CN109490861B - Blast furnace burden line extraction method - Google Patents
Blast furnace burden line extraction method Download PDFInfo
- Publication number
- CN109490861B CN109490861B CN201811268646.2A CN201811268646A CN109490861B CN 109490861 B CN109490861 B CN 109490861B CN 201811268646 A CN201811268646 A CN 201811268646A CN 109490861 B CN109490861 B CN 109490861B
- Authority
- CN
- China
- Prior art keywords
- blast furnace
- time sequence
- stockline
- radar
- signal time
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Computer Networks & Wireless Communication (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention provides a blast furnace burden line extraction method which can improve the accuracy, the generalization performance and the real-time performance of blast furnace burden line extraction. The method comprises the following steps: collecting a field radar echo signal of the blast furnace, wherein radar waves emitted by a radar system can cover all areas of the radial charge level of the blast furnace; generating a signal time sequence spectrogram of the radar echo signal; performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result; constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database; the method comprises the steps of collecting blast furnace radar echo signals to be detected in real time, generating a signal time sequence spectrogram, and carrying out stockline segmentation and extraction on the signal time sequence spectrogram according to a trained modular full convolution network model. The invention relates to the field of blast furnace burden line monitoring.
Description
Technical Field
The invention relates to the field of blast furnace burden line monitoring, in particular to a blast furnace burden line extraction method.
Background
In the iron-making industry, the measurement and the optimized control of the blast furnace burden surface to adjust the burden distribution are the keys of improving the production efficiency, saving energy and reducing emission, the accurate and real burden surface information is obtained, the adjustment of the burden distribution strategy of the blast furnace is facilitated, the reasonable distribution of the materials in the furnace and the gas flow in the furnace is ensured, and therefore the method has important significance in accurately obtaining the shape information of the burden surface in the blast furnace in real time. At present, a high-frequency microwave radar is widely used in blast furnace burden surface detection as an ideal measuring sensor in a blast furnace complex environment. And finally, calculating the distance value of the stockline to be measured according to the linear relation between the peak spectral line number corresponding to the echo of the feeding surface of the frequency spectrum and the radar measurement distance.
However, the energy distribution of the echo signals acquired on site on the frequency spectrum mainly comprises a distribution chute, a cross temperature measuring device, real material echoes and strong background noise, and belongs to nonlinear and non-stable signals, and the real material surface echoes are often covered in the material surface echoes and are difficult to identify. The basic mode of the existing radar signal processing algorithm is to obtain rough estimation of signal frequency according to discrete Fourier transform and then use a discrete spectrum correction algorithm for fine estimation. The mode is established on the basis of peak searching, has high requirements on radar echo signals, independently processes each group of radar signals, lacks deeper mining of regular information of continuous change of real charge level hidden among each group of radar signals, and has serious information leakage. In the face of severe environment of a blast furnace, the problems of stockline missing, local jumping and the like easily occur in the conventional algorithm.
Disclosure of Invention
The invention aims to provide a blast furnace burden line extraction method to solve the problems that a burden line is easy to lose and jump locally when the burden line is detected in the prior art.
In order to solve the above technical problem, an embodiment of the present invention provides a blast furnace burden line extraction method, including:
collecting a field radar echo signal of the blast furnace, wherein radar waves emitted by a radar system can cover all areas of the radial charge level of the blast furnace;
generating a signal time sequence spectrogram of the radar echo signal;
performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result;
constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database;
the method comprises the steps of collecting blast furnace radar echo signals to be detected in real time, generating a signal time sequence spectrogram, and carrying out stockline segmentation and extraction on the signal time sequence spectrogram according to a trained modular full convolution network model.
Further, the collecting the radar echo signals on the blast furnace site comprises:
according to a bell-less chute rotary distribution mode adopted by the blast furnace and the installation principle of a radar system, the corresponding radar system is arranged on the top of the blast furnace to meet the requirement of full coverage of the radial charge level of the blast furnace;
and radar echo signals are continuously collected in real time in the production process of the blast furnace.
Further, the generating a signal timing spectrogram of the radar return signal comprises:
and carrying out time-frequency transformation, normalization, time sequence expansion and gray level mapping on the collected radar echo signals to generate a signal time sequence spectrogram of the radar echo signals.
Further, the performing time-frequency transformation, normalization, time sequence expansion and gray scale mapping on the collected radar echo signals to generate a signal time sequence spectrogram of the radar echo signals includes:
performing time-frequency conversion on the radar echo signal through fast Fourier transform to obtain an amplitude frequency spectrum of the radar echo signal;
according to the distance between the furnace top radar and the charge level, carrying out interval interception on the amplitude frequency spectrum of each group of radar echo signals, and carrying out normalization processing on the amplitude;
according to the continuity of radar echo signals on a time sequence, taking continuous M groups of signal spectrum data after normalization processing as a unit, and forming a two-dimensional matrix by time sequence arrangement;
and carrying out gray mapping on the two-dimensional matrix to generate a signal time sequence frequency spectrum gray map.
Further, after the generating the signal timing spectrum gray-scale map, the method further comprises:
and stretching the height of the signal time sequence frequency spectrum gray-scale map by cubic interpolation.
Further, the data enhancing the signal time series spectrogram comprises:
and performing data enhancement on the generated signal time sequence spectrogram through horizontal inversion.
Further, the obtaining of the stockline labeling result includes:
acquiring a pixel-by-pixel labeling result of a stockline in a signal time sequence spectrogram after data enhancement, wherein the labeling content comprises: the change trend of the stockline, and floating, double-layer stockline and continuous fault on the stockline;
and converting pixel values which are greater than or equal to the preset pixel threshold value in the marked picture into first identifications and converting pixel values which are smaller than the preset pixel threshold value into second identifications through the preset pixel threshold value, and generating first identification-second identification label pictures which are in one-to-one correspondence with the signal time sequence frequency spectrogram.
Further, the building the modular full-convolution network model includes:
constructing a modular full-convolution network model comprising a feature extraction module, a feature fusion module and a feature decoding module; wherein the content of the first and second substances,
the characteristic extraction module is formed by crossing a plurality of convolution units and pooling layers and is used for learning the stockline characteristics in the input signal time sequence spectrogram through the convolution units of different layers to obtain characteristic graphs of different layers and different sizes;
the feature fusion module is used for fusing the feature graphs of different levels and different sizes extracted by the feature extraction module;
and the characteristic decoding module is used for carrying out up-sampling operation on the characteristic graph to obtain network output with the same size as the label picture, changing the number of output channels to keep the number of the output channels consistent with the number of the image categories, and obtaining a stockline segmentation result.
Further, the stockline segmentation database includes: a training set, a verification set and a test set;
the training of the modular full convolution network model according to the established stockline segmentation database comprises the following steps:
training the modular full convolution network model using the training set;
optimizing and adjusting hyper-parameters of the modular full convolution network model by utilizing the verification set;
and testing and evaluating the performance of the modular full convolution network model by using the test set.
Further, the radar system includes: distributed array radar, scanning radar, and/or phased array radar.
The technical scheme of the invention has the following beneficial effects:
in the scheme, a radar echo signal of a blast furnace site is collected, wherein radar waves emitted by a radar system can cover all areas of a radial charge level of the blast furnace; generating a signal time sequence spectrogram of the radar echo signal, wherein the signal time sequence spectrograms of different radars in different time periods are different; performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result; constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database; the method comprises the steps of collecting echo signals of the blast furnace radar to be detected in real time, generating a signal time sequence frequency spectrogram, carrying out stockline segmentation and extraction on the signal time sequence frequency spectrogram according to a trained modularized full convolution network model, and realizing real-time detection of the variation trend of the blast furnace stockline in the production process. Therefore, the change rule of the stockline in each radial region in the continuous time range in the production process of the blast furnace is completely reflected through the stockline segmentation database, the same model is trained through the multi-radar signal time sequence spectrogram, the generalization performance of the model can be effectively improved, and the extraction accuracy, the generalization performance and the real-time performance of the blast furnace stockline are improved.
Drawings
FIG. 1 is a schematic flow chart of a blast furnace burden line extraction method provided by an embodiment of the present invention;
FIG. 2 is a schematic flow chart illustrating a detailed process of a blast furnace burden line extraction method according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an eight-point distributed array radar installation provided by an embodiment of the present invention;
FIG. 4 is a time domain diagram of a collected blast furnace radar signal provided by an embodiment of the present invention;
FIG. 5 is a schematic diagram of a generated signal timing spectrum according to an embodiment of the present invention;
FIG. 6 is a schematic view of a part of a material line according to an embodiment of the present invention;
FIG. 7 is a schematic diagram illustrating a stockline labeling result according to an embodiment of the present invention;
FIG. 8 is a schematic structural diagram of a modular full-convolution network model according to an embodiment of the present invention;
fig. 9 is a schematic diagram illustrating a stockline segmentation result according to an embodiment of the present invention;
fig. 10 is a schematic diagram illustrating comparison of stockline detection according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.
The invention provides a blast furnace burden line extraction method aiming at the problems that the existing burden line is easy to lose and jump locally when the burden line is detected.
As shown in fig. 1, a blast furnace burden line extraction method provided by an embodiment of the present invention includes:
s101, collecting a field radar echo signal of the blast furnace, wherein radar waves emitted by a radar system can cover all areas of a radial charge level of the blast furnace;
s102, generating a signal time sequence spectrogram of the radar echo signal;
s103, performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result;
s104, constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database;
and S105, collecting the blast furnace radar echo signals to be detected in real time, generating a signal time sequence spectrogram, and performing stockline segmentation and extraction on the signal time sequence spectrogram according to the trained modular full convolution network model.
According to the blast furnace burden line extraction method, on-site radar echo signals of the blast furnace are collected, wherein radar waves emitted by a radar system can cover all areas of the radial burden surface of the blast furnace; generating a signal time sequence spectrogram of the radar echo signal, wherein the signal time sequence spectrograms of different radars in different time periods are different; performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result; constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database; the method comprises the steps of collecting echo signals of the blast furnace radar to be detected in real time, generating a signal time sequence frequency spectrogram, carrying out stockline segmentation and extraction on the signal time sequence frequency spectrogram according to a trained modularized full convolution network model, and realizing real-time detection of the variation trend of the blast furnace stockline in the production process. Therefore, the change rule of the stockline in each radial region in the continuous time range in the production process of the blast furnace is completely reflected through the stockline segmentation database, the same model is trained through the multi-radar signal time sequence spectrogram, the generalization performance of the model can be effectively improved, and the accuracy, the generalization performance and the real-time performance of the blast furnace stockline extraction are improved.
