthening component in the ponding detection network after debugging, so as to obtain a comparison characterization vector set, wherein the comparison characterization vector set comprises the previously deployed comparison characterization vectors.
4. A method according to claim 3, wherein annotating each of the plurality of anchor ultrasound signal distribution tensor samples according to the base characterization matrix to obtain corresponding sample prior markers comprises:
for the plurality of anchor ultrasonic signal distribution tensor samples, respectively carrying out the following processing:
For one anchor ultrasonic signal distribution tensor sample, based on the second characterization vector extraction component, respectively extracting characterization vectors of each anchor ultrasonic signal matrix in the one anchor ultrasonic signal distribution tensor sample to obtain corresponding sample characterization vectors;
Determining a composition priori mark of the corresponding anchor ultrasonic signal matrix according to the obtained commonality evaluation coefficient between each sample characterization vector and the basic characterization matrix, wherein each composition priori mark characterizes the confidence degree of the corresponding anchor ultrasonic signal matrix as a ponding anchor ultrasonic signal matrix;
and obtaining the sample prior mark of the anchor ultrasonic signal distribution tensor sample according to the obtained each composition prior mark.
5. The method of claim 3, wherein the extracting the characterization vector from the first characterization vector extracting component in the water accumulation detection network after the debugging is completed, respectively extracting the characterization vector from each water accumulation-free ultrasonic signal data, and performing a distribution information strengthening operation on the matrix characterization vector obtained by the extracting based on the distribution information strengthening component in the water accumulation detection network after the debugging is completed, to obtain a comparison characterization vector set, and the method comprises:
According to the first characterization vector extraction component, performing characterization vector extraction on each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data respectively, and performing distribution information reinforcement operation on the extracted matrix characterization vector based on the distribution information reinforcement component in the ponding detection network after debugging is completed to obtain an iteration characterization vector set, wherein the iteration characterization vector set comprises iteration characterization vectors respectively corresponding to each anchor ultrasonic signal matrix of each ponding-free ultrasonic signal data;
And determining partial iterative characterization vectors in the iterative characterization vector set based on the commonality evaluation coefficients between every two iterative characterization vectors in the iterative characterization vector set, so as to obtain the comparison characterization vector set.
6. The method of claim 5, wherein the determining a portion of the iterative token vectors in the iterative token vector set based on the commonality assessment coefficients between each two iterative token vectors in the iterative token vector set, resulting in the collation token vector set, comprises:
Determining one or more iterative characterization vectors from the iterative characterization vector set, and constructing a basic set for the contrast characterization vector set;
Iteratively optimizing the set of control token vectors; wherein, in one round of optimization, the following processes are performed:
For each iteration characterization vector in the iteration characterization vector set, respectively determining a corresponding characterization vector doublet, wherein each characterization vector doublet comprises one iteration characterization vector and a comparison characterization vector with the largest common evaluation coefficient with the one iteration characterization vector in the comparison characterization vector set;
for each collation token vector contained in each token vector doublet, adding the iteration token vector with the smallest commonality evaluation coefficient with each collation token vector to the collation token vector set.
7. The method according to any one of claims 1 to 6, wherein based on the distribution information strengthening component after the debugging is completed, according to the distribution condition of each of the anchor ultrasonic signal matrices in the anchor ultrasonic signal distribution tensor, respectively performing a distribution information strengthening operation on the corresponding matrix characterization vector to obtain a corresponding strengthening characterization vector including the distribution condition, including:
Based on the distributed information strengthening assembly after debugging, the following treatments are respectively carried out on each anchor ultrasonic signal matrix:
for an anchor ultrasonic signal matrix, obtaining a corresponding sequence number representation vector according to the sequence number distribution condition of the anchor ultrasonic signal matrix in the anchor ultrasonic signal distribution tensor;
And obtaining a corresponding strengthening characterization vector according to the matrix characterization vector and the serial number characterization vector of the anchor ultrasonic signal matrix.
8. The method of any one of claims 1 to 6, wherein extracting the characterization vector of each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector, includes:
carrying out multi-round characterization vector extraction on the anchor ultrasonic signal distribution tensor; wherein each round performs the following processes:
According to a preset filtering matrix, filtering characterization vector extraction is carried out on the input of the current wheel characterization vector extraction, and a filtering characterization vector is obtained; wherein, when the first round of characterization vector extraction is performed, the input is the distribution tensor of the anchor ultrasonic signal, and when the other rounds of characterization vector extraction is performed, the input is the output of the previous round of characterization vector extraction;
Performing cross-layer identity connection according to the filtering characterization vector and the input to obtain a fusion characterization vector;
if the current wheel representation vector extraction is the last wheel representation vector extraction, taking the fusion representation vector as the output of the current wheel representation vector extraction;
and if the current wheel representation vector extraction is not the last wheel representation vector extraction, performing feature adaptive dimension reduction on the fusion representation vector, and taking the obtained dimension reduction representation vector as the output of the current wheel representation vector extraction.
9. The method of any one of claims 1 to 6, wherein extracting the characterization vector of each anchor ultrasonic signal matrix included in the anchor ultrasonic signal distribution tensor to obtain a corresponding matrix characterization vector, includes:
obtaining corresponding target reference bridge anchor data from a plurality of reference bridge anchor data which are deployed in advance according to the bridge tag of the bridge anchor;
According to preset anchor distribution information in the target reference bridge anchor data, carrying out data alignment on the anchor ultrasonic signal distribution tensor to obtain an alignment matrix;
according to a preset matrix area in the target reference bridge anchor data, obtaining a matrix area partition where the bridge anchor is located in the alignment matrix;
And extracting characterization vectors of each anchor ultrasonic signal matrix contained in the matrix region blocks to obtain corresponding matrix characterization vectors.
10. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, wherein the processor implements the steps of the method of any one of claims 1 to 9 when the program is executed.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410413990.5A CN118010848B (en) | 2024-04-08 | 2024-04-08 | Intelligent anchorage device ponding detection method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410413990.5A CN118010848B (en) | 2024-04-08 | 2024-04-08 | Intelligent anchorage device ponding detection method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN118010848A true CN118010848A (en) | 2024-05-10 |
CN118010848B CN118010848B (en) | 2024-06-11 |
Family
ID=90950218
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410413990.5A Active CN118010848B (en) | 2024-04-08 | 2024-04-08 | Intelligent anchorage device ponding detection method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN118010848B (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112629617A (en) * | 2020-12-15 | 2021-04-09 | 华中科技大学 | Method and system for detecting accumulated water in cable based on magnetostrictive guided waves |
CN113128473A (en) * | 2021-05-17 | 2021-07-16 | 哈尔滨商业大学 | Underground comprehensive pipe gallery-oriented inspection system, method, equipment and storage medium |
CN116773652A (en) * | 2023-04-20 | 2023-09-19 | 北京市政路桥管理养护集团有限公司 | Bridge modal parameter testing method and system |
CN116818892A (en) * | 2023-06-09 | 2023-09-29 | 中国电建集团贵阳勘测设计研究院有限公司 | Underground cavern side roof arch sound wave testing device |
CN117392859A (en) * | 2023-10-23 | 2024-01-12 | 四川遂广遂西高速公路有限责任公司 | Traffic safety management system and method based on expressway drainage structure |
-
2024
- 2024-04-08 CN CN202410413990.5A patent/CN118010848B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112629617A (en) * | 2020-12-15 | 2021-04-09 | 华中科技大学 | Method and system for detecting accumulated water in cable based on magnetostrictive guided waves |
CN113128473A (en) * | 2021-05-17 | 2021-07-16 | 哈尔滨商业大学 | Underground comprehensive pipe gallery-oriented inspection system, method, equipment and storage medium |
CN116773652A (en) * | 2023-04-20 | 2023-09-19 | 北京市政路桥管理养护集团有限公司 | Bridge modal parameter testing method and system |
CN116818892A (en) * | 2023-06-09 | 2023-09-29 | 中国电建集团贵阳勘测设计研究院有限公司 | Underground cavern side roof arch sound wave testing device |
CN117392859A (en) * | 2023-10-23 | 2024-01-12 | 四川遂广遂西高速公路有限责任公司 | Traffic safety management system and method based on expressway drainage structure |
Non-Patent Citations (2)
Title |
---|
WANG HUI 等: "Two-Dimensional FEM Analysis of Influencing Factors on Ultimate Pull-Out Capacity of Vertically Loaded Anchor", JOURNAL OF TIANJIN UNIVERSITY, vol. 43, no. 11, 15 November 2020 (2020-11-15), pages 964 - 70 * |
王翼超: "基于多元检测数据的大跨径桥梁索体系评估方法研究", 中国优秀硕士学位论文全文数据库工程科技Ⅱ辑, no. 01, 15 January 2024 (2024-01-15), pages 034 - 553 * |
Also Published As
Publication number | Publication date |
---|---|
CN118010848B (en) | 2024-06-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112784881B (en) | Network abnormal flow detection method, model and system | |
CN111785329B (en) | Single-cell RNA sequencing clustering method based on countermeasure automatic encoder | |
CN111753881B (en) | Concept sensitivity-based quantitative recognition defending method against attacks | |
CN111967343A (en) | Detection method based on simple neural network and extreme gradient lifting model fusion | |
CN110213244A (en) | A kind of network inbreak detection method based on space-time characteristic fusion | |
CN113297572B (en) | Deep learning sample-level anti-attack defense method and device based on neuron activation mode | |
Ferdous et al. | Machine learning approach towards satellite image classification | |
CN113569657A (en) | Pedestrian re-identification method, device, equipment and storage medium | |
CN117407772A (en) | Method and system for classifying training multi-element time sequence data by supervising and comparing learning network model | |
CN113591962B (en) | Network attack sample generation method and device | |
CN117349743A (en) | Data classification method and system of hypergraph neural network based on multi-mode data | |
Celik et al. | Change detection without difference image computation based on multiobjective cost function optimization | |
CN118010848B (en) | Intelligent anchorage device ponding detection method and system | |
CN115643153A (en) | Alarm correlation analysis method based on graph neural network | |
Gao et al. | Feature tracking for target identification in acoustic image sequences | |
CN112232329A (en) | Multi-core SVM training and alarming method, device and system for intrusion signal recognition | |
Trentin et al. | Unsupervised nonparametric density estimation: A neural network approach | |
CN115546521B (en) | Point matching method based on key point response constraint | |
CN116861175B (en) | Operation track correction method based on neural network | |
CN118070107B (en) | Deep learning-oriented network anomaly detection method, device, storage medium and equipment | |
CN117874521A (en) | Variable distribution stability judging method suitable for lifting tree model | |
Pandey et al. | Research Article A Metaheuristic Autoencoder Deep Learning Model for Intrusion Detector System | |
CN118312851A (en) | Die casting defect diagnosis method based on improved ant lion colony algorithm and efficient MLP model training | |
CN117909881A (en) | Fault diagnosis method and device for multi-source data fusion pumping unit | |
CN117375970A (en) | Network intrusion detection method, system and equipment for power system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |