CN114169374A - Cable-stayed bridge stay cable damage identification method and electronic equipment - Google Patents

Cable-stayed bridge stay cable damage identification method and electronic equipment Download PDF

Info

Publication number
CN114169374A
CN114169374A CN202111513368.4A CN202111513368A CN114169374A CN 114169374 A CN114169374 A CN 114169374A CN 202111513368 A CN202111513368 A CN 202111513368A CN 114169374 A CN114169374 A CN 114169374A
Authority
CN
China
Prior art keywords
cable
data
stay cable
stay
model
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
Application number
CN202111513368.4A
Other languages
Chinese (zh)
Other versions
CN114169374B (en
Inventor
陈建文
赵军产
王宇
张健
向浩楠
王敬童
董倩
姜永滚
欧阳彬
曾可涵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hunan University of Technology
Original Assignee
Hunan University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hunan University of Technology filed Critical Hunan University of Technology
Priority to CN202111513368.4A priority Critical patent/CN114169374B/en
Publication of CN114169374A publication Critical patent/CN114169374A/en
Application granted granted Critical
Publication of CN114169374B publication Critical patent/CN114169374B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Signal Processing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method for identifying damage of a stay cable of a cable-stayed bridge and electronic equipment, relates to the technical field of stay cable damage identification, and is used for solving the problem that the identification precision of the stay cable damage of the cable-stayed bridge in the prior art is not high enough due to the fact that the problem that large-scale working condition data are not classified and unbalanced in the prior art is not solved effectively, and the method comprises the following steps: collecting the stay cable data of the cable-stayed bridge; and (3) processing the stay cable data: performing characteristic engineering on the balanced stay cable data, and extracting the characteristics of the stay cable data; constructing a model, and combining the convolutional neural network with the long-short term memory network to form a one-dimensional convolutional long-short term memory model; and (3) optimizing the model, namely inputting the extracted stay cable data characteristics into the constructed model, and optimizing the model through a combination of characteristic engineering, a cross entropy loss function and a hyper-parameter. According to the technical scheme, the problem of unbalanced classification of large-scale working condition data can be effectively solved, so that the identification precision of the damage of the stay cable of the cable-stayed bridge can be improved.

Description

Cable-stayed bridge stay cable damage identification method and electronic equipment
Technical Field
The invention relates to the technical field of stay cable damage identification, in particular to a stay cable damage identification method for a cable-stayed bridge and electronic equipment.
Background
Traffic is an economic life line of society, and a bridge is an important junction of a traffic trunk line, so the health and safety of the bridge have great significance for the steady development of national economy, due to the continuous degradation of structural materials and the influence of environmental factors (such as wind load, temperature load and the like) and human factors (such as design, construction, management and the like), the bridge can generate defects and age during service, so that all parts of the structure are damaged before the design age, if the damage of the structure cannot be found in time and proper maintenance and repair are carried out, the damage is accumulated in the bridge structure, and the resistance of the bridge to normal load and environmental action is reduced, therefore, the service life of the bridge is shortened, and a large amount of capital of the current bridge investment is counted to be used for maintaining and updating the bridge, so that the monitoring and evaluation of the bridge structure by adopting the structure health monitoring technology is particularly important.
The material and the structural performance of the cable-stayed bridge can be degraded along with the lengthening of the service time in the service process of the cable-stayed bridge, if the cable-stayed bridge generates structural damage, the structural damage can seriously affect the life health and social economy of people, so that the unscheduled maintenance and detection of the cable-stayed bridge in the service process of the bridge are necessary, the cable-stayed bridge mainly comprises a pylon, a girder and a stay cable, wherein the stay cable is used as a connecting medium of the pylon and the girder and becomes a key component in the whole system of the cable-stayed bridge, and the stay cable is used as a medium between the pylon and the girder and bears the most of the force of the cable-stayed bridge system, so that the health of the stay cable can be reflected by the force born by the stay cable, compared with the pylon and the girder, the stay cable is more easily damaged structurally, and the efficient and accurate damage condition identification of the stay cable is very important.
The traditional damage identification problem takes a related intelligent algorithm as a reference learning device, such as a genetic algorithm, a particle swarm algorithm and the like, meanwhile, a sensor technology is widely applied to the technical field of engineering, the large-scale data storage device, the improvement of computer performance and the development of artificial intelligence jointly promote the acquisition of large-scale engineering data and the application of a machine learning theory, so that how to combine the machine learning theory with the cable-stayed bridge working condition identification problem from the data mining angle is very worthy of research by combining the machine learning working condition identification algorithm for identifying the cable-stayed cable damage working condition.
The stay cable in the large-span cable-stayed bridge system is likely to be influenced by other factors such as time, temperature, wind speed and the like in the using process, the stay cable needs to be periodically detected to ensure the normal operation of the cable-stayed bridge system, the existing stay cable detection method has the problems of low efficiency, low safety coefficient, excessively low learning convergence speed and low learning efficiency, the stay cable working condition identification algorithm belongs to the typical abnormal detection problem and has the difficulty of unbalanced data distribution, and the existing cable-stayed bridge damage identification algorithm is mainly researched for unbalanced learning in the fields of data processing, algorithm design, deep learning theoretical application, hyper-parameter optimization, model evaluation indexes and the like.
The method mainly has the following problems in the prior art for identifying the damage of the stay cable:
firstly, the format of an unbalanced learning research data object of the stay cable damage identification problem needs to be expanded, the existing algorithm is concentrated on researching the unbalanced learning problem under a certain data format, the detection of a resampling technology under various types of data is lacked, and the existing research data volume is not enough to verify the stability of the unbalanced learning under the resampling technology;
secondly, the design of an embedded algorithm of an unbalanced learning theory for identifying the damage of the stay cable needs to be supplemented, the resampling technical process and the subsequent classification learning process in the existing unbalanced learning frame are mostly independent individuals, the overall unbalanced learning frame is easy to split, the existing embedded resampling algorithm has the problems of fragmentation, heavy data processing, light theory deduction and the like, and the embedded design of a damage identification system is lacked;
and thirdly, optimizing a hyper-parameter set of the stay cable damage identification classification algorithm based on the unbalanced learning theory. Most of the existing unbalanced learning technologies lack constraint conditions, the specific optimization direction of the technology cannot be judged, and the final model performance is influenced by hyper-parameters in an integral learning framework;
and fourthly, reconstructing an evaluation system of the stay cable damage identification algorithm based on the unbalanced learning. The existing evaluation index system for the unbalanced learning is single, and most of the existing evaluation index systems for the unbalanced learning use G-means as the evaluation index for the unbalanced learning. And constructing an inclined cable working condition learning evaluation system under the comprehensive unbalanced state by integrating the characteristics of abnormal detection problems and unbalanced learning and specific research data types.
Generally speaking, the data of the damage working condition of the stay cable is far lower than the data of the healthy working condition, and the problem of unbalanced classification of the data of the large working condition is not effectively solved in the prior art, so that the identification precision of the damage of the stay cable of the cable-stayed bridge in the prior art is not high enough.
Disclosure of Invention
The invention aims to: the invention provides a method for identifying damage of a stay cable of a cable-stayed bridge and electronic equipment, aiming at solving the problem that the identification precision of the prior art for the damage of the stay cable of the cable-stayed bridge is not high enough because the problem of unbalanced classification of large-scale working condition data is not effectively solved in the prior art.
The invention specifically adopts the following technical scheme for realizing the purpose:
a method for identifying damage of a stay cable of a cable-stayed bridge comprises the following steps:
s1, collecting the stayed-cable data of the cable-stayed bridge: acquiring acceleration data of each stay cable through a plurality of sensors, and simulating working condition data of the stay cable under different working condition states by using ANSYS finite element analysis software according to the acquired acceleration data;
the balanced stay cable data can be processed in the following mode in the characteristic engineering, and the method comprises the following steps:
dividing the stay cable data into damaged data and undamaged data;
respectively calculating the similarity between the characteristic points of each stay cable corresponding to the damaged data and the undamaged data to obtain a similar matrix, and establishing an adjacent matrix through the similar matrix and a threshold value of a coefficient in the similar matrix;
analyzing the relation among the characteristic points of the stay cables, selecting a threshold value of a coefficient in a similar matrix to respectively generate a minimum spanning tree and a node graph, if the similarity of the characteristic points of the two stay cables is greater than the threshold value, connecting the characteristic points of the two stay cables, and representing the characteristic points of the two stay cables by 1 in an adjacent matrix; if the similarity of the characteristic points of the two stay cables is smaller than the threshold value, the characteristic points of the two stay cables are not connected and are represented by 0 in the adjacent matrix;
respectively calculating the node degrees of the damaged data and the undamaged data by using the generated minimum spanning tree;
respectively calculating the node degree of the damaged data and the node degree of the undamaged data;
averaging the variation values of the node degrees of the damaged data and the node degrees of the undamaged data respectively to obtain an average value of the node degree variation, sequencing the node degree variation from large to small, and extracting the feature importance of the stay cable feature points with the node degree variation exceeding the average value.
Putting undamaged data into a set A, putting damaged data into a set B, and calculating a similar matrix of the characteristic points of the stay cables of the damaged data in the set A and a similar matrix of the characteristic points of the stay cables of the damaged data in the set B, wherein the coefficient of the similar matrix between the characteristic points of the stay cables is calculated as follows:
Figure BDA0003403333050000041
wherein
Figure BDA0003403333050000042
m represents the node data of the characteristic point of the nth 1 stay cables in the mth sample,
Figure BDA0003403333050000043
the sample mean value of the characteristic point of the nth 1 stay cables is shown,
Figure BDA0003403333050000044
the coefficient of the similarity matrix between the stay cable characteristic point n1 and the stay cable characteristic point n2 is shown, and M represents the number of samples.
And S2, processing the stay cable data: resampling the acquired unbalanced stay cable data to convert the unbalanced stay cable data into balanced stay cable data, and performing characteristic engineering on the balanced stay cable data to extract the characteristics of the stay cable data;
s3, constructing a model, and combining the convolutional neural network with the long-short term memory network to form a one-dimensional convolutional long-short term memory model;
the model construction comprises the following steps:
repeatedly compressing and extracting the stay cable characteristics through the convolution pooling operation of the one-dimensional convolution neural network to obtain characteristic information;
and learning and identifying the feature information after the convolution pooling through a long-term and short-term memory network to construct a model.
The convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the method comprises the following steps that stay cable data enter an input layer after being subjected to standardization processing, feature compression is carried out on a convolution layer after being processed, and convolution operation is carried out on the obtained stay cable data through a convolution kernel;
after the stay cable data is subjected to convolution operation, performing pooling operation to extract characteristic information in the stay cable data;
performing Flatten expansion on the stay cable data after convolution pooling, wherein the Flatten expansion is used as the input of a full connection layer, and the full connection layer integrates the local information after convolution pooling;
after convolution pooling and processing of the full connection layer, a logic value of the working condition state is obtained, the working condition state is divided by the logic value, and a network output layer is constructed through a softmax function.
The output of each layer in the convolutional neural network is set as the convolution result of a plurality of stayed-cable characteristics, and the output signal after the convolution operation is as follows:
Figure BDA0003403333050000051
wherein the content of the first and second substances,
Figure BDA0003403333050000052
represents the weights of the/layer ith order convolution kernels,
Figure BDA0003403333050000053
expressed as the offset, x, of the convolverl(j) Represents the jth local area of the l layer,
Figure BDA0003403333050000054
representing the output signal after the convolution operation;
extracting feature information using a max-pooling operation, the max-pooling using a maximum value of the pooled region as the extracted target information, expressed as follows:
Figure BDA0003403333050000055
wherein the content of the first and second substances,
Figure BDA0003403333050000056
represents the ith characteristic value in t neurons of the l layer, and t epsilon[(j-1)W，jW]W is the width of the pooling zone,
Figure BDA0003403333050000057
represents the value of (l +1) layer neurons;
the expression of the softmax function is:
Figure BDA0003403333050000058
wherein z isp(j) A logic value of the j-th neuron of the output layer, M represents the total number of classes, zp(M) denotes the logic value of the mth neuron in the output layer, j ═ 1, 2.
S4: and (3) optimizing the model, namely inputting the extracted data characteristics of the stay cable into the constructed model, and optimizing the model through characteristic engineering, a cross entropy loss function and a hyper-parameter set until the best model is trained.
In the model optimization, the model is optimized through characteristic engineering, and the method comprises the following steps:
constructing new characteristics, and performing characteristic engineering exploration by using polynomial characteristic conversion preprocessed in a sklern module, wherein the expression of the processing of the stay cable characteristic points is as follows:
Figure BDA0003403333050000061
in the formula, the highest order of the transformation process of the characteristic points of the stay cables is limited to 2, and the newly generated characteristic matrix is formed by combining polynomials of all the characteristic points of the stay cables with the order less than or equal to 2, wherein x'1，x′n，x′1，
Figure BDA0003403333050000062
All represent the characteristic points of the stay cables;
and extracting features, namely extracting the features of the stay cable feature points through correlation analysis and variance analysis.
In the model optimization, the optimization of the model through the cross entropy loss function comprises the following steps:
introduction of an imbalance weight factor alpha in a cross entropy loss functiontThe cross entropy loss function with the weighting factor is:
C(pt)＝-αtlog(pt)，t∈[0，1]
wherein p istFor the prediction probability, the subscript t represents the sample class, 0 is the majority class, and 1 is the minority class;
adding a modulation factor (1-p) to the cross-entropy loss function to which the weight factor is addedt)γNamely:
Lf＝-αt(1-pt)γlog(pt)
when a sample is misclassified or predicted probability ptWhen the modulation factor approaches 1, the cross entropy loss function is not affected, and the probability p is predictedtWhen approaching 1, the modulation factor tends to 0, LfRepresenting the cross entropy loss function of the added modulation factor.
In the model optimization, the optimization of the model through the hyper-parameters comprises the following steps:
defining a hyper-parameter: defining four hyper-parameter objects of a one-dimensional convolutional neural network, a long-term and short-term memory network, a full connection layer and an optimizer;
acquiring a hyper-parameter: using an evolution strategy in a superactive module as an optimization algorithm, using a Focal Loss function as a constraint, setting the mutation rate to be 0.5, the cross rate to be 0.5 and the evolution times to be 100, and finally obtaining the optimal hyper-parameters of the one-dimensional convolutional neural network, the long-short term memory network, the full-link layer and the optimizer;
and introducing the model for retraining, adding batch standardization items into the model to obtain a loss function under the optimal hyper-parameter set and the change trend of the model precision, and finally obtaining the optimal model.
An electronic device comprising storage and one or more processors;
storage means for storing one or more programs;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement a cable-stayed bridge cable damage identification method as described in an embodiment.
The invention has the following beneficial effects:
according to the method, the unbalanced stay cable data are converted into balanced stay cable data through resampling, then the balanced stay cable data are subjected to feature extraction, and finally the model is optimized through feature engineering, a cross entropy loss function and a hyper-parameter set combination to obtain an optimal model, so that the problem of large-scale working condition data classification unbalance can be effectively solved, and the identification precision of the stay cable damage of the cable-stayed bridge can be improved.
Drawings
FIG. 1 is a flow chart of a method for identifying damage to a stay cable of a cable-stayed bridge according to the present invention;
FIG. 2 is a schematic structural diagram of a subject of the invention;
FIG. 3 is a schematic diagram of a one-dimensional convolution long and short term memory model according to the present invention;
FIG. 4 is a simplified graph of the trend of loss and model accuracy using a one-dimensional convolutional neural network in accordance with the present invention;
FIG. 5 is a simplified training trend for the present invention using an unoptimized one-dimensional convolution long and short term memory model;
FIG. 6 is a block diagram of the model optimization of the present invention;
FIG. 7 is a schematic diagram of loss and accuracy trend of the optimal model of the present invention;
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Example 1
As shown in fig. 1, the present embodiment provides a method for identifying damage to a stay cable of a cable-stayed bridge, including the following steps:
s1, collecting the stayed-cable data of the cable-stayed bridge: acquiring acceleration data of each stay cable through a plurality of sensors, and simulating working condition data of the stay cable under different working condition states by using ANSYS finite element analysis software according to the acquired acceleration data;
the working condition data under different working condition states refer to stay cable data with different damage degrees, the stay cable working condition data are derived from ANSYS finite element analysis software, and a data set under different working condition states is simulated by constructing a real cable-stayed bridge finite element model. For example, fig. 2 shows a constructed cable-stayed bridge unit, which is symmetric to the left and right in front and rear directions as shown in fig. 2, and therefore 1/4 is taken as a research object, cables of the research object are coded according to 1-27, 12 sensors are arranged on each cable to collect acceleration data of x-axis, y-axis and z-axis of the cable, data nodes are coded into node01-node11, wherein node0 represents a distance of 2 meters along the cable direction near a main beam, and the rest nodes represent positions 1/14, 1/12, 1/10, 1/8, 1/6, 3/14, 1/4, 3/10, 5/12 and 1/2 of the cable (0 point on one side of a bridge deck), and interference factors such as measurement errors, temperature loads and wind loads which may exist in reality are considered during a numerical simulation process, therefore, white noise excitation is introduced in the simulation process, the white noise excitation in each working condition state is different, after the numerical simulation process is completed, a pandas module is used for data preprocessing, all working condition data are integrated, the processed stay cable data are characterized by z-axis acceleration in each node, the data labels are whether the stay cable is damaged (damage) and the damage degree (level), if the stay cable is damaged, the stay cable data are marked as 1, and the stay cable data are marked according to the corresponding damage level, the experiment only relates to the stay cable data characteristics and whether the stay cable is damaged, and therefore the problem of binary classification can be considered.
The feature engineering is a process of extracting features from original data, the features can well describe the data, performance of a model established by the features on unknown data can be optimal (or close to optimal), the processing flow of the feature engineering is to firstly remove useless features, then remove redundant features such as collinear features, generate new features by using existing features, conversion features, features in contents and other data sources, then convert the features (digitalization, category conversion, normalization and the like), and finally process the features (abnormal values, maximum values, minimum values, missing values and the like) so as to meet the use of the model.
Specifically, the balanced stay cable data can be processed in the following way in the characteristic engineering, and the method comprises the following steps:
dividing the stay cable data into damaged data and undamaged data;
respectively calculating the similarity between the characteristic points of each stay cable corresponding to the damaged data and the undamaged data to obtain a similar matrix, and establishing an adjacent matrix through the similar matrix and a threshold value of a coefficient in the similar matrix;
taking the nodes of all sensors of a primary sample as one node of the network, for example, 12 nodes per cable, 1/4 of a cable-stayed bridge total 27 cables, 12 × 27 ═ 324 nodes, putting all samples of each node into a one-dimensional vector, and obtaining a vector sample:
n1＝[24，15，43，........,79]
where n1 denotes the 1 st node, the number of samples X is related to the data sampled; putting undamaged data into a set A, putting damaged data into a set B, and calculating a similar matrix of the characteristic points of the stay cables of the damaged data in the set A and a similar matrix of the characteristic points of the stay cables of the damaged data in the set B, wherein the coefficient of the similar matrix between the characteristic points of the stay cables is calculated as follows:
Figure BDA0003403333050000091
wherein
Figure BDA0003403333050000092
m represents the node data of the characteristic point of the nth 1 stay cables in the mth sample,
Figure BDA0003403333050000093
indicates the nth 1 diagonal drawRetrieving the sample mean of the feature points,
Figure BDA0003403333050000094
the coefficient of the similarity matrix between the stay cable characteristic point n1 and the stay cable characteristic point n2 is shown, and M represents the number of samples.
Analyzing the relation among the characteristic points of the stay cables, selecting a threshold value of a coefficient in a similar matrix to respectively generate a minimum spanning tree and a node graph, if the similarity of the characteristic points of the two stay cables is greater than the threshold value, connecting the characteristic points of the two stay cables, and representing the characteristic points of the two stay cables by 1 in an adjacent matrix; if the similarity of the characteristic points of the two stay cables is smaller than the threshold value, the characteristic points of the two stay cables are not connected and are represented by 0 in the adjacent matrix;
respectively calculating the node degrees of the damaged data and the undamaged data by using the generated minimum spanning tree; the node graph can be obtained by using the adjacency matrix, the minimum spanning tree is generated by using a Kruskal algorithm, the node degree refers to how many edges connected with the nodes are, for example, node n1 has 7 nodes connected with it, and the node degree is 7.
Respectively calculating the change values of the node degrees of the damaged data and the node degrees of the undamaged data, namely, making a difference with the node degree sequences of two adjacent networks, and finding out the change of each node degree;
averaging the variation values of the node degrees of the damaged data and the node degrees of the undamaged data respectively to obtain an average value of the node degree variation, sequencing the node degree variation from large to small, and extracting the feature importance of the stay cable feature points with the node degree variation exceeding the average value.
And S2, processing the stay cable data: resampling the acquired unbalanced stay cable data to convert the unbalanced stay cable data into balanced stay cable data, and performing characteristic engineering on the balanced stay cable data to extract the characteristics of the stay cable data;
the sampling technology can be divided into two types of oversampling technology and undersampling technology, wherein the undersampling technology mainly processes a plurality of types of sample sets to balance the majority of sample sets with a few types of sample sets, the oversampling technology balances a data set by processing a few types of samples, in order to compare the concrete performances of different sampling technologies in identifying the damage of the stay cable, the invention uses sampling technologies such as SMOTE, random sampling, NearMiss and the like to balance the data set, evaluates the performance of the balanced data set by a reference classifier, tests a stay cable working condition identification model by using model evaluation indexes in an unbalanced learning problem, the undersampling technology has more advantages compared with the oversampling technology, the oversampling technology needs to generate a few types of sample families and can influence the authenticity of data to a certain extent, so the application selects the NearMiss undersampling technology to process the stay cable data set on the basis of combining a K nearest neighbor algorithm, after sampling is completed, the sample set is divided into a training set and a test set according to the proportion, for example, the training set and the test set can be divided into 8: 2.
S3, constructing a model, and combining the convolutional neural network with the long-short term memory network to form a one-dimensional convolutional long-short term memory model;
the model construction comprises the following steps:
as shown in fig. 3, the stay cable features are repeatedly compressed and extracted through the convolution pooling operation of the one-dimensional convolution neural network (1D-CNN) to obtain feature information;
learning and identifying the feature information after the convolution pooling through a long-short term memory network (LSTM) to construct a one-dimensional convolution long-short term memory model (1D-CNN-LSTM);
the convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the method comprises the steps that stay cable data enter an input layer after standardized processing, feature compression is carried out on a convolutional layer after the stay cable data are processed, convolution operation is carried out on the obtained stay cable data through a convolution kernel, certain layering is achieved, feature extraction is carried out on a bottom layer network, a high-rise network is responsible for extracting more complex and abstract relevant features, the convolution neural network can automatically carry out feature compression and extraction through convolution pooling operation, a corresponding recognition model is established focusing on local feature information, and the calculation rate can be greatly improved.
Firstly, introducing an activation function to perform nonlinear conversion, setting the output of each layer in a convolutional neural network as a convolution result of a plurality of stay cable characteristics, wherein the output signal after the convolution operation is as follows:
Figure BDA0003403333050000111
wherein the content of the first and second substances,
Figure BDA0003403333050000112
represents the weights of the/layer ith order convolution kernels,
Figure BDA0003403333050000113
expressed as the offset, x, of the convolverl(j) Represents the jth local area of the l layer,
Figure BDA0003403333050000114
representing the output signal after the convolution operation;
after the convolution signals of the stay cables are obtained, the signals are subjected to nonlinear transformation by the activation function, so that the difference between the characteristic information can be amplified, the deep learning model is troubled by the problems of gradient explosion and gradient disappearance, the phenomenon can be relieved by activating by using the relu function, and the time complexity of the algorithm is effectively reduced.
After the stay cable data are subjected to convolution operation, pooling operation is performed to extract characteristic information in the stay cable data, a large sample working condition matrix is converted into a small sample working condition matrix, the calculated amount is reduced, overfitting can be effectively avoided, the characteristic information is extracted by utilizing the maximum pooling operation, the maximum pooling area is used as the extracted target information, and the expression is as follows:
Figure BDA0003403333050000115
wherein the content of the first and second substances,
Figure BDA0003403333050000116
represents the ith characteristic value in t neurons of the l layers, and t is the [ (j-1) W, jW]W is the width of the pooling zone,
Figure BDA0003403333050000117
represents the value of (l +1) layer neurons; the maximum pooling operation extracts a sensitive characteristic value set, important characteristic information can be automatically extracted after the convolution pooling operation, and the operation can be repeatedly carried out to realize the compression of data dimensionality when the data dimensionality is too high.
Performing Flatten expansion on the stay cable data after convolution pooling, wherein the Flatten expansion is used as the input of a full connection layer, and the full connection layer integrates the local information after convolution pooling; it can be understood as a traditional neural network model:
Figure BDA0003403333050000121
Figure BDA0003403333050000122
represents a weight matrix between the t-th neuron corresponding to the l-layer characteristic i and the j-th neuron of the l +1 layer, zl+1(j) A logic value representing the j-th neuron of the l +1 layer; bjRepresenting a bias coefficient;
Figure BDA0003403333050000123
for the output value of the neuron in the previous layer, f takes a relu function.
After convolution pooling and processing of the full connection layer, a logic value of the working condition state is obtained, the working condition state is divided by the logic value, and a network output layer is constructed through a softmax function.
The expression of the softmax function is:
Figure BDA0003403333050000124
wherein z isp(j) A logic value of the j-th neuron of the output layer, M represents the total number of classes, zp(M) denotes the logic value of the mth neuron in the output layer, and j is 1, 2, … M.
The 1D-CNN only executes one-dimensional convolution, can be suitable for the real-time fault detection of the stay cable, has the capability of working after manual feature extraction, and most importantly can provide stable identification capability under the conditions of limited size of a training set and limited back propagation times; the learning framework of the 1D-CNN-LSTM is used as the extension of the 1D-CNN, namely an LSTM unit structure is added between the last pooling layer and the full connection layer, feedforward in a network and dynamic interaction of internal information are selectively realized through the LSTM, the LSTM is used as a special circulating neural network structure, not only can the overfitting phenomenon generated by the network structure be relieved, but also the risks of gradient explosion and gradient disappearance can be avoided to a certain extent, the LSTM is embedded into the 1D-CNN network structure, the whole learning framework carries out extraction on characteristic information of a stay cable through multiple convolution pooling operations, and the LSTM unit structure is embedded between the pooling layer and the full connection layer, and the specific whole network structure is shown in figure 3; as can be seen from fig. 3, the convolutional layer and the pooling layer exist in the overall network structure, the dotted line connection between the units indicates that the pooling window is 2, the step length is 1, there are multiple convolutional pooling operations to compress and extract the characteristic data of the stay cable, the characteristic information array format obtained after multiple convolutional pooling operations is to be reshaped, the reshaped information is input to the recurrent neural network part, it is worth noting that multiple LSTM units also exist in the overall network, the working condition recognition algorithm capability is improved by learning the pre-and post-dependency relationship, the information output by the recurrent network part is to be unfolded by Flatten, then the information enters the fully-connected layer part to be learned, the fully-connected layer gradually distinguishes the characteristic information corresponding to the stay cable working condition through the activation function, and is connected with the output layer of the overall network, specifically:
zl＝wl·zl-1+bl
and processing the output information of the network structure by adopting softmax, and outputting stay cable working condition prediction information after receiving the softmax processing by the characteristic information processed by the full connection layer.
After the model is built, a test is carried out on the built model, and the specific test is as follows:
A1D-CNN (one-dimensional convolutional neural network) model is built through keras for training, a binary cross entropy function used by a loss function is used, an optimizer is adam, a learning rate parameter is default, and the built CNN network parameter is shown in table 1.
TABLE 11D-CNN convolution pooling partial parameters
Figure BDA0003403333050000131
Table 1 shows convolution pooling parameters used in the experiment, a Training batch of the model is 100, a regularization term and an early stopping method are not used in the Training process, and a variation trend of the overall Accuracy of the Loss function and the model is finally obtained, fig. 4 shows a variation trend of the Loss function and the model Accuracy of 1D-CNN in the Training process, in fig. 4, significance/Test Loss represents verification/Test set Loss, Training Loss represents Training set Loss, significance/Test Accuracy represents verification/Test set Accuracy, Training Accuracy represents Training set Accuracy, and Epochs represents the number of iterations, so that it can be known that there are two downward trends in the overall Training process, the first time is 0-20, and the second time is 20-100. After training, stabilizing the loss functions of the training set and the test set to be about 0.35, wherein the overall accuracy of the training set and the test set is less than 0.86, carrying out inspection on the test set, and carrying out detailed evaluation by using a confusion matrix and a classification learning report, wherein TP, FP, TN and FN in the confusion matrix are 245628, 65853, 204476 and 24701 respectively; table 2 shows the classification learning report, using the evaluation indexes of precision, recall and f1 score, wherein the precision can reflect the correct prediction ability of the model for each class, the recall measures how many samples are classified correctly, and the f1 score is the harmonic mean of precision and recall.
TABLE 21D-CNN Classification learning report
Label (R) Precision ratio Recall rate f1 fraction Number of samples
Healthy condition 0.79 0.91 0.84 270329
Damage condition 0.89 0.76 0.82 270329
From table 2, it can be seen that the data sets of the healthy condition and the damaged condition have reached a balance, the precision ratio of the healthy condition is 0.79, the recall ratio is 0.91, and the f1 score is 0.84; the precision of the damage condition is 0.89, the recall ratio is 0.76, the f1 score is 0.82, from the prediction accuracy of the traditional evaluation index, the recognition capability of the 1D-CNN model to the damage condition is better than that of the healthy condition, but the recall ratio result is not ideal at the moment, and in the imbalance problem, the traditional classification evaluation index may not be applicable, so Kubat and the like propose an evaluation index (g-mean) specially used for solving the imbalance condition, the index not only pays attention to the recall ratio condition of a minority class and a majority class, but also considers the specificity condition (TNR) in the actual classification, and the index is expressed by a mathematical formula as follows:
Figure BDA0003403333050000141
besides the g-mean evaluation index, ROC curves and AUC (false positive rate) are introduced to perform comprehensive measurement of the model, wherein the ROC curves are all called a receiver working characteristic curve and are composed of TPR (true positive rate) and FPR (false positive rate), the optimal model in the ROC curves is represented by TPR (thermal processing temperature value) 1 and FPR (false positive rate) 0, the ROC curve is infinitely close to the upper left corner, the AUC is the area enclosed by the ROC curve and the lower region, the greater the AUC value is, the better the model phenotype is, the basic evaluation of the model has been performed in the classification learning report, and the comprehensive evaluation of the model is performed through g-mean and AUC, the g-mean is 0.8548 and the AUC value is 0.8325.
Then, continuing to perform the test by using the constructed 1D-CNN-LSTM (one-dimensional convolution long-short term memory model), wherein the parameters of the convolution pooling operation part are consistent with the parameters of the 1D-CNN, the number of LSTM units of the recurrent neural network part is 4, and the parameters of each unit can be seen in a table 3; in the experiment, a 1D-CNN-LSTM network is built according to the parameters in the table 1 and the table 3, and then the training of the recognition algorithm is carried out, and finally the training result is obtained.
TABLE 3LSTM cell parameters
Unit cell Activating a function Whether to return a sequence
LSTM1 128 tanh TRUE
LSTM2 256 tanh TRUE
LSTM3 256 tanh TRUE
LSTM4 512 tanh TRUE
FIG. 5 is a Loss function and a model precision variation trend of 1D-CNN-LSTM in the Training process, wherein Validation/Test Loss represents verification/Test set Loss, Training Loss represents Training set Loss, Validation/Test Accuracy represents verification/Test set precision, Training Accuracy represents Training set precision, Epochs represents iteration times, and it can be known that the Loss function of the Training set is stabilized at about 0.32 and the overall precision of the Training set is stabilized at about 0.86 in FIG. 5; the loss function of the test set floats between 0.32 and 0.70, the overall precision floats in the area of 0.50 to 0.86, the overall training precision of the model is 0.8650 after training is finished, and the loss of the training set is 0.3262; the overall accuracy of the test set was 0.8567 with a loss of 0.3474; the results in fig. 5 show that the 1D-CNN-LSTM recognition model has faster convergence speed, and does not have the multiple convergence phenomenon in fig. 4, and is evaluated by the classification learning report and the confusion matrix, where TP, FP, TN, and FN of the confusion matrix are 232533, 59705, 210624, and 17796, respectively, and compared with the confusion matrix in the 1D-CNN state, the number of TPs and TNs that are predicted correctly is increased, and there is a certain decrease in FP and FN that are predicted incorrectly, as shown in the 1D-CNN-LSTM classification learning report in table 4.
TABLE 41D-CNN-LSTM Classification learning report
Label (R) Precision ratio Recall rate f1 fraction Number of samples
Healthy condition 0.81 0.93 0.87 270329
Damage condition 0.91 0.78 0.84 270329
Table 4 classification learning reports accuracy, recall, f1 scores for health conditions were raised to 0.81, 0.93, and 0.87, respectively; the precision rate, the recall rate and the f1 score of the damage working condition are respectively increased to 0.91, 0.78 and 0.84, and a g-mean and an AUC evaluation model are further used to know that the g-mean is increased from 0.8548 to 0.8703 and the AUC is also increased from 0.8325 to 0.8566 at the moment, and the result of each evaluation index in the evaluation system is better than that of a single CNN model according to the result of a comprehensive evaluation system, so that the 1D-CNN-LSTM can improve the stay cable working condition recognition capability to a certain extent, but needs to be improved, and further training and optimization of the models are needed.
S4: and (3) optimizing the model, inputting the extracted data characteristics of the stay cable into the constructed model, optimizing the model through a characteristic engineering, a cross entropy loss function and a hyper-parameter set, setting the training epoch as 100, adding an L1 regularization item, and monitoring the loss of the test set by using an early-stop method until the best model is trained.
As shown in fig. 6, in the model optimization, the optimization of the model by the feature engineering includes the following steps:
the new characteristics are constructed, the imbalance problem is explored from the data angle, certain learning difficulty can be relieved, and the characteristic engineering can also play a certain influence role on the experimental result. Aiming at the characteristic data of the stay cable, the characteristic engineering exploration is carried out by using polynomial characteristic conversion preprocessed in a sklern module in an experiment, and the expression of the stay cable characteristic point processing is as follows:
Figure BDA0003403333050000161
in the formula, the highest order limit of the transformation process of the stay cable characteristic points is 2, the newly generated characteristic matrix is formed by combining all the stay cable characteristic point polynomials which are less than or equal to the order 2, the order is not too high in the process, and otherwise, an overfitting phenomenon can be caused; wherein x'1，x′n，x1，
Figure BDA0003403333050000162
All represent the characteristic points of the stay cables; for example, the characteristic dimension of the data of the stay cable is 12, the data is composed of the z-direction acceleration of 12 sensors on the stay cable, and the characteristic dimension is changed into 91 after polynomial conversion.
Extracting features, wherein the dimension of the converted polynomial features is far higher than the original dimension, so that the feature extraction is carried out on the stay cable feature points through correlation analysis and variance analysis; the extracted features can be used for establishing a stay cable damage identification model, the stay cable damage identification model is established based on the extracted features, the overall precision of the model is taken as an example, no processing is carried out after the polynomial change, the effect of directly carrying out training is poor, and the precision of a test set is about 0.5; after the features are extracted through analysis of variance, the precision of the training set and the precision of the testing set respectively reach 0.8584 and 0.8497; after the characteristics with high autocorrelation are eliminated through correlation analysis, the precision of the training set and the precision of the testing set respectively reach 0.8681 and 0.8624, and the stay cable polynomial characteristic set is extracted by using the correlation coefficient, so that a relatively ideal effect is achieved, but the obtained result is inferior to the recognition model established by the basic characteristics.
TABLE 5 polynomial change Classification learning report
Figure BDA0003403333050000171
