Electrically large complex target identification method based on gated recursive residual network
Through the method based on the gated recursive residual network, combined with the residual convolutional network and the GRU gated loop unit, the problem of unstable feature extraction in the traditional radar target extraction method is solved, and high-precision and fast recognition of complex large objects is achieved. , suitable for practical engineering applications.
Patent Information
- Application Number
- CN202411020491.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-05
AI Technical Summary
Traditional radar target extraction methods rely on radar system performance and researchers' technical experience, resulting in unstable feature extraction and affecting the accuracy of target recognition, especially in the recognition of complex target recognition by TVU.
Using a method based on a gated recursive residual network, combining the residual convolution network and the GRU gated loop unit, the spatial scattering information and the gated loop unit model the time series information through a one-dimensional convolution kernel, and using jump connections and self-directed loop units The attention mechanism layer dynamically allocates feature weights to form the final feature representation of the complex goals of TVU.
It realizes high-precision, fast and convenient identification of complex objects in large and large quantities, avoids gradient explosion and gradient disappearance problems, reduces network training time and parameter amount, and is suitable for practical engineering applications.
Smart Images

Figure CN118897978A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for automatically identifying electrically large objects according to electromagnetic scattering characteristics, and in particular to a method for identifying electrically large complex targets based on a gated recursive residual network. Background Art
[0002] As an indispensable and important equipment in the military field, the intelligence of radar is crucial for perceiving the ever-changing battlefield environment. With the complexity of modern warfare environment and the diversification of hostile targets and tasks, how to accurately identify targets in harsh environments is the key to radar detection. One-dimensional range image can better reflect the scattering characteristics and geometric structure information of radar targets. It has the advantages of simple imaging, easy acquisition, small data volume, and low computational complexity. It is especially suitable for the recognition and extraction of electrically large and complex targets, and has high research value and application prospects.
[0003] The traditional radar target extraction method needs to extract the features of the one-dimensional range image first, build the feature data of different targets into a feature database, and classify them in the classifier. The extraction of feature parameters not only depends on the performance of the radar system, but also on the technology and experience of the researchers, resulting in unstable extracted features. The quality of the classifier also directly affects the accuracy of target recognition and extraction. Therefore, a more effective target detection method is urgently needed to accurately and quickly extract radar target echo signals.
[0004] Based on the above background, the present invention proposes a method for identifying large electrical objects based on a gated recursive residual network. Among them, the residual convolution network can use a one-dimensional convolution kernel and a residual block to effectively capture the scattered information in each distance unit, while minimizing the number of parameters; the GRU residual gated recurrent unit network module aggregates the scattered information along the spatial dimension to construct an object feature representation. This model fully mines the temporal and spatial feature information of the one-dimensional range image data from two modules, the residual convolution and the residual gated recurrent unit, in parallel, uses a fusion module to splice the two parts of the temporal and spatial information, and dynamically allocates the fusion feature weights through the self-attention mechanism layer to form the final feature representation. Both the network convolution block and the gated recurrent unit use jump connections, which effectively avoids problems such as gradient explosion and gradient disappearance. Without the help of expert experience and complex mechanism knowledge, it can achieve high-precision, fast and convenient recognition of large electrical complex objects. Summary of the invention
[0005] In response to the above needs, the present invention aims to provide a method for recognizing electrically large complex targets by combining residual networks and gated recurrent units. In the ResNet-GRU model, the residual network uses one-dimensional convolution kernels and residual blocks to effectively capture the spatial scattering information within each distance unit; the gated recurrent unit network models the data time series related information, and selectively memorizes the data information through the gating mechanism. On this basis, a skip connection mechanism is introduced to avoid problems such as gradient explosion and gradient disappearance; taking into account the complementarity between different features, the fusion module fuses the deep high-level features mined by the convolution kernel and the residual module with the time series information features aggregated by the gated recurrent unit, and dynamically allocates the fused feature weights through the self-attention mechanism layer to form the final feature representation of the target.
[0006] In order to achieve the above object, the technical solution steps of the present invention are:
[0007] A method for recognizing complex targets based on a gated recursive residual network comprises the following steps:
[0008] Step S1, obtaining geometric shape data of a complex target of a TVU through surveying and mapping, and establishing a fine geometric model of the complex target of a TVU;
[0009] Step S2, meshing the fine geometric model;
[0010] Step S3, using a high-frequency approximation algorithm on the divided grid to quickly obtain a one-dimensional range image database of electrically large and complex targets, wherein the one-dimensional range image database contains one-dimensional range image data, and obtains a dynamic distribution map of the electric field;
[0011] Step S4, preprocessing the one-dimensional range image data obtained in step S3;
[0012] Step S5, establishing a ResNet-GRU neural network structure, wherein the ResNet-GRU neural network includes a residual convolutional network, a GRU gated recurrent unit network, a fusion module and a self-attention layer; mining the one-dimensional range image data of the complex target of the TVU in parallel from the residual convolutional network and the residual gated recurrent unit module; the fusion module fuses the deep high-level features mined by the residual convolutional network and the time series information features aggregated by the residual gated recurrent unit; dynamically allocating the fusion feature weights through the self-attention mechanism layer to form the final feature representation of the complex target of the TVU;
[0013] Step S6, training the constructed ResNet-GRU neural network: input the preprocessed one-dimensional range image database into the ResNet-GRU neural network, and train it by using the Adam gradient descent algorithm and the SGD gradient descent algorithm, using a variable learning rate to optimize the cross loss function, and adding dropout in the training process to prevent overfitting;
[0014] Step S7, classification of electrically large complex targets: input the one-dimensional range image data of the electrically large complex object to be tested into the trained ResNet-GRU neural network, and output the classification name of the electrically large complex object to be tested.
[0015] Preferably, the specific steps in step S3 are:
[0016] S31: Select corresponding grids for segmentation according to different complex targets of electrical engineering;
[0017] S32: The antenna transmits a linear frequency modulation signal, and the probe receives an echo signal;
[0018] S33: Pulse compression is performed on the echo signal to obtain one-dimensional range image data, and at the same time, a dynamic electric field map is obtained through a time domain solver.
[0019] Preferably, the specific steps of step S4 are:
[0020] S41: Use the maximum and minimum normalization method to eliminate the amplitude sensitivity of the one-dimensional range image data;
[0021] S42: The centroid alignment method is used to eliminate the translation sensitivity of one-dimensional range image data.
[0022] Preferably, the residual convolutional network of the ResNet module in step S5 includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer, and the number of channels is 16, 32, 64, and 128 respectively; wherein the first convolutional layer performs a convolution operation and a maximum pooling operation; the second convolutional layer, the third convolutional layer, and the fourth convolutional layer are all composed of two first residual blocks, and each first residual block is composed of two layers of convolutional networks:
[0023] The preprocessed one-dimensional range image data is input into the first convolutional layer, and the formula is as follows (1):
[0024]
[0025] Where: represents the jth weight of the i-th convolution kernel, represents the jth convolved local area r in the mth convolutional layer, * is the convolution operation, and m is 1; where:
[0026] Then, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer are used to extract deep high-level features. The formula is as follows (2):
[0027] y l =h(x l )+F(x l , w l ) (2)
[0028] xl+1 =f(y l ) (3)
[0029] Where x l is the input of the lth convolutional layer, that is, y in formula (1) m(i,j) , w l is the weight of the first residual block, F(x l ,w l ) is the function module of the first residual block, f is the activation function ReLU, h(x l ) is x l Convolutional identity mapping of ;
[0030] S2: GRU gated recurrent unit network introduces an update gate z in the hidden layer t and reset gate r t The two gate controllers are used as memory units, and the calculation steps are:
[0031] z t =σ(w z ·[h t-1 ,x t ]) (4)
[0032] r t =σ(w z ·[h t-1 , x t ]) (5)
[0033]
[0034] In the formula, σ is the activation function, w z and w z They are update gate and reset gate r respectively. t The weight of is the hidden layer, h t is the extracted time series features;
[0035] S3: The deep high-level features are fused with the time series features to obtain the fused features, and then the fused features are input into the self-attention layer.