In an embodiment of the aforementioned blast furnace burden line extraction method, further, the acquiring a blast furnace field radar echo signal includes:
according to a bell-less chute rotary distribution mode adopted by the blast furnace and the installation principle of a radar system, the corresponding radar system is arranged on the top of the blast furnace to meet the requirement of full coverage of the radial charge level of the blast furnace;
and radar echo signals are continuously collected in real time in the production process of the blast furnace.
In an embodiment of the aforementioned method for extracting a blast furnace burden line, further, the generating a signal time-series spectrogram of the radar echo signal includes:
and carrying out time-frequency transformation, normalization, time sequence expansion and gray level mapping on the collected radar echo signals to generate a signal time sequence spectrogram of the radar echo signals.
In a specific embodiment of the above method for extracting a blast furnace burden line, further, the performing time-frequency transformation, normalization, time sequence expansion and gray scale mapping on the collected radar echo signal, and generating a signal time sequence spectrogram of the radar echo signal includes:
performing time-frequency conversion on the radar echo signal through fast Fourier transform to obtain an amplitude frequency spectrum of the radar echo signal;
according to the distance between the furnace top radar and the charge level, carrying out interval interception on the amplitude frequency spectrum of each group of radar echo signals, and carrying out normalization processing on the amplitude, wherein the interception range is N;
according to the continuity of radar echo signals on a time sequence, taking continuous M groups of signal spectrum data after normalization processing as a unit, and forming a two-dimensional matrix by time sequence arrangement, wherein the size of the two-dimensional matrix is N x M;
and carrying out gray mapping on the two-dimensional matrix to generate a signal time sequence frequency spectrum gray map.
In an embodiment of the aforementioned method for extracting a blast furnace burden line, further, in order to more finely embody details of a change of a burden line in the signal time-series spectrogram and facilitate fine labeling, after the signal time-series spectrogram gray-scale map is generated, the method further includes:
and stretching the height of the signal time sequence frequency spectrum gray-scale map by cubic interpolation.
In the embodiment, according to the radar ranging principle, the distance to be measured of the stockline is in direct proportion to the number of the spectrum peak value spectral line, and after the distance to be measured of the stockline is converted into the image, the distance of the stockline in the image is in direct proportion to the height of the image; therefore, the height of the signal time sequence frequency spectrum gray level image can be stretched by adopting the cubic interpolation, for example, the image height can be stretched by one time, and the width is unchanged, so that the change details of the material line can be displayed more clearly on the premise of ensuring that the distance of the material line is unchanged, and meanwhile, the marking is convenient.
In an embodiment of the aforementioned blast furnace burden line extraction method, further, the performing data enhancement on the signal time series spectrogram includes:
and performing data enhancement on the generated signal time sequence spectrogram through horizontal inversion.
In a specific embodiment of the aforementioned method for extracting a blast furnace burden line, further, the acquiring a burden line marking result includes:
acquiring a pixel-by-pixel labeling result of a stockline in a signal time sequence spectrogram after data enhancement, wherein the labeling content comprises: the change trend of the stockline, and floating, double-layer stockline and continuous fault on the stockline;
and converting pixel values which are greater than or equal to the preset pixel threshold value in the marked picture into first identifications and converting pixel values which are smaller than the preset pixel threshold value into second identifications through the preset pixel threshold value, and generating first identification-second identification label pictures which are in one-to-one correspondence with the signal time sequence frequency spectrogram.
In this embodiment, through a preset pixel threshold, a pixel value greater than or equal to a preset pixel threshold (i.e., corresponding to a wire) in the labeled picture may be converted into a first identifier (e.g., 0), and a pixel value smaller than the preset pixel threshold (corresponding to a background) may be converted into a second identifier (e.g., 1), so as to generate 0-1 labeled pictures corresponding to the signal timing spectrogram one by one, so that in network training, a network loss function is calculated through the labeled pictures and a network output result, and a network parameter is updated.
In an embodiment of the aforementioned blast furnace burden line extraction method, further, the building a modular full-volume network model includes:
the characteristic extraction module is formed by crossing a plurality of convolution units and pooling layers and is used for learning the stockline characteristics in the input signal time sequence spectrogram through the convolution units of different layers to obtain characteristic graphs of different layers and different sizes; assuming that there are 5 convolution units, local features such as edges, textures and shapes are learned by shallow convolution units (e.g., a first convolution unit, a second convolution unit, a third convolution unit and a fourth convolution unit), and then shallow features are fused by deep convolution units (e.g., a fifth convolution unit) on the basis of the local features, so that more abstract and more essential features are learned;
the feature fusion module is used for fusing feature graphs of different levels and different sizes extracted by the feature extraction module so as to learn the corresponding relation between the spatial positions and the channels of different shallow features and deep features through the added convolutional layers, so that the loss of deep abstract features on target details is compensated through the fusion of the shallow features in the decoding process of the feature graphs; the difference of the shallow feature and the deep feature on the space position and the channel leads the difference of the feature map fusion mode to have larger influence on the performance of the network model;
and the characteristic decoding module is used for carrying out up-sampling operation on the characteristic graph to obtain network output with the same size as the label picture, changing the number of output channels to keep the number of the output channels consistent with the number of the image categories, and obtaining a stockline segmentation result.
In this embodiment, the constructed modular full-convolution network model structure includes: the device comprises a feature extraction module, a feature fusion module and a feature decoding module; wherein the content of the first and second substances,
a feature extraction module: selecting a convolutional neural network as a basic model of a feature extraction module; removing a full connection layer and an output layer in the convolutional neural network; the convolutional neural network comprises five convolutional units, each convolutional unit consists of a plurality of convolutional layers, 1 correction linear unit and 1 batch normalization unit, and specifically: the front two convolution units are formed by connecting two convolution layers in series, the rear three convolution units are formed by connecting three convolution layers in series, and a correction linear unit and a batch normalization unit are arranged behind each convolution layer; determining the size of a convolution kernel in the convolution neural network; the five convolution units in the convolution neural network are alternately connected with the pooling layer, maximum pooling is adopted, the size of a pooling window is 2 x 2, and the sliding step length is 2 x 2;
in the embodiment, due to the fact that the linear model has insufficient expression capacity, the use of the linear unit is corrected, and the nonlinear characteristic is introduced into the network, so that the modular full convolution network model can represent nonlinear complex arbitrary function mapping between input and output, and the method has the advantages of being high in calculation speed, simple in derivation, capable of achieving single-side inhibition, sparse activation, wider in excitation boundary, free of gradient dispersion and the like; the batch normalization unit can accelerate the network training convergence speed by adjusting the change of the distribution of the interlayer data, prevent overfitting and improve the generalization capability of the network.
A feature fusion module: setting a fusion mode among the feature graphs as splicing; before feature map fusion, changing the number of feature map channels through convolution layers with convolution kernel size of 1 x 1; after feature map fusion, changing the number of feature map channels by convolution layers with convolution kernel size of 1 x 1;
a feature decoding module: sequentially adding three convolutional layers after the fifth pooling layer pool5, setting a convolution kernel of the first convolutional layer with the same size as the output feature map of the pool5, setting the size of the convolution kernels of the subsequent two convolutional layers to be 1 x 1, and sequentially reducing the number of channels of the feature map to make the number of the final feature map channels be the number of categories to obtain a feature map score 5; performing double up-sampling on the score5 through deconvolution to obtain a pool5 layer feature decoding result score _ up 5; fusing the result of the feature map of the fourth pooling layer pool4 after 1 × 1 convolution operation with score _ up5, changing the number of the feature map channels after fusion to be consistent with pool4 through 1 × 1 convolution operation, and performing double upsampling through an deconvolution layer to obtain a feature decoding result score _ up4 of a pool4 layer; and fusing the result of the feature map of the third pooling layer pool3 after 1 × 1 convolution operation with score _ up4, changing the number of channels of the feature map after fusion to be consistent with pool3 through the convolution layer with convolution kernel of 1 × 1, and performing octave upsampling through the deconvolution layer to obtain a final stockline segmentation result, wherein the number of channels is consistent with the number of categories.
In this embodiment, a modular full-convolution network model loss function is also set as a cross entropy loss function, so that parameter updating is performed in the training process of the modular full-convolution network model, and an L2 parameter regular term is added;
s42, the working principle of training the modular full convolution network model comprises the following steps: initializing parameters of a characteristic extraction module, selecting parameters of a convolutional neural network model pre-trained on ImageNet, and initializing parameters of other modules of the model by adopting truncated normal distribution; optimizing model parameters by adopting a batch gradient descent method; initial learning rate of 10-4The update strategy is a reduction of 10 times per 5000 iterations; in the network optimization process, on the verification set according to the network modelThe performance index adjusts the hyper-parameter of the performance index, wherein the hyper-parameter comprises: the number of samples (batch _ size), learning rate (learning _ rate), number of network iterations, etc. are included in each batch.
In an embodiment of the aforementioned blast furnace burden line extraction method, further, the burden line division database includes: a training set, a verification set and a test set;
the training of the modular full convolution network model according to the established stockline segmentation database comprises the following steps:
training the modular full convolution network model using the training set;
optimizing and adjusting hyper-parameters of the modular full convolution network model by utilizing the verification set;
and testing and evaluating the performance of the modular full convolution network model by using the test set.
In this embodiment, the training set is a data set used for training the modular full-convolution network model, the verification set is a data set used for optimizing and adjusting hyper-parameters of the modular full-convolution network model in the training process, and the test set is a data set used for testing the performance of the modular full-convolution network model and performing performance evaluation.
In a specific embodiment of the aforementioned blast furnace burden line extraction method, further, the radar system includes: distributed array radar, scanning radar, and/or phased array radar.
The blast furnace burden line extraction method is suitable for continuous and long-time accurate monitoring of blast furnace burden line changes in the production process of a blast furnace, and is beneficial to optimizing and adjusting a material distribution strategy, improving production efficiency and realizing energy conservation and emission reduction; the applicable radar system is not limited to distributed array radar, scanning radar and/or phased array radar, and is applicable to all occasions where stockline tracking of radar in a continuous time range is utilized.
The blast furnace burden line extraction method is not limited to a blast furnace burden line extraction task, has universality in various industrial radar scenes applied to target distance detection, and is fundamentally characterized in that the method combines a modularized full convolution network model and a signal time sequence spectrogram, completes traditional signal processing in an image segmentation mode, and meets the requirements of accuracy, generalization and instantaneity.