As shown in table 5, a polynomial change classification learning report is recorded, and table 5 shows that the identification algorithm test set established by the above three feature sets shows that the precision, recall and f1 scores of the health condition after the correlation coefficient processing are 0.80, 0.90 and 0.83, the damage condition is 0.89, 0.76 and 0.83 respectively, and the g-mean and AUC are 0.8498 and 0.8415 respectively; however, combining the results of Table 5, building a new set of features using polynomial transformations does not result in an optimal model, and therefore the model continues to be optimized.
In the model optimization, the optimization of the model through the cross entropy loss function comprises the following steps:
the Focal loss function is often used for an extremely unbalanced scene, the loss function is evolved based on cross entropy, and when the unbalanced problem is solved, an unbalanced weight factor alpha needs to be introducedtThe imbalance ratio is defined as the minority class samples divided by the majority class samples, and the imbalance ratio can be generally used as a weighting factor of the weight cross entropy, so that the cross entropy loss function with the weighting factor can be obtained as:
C(pt)＝-αtlog(pt)，t∈[0，1]
wherein p istFor the prediction probability, the subscript t represents the sample class, 0 is the majority class, and 1 is the minority class;
cross entropy loss upon addition of weight factorsAdding a modulation factor (1-p) to the loss functiont)γNamely:
Lf＝-αt(1-pt)γlog(pt)
when a sample is misclassified or predicted probability ptWhen the modulation factor approaches 1, the cross entropy loss function is not affected, and the probability p is predictedtWhen approaching 1, the modulation factor tends to 0, LfRepresents a cross entropy loss function of the added modulation factor; the sample weight loss for better classification performance will be reduced. Secondly, the value of gamma will affect the weight performance of different classes of samples in the loss function, i.e. gamma can reduce the weight performance of most classes of samples in the loss function and increase the weight value of loss in a few classes. In summary, the Focal loss function expands the range of accepting low loss, and the learner increases the attention to the misclassification of a few classes of samples, thereby constructing a cost-sensitive learning process under the unbalanced learning framework.
Specifically, the foregoing experiment uses a binary cross entropy loss function, which is replaced by a Focal loss function, and traverses the values of α and γ to find an optimal parameter value, there is a basic assumption in the traversal process, that is, the optimal parameter set exists in the traversal space, the traversal process is similar to the grid search in the hyper-parameter optimization, and it can be known from the Focal loss that the gamma is 0 and will be consistent with the cross entropy loss, so the search space of γ is [0, 3], the step size is 0.5, the search space of α is defined as [0.01, 0.31], and the step size is 0.05; the parameter setting comprises a weight cross entropy function form, and simultaneously, the change of gamma and alpha parameters in a Focal loss function can cause the change of a loss value, so that the main investigation object is the identification precision of a test set. Table 6 shows the results of the Focal loss function traversal.
TABLE 6Focal loss function traversal results
alpha gamma Testing precision transformer alpha gamma Test accuracy
0.01 0 0.868397 0.16 2 0.855256
0.01 0.5 0.870277 0.16 2.5 0.854185
0.01 1 0.870253 0.16 3 0.855691
0.01 1.5 0.871355 0.21 0 0.854065
0.01 2 0.867609 0.21 0.5 0.853105
0.01 2.5 0.865784 0.21 1 0.852768
0.01 3 0.865506 0.21 1.5 0.855143
0.06 0 0.866363 0.21 2 0.852905
0.06 0.5 0.865125 0.21 2.5 0.85526
0.06 1 0.863773 0.21 3 0.852552
0.06 1.5 0.862924 0.26 0 0.853776
0.06 2 0.861683 0.26 0.5 0.852082
0.06 2.5 0.860642 0.26 1 0.85304
0.06 3 0.86084 0.26 1.5 0.853728
0.11 0 0.859749 0.26 2 0.853834
0.11 0.5 0.859144 0.26 2.5 0.853532
0.11 1 0.858347 0.26 3 0.851982
0.11 1.5 0.8586 0.31 0 0.823776
0.11 2 0.857069 0.31 0.5 0.842082
0.11 2.5 0.857154 0.31 1 0.83304
0.11 3 0.857287 0.31 1.5 0.823728
0.16 0 0.854905 0.31 2 0.843834
0.16 0.5 0.855822 0.31 2.5 0.833532
0.16 1 0.855639 0.31 3 0.841982
0.16 1.5 0.85649
The training precision is better when alpha is 0.01 and beta is 1.5, the training precision reaches 0.8713, and compared with the cross entropy loss, the training precision is improved to a certain extent, so that the Focal loss is selected as the integral constraint.
In the model optimization, the optimization of the model through the hyper-parameters comprises the following steps:
the previous experiments have carried out the improvement of the stay cable identification algorithm in three aspects of data level, algorithm framework design and cost-sensitive learning, and in the part, the evolution strategy is used for optimizing the hyper-parameter set involved in the learning framework. The evolution strategy carries out variation and combination on the best individuals in the population for a plurality of generations, the algorithm directly uses built-in actual values for variation and combination, and the genetic algorithm can continue to operate after encoding the actual values. The method is different from a parameter set which can be obtained by self in the learning process, and the hyper-parameters are special sets which need to be set manually before the learning begins, such as an optimizer, a learning rate, training batches and the like in the learning process, and can be specifically divided into three types of hyper-parameters of classification, dispersion and continuity. The hyper-parameter set of the learning framework can influence the generalization ability of the whole model, the hyper-parameter sets of the 1D-CNN framework part, the LSTM framework part and the full connection layer part are optimized, the optimal hyper-parameter set is assumed to exist in a defined search space, and the hyper-parameters are obtained after the hyper-parameters are defined in an experiment.
Defining a hyper-parameter: defining four hyper-parameter objects of a one-dimensional convolutional neural network, a long-term and short-term memory network, a full connection layer and an optimizer;
acquiring a hyper-parameter: using an evolution strategy in a superactive module as an optimization algorithm, using a Focal Loss function as a constraint, setting the mutation rate to be 0.5, the cross rate to be 0.5 and the evolution times to be 100, and finally obtaining the optimal hyper-parameters of the one-dimensional convolutional neural network, the long-short term memory network, the full-link layer and the optimizer, wherein the specific results are shown in table 7;
TABLE 7 optimal hyper-parameter set for evolutionary strategies
Figure BDA0003403333050000211
The experiment is to reduce the overfitting risk of the damage working condition of the inclined stay cable under the optimal hyper-parameter set.
Carrying in a model for retraining, adding a batch normalization term (batch _ normalization) into the model to obtain a variation trend of a Loss function and model precision under an optimal hyper-parameter set, wherein FIG. 7 shows the variation trend of the Loss function and the model precision in the Training process, Validation/Test Loss in FIG. 7 shows the Loss of a verification/Test set, Training Loss shows the Loss of the Training set, Validation/Test Accuracy shows the precision of the verification/Test set, Training Accuracy shows the precision of the Training set, Ephsoc shows the iteration times, and since the limiting model of an early stop method is converged in 48 Epochs, the Loss function value of the Training set is known to be reduced to 0.0012, and the overall precision is increased to 0.8745; the test set loss function value falls to 0.0012 with an accuracy of 0.8721, where TP, FP, TN, FN of the confusion matrix are 256467, 56575, 213751, and 13862, respectively. Classification learning report results display of Table 8
TABLE 8 optimal hyper-parameter set classification learning report
Label (R) Precision ratio Recall rate f1 fraction Number of samples
Healthy condition 0.83 0.95 0.88 270329
Damage condition 0.94 0.80 0.86 270329
As can be seen from Table 8, the precision, recall and f1 scores for the health condition of the test set were 0.83, 0.95, 0.88, and the loss condition was 0.94, 0.80, 0.86, respectively, at which the g-mean and AUG values were 0.8816 and 0.8700, respectively.
Table 9 improves the recognition method from the perspective of algorithm design, feature engineering, loss function and hyper-parameter set, etc., and compares it with the recognition capability of the 1D-CNN model.
TABLE 9 identification method comparison
Figure BDA0003403333050000221
The second column of table 9 is optimized from the perspective of algorithm design, and introduces an LSTM unit to memorize the overall network information; the third column starts from feature engineering, performs feature selection by using a correlation coefficient after polynomial feature transformation, performs feature selection by using an experiment from two angles of the correlation coefficient and variance analysis after feature transformation, and only shows an optimal result in a table; the fourth column optimizes the loss function and the hyper-parameter set. After the final stay cable damage identification algorithm passes through the operation flow, g-mean and AUG are increased from 0.8548 and 0.8325 to 0.8816 and 0.8700, the precision ratio of the health condition is increased from 0.79 to 0.83, and the damage condition is increased from 0.89 to 0.94.
The optimal model can be obtained by optimizing the model through the combination of the characteristic engineering, the cross entropy loss function and the hyper-parameter. The method for identifying the damage of the whole stay cable converts unbalanced stay cable data into balanced stay cable data by resampling, extracts features of the balanced stay cable data, and optimizes the model through feature engineering, cross entropy loss function and hyper-parameter set combination to obtain an optimal model, so that the problem of unbalanced classification of large-scale working condition data can be effectively solved, and the precision of identifying the damage of the stay cable of the cable-stayed bridge can be improved.
Example 2
The embodiment provides an electronic device, comprising a storage device and one or more processors;
storage means for storing one or more programs; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the cable-stayed bridge cable damage identification method as described in embodiment 1.
The electronic equipment converts unbalanced stay cable data into balanced stay cable data by resampling, extracts features of the balanced stay cable data, and optimizes the model through feature engineering, cross entropy loss function and hyper-parameter set combination to obtain an optimal model, so that the problem of large-scale working condition data classification imbalance can be effectively solved, and the accuracy of identification of damage of the stay cable of the cable-stayed bridge can be improved.