[0036] Preferably, in step S5,
[0037] The GRU gated recurrent unit network includes three GRU modules. A skip connection is introduced in the second GRU module. A normalization layer and a RuLU activation function are added between the second GRU module and the third GRU module. Assuming that the output of the last layer of the skip-connected GRU module is y, the output y of the second residual block is R It is expressed as:
[0038] y R =relu(BNγ,β (y)+g(x t )) (8)
[0039] Where relu(·) represents the relu activation function; BN(·) is the batch normalization function; γ and β represent two learnable variables in the function; g(·) is the adjustment function that makes x t With h t have the same dimensions.
[0040] Preferably, in step S5,
[0041] Pass the fused features to the self-attention layer: Calculate the query vector Q and key vector K in the self-attention based on the fusion parameters, and the fusion feature weight calculation formula with the input is:
[0042]
[0043] Where d is the hidden dimension.
[0044] Preferably, the calculation formulas of Q and K are:
[0045] q i =W Q ·x i (10)
[0046] k i =W k ·x i (11)
[0047] Where W Q and W k is the projection matrix.
[0048] Preferably, in step S2, the frequency range of the electromagnetic wave is 2 GHz to 4 GHz, the antenna pitch excitation direction is 90°, and the antenna azimuth excitation direction is 0 to 360°, with an interval of 1°.
[0049] Preferably, in step S6, the one-dimensional range image data input to the ResNet-GRU neural network is a complex combination of amplitude and phase.
[0050] Preferably, in step S7, the one-dimensional range image data to be identified is input into the network that has been fully trained in step S6, and the category of the large complex target is output.
[0051] Beneficial effects:
[0052] The present invention has the following advantages: (1) the time domain electric field waveform and one-dimensional distance image are given, that is, a one-dimensional distance image database of electrically large complex objects is established to characterize the electromagnetic scattering characteristics of the surface of electrically large complex objects; (2) the residual convolution network in the ResNet-GRU model uses a one-dimensional convolution kernel and a residual block to effectively capture the scattering information in each distance unit; the residual gated recurrent unit network models the data time series related information, and introduces a gating mechanism to selectively memorize data information; considering the complementarity between different features, the fusion module fuses the deep high-level features captured by the convolution kernel and the residual module with the time series information features aggregated by the gated recurrent unit, and dynamically allocates the fused feature weights through the self-attention mechanism layer to form the final feature representation of the target. The combined network effectively avoids the problems of gradient vanishing and gradient explosion that occur in traditional deep networks, and can make full use of sequence information, greatly reducing the network training time while ensuring training accuracy, and realizing network lightweighting. (3) The deep learning algorithm provided by the present invention can realize a fast, high-precision, maintainable and convenient electrical large complex object recognition method without relying on expert experience and complex mechanism knowledge. The model has a small number of parameters and high recognition accuracy, and is suitable for practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 The present invention provides a flow chart of a method for recognizing complex targets based on a gated recursive residual network.
[0054] Figure 2 A geometric model of a car, which is an electrically large complex object, is provided in an embodiment of the present invention.
[0055] Figure 3 It is the one-dimensional range image data of the car in the embodiment of the present invention.
[0056] Figure 4 Graph showing the time domain electric field distribution of a car in an embodiment of the present invention.
[0057] Figure 5 It is an embodiment of the present invention to verify the correctness of simulation and test data;
[0058] Figure 6 The embodiment of the present invention provides a ResNet-GRU network structure;
[0059] Figure 7 This is a GRU network diagram in an embodiment of the present invention;
[0060] Figure 8 This is a network diagram of the self-attention layer in an embodiment of the present invention. DETAILED DESCRIPTION
[0061] The following is based on Figures 1 to 8 , specifically describe the preferred embodiments of the present invention.
