Phased array radar working mode identification method based on attention mechanism and feature fusion

By using DenseNet neural network based on attention mechanism and feature fusion in phased array radar recognition, combined with the CBAM module, the problem of low radar operating mode recognition accuracy is solved, and more efficient feature extraction and recognition performance is achieved.

CN118940103APending Publication Date: 2024-11-12HOHAI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410963578.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The working mode of phased array radar is flexible and varied, the existing recognition methods have low recognition accuracy, and it is difficult for deep learning to pay attention to the connection between various features of the radar signal. Insufficient feature parameters affect the recognition effect.

Method used

Using an attention mechanism and feature fusion method, the DenseNet neural network is used to add the CBAM module to enhance the feature retention of key information, suppress noise and irrelevant information through feature enhancement, extraction, fusion and attention mechanisms, and improve identification performance.

Benefits of technology

It improves the accuracy and performance of radar pattern recognition, can pay more effectively to the importance of channels and space in the feature map, and enhances the feature extraction and recognition capabilities of radar signals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940103A_ABST
    Figure CN118940103A_ABST
Patent Text Reader

Abstract

The invention discloses a phased array radar working mode recognition method based on an attention mechanism and feature fusion, and the method comprises the steps: carrying out the feature enhancement of a radar signal pulse sequence according to the obtained radar signal pulse sequence data, and scaling a part with an overlarge TOA (Time of Arrival) difference value in the radar signal pulse sequence; feature extraction is carried out on the radar signal pulse sequence, and images corresponding to all feature parameters of the processed radar pulse sequence are generated; feature fusion is carried out on the radar signal pulse sequence, and images drawn by all feature parameters are fused into one image; and inputting the radar signal fusion image into a DenseNet neural network based on an attention mechanism, and after the radar signal fusion image passes through a multi-layer network, outputting the radar signal fusion image through a Softmax classifier to obtain a classification result. According to the method, improvement is carried out on the basis of Densenet, the CBAM module is added, attention features can be enhanced, key information is reserved, noise and irrelevant information are restrained, the radar mode recognition capability is improved, and the recognition performance is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of radar design, and relates to phased array radar working mode recognition, and specifically to a phased array radar working mode recognition method based on attention mechanism and feature fusion. Background Art

[0002] The flexible and changeable working modes of phased array radars pose a severe challenge to the recognition methods of phased array radar working modes. Common radar working mode recognition methods include radar phrases, deep learning and clustering methods. The use of radar phrases is to represent the intercepted pulse stream as radar words, radar phrases, and radar sentences, and match the obtained radar phrases with the radar phrases in the knowledge base to realize the recognition of the working mode. The deep learning and clustering methods are based on an in-depth analysis of the differences between the various features of different working modes, and use deep learning or clustering algorithms to realize the recognition of the working mode of phased array radars.

[0003] However, the radar word and radar phrase methods use insufficient feature parameters, which will lead to unsatisfactory recognition results. The working mode of phased array radar is flexible and changeable, and the radar signal changes in various working modes are highly diverse, which will seriously affect the results of clustering methods. It is difficult to focus on the connection between the various features of radar signals using only deep learning methods. Summary of the invention

[0004] Purpose of the invention: In order to overcome the problem that the flexible and changeable working modes of phased array radars in the prior art bring great difficulties to the recognition method of phased array radar working modes, and the recognition accuracy needs to be improved, a phased array radar working mode recognition method based on attention mechanism and feature fusion is provided, and the basic deep learning neural network used is DenseNet (densely connected neural network). And improvements are made on the basis of Densenet, and a CBAM module (channel and spatial attention mechanism module) is added therein, which can enhance the attention feature to retain key information while suppressing noise and irrelevant information, thereby improving the radar pattern recognition ability and recognition performance.

[0005] Technical solution: To achieve the above purpose, the present invention provides a phased array radar working mode recognition method based on attention mechanism and feature fusion, comprising the following steps:

[0006] S1: Based on the acquired radar signal pulse sequence data, the radar signal pulse sequence is feature enhanced, and the part of the radar signal pulse sequence with a large TOA difference between the front and rear arrival times is scaled;

[0007] S2: extracting features from the radar signal pulse sequence and generating images corresponding to each characteristic parameter of the processed radar pulse sequence;

[0008] S3: performing feature fusion on the radar signal pulse sequence, fusing the images drawn by the characteristic parameters in step S2 into one image to obtain a radar signal fusion image;

[0009] S4: The radar signal fusion image is input into the DenseNet neural network based on the attention mechanism. After passing through multiple layers of the network, the classification result is obtained through the Softmax classifier output.

[0010] Furthermore, in step S1, feature enhancement is performed on the received multiple long-duration radar signal pulse sequences, specifically:

[0011] In the input radar signal pulse sequence, multiple fixed-time data are randomly intercepted and scaled for preprocessing. The arrival time TOA processing formula is shown in the following formula (1):

[0012] TOA(i)=TOA(i)-coef*diff (1)

[0013] Among them, coef is the scaling factor, diff is the current excessive arrival time TOA difference, the size of coef and diff is generally determined based on empirical values, and TOA(i) is the arrival time of a pulse sequence after two pulses with excessive first-order difference.

[0014] Furthermore, the feature extraction in step S2 refers to plotting the radar signal pulse sequence into an image with the arrival time TOA as the abscissa and each feature of the sequence as the ordinate. The generated image is in RGB mode, with a white background and black features.

[0015] Furthermore, the feature fusion in step S3 refers to drawing the two-dimensional graphs of the various features of the radar pulse sequence and the arrival time TOA in different colors in the same image, so that all the features of the intercepted radar pulse signal can be included in one image; the image of the original radar pulse sequence after feature extraction can only reflect some features of the radar signal, and cannot reflect all feature information in the sequence;

[0016] A feature fusion operation refers to fusing all the single-dimensional feature maps extracted from a segment of intercepted radar pulse sequence features into one map in sequence.

[0017] Furthermore, the feature fusion in step S3 specifically includes the following steps:

[0018] A1: First, convert the single-dimensional feature image after feature extraction into a grayscale image. The specific value of the pixel after grayscale processing is shown in formula (2):

[0019] Gray=0.2989×R+0.5870×G+0.1140×B (2)

[0020] Among them, the single-dimensional feature image is stored in 24-bit RGB mode; R, G, B are the brightness values ​​of the pixel in the red, green, and blue channels, with a value range of 0-255, all of which are integers, indicating the brightness of each channel; after grayscale conversion, the value range of the pixel is 0-255, indicating different grayscale levels, and the larger the value, the brighter it is;

[0021] A2: Perform color inversion on the grayscale processed image, convert black pixels into white pixels and white pixels into black pixels, and use them as the subsequent mask image. The color inversion formula is shown in formula (3):

[0022] t_pixel = 255 - s_pixel (3)

[0023] Among them, t_pixel is the pixel value after color inversion, and s_pixel is the pixel value before color inversion;

[0024] A3: Set the color used to represent the features in each single-dimensional feature map; the current feature fusion single-dimensional feature map is F-1, and create two images with the same size as the image after feature extraction, one with all pixels white and the other with all pixels of the color of the feature in F-1;

[0025] A4: Merge each single-dimensional feature map into one image in turn. Specifically, paste the feature color image in step A3 onto the newly created white image in step A3, and use a mask to make it transparent. The mask image is usually a grayscale image, in which each pixel value represents the transparency of the corresponding pixel. The value range is 0-255, where 0 means completely transparent and 255 means completely opaque.