Claims (10)

1. A method for identifying damage of a stay cable of a cable-stayed bridge is characterized by comprising the following steps:
collecting the stay cable data of the cable-stayed bridge: acquiring acceleration data of each stay cable through a plurality of sensors, and simulating working condition data of the stay cable under different working condition states by using ANSYS finite element analysis software according to the acquired acceleration data;
and (3) processing the stay cable data: resampling the acquired unbalanced stay cable data to convert the unbalanced stay cable data into balanced stay cable data, and performing characteristic engineering on the balanced stay cable data to extract the characteristics of the stay cable data;
constructing a model, and combining the convolutional neural network with the long-short term memory network to form a one-dimensional convolutional long-short term memory model;
and (3) optimizing the model, namely inputting the extracted data characteristics of the stay cable into the constructed model, and optimizing the model through characteristic engineering, a cross entropy loss function and a hyper-parameter set until the best model is trained.
2. The method for identifying the damage of the stayed cable of the cable-stayed bridge according to the claim 1, characterized in that the method for characterizing the balanced stayed cable data comprises the following steps:
dividing the stay cable data into damaged data and undamaged data;
respectively calculating the similarity between the characteristic points of each stay cable corresponding to the damaged data and the undamaged data to obtain a similar matrix, and establishing an adjacent matrix through the similar matrix and a threshold value of a coefficient in the similar matrix;
analyzing the relation among the characteristic points of the stay cables, selecting a threshold value of a coefficient in a similar matrix to respectively generate a minimum spanning tree and a node graph, if the similarity of the characteristic points of the two stay cables is greater than the threshold value, connecting the characteristic points of the two stay cables, and representing the characteristic points of the two stay cables by 1 in an adjacent matrix; if the similarity of the characteristic points of the two stay cables is smaller than the threshold value, the characteristic points of the two stay cables are not connected and are represented by 0 in the adjacent matrix;
respectively calculating the node degrees of the damaged data and the undamaged data by using the generated minimum spanning tree;
respectively calculating the node degree of the damaged data and the node degree of the undamaged data;
averaging the variation values of the node degrees of the damaged data and the node degrees of the undamaged data respectively to obtain an average value of the node degree variation, sequencing the node degree variation from large to small, and extracting the feature importance of the stay cable feature points with the node degree variation exceeding the average value.
3. The method for identifying the damage of the stayed-cable bridge cable according to claim 2, characterized in that the undamaged data are put into a set A, the damaged data are put into a set B, and a similarity matrix of the damaged data stayed-cable characteristic points in the set A and a similarity matrix of the damaged data stayed-cable characteristic points in the set B are calculated, wherein the similarity matrix coefficient between the stayed-cable characteristic points is calculated as follows:
Figure FDA0003403333040000021
wherein
Figure FDA0003403333040000022
Node data representing the characteristic point of the (n) 1 th stay cable in the (m) th sample,
Figure FDA0003403333040000023
the sample mean value of the characteristic point of the nth 1 stay cables is shown,
Figure FDA0003403333040000024
the coefficient of the similarity matrix between the stay cable characteristic point n1 and the stay cable characteristic point n2 is shown, and M represents the number of samples.
4. The method for identifying the damage of the stayed-cable of the cable-stayed bridge according to claim 1, wherein the model building comprises the following steps:
repeatedly compressing and extracting the stay cable characteristics through the convolution pooling operation of the one-dimensional convolution neural network to obtain characteristic information;
and learning and identifying the feature information after the convolution pooling through a long-term and short-term memory network to construct a model.
5. The method for identifying the damage of the stayed-cable bridge of claim 4, wherein the convolutional neural network is composed of an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer; the method comprises the following steps that stay cable data enter an input layer after being subjected to standardization processing, feature compression is carried out on a convolution layer after being processed, and convolution operation is carried out on the obtained stay cable data through a convolution kernel;
after the stay cable data is subjected to convolution operation, performing pooling operation to extract characteristic information in the stay cable data;
performing Flatten expansion on the stay cable data after convolution pooling, wherein the Flatten expansion is used as the input of a full connection layer, and the full connection layer integrates the local information after convolution pooling;
after convolution pooling and processing of the full connection layer, a logic value of the working condition state is obtained, the working condition state is divided by the logic value, and a network output layer is constructed through a softmax function.
6. The method for identifying damage to a stay cable of a cable-stayed bridge according to claim 5, wherein the output of each layer of the convolutional neural network is set as a convolution result of a plurality of stay cable characteristics, and the output signal after the convolution operation is:
Figure FDA0003403333040000031
wherein the content of the first and second substances,
Figure FDA0003403333040000032
represents the weights of the/layer ith order convolution kernels,
Figure FDA0003403333040000033
expressed as the offset, x, of the convolverl(j) Represents the jth local area of the l layer,
Figure FDA0003403333040000034
representing the output signal after the convolution operation;
extracting feature information using a max-pooling operation, the max-pooling using a maximum value of the pooled region as the extracted target information, expressed as follows:
Figure FDA0003403333040000035
wherein the content of the first and second substances,
Figure FDA0003403333040000036
represents the ith characteristic value in t neurons of the l layers, and t is the [ (j-1) W, jW]W is the width of the pooling zone,
Figure FDA0003403333040000037
represents the value of (l +1) layer neurons;
the expression of the softmax function is:
Figure FDA0003403333040000038
wherein z isp(j) A logic value of the j-th neuron of the output layer, M represents the total number of classes, zp(M) denotes the logic value of the mth neuron in the output layer, j ═ 1, 2.
7. The method for identifying the damage of the stayed cable of the cable-stayed bridge according to the claim 1, wherein in the model optimization, the model is optimized through the characteristic engineering, and the method comprises the following steps:
constructing new characteristics, and performing characteristic engineering exploration by using polynomial characteristic conversion preprocessed in a sklern module, wherein the expression of the processing of the stay cable characteristic points is as follows:
Figure FDA0003403333040000039
in the formula, the highest order of the transformation process of the characteristic points of the stay cables is limited to 2, and the newly generated characteristic matrix is formed by combining polynomials of all the characteristic points of the stay cables with the order less than or equal to 2, wherein x'1，x′n，x1，
Figure FDA0003403333040000041
Uniform expression diagonal drawSearching for characteristic points;
and extracting features, namely extracting the features of the stay cable feature points through correlation analysis and variance analysis.
8. The method for identifying the damage of the stayed-cable of the cable-stayed bridge according to the claim 1, wherein the step of optimizing the model by the cross entropy loss function in the model optimization comprises the following steps:
introduction of an imbalance weight factor alpha in a cross entropy loss functiontThe cross entropy loss function with the weighting factor is:
C(pt)＝-αtlog(pt)，t∈[0，1]
wherein p istFor the prediction probability, the subscript t represents the sample class, 0 is the majority class, and 1 is the minority class;
adding a modulation factor (1-p) to the cross-entropy loss function to which the weight factor is addedt)γNamely:
Lf＝-αt(1-pt)γlog(pt)
when a sample is misclassified or predicted probability ptWhen the modulation factor approaches 1, the cross entropy loss function is not affected, and the probability p is predictedtWhen approaching 1, the modulation factor tends to 0, LfRepresenting the cross entropy loss function of the added modulation factor.
9. The method for identifying damage to a stay cable of a cable-stayed bridge according to claim 8, wherein the optimization of the model by the hyper-parameters comprises the following steps:
defining a hyper-parameter: defining four hyper-parameter objects of a one-dimensional convolutional neural network, a long-term and short-term memory network, a full connection layer and an optimizer;
acquiring a hyper-parameter: using an evolution strategy in a superactive module as an optimization algorithm, using a Focal Loss function as a constraint, setting the mutation rate to be 0.5, the cross rate to be 0.5 and the evolution times to be 100, and finally obtaining the optimal hyper-parameters of the one-dimensional convolutional neural network, the long-short term memory network, the full-link layer and the optimizer;
and introducing the model for retraining, adding batch standardization items into the model to obtain a loss function under the optimal hyper-parameter set and the change trend of the model precision, and finally obtaining the optimal model.
10. An electronic device comprising storage and one or more processors;
storage means for storing one or more programs;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method for identifying damage to a cable-stayed bridge cable of any one of claims 1-9.
CN202111513368.4A 2021-12-10 2021-12-10 Cable-stayed bridge stay cable damage identification method and electronic equipment Active CN114169374B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111513368.4A CN114169374B (en) 2021-12-10 2021-12-10 Cable-stayed bridge stay cable damage identification method and electronic equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111513368.4A CN114169374B (en) 2021-12-10 2021-12-10 Cable-stayed bridge stay cable damage identification method and electronic equipment