[0062] like Figure 1 As shown, the present invention provides a method for recognizing complex targets based on a gated recursive residual network, comprising the following steps:
[0063] Step S1: obtain geometric shape data of large and complex objects through surveying and mapping, and establish a fine geometric model database of large and complex objects. Take a car as an example. Figure 2 The traffic scene model database includes a bus, an electric bicycle, five different types of cars and a traffic light;
[0064] Step S2, meshing the fine geometric model, the electromagnetic wave frequency range is 2 GHz to 4 GHz, the antenna pitch excitation direction is 90°, the antenna azimuth excitation direction is 0 to 360°, and the interval is 1°;
[0065] Step S3: Use a high-frequency approximation algorithm on the divided grid to quickly obtain a one-dimensional range image database of a large and complex target scattering, such as Figure 3 , the dynamic distribution diagram of the electric field is obtained by solving in the time domain, such as Figure 4 ;
[0066] Step S4: preprocess the one-dimensional range image database obtained in step S3. Use the maximum and minimum normalization method to reduce the amplitude sensitivity of the one-dimensional range image. The formula is as follows:
[0067]
[0068] x=[x1,x2,x3…x n ] is a frame of data.
[0069] The centroid alignment method is used to eliminate the translation sensitivity of the one-dimensional range image. The formula is as follows:
[0070]
[0071] W is the data center, x n For x = [x1, x2, x3…x n ] is the amplitude value of the nth distance unit, and N is the total number of distance units in a frame of data. After calculating the center of gravity, the one-dimensional distance image data is translated, and the translation length d is:
[0072]
[0073] Verify the correctness of electromagnetic simulation, such as Figure 5 ;
[0074] Step S5: Establish a ResNet-GRU neural network structure, wherein the ResNet-GRU neural network includes a residual convolutional network, a residual gated recurrent unit, a fusion module and a self-attention layer network, such as Figure 6 ; Parallel mining of target one-dimensional range image data feature information from residual convolutional network and residual gated recurrent unit modules;
[0075] The residual convolutional network of the ResNet module includes the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer, with the number of channels being 16, 32, 64, and 128 respectively; the first convolutional layer performs convolution and maximum pooling operations, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer are each composed of two residual blocks, each of which is composed of two layers of convolutional networks. The feature matrices of the two branches of the residual block are added together and then output through the activation function;
[0076] The initial information of the one-dimensional range image of the complex target is input into the residual convolutional network, and the formula is as follows:
[0077]
[0078] Where: represents the jth weight of the i-th convolution kernel, represents the jth convolved local area r in the lth convolutional layer, * is the convolution operation;
[0079] Then, the residual convolution module in the ResNet module is used to process the information, deeply extract the one-dimensional distance image information, and further explore the data connection. The formula is as follows:
[0080] y l =h(x l )+F(x l ,w l )
[0081] x 1+1 =f(y l )
[0082] Where x l is the input of the lth layer in the residual unit, w l is the weight of the residual unit, F(x l ,w l ) is the function module of the residual unit, f is the activation function ReLU, h(x l ) is x l Convolutional identity map of .
[0083] The GRU gated recurrent unit network includes three GRU modules, the number of hidden units is 64, a skip connection is introduced in the second GRU gated recurrent unit network, and a normalization layer and a RuLU activation function are added between the second GRU and the third GRU. The calculation formula is as follows:
[0084] z t =σ(w z ·[h t-1 ,x t ])
[0085] r t =σ(w z ·[h t-1 ,x t ])
[0086]
[0087]
[0088] Assuming that the output of the last layer of the skip-connected GRU is y, the output of the residual block is y R It can be expressed as:
[0089] y R =reluu(BN γ,β (y)+g(x t ))
[0090] Where relu(·) represents the relu activation function; BN(·) is the batch normalization function; γ and β represent two learnable variables in the function; g(·) is the adjustment function that makes x t With h t have the same dimensions.