In summary, the technical scheme of the invention has the following beneficial effects:
1) establishing a stockline segmentation database for extracting blast furnace stocklines, wherein the stockline segmentation database consists of a blast furnace field radar signal time sequence spectrogram and a labeling result thereof, can completely reflect the change rule of the stocklines in each radial region in a continuous time range in the production process of a blast furnace, has different signal time sequence spectrograms of different radars in different time periods, and provides data support for extracting the blast furnace stocklines;
2) aiming at the target of accurate extraction of a blast furnace burden line, a modular full convolution network model comprising a feature extraction module, a feature fusion module and a feature decoding module is constructed, a training set is adopted to train parameters of the modular full convolution network model, a verification set is adopted to optimize and adjust the parameters of the modular full convolution network model, a multi-radar signal time sequence spectrogram is used for training the same model, the generalization performance of the model can be effectively improved, and the problems of stockline missing, local jumping and the like easily occurring in the blast furnace environment by the existing algorithm are solved.
For better understanding of the invention, the blast furnace burden line extraction method according to the embodiment of the invention is explained based on an eight-point distributed array radar system:
the embodiment of the invention consists of two parts:
the first part is the establishment of a stockline segmentation database, which comprises data acquisition, image generation, data enhancement and data annotation and provides data support for subsequent algorithm design;
the second part is to construct a modular full convolution network model, train the parameters of the modular full convolution network model through a training set, optimize and adjust the hyper-parameters of the modular full convolution network model through a verification set, and test the performance of the modular full convolution network model on a test set.
According to the embodiment of the invention, under an Ubuntu16.04 operating system, a TensorFlow deep learning framework is built on a GPU hardware platform based on Nvidia GTX 1080Ti, and the training and testing of a modular full convolution network model are completed.
As shown in fig. 2, the blast furnace burden line extraction method according to this embodiment may include the following specific steps:
s1, collecting radar echo signals of the blast furnace field eight-point distributed array radar system;
s2, carrying out time-frequency transformation, normalization, time sequence expansion and gray mapping on the collected radar echo signals to generate a signal time sequence spectrogram;
s3, performing data enhancement on the generated signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result;
s4, constructing a modular full convolution network model comprising a feature extraction module, a feature fusion module and a feature decoding module, and training the modular full convolution network model according to the stockline segmentation database;
and S5, collecting the blast furnace radar echo signals to be detected in real time, generating a signal time sequence frequency spectrogram, and performing stockline segmentation and extraction on the signal time sequence frequency spectrogram according to the trained modular full convolution network model to realize real-time detection on the variation trend of the blast furnace stockline.
In this embodiment, the acquiring radar echo signals of the on-site eight-point distributed array radar system of the blast furnace includes:
according to a bell-less chute rotary material distribution mode adopted by a blast furnace and an installation principle of a distributed radar system, the eight-point distributed array radar system is arranged on the top of the blast furnace to meet the requirement of radial annular full coverage of a radar charge level, the installation schematic diagram is shown in figure 3, and ① - ⑧ in figure 3 shows eight radars in the eight-point distributed array radar system;
in this embodiment, the performing time-frequency transformation, normalization, time sequence expansion and gray scale mapping on the collected radar echo signals to generate a signal time sequence spectrogram includes:
as can be seen from FIG. 4, a single radar echo signal contains 1024 data points, the radar echo signal is subjected to time-frequency conversion through fast Fourier transform to obtain an amplitude spectrum of the radar echo signal, and the amplitude spectrum can be expressed as [ x ] by a vector1,x2,…x512]T(ii) a According to the limited distance between the furnace top radar and the charge level, carrying out interval interception on the amplitude frequency spectrum of each group of radar echo signals, and expressing the vector after interception as [ x [ ]1,x2,…x128]TNormalizing the amplitude; according to the continuity and time interval of radar echo signals on a time sequence, taking 256 groups of signal spectrum data after normalization processing as a unit, and forming a two-dimensional matrix in a time sequence arrangement mode, wherein the size of the two-dimensional matrix is 128 × 256; performing gray mapping on the two-dimensional matrix to generate a signal time sequence frequency spectrum gray map; in order to more finely reflect the change details of the material lines in the signal time sequence spectrogram and facilitate fine marking, the size of the signal time sequence spectrogram is adjusted to 256 × 256 by cubic interpolation, wherein the signal time sequence spectrogram of the 1#, 6#, and 7# radars is shown in fig. 5;
in this embodiment, the data enhancing the generated signal timing spectrogram and obtaining the stockline annotation result, and the establishing the stockline segmentation database according to the signal timing spectrogram after the data enhancing and the stockline annotation result includes:
performing data enhancement on the generated signal time sequence spectrogram through horizontal turnover; manually marking the stocklines in all the signal time sequence frequency spectrograms after data enhancement pixel by pixel, wherein the marking content comprises the following steps: according to the overall change trend of the stockline and local change details such as micro floating, double-layer stocklines and continuous weak faults on the stockline, a labeling picture is converted to generate a 0-1 label through a preset pixel threshold, and the local change details of the stockline are shown in fig. 6, wherein (a) is micro floating, (b) is double-layer stocklines, and (c) is continuous weak faults, and a No. 6 radar stockline labeling picture is shown in fig. 7; in the embodiment, 1#, 6#, and 7# radar echo signals are selected to establish a stockline segmentation database, which comprises 5692 signal time sequence frequency spectrograms in total, 75% of the signal time sequence spectrograms are selected as a training set, 10% of the signal time sequence spectrograms are selected as a verification set, 15% of the signal time sequence spectrograms are selected as a test set after the sequence is disturbed, and the three data sets do not comprise the same pictures;
in this embodiment, the structure of the modular full-convolution network model constructed in S4 includes: the device comprises a feature extraction module, a feature fusion module and a feature decoding module; wherein the content of the first and second substances,