Publications (2)

Publication Number Publication Date
CN114169374A true CN114169374A (en) 2022-03-11
CN114169374B CN114169374B (en) 2024-02-20

Family

ID=80485759

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111513368.4A Active CN114169374B (en) 2021-12-10 2021-12-10 Cable-stayed bridge stay cable damage identification method and electronic equipment

Country Status (1)

Country Link
CN (1) CN114169374B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114595733A (en) * 2022-05-10 2022-06-07 山东大学 Bridge inhaul cable broken wire signal identification method and system based on long-term and short-term memory network
CN114969934A (en) * 2022-05-31 2022-08-30 湖南工商大学 Stay cable damage degree identification method and model construction method
CN116881819A (en) * 2023-09-07 2023-10-13 成都理工大学 Stay cable working state monitoring method based on isolated forest
CN117076889A (en) * 2023-09-04 2023-11-17 中铁建大桥工程局集团南方工程有限公司 Three-main truss steel truss girder construction monitoring method and system, storage medium and electronic equipment

Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106953768A (en) * 2017-04-13 2017-07-14 西安电子科技大学 A kind of network reliability model and mixing intelligent optimizing method
CN108009710A (en) * 2017-11-19 2018-05-08 国家计算机网络与信息安全管理中心 Node test importance appraisal procedure based on similarity and TrustRank algorithms
CN108108766A (en) * 2017-12-28 2018-06-01 东南大学 Driving behavior recognition methods and system based on Fusion
CN110555717A (en) * 2019-07-29 2019-12-10 华南理工大学 method for mining potential purchased goods and categories of users based on user behavior characteristics
CN110633855A (en) * 2019-09-17 2019-12-31 江南大学 Bridge health state detection and management decision making system and method
CN110807562A (en) * 2020-01-07 2020-02-18 杭州鲁尔物联科技有限公司 Regional bridge risk prediction method and system
CN111044617A (en) * 2019-11-28 2020-04-21 东南大学 Inhaul cable damage identification method based on deep learning and acoustic emission technology
CN111340292A (en) * 2020-02-26 2020-06-26 福州大学 Integrated neural network PM2.5 prediction method based on clustering
WO2020156348A1 (en) * 2019-01-31 2020-08-06 青岛理工大学 Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network
CN111553003A (en) * 2020-04-03 2020-08-18 中交第二航务工程局有限公司 Visual large-span cable-stayed bridge cable evaluation method based on data driving
CN111639678A (en) * 2020-05-08 2020-09-08 西安工业大学 INS/GPS integrated navigation fault detection and diagnosis method based on integrated neural network
CN111783212A (en) * 2020-07-09 2020-10-16 长沙理工大学 Typical damage identification method for cable-stayed bridge
CN111860188A (en) * 2020-06-24 2020-10-30 南京师范大学 Human body posture recognition method based on time and channel double attention
WO2021082809A1 (en) * 2019-10-29 2021-05-06 山东科技大学 Training optimization method for foreign exchange time series prediction
CN113257406A (en) * 2021-04-30 2021-08-13 中国人民解放军总医院第一医学中心 Disaster rescue triage and auxiliary diagnosis method based on intelligent glasses
CN113468635A (en) * 2021-06-08 2021-10-01 桂林理工大学 Stay cable force identification method based on vibration signal and cyclic neural network
CN113558638A (en) * 2021-07-05 2021-10-29 杭州电子科技大学 Vine Copula-based brain muscle coupling model construction method

Patent Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106953768A (en) * 2017-04-13 2017-07-14 西安电子科技大学 A kind of network reliability model and mixing intelligent optimizing method
CN108009710A (en) * 2017-11-19 2018-05-08 国家计算机网络与信息安全管理中心 Node test importance appraisal procedure based on similarity and TrustRank algorithms
CN108108766A (en) * 2017-12-28 2018-06-01 东南大学 Driving behavior recognition methods and system based on Fusion
WO2020156348A1 (en) * 2019-01-31 2020-08-06 青岛理工大学 Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network
CN110555717A (en) * 2019-07-29 2019-12-10 华南理工大学 method for mining potential purchased goods and categories of users based on user behavior characteristics
CN110633855A (en) * 2019-09-17 2019-12-31 江南大学 Bridge health state detection and management decision making system and method
WO2021082809A1 (en) * 2019-10-29 2021-05-06 山东科技大学 Training optimization method for foreign exchange time series prediction
CN111044617A (en) * 2019-11-28 2020-04-21 东南大学 Inhaul cable damage identification method based on deep learning and acoustic emission technology
CN110807562A (en) * 2020-01-07 2020-02-18 杭州鲁尔物联科技有限公司 Regional bridge risk prediction method and system
CN111340292A (en) * 2020-02-26 2020-06-26 福州大学 Integrated neural network PM2.5 prediction method based on clustering
CN111553003A (en) * 2020-04-03 2020-08-18 中交第二航务工程局有限公司 Visual large-span cable-stayed bridge cable evaluation method based on data driving
CN111639678A (en) * 2020-05-08 2020-09-08 西安工业大学 INS/GPS integrated navigation fault detection and diagnosis method based on integrated neural network
CN111860188A (en) * 2020-06-24 2020-10-30 南京师范大学 Human body posture recognition method based on time and channel double attention
CN111783212A (en) * 2020-07-09 2020-10-16 长沙理工大学 Typical damage identification method for cable-stayed bridge
CN113257406A (en) * 2021-04-30 2021-08-13 中国人民解放军总医院第一医学中心 Disaster rescue triage and auxiliary diagnosis method based on intelligent glasses
CN113468635A (en) * 2021-06-08 2021-10-01 桂林理工大学 Stay cable force identification method based on vibration signal and cyclic neural network
CN113558638A (en) * 2021-07-05 2021-10-29 杭州电子科技大学 Vine Copula-based brain muscle coupling model construction method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
LEI FU等: "Damage Identification of Long-Span Bridges Using the Hybrid of Convolutional Neural Network and Long Short-Term Memory Network", 《ALGORITHMS》, vol. 14, no. 6, pages 1 - 20 *
MARIANO BEGUERISSE-DFAZ等: "Finding role communities in directed networks using Role-Based Similarity, Markov Stability and the Relaxed Minimum Spanning Tree", 《2013 IEEE GLOBAL CONFERENCE ON SIGNAL AND INFORMATION PROCESSING》, pages 937 - 940 *
侯立群: "大型斜拉桥基于健康监测的模型修正、损伤诊断与预警方法", 《中国博士学位论文全文数据库 工程科技II辑》, no. 2011, pages 034 - 17 *
皇鹏飞: "基于CNN-LSTM混合架构神经网络的桥梁损伤识别方法研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》, no. 2021, pages 034 - 407 *
高浩: "复杂网络中重要节点辨识方法研究", 《中国优秀硕士学位论文全文数据库 基础科学辑》, no. 2021, pages 002 - 50 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114595733A (en) * 2022-05-10 2022-06-07 山东大学 Bridge inhaul cable broken wire signal identification method and system based on long-term and short-term memory network
CN114595733B (en) * 2022-05-10 2024-04-02 山东大学 Bridge inhaul cable broken wire signal identification method and system based on long-short-term memory network
CN114969934A (en) * 2022-05-31 2022-08-30 湖南工商大学 Stay cable damage degree identification method and model construction method
CN114969934B (en) * 2022-05-31 2023-09-05 湖南工商大学 Stay cable damage degree identification method and model construction method
CN117076889A (en) * 2023-09-04 2023-11-17 中铁建大桥工程局集团南方工程有限公司 Three-main truss steel truss girder construction monitoring method and system, storage medium and electronic equipment
CN117076889B (en) * 2023-09-04 2024-05-31 中铁建大桥工程局集团南方工程有限公司 Three-main truss steel truss girder construction monitoring method and system, storage medium and electronic equipment
CN116881819A (en) * 2023-09-07 2023-10-13 成都理工大学 Stay cable working state monitoring method based on isolated forest
CN116881819B (en) * 2023-09-07 2023-11-14 成都理工大学 Stay cable working state monitoring method based on isolated forest

Also Published As

Publication number Publication date
CN114169374B (en) 2024-02-20

Similar Documents

Publication Publication Date Title
CN114169374B (en) Cable-stayed bridge stay cable damage identification method and electronic equipment
CN110533631B (en) SAR image change detection method based on pyramid pooling twin network
CN112070128A (en) Transformer fault diagnosis method based on deep learning
US20220261655A1 (en) Real-time prediction method for engine emission
CN111144552B (en) Multi-index grain quality prediction method and device
CN114563150B (en) Bridge health online detection module generation method, detection method, tool box and device
CN113901977A (en) Deep learning-based power consumer electricity stealing identification method and system
CN112464996B (en) LSTM-XGBoost-based intelligent power grid intrusion detection method
CN115618732B (en) Nuclear reactor digital twin key parameter autonomous optimization data inversion method
CN114676742A (en) Power grid abnormal electricity utilization detection method based on attention mechanism and residual error network
CN111753776B (en) Structural damage identification method based on echo state and multi-scale convolution combined model
CN111915022B (en) Gaussian process method and device for rapidly identifying stability coefficient of sliding karst dangerous rock
CN113011660A (en) Air quality prediction method, system and storage medium
CN113420795A (en) Mineral spectrum classification method based on void convolutional neural network
CN111461923A (en) Electricity stealing monitoring system and method based on deep convolutional neural network
CN113987910A (en) Method and device for identifying load of residents by coupling neural network and dynamic time planning
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN115759343A (en) E-LSTM-based user electric quantity prediction method and device
CN113935413A (en) Distribution network wave recording file waveform identification method based on convolutional neural network
LUO et al. GRU neural network-based method for box girder crack damage detection
CN115598459A (en) Power failure prediction method for 10kV feeder line fault of power distribution network
CN111177975A (en) Aviation equipment availability prediction method based on artificial intelligence
Zain Al-Thalabi et al. Modeling and prediction using an artificial neural network to study the impact of foreign direct investment on the growth rate/a case study of the State of Qatar
CN114254828B (en) Power load prediction method based on mixed convolution feature extractor and GRU
CN111291490B (en) Nonlinear mapping intelligent modeling method for structure multi-scale heterogeneous response

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