[0091] The fusion module fuses the deep high-level features mined by the residual convolutional network and the time series information features aggregated by the residual gated recurrent unit; dynamically allocates the fusion feature weights through the self-attention mechanism layer to form the final feature representation of the target, and calculates the query vector Q and key vector K in the self-attention based on the fusion parameters. The calculation formula of the fusion feature weight with the input is:
[0092]
[0093] q i =W Q ·x i
[0094] k i =W k ·x i
[0095] Where W Q and Wk is the projection matrix and d is the hidden dimension.
[0096] Step S6, training the constructed ResNet-GRU neural network: input the preprocessed one-dimensional range image database into the ResNet-GRU neural network, and train it using the Adam gradient descent algorithm and the SGD gradient descent algorithm, using a variable learning rate to optimize the cross loss function, with an initial learning rate of 0.1, which is reduced by half every 20 rounds, and adding dropout in the training process to prevent overfitting;
[0097] Step S7, classification of electrically large complex targets: input the one-dimensional range image data of the electrically large object to be tested into the trained ResNet-GRU neural network, and output various classification names of electrically large complex targets.
[0098] The present invention has the following advantages: (1) It provides a time-domain electric field waveform and a one-dimensional distance image, that is, a one-dimensional distance image database of complex objects is established to characterize the electromagnetic scattering characteristics of the surface of complex objects; (2) The residual network in the ResNet-GRU model uses a one-dimensional convolution kernel and a residual block to effectively capture the spatial scattering information in each distance unit; the gated recurrent unit network models the data time series related information, and introduces a gating mechanism to selectively memorize the data information; considering the complementarity between different features, the fusion module fuses the deep high-level features mined by the convolution kernel and the residual module with the time series information features aggregated by the gated recurrent unit, and dynamically allocates the fusion feature weights through the self-attention mechanism layer to form the final feature representation of the target. (3) The present invention provides a fast, high-precision, maintainable and convenient method for recognizing complex objects based on a deep learning algorithm without the help of expert experience and complex mechanism knowledge. The model has the characteristics of small size, fast speed and strong portability, and is suitable for application in practical engineering applications.
[0099] The advantages of the present invention over other existing network structures are shown in the following table:
[0100] network Parameter quantity Average recognition rate VGG19 29.6M 92.63% ResNet18 3.7M 94.16% The present invention 0.74M 97.46%
[0101] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.
Claims
1. A method for recognizing complex targets based on gated recursive residual networks, characterized in that: The following steps are involved: Step S1, obtaining geometric shape data of a complex target of a TVU through surveying and mapping, and establishing a fine geometric model of the complex target of a TVU; Step S2, meshing the fine geometric model; Step S3, using a high-frequency approximation algorithm on the divided grid to quickly obtain a one-dimensional range image database of electrically large and complex targets, wherein the one-dimensional range image database contains one-dimensional range image data, and obtains a dynamic distribution map of the electric field; Step S4, preprocessing the one-dimensional range image data obtained in step S3; Step S5, establishing a ResNet-GRU neural network structure, wherein the ResNet-GRU neural network includes a residual convolutional network, a GRU gated recurrent unit network, a fusion module and a self-attention layer; mining the one-dimensional range image data of the complex target of the TVU in parallel from the residual convolutional network and the residual gated recurrent unit module; the fusion module fuses the deep high-level features mined by the residual convolutional network and the time series information features aggregated by the residual gated recurrent unit; dynamically allocating the fusion feature weights through the self-attention mechanism layer to form the final feature representation of the complex target of the TVU; Step S6, training the constructed ResNet-GRU neural network: input the preprocessed one-dimensional range image database into the ResNet-GRU neural network, and train it by using the Adam gradient descent algorithm and the SGD gradient descent algorithm, using a variable learning rate to optimize the cross loss function, and adding dropout in the training process to prevent overfitting; Step S7, classification of electrically large complex targets: input the one-dimensional range image data of the electrically large complex object to be tested into the trained ResNet-GRU neural network, and output the classification name of the electrically large complex object to be tested.
2. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that: The specific steps in step S3 are: S31: Select corresponding grids for segmentation according to different complex targets of electrical engineering; S32: The antenna transmits a linear frequency modulation signal, and the probe receives an echo signal; S33: Pulse compression is performed on the echo signal to obtain one-dimensional range image data, and at the same time, a dynamic electric field map is obtained through a time domain solver.
3. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that: The specific steps of step S4 are: S41: Use the maximum and minimum normalization method to eliminate the amplitude sensitivity of the one-dimensional range image data; S42: The centroid alignment method is used to eliminate the translation sensitivity of one-dimensional range image data.
4. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that: The residual convolutional network of the ResNet module in step S5 includes the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer, and the number of channels is 16, 32, 64, and 128 respectively; wherein the first convolutional layer performs convolution operation and maximum pooling operation; the second convolutional layer, the third convolutional layer, and the fourth convolutional layer are all composed of two first residual blocks, and each first residual block is composed of two layers of convolutional networks: The preprocessed one-dimensional range image data is input into the first convolutional layer, and the formula is as follows (1): Where: represents the jth weight of the i-th convolution kernel, represents the jth convolved local area r in the mth convolutional layer, * is the convolution operation, and m is 1; where: Then, the second convolutional layer, the third convolutional layer, and the fourth convolutional layer are used to extract deep high-level features. The formula is as follows (2): y l =h(x l )+F(x l ,w l ) (2) x l+1 =f(y l ) (3) Where x l is the input of the lth convolutional layer, that is, y in formula (1) m(i,j) , w l is the weight of the first residual block, F(x l ,w l ) is the function module of the first residual block, f is the activation function ReLU, h(x l ) is x l Convolutional identity mapping of ; S2: GRU gated recurrent unit network introduces an update gate z in the hidden layer t and reset gate r t The two gate controllers are used as memory units, and the calculation steps are: z t =σ(w z ·[h t-1 ,x t ]) (4) r t =σ(w z ·[h t-1 ,x t ]) (5) In the formula, σ is the activation function, w z and w z They are update gate and reset gate r respectively. t The weight of is the hidden layer, h t is the extracted time series features; S3: The deep high-level features are fused with the time series features to obtain the fused features, and then the fused features are input into the self-attention layer.
5. The method for recognizing complex targets based on gated recursive residual networks according to claim 4, characterized in that: In step S5, The GRU gated recurrent unit network includes three GRU modules. A skip connection is introduced in the second GRU module. A normalization layer and a RuLU activation function are added between the second GRU module and the third GRU module. Assuming that the output of the last layer of the skip-connected GRU module is y, the output y of the second residual block is R It is expressed as: and R =relu(BN γ,β (y)+g(x t )) (8) Where relu(·) represents the relu activation function; BN(·) is the batch normalization function; γ and β represent two learnable variables in the function; g(·) is the adjustment function that makes x t With h t have the same dimensions.
6. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that In step S5, Pass the fused features to the self-attention layer: Calculate the query vector Q and key vector K in the self-attention based on the fusion parameters, and the fusion feature weight calculation formula with the input is: Where d is the hidden dimension.
7. The method for recognizing electrically large complex targets based on a gated recursive residual network as claimed in claim 6, characterized in that: The calculation formulas of Q and K are: q i =W Q ·x i (10) k i =W k ·x i (11) Where W Q and W k is the projection matrix.
8. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that: In step S2, the frequency range of the electromagnetic wave is 2 GHz to 4 GHz, the antenna elevation excitation direction is 90°, and the antenna azimuth excitation direction is 0 to 360° with an interval of 1°.
9. The method for identifying complex targets based on gated recursive residual networks according to claim 1, characterized in that: In step S6, the one-dimensional range image data input to the ResNet-GRU neural network is a complex combination of amplitude and phase.
10. The method for recognizing complex targets based on gated recursive residual networks according to claim 1, characterized in that: In step S7, the one-dimensional range image data to be identified is input into the network that has been fully trained in step S6, and the category of the complex target is output.
Citation Information
Patent Citations
Network anomaly detection method based on deep neural network
CN113516228A
Deep neural network target identification method based on radar scene
CN115902879A