a feature extraction module: selecting VGG16 as a basic model of the feature extraction module; removing the fully connected layer and the output layer in the VGG 16; the VGG16 comprises five convolution units, each convolution unit comprises a plurality of convolution layers, 1 correction linear unit and 1 batch normalization unit, and specifically: the front two convolution units are formed by connecting two convolution layers in series, the rear three convolution units are formed by connecting three convolution layers in series, and a correction linear unit and a batch normalization unit are arranged behind each convolution layer; determining the convolution kernel size in the convolution neural network to be (3, 6); the five pooling layers in the convolutional neural network are alternately connected with the convolution unit, maximum pooling is adopted, the size of a pooling window is 2 x 2, and the sliding step length is 2 x 2;
a feature fusion module: before feature map fusion, adding a convolution layer with convolution kernel size of 1 x 1 to keep the number of channels of the two feature maps consistent; after feature map fusion, reducing the number of feature map channels by half by convolution layers with convolution kernel size 1 x 1; the fusion mode among the characteristic graphs is splicing, and two characteristic graphs to be fused are respectively set as f1＝f(x1,x2,…,xn) And f2＝f(y1,y2,…,ym) Wherein x isi，yiRepresenting a single channel in the feature diagram, wherein n and m are the number of channels, and the splicing and fusion mode can be represented as follows:
fconcat＝f(x1,x2,…,xn,y1,y2,…ym)
a feature decoding module: sequentially adding three convolutional layers after the fifth pooling layer maxpool5, setting a convolution kernel with the same size of the first convolutional layer as the output feature map of maxpool5, sequentially reducing the number of channels of the feature map to enable the final number of channels of the feature map to be the number of image number categories to obtain a feature map score5, wherein the size of the convolution kernel of the subsequent two convolutional layers is 1 x 1; performing double up-sampling on the score5 through deconvolution to obtain a maxpool5 layer feature decoding result score _ up 5; fusing the result of 1 × 1 convolution operation on the feature map of the fourth pooling layer maxpool4 with score _ up5, changing the number of channels of the feature map after fusion to be consistent with maxpool4 through 1 × 1 convolution operation, and performing double up-sampling through an anti-convolution layer to obtain a feature decoding result score _ up4 of the maxpool4 layer; fusing the result of 1 × 1 convolution operation on the feature map of the third pooling layer maxpool3 with score _ up4, changing the number of channels of the feature map after fusion to be consistent with maxpool3 through the convolution layer with convolution kernel of 1 × 1, and performing octave up-sampling through the deconvolution layer to obtain the final stockline segmentation result, wherein the number of channels is 2, the channels respectively correspond to stocklines and background, and the structure diagram of the complete network model is shown in fig. 8, wherein a is a feature extraction module, b is a feature fusion module, and c is a feature decoding module; in fig. 8, conv1, conv2, conv3, conv4 and conv5 respectively represent a first convolution unit, a second convolution unit, a third convolution unit, a fourth convolution unit and a fifth convolution unit, and each convolution unit is composed of a plurality of convolution layers, a modified linear unit and a batch normalization unit; maxpool1, maxpool2, maxpool3, maxpool4 and maxpool5 respectively represent a first pooling layer, a second pooling layer, a third pooling layer, a fourth pooling layer and a fifth pooling layer, and the pooling mode of each pooling layer is maximum pooling; fuse is a feature fusion layer, and corresponding feature maps are spliced; deconv is a deconvolution layer, and performs an upsampling operation on the feature map.
In this implementation, a modular full-convolution network model loss function is set as a cross entropy loss function, and an L2 parameter regular term is added, where the loss function can be expressed as:
where m is the number of training set samples, yiTrue distribution of pixel classes, yi' is the predicted distribution of pixel classes, and ω is the model convolution kernel parameter.
In this embodiment, the loss function loss includes an L2 parameter regular term, and the loss also includes a part of an L2 regular term when performing partial derivation on a specific parameter, where the pixel category refers to: the material line and the background.
In this embodiment, the working principle of training the modular full-convolution network model includes: initializing parameters of a feature extraction module, selecting pre-trained VGG16 model parameters (parameters before pool 5) on ImageNet, and initializing parameters of other modules of the model by adopting truncated normal distribution; optimizing network model parameters by adopting a batch gradient descent method, after calculating a loss function, performing back propagation according to error loss, and adjusting a parameter omega according to the following formula:
wherein, ω isnewAnd ωoldBefore and after updating weight parameters; ε is the learning rate, the initial learning rate is 10-4The update strategy is a reduction of 10 times per 5000 iterations; performing 15000 times of network iteration training, calculating a loss function once after each iteration, and updating parameters; in the iteration process, if the error of the verification set gradually becomes larger from the gradual reduction, the network is considered to be over-fitted, the training is terminated, and the hyper-parameters are adjusted until the network is optimal; after training, sending the test set into a network model for stockline extraction, performing forward propagation only once in the extraction process, and performing pixel conversion on a network output result to visualize the extracted stockline, wherein the test set segmentation result is shown in fig. 9;
in this embodiment, the real-time collection of blast furnace radar echo signals to be detected generates a signal timing sequence spectrogram, and according to a trained modularized full convolution network model, the signal timing sequence spectrogram is subjected to stockline segmentation and extraction, and the realization of the real-time detection of the variation trend of the blast furnace stockline includes:
in the production process of the blast furnace, radar echo signals are collected in real time through an eight-point distributed array radar system to generate a signal time sequence spectrogram, the signal time sequence spectrogram is sent to a trained network model, stocklines in the network model are segmented and extracted, the change trend of the blast furnace stocklines in the production process is detected in real time, a comparison result with the existing algorithm is shown in fig. 10, wherein gray represents stocklines extracted through a Finite Impulse Response (FIR) band-pass filter combined with Fast Fourier Transform (FFT) and ChirpZ (ChirpZ is a frequency spectrum refining algorithm) frequency spectrum correction algorithm, and black represents stocklines extracted by the method.