[0026] Each single-dimensional feature image processed by step A1 and step A2 is used as a mask image in turn; in the processed single-dimensional feature image, the background pixel value is 0, and the feature pixel value is 255; in step A3, the pixel value of the background coordinate of the single-dimensional feature image in the feature color image will not be copied to the white image, and the pixel value of the feature coordinate of the single-dimensional feature image in the feature color image will be completely copied to the pixel of the corresponding coordinate in the white image; the formula of the weight α at each coordinate of the mask image is shown in formula (4):

[0027] α=mask[src_x,src_y] / 255 (4)

[0028] α is the weight value of the mask image, mask[src_x,src_y] is the pixel value of the mask image at the coordinate (src_x,src_y), src_pixel is the pixel value at the corresponding position of the characteristic color image to be pasted, and tar_pixel is the pixel value at the original corresponding position of the white image to be pasted;

[0029] The formula for the pixel value in the final image is shown in formula (5):

[0030] final_pixel=α*src_pixel+(1-α)*tar_pixel (5)

[0031] After the above operations, each feature map after the radar signal pulse sequence feature extraction of the intercepted time period is merged into one image in turn; the background color of the fused image is white, pixels of different colors represent different characteristics of the radar signal pulse sequence, the length of the data contains information about the number of coherently processed pulses, the multi-segment data contains information about the number of wave position residences, and the intervals in the data contain echo information; the obtained image is used as the input data of the neural network.

[0032] Furthermore, the DenseNet neural network based on the attention mechanism in step S4 includes three main parts: DenseBlock, Transition layer and CBAM module. The construction method of the DenseNet neural network is:

[0033] The existing DenseNet neural network connects all the previous layers as input and has the characteristic of feature reuse. However, the input of each layer in DenseBlock and the Transition layer will compress the channel and size of the feature map, which may lose some key information. In order to emphasize the channel and spatial features in the feature map and reduce the loss of key feature information, the present invention improves DenseNet and adds CBAM (channel and space attention mechanism) before the last layer of DenseBlock to facilitate the next layer of DenseBlock to extract features, improve the accuracy of radar working mode recognition, and form a DenseNet network based on the attention mechanism.

[0034] Furthermore, the operation process of the DenseNet neural network based on the attention mechanism in step S4 is:

[0035] The image first passes through the convolution layer to obtain the feature image, which is input into the three-layer DenseBlock and Transition layers; the calculation formula of the convolution layer is shown in formula (6):

[0036] a l =f(z l)=f(a l-1 *W l +b l ) (6)

[0037] Among them, W l and b l They represent the corresponding weights and biases in the lth convolutional layer, and f is the ReLU activation function, which is mainly used to solve the overfitting problem and the gradient disappearance phenomenon.

[0038] The pooling layer in the DenseBlock module is also called the downsampling layer. It is mainly responsible for controlling the network parameters and calculation amount, preventing overfitting, and improving the network's model generalization ability and training efficiency. The pooling operation formula is shown in formula (7):

[0039] a l =f(down(a l-1 )+b l ) (7)

[0040] Among them, b l represents the bias of the lth pooling layer, and down(·) represents the pooling function.

[0041] Dense Block connects the output of all previous layers with the output of the previous layer, so there are a total of (L+1)*L / 2(1+2+...+L)*L=(L+1)*L / 2 connections; the DenseBlock module connects all previous layers as input, where H(.) represents a nonlinear combination function, which is a combination operation. The feature transfer method is to directly concat the features of all previous layers and pass them to the next layer, instead of having an arrow pointing to all the subsequent layers in the previous layer, as shown in formula (8):

[0042] x l =H l ([x0,x0,...x l-1 ]) (8)

[0043] The nonlinear combination function in the above DenseBlock refers to the combination of BN+ReLU+3x3 Conv, although only k new feature maps are generated for each layer. There may be many layers in DenseBlock. For example, when there are l layers, the input of the last layer can reach l*k layers, which will be very large. Therefore, before performing the 3×3 convolution, a 4*k 1×1 convolution kernel is used to reduce the number of input feature maps to 4*k. This design is particularly effective for DenseBlock. Therefore, the nonlinear combination function of the present invention also adopts the structure of BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv;