The results of the above embodiments show that the method of the present invention can accurately and continuously monitor the blast furnace burden line in real time in the severe environment of the blast furnace, and the performance of the method is consistent in different radars and different times, and the generalization performance is good.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.
While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (10)
1. A blast furnace burden line extraction method, characterized by comprising:
collecting a field radar echo signal of the blast furnace, wherein radar waves emitted by a radar system can cover all areas of the radial charge level of the blast furnace;
generating a signal time sequence spectrogram of the radar echo signal;
performing data enhancement on the signal time sequence spectrogram, acquiring a stockline marking result, and establishing a stockline segmentation database according to the signal time sequence spectrogram after the data enhancement and the stockline marking result;
constructing a modular full-convolution network model, and training the modular full-convolution network model according to the established stockline segmentation database;
the method comprises the steps of collecting blast furnace radar echo signals to be detected in real time, generating a signal time sequence spectrogram, and carrying out stockline segmentation and extraction on the signal time sequence spectrogram according to a trained modular full convolution network model.
2. The blast furnace burden line extraction method of claim 1, wherein said collecting blast furnace site radar echo signals comprises:
according to a bell-less chute rotary distribution mode adopted by the blast furnace and the installation principle of a radar system, the corresponding radar system is arranged on the top of the blast furnace to meet the requirement of full coverage of the radial charge level of the blast furnace;
and radar echo signals are continuously collected in real time in the production process of the blast furnace.
3. The blast furnace burden line extraction method of claim 1, wherein said generating a signal time series spectrogram of said radar return signal comprises:
and carrying out time-frequency transformation, normalization, time sequence expansion and gray level mapping on the collected radar echo signals to generate a signal time sequence spectrogram of the radar echo signals.
4. The blast furnace burden line extraction method of claim 3, wherein the performing time-frequency transformation, normalization, time sequence expansion and gray scale mapping on the collected radar echo signals to generate a signal time sequence spectrogram of the radar echo signals comprises:
performing time-frequency conversion on the radar echo signal through fast Fourier transform to obtain an amplitude frequency spectrum of the radar echo signal;
according to the distance between the furnace top radar and the charge level, carrying out interval interception on the amplitude frequency spectrum of each group of radar echo signals, and carrying out normalization processing on the amplitude;
according to the continuity of radar echo signals on a time sequence, taking continuous M groups of signal spectrum data after normalization processing as a unit, and forming a two-dimensional matrix by time sequence arrangement;
and carrying out gray mapping on the two-dimensional matrix to generate a signal time sequence frequency spectrum gray map.
5. The blast furnace burden line extraction method of claim 4, wherein after said generating a signal time series frequency spectrum gray scale map, said method further comprises:
and stretching the height of the signal time sequence frequency spectrum gray-scale map by cubic interpolation.
6. The blast furnace burden line extraction method of claim 1, wherein said data enhancing the signal time series spectrogram comprises:
and performing data enhancement on the generated signal time sequence spectrogram through horizontal inversion.
7. The blast furnace burden line extraction method of claim 1, wherein the obtaining of the burden line labeling result comprises:
acquiring a pixel-by-pixel labeling result of a stockline in a signal time sequence spectrogram after data enhancement, wherein the labeling content comprises: the change trend of the stockline, and floating, double-layer stockline and continuous fault on the stockline;
and converting pixel values which are greater than or equal to the preset pixel threshold value in the marked picture into first identifications and converting pixel values which are smaller than the preset pixel threshold value into second identifications through the preset pixel threshold value, and generating first identification-second identification label pictures which are in one-to-one correspondence with the signal time sequence frequency spectrogram.
8. The blast furnace burden line extraction method of claim 7, wherein said constructing a modular full convolution network model comprises:
constructing a modular full-convolution network model comprising a feature extraction module, a feature fusion module and a feature decoding module; wherein the content of the first and second substances,
the characteristic extraction module is formed by crossing a plurality of convolution units and pooling layers and is used for learning the stockline characteristics in the input signal time sequence spectrogram through the convolution units of different layers to obtain characteristic graphs of different layers and different sizes;
the feature fusion module is used for fusing the feature graphs of different levels and different sizes extracted by the feature extraction module;
and the characteristic decoding module is used for carrying out up-sampling operation on the characteristic graph to obtain network output with the same size as the label picture, changing the number of output channels to keep the number of the output channels consistent with the number of the image categories, and obtaining a stockline segmentation result.
9. The blast furnace burden line extraction method of claim 1, wherein the burden line segmentation database comprises: a training set, a verification set and a test set;
the training of the modular full convolution network model according to the established stockline segmentation database comprises the following steps:
training the modular full convolution network model using the training set;
optimizing and adjusting hyper-parameters of the modular full convolution network model by utilizing the verification set;
and testing and evaluating the performance of the modular full convolution network model by using the test set.
10. The blast furnace burden line extraction method of claim 1, wherein the radar system comprises: distributed array radar, scanning radar, and/or phased array radar.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268646.2A CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268646.2A CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109490861A CN109490861A (en) | 2019-03-19 |
| CN109490861B true CN109490861B (en) | 2020-06-02 |
Family
ID=65693310
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811268646.2A Active CN109490861B (en) | 2018-10-29 | 2018-10-29 | Blast furnace burden line extraction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109490861B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110647891B (en) * | 2019-09-17 | 2023-01-24 | 上海仪电（集团）有限公司中央研究院 | CNN (convolutional neural network) -based automatic extraction method and system for time sequence data characteristics of self-encoder |
| CN110819751B (en) * | 2020-01-09 | 2020-04-10 | 江苏金恒信息科技股份有限公司 | Blast furnace burden line obtaining method and device based on radar data processing |
| CN111449644A (en) * | 2020-03-19 | 2020-07-28 | 复旦大学 | Bioelectricity signal classification method based on time-frequency transformation and data enhancement technology |
| CN111784642B (en) * | 2020-06-10 | 2021-12-28 | 中铁四局集团有限公司 | Image processing method, target recognition model training method and target recognition method |
| CN113627283B (en) * | 2021-07-23 | 2024-05-24 | 中冶南方工程技术有限公司 | Material level measuring method based on radar echo signals, terminal equipment and storage medium |
| CN114136194B (en) * | 2021-10-12 | 2024-06-14 | 江苏丰尚智能科技有限公司 | Method, device, monitoring equipment and storage medium for monitoring volume of materials in warehouse |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107993215A (en) * | 2017-11-27 | 2018-05-04 | 象辑知源（武汉）科技有限公司 | A kind of weather radar image processing method and system |
| CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
| CN108062753A (en) * | 2017-12-29 | 2018-05-22 | 重庆理工大学 | The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study |
| CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
| CN108647568A (en) * | 2018-03-30 | 2018-10-12 | 电子科技大学 | Grassland degeneration extraction method based on full convolutional neural networks |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102014200928A1 (en) * | 2014-01-20 | 2015-07-23 | Tmt Tapping Measuring Technology Sàrl | Device for determining the topography of the Möller surface in a shaft furnace |
-
2018
- 2018-10-29 CN CN201811268646.2A patent/CN109490861B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
| CN107993215A (en) * | 2017-11-27 | 2018-05-04 | 象辑知源（武汉）科技有限公司 | A kind of weather radar image processing method and system |
| CN108062753A (en) * | 2017-12-29 | 2018-05-22 | 重庆理工大学 | The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study |
| CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
| CN108647568A (en) * | 2018-03-30 | 2018-10-12 | 电子科技大学 | Grassland degeneration extraction method based on full convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| Radar Data Processing of Blast Furnace Stock-Line Based on Spatio-Temporal Data Association;Hou Qingwen 等;《2015 34TH CHINESE CONTROL CONFERENCE (CCC)》;20151231;全文 * |
| 高炉雷达料面测量信号处理***改进;陈先中 等;《北京科技大学学报》;20110228;第33卷(第2期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109490861A (en) | 2019-03-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109490861B (en) | Blast furnace burden line extraction method | |
| CN111583214B (en) | Sea surface wind speed inversion method based on RBF neural network and based on marine radar image | |
| CN110969088B (en) | Remote sensing image change detection method based on significance detection and deep twin neural network | |
| CN108875905B (en) | Direct inversion method for visibility function of atmospheric temperature and humidity profile | |
| CN110456355B (en) | Radar echo extrapolation method based on long-time and short-time memory and generation countermeasure network | |
| CN102110227B (en) | Compound method for classifying multiresolution remote sensing images based on context | |
| CN105139395A (en) | SAR image segmentation method based on wavelet pooling convolutional neural networks | |
| CN110702792B (en) | Alloy tissue ultrasonic detection classification method based on deep learning | |
| CN105469359A (en) | Locality-constrained and low-rank representation based human face super-resolution reconstruction method | |
| CN113780242A (en) | Cross-scene underwater sound target classification method based on model transfer learning | |
| CN111624606B (en) | Radar image rainfall identification method | |
| CN114758230A (en) | Underground target body classification and identification method based on attention mechanism | |
| CN114970378A (en) | Sea clutter sample library construction method based on GAN network | |
| CN117009916A (en) | Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning | |
| CN113468804B (en) | Underground pipeline identification method based on matrix bundles and deep neural network | |
| Xu et al. | Detection method of tunnel lining voids based on guided anchoring mechanism | |
| CN117574202A (en) | Radar antenna scanning mode identification method based on multi-time sequence feature fusion | |
| CN114814776B (en) | PD radar target detection method based on graph attention network and transfer learning | |
| CN115598714B (en) | Time-space coupling neural network-based ground penetrating radar electromagnetic wave impedance inversion method | |
| CN117351321A (en) | Single-stage lightweight subway lining cavity recognition method and related equipment | |
| CN116935213A (en) | Knowledge distillation-based light SAR image target detection method | |
| CN114021422B (en) | Underground structure internal defect identification method based on cross-hole radar and deep learning | |
| CN113627480B (en) | Polarization SAR image classification method based on reinforcement learning | |
| CN107341798A (en) | High Resolution SAR image change detection method based on global local SPP Net | |
| CN111965601A (en) | Underwater sound source passive positioning method based on nuclear extreme learning machine |
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 |