[0044] The BN layer in the DenseBlock module is the abbreviation of BatchNormalization, which is a regularization method in deep learning. It can accelerate the convergence speed of the network, suppress overfitting, and improve generalization ability. The BN layer needs to calculate a minibatch input feature (x i ) and then calculate the mean μ and variance δ of all elements in x i Subtract the mean and divide by the standard deviation, and finally use the learnable parameters γ and β to perform an affine transformation to get the final BN output y i The specific formulas are shown in (9), (10), (11), and (12):

[0045]

[0046]

[0047]

[0048]

[0049] The Transition module connects two adjacent DenseBlocks, and reduces the size of the feature map through the pooling layer to integrate the features obtained from the previous DenseBlock; that is, a Transition layer is used between each Dense Block for downsampling transition connection, including BN, ReLU, 1×1 convolution, and 2×2 average pooling operations. The number of channels is reduced through a 1×1 convolution layer, and the average pooling layer with a stride of 2 is used to halve the height and width.

[0050] Furthermore, in step S4, an attention mechanism is added to the DenseNet neural network, specifically, a CBAM module is added after the three-layer Dense Block and Transition layers. The CBAM module includes a channel attention mechanism module and a spatial attention mechanism module. The CBAM module pays attention to the channel and space of the feature map. The attention maps output by the model will be multiplied by the input feature map and then used as the output of the module. The output of the CBAM module is shown in formulas (13) and (14):

[0051]

[0052]

[0053] Among them, F is the input feature map before the attention mechanism module, F' is the output of the channel attention mechanism module, and M cTo extract channel features, F" is the output of the spatial attention mechanism module. s To extract spatial features.

[0054] In the attention mechanism, the channel attention mechanism first performs a squeeze operation to compress the spatial dimension, that is, each two-dimensional feature map becomes a real number, and the number of feature channels remains unchanged; then an excitation operation is performed to generate weights for each feature channel through parameters ω, ω. ω is learned to explicitly model the correlation between feature channels. After obtaining the weight of each feature channel, the weight is applied to each original feature channel. The specific calculation of the channel attention mechanism module is shown in formula (15):

[0055] M c (F)=δ(MLP(AvgPool(F))+MLP(MaxPool(F))) (15)

[0056] The spatial attention mechanism is to find the most important part of the network for processing. First, the channel itself is reduced in dimension to obtain the maximum pooling and mean pooling results respectively, and then spliced ​​into a feature map. Then a convolutional layer is used for learning to obtain the result of the spatial attention mechanism. The specific calculation formula of the spatial attention module is shown in formula (16):

[0057]

[0058] Add the feature map F" after the channel attention mechanism and the spatial attention mechanism to the original input F to obtain the final feature map and output it.

[0059] Furthermore, after passing through the CBAM module in step S4, it passes through a Dense Block and is sent to the Softmax layer to obtain the output result:

[0060] The feature map after the CBAM module passes through a Dense Block again and is sent to the pooling layer and the Softmax layer for classification. The Softmax calculation formula is shown in formula (17):

[0061]

[0062] Among them, i represents the i-th neuron, so the Softmax layer has a total of N neurons, so the total number of classifications of the network is N.

[0063] The present invention mainly identifies six working modes, namely, velocity search (VS) mode, tracking while searching (TWS), ranging while searching (RWS), search and tracking mode (TAS), single target tracking (STT), and multi-target tracking mode (MTT). When the data of the phased array radar signal pulse sequence is obtained, it is used to identify which working mode the radar radiation source is in.

[0064] The CBAM module designed by the present invention includes a channel attention mechanism and a spatial attention mechanism. This model is lightweight and universal, can be added to any CNN, and can be trained end-to-end with the network. CBAM achieves dual refinement of input features by combining channel attention and spatial attention. This design enables the model to simultaneously focus on which channels and which spatial positions are meaningful, thereby improving the model's representation ability and decision accuracy. CBAM can dynamically adjust the importance of each channel and spatial position in the feature map according to task requirements and content context. This adaptive recalibration mechanism helps the model better focus on key features and improve recognition performance.

[0065] In the present invention, because the input data is a fused image of a radar signal pulse sequence, the image will contain multi-dimensional features of the radar pulse sequence. Adding the CBAM module to the network can deeply mine the various features in the fused image, increase the degree of attention to effective channels and spaces in the feature map, enhance the attention feature to retain key information, and suppress noise and irrelevant information.

[0066] Beneficial effects: Compared with the existing recognition method that directly uses radar pulse sequences, the phased array radar working mode recognition based on the attention mechanism and feature fusion of the present invention not only establishes connections between various features and amplifies the features of the radar pulse sequence in the image, but also allows a large amount of feature information to be contained in one image, which is more conducive to feature extraction by the neural network. The neural network that integrates DenseNet and the attention mechanism has the advantages of reducing gradient disappearance, strengthening feature transfer, encouraging feature reuse, and a smaller number of parameters. It can also focus on important features and suppress unnecessary features, thereby improving the radar pattern recognition capability and recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the process of the present invention;

[0068] Figure 2 The network structure of DenseNet based on the attention mechanism of the present invention;

[0069] Figure 3 This is a comparison chart of the recognition accuracy of the method of the present invention and the working mode recognition method based on the BP neural network;

[0070] Figure 4 is the input image of the network after feature fusion in the present invention;

[0071] Figure 5 The following is a diagram showing the characteristic parameters of the six phased array radar working modes extracted from radar signals;

[0072] Figure 6 A 2D image of time of arrival (TOA)-amplitude (PA) drawn before feature enhancement;

[0073] Figure 7 The two-dimensional image of arrival time (TOA)-amplitude (PA) is drawn after feature enhancement. DETAILED DESCRIPTION

[0074] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0075] The present invention provides a phased array radar working mode recognition method based on attention mechanism and feature fusion, such as Figure 1 As shown, it includes the following steps:

[0076] S1: Based on the acquired radar signal pulse sequence data, the radar signal pulse sequence is feature enhanced, and the part of the radar signal pulse sequence with a large TOA difference between the front and rear arrival times is scaled;

[0077] S2: extracting features from the radar signal pulse sequence and generating images corresponding to each characteristic parameter of the processed radar pulse sequence;

[0078] S3: performing feature fusion on the radar signal pulse sequence, fusing the images drawn by the characteristic parameters in step S2 into one image to obtain a radar signal fusion image;

[0079] S4: The radar signal fusion image is input into the DenseNet neural network based on the attention mechanism. After passing through multiple layers of the network, the classification result is obtained through the Softmax classifier output.

[0080] Based on the above method flow, steps S1 to S4 are described in detail below:

[0081] In step S1, the six working modes of speed search (VS), tracking while searching (TWS), ranging while searching (RWS), searching and tracking (TAS), single target tracking (STT), and multi-target tracking (MTT) are identified. In the case of obtaining the data of the phased array radar signal pulse sequence, it is used to identify which of the above working modes the radar radiation source is in; in step S1, the characteristics of the received multiple long-term radar signal pulse sequences are enhanced, specifically:

[0082] In the input radar signal pulse sequence, multiple fixed-time data are randomly intercepted and scaled for preprocessing. The arrival time TOA processing formula is shown in the following formula (1):

[0083] TOA(i)=TOA(i)-coef*diff (1)

[0084] Among them, coef is the scaling factor, diff is the current excessive arrival time TOA difference, the size of coef and diff is generally determined based on experience, and TOA(i) is the arrival time of a pulse sequence after two pulses with excessive first-order difference;

[0085] In this embodiment, the image of arrival time (TOA)-amplitude (PA) is taken as an example. Figure 6 is the image before scaling. Figure 7 It is a zoomed image, where the horizontal axis is the arrival time (TOA) and the vertical axis is the amplitude (PA).

[0086] The feature extraction in step S2 refers to plotting the radar signal pulse sequence into an image with the arrival time TOA as the horizontal axis and the various features of the sequence as the vertical axis. The generated image is in RGB mode, with a white background and black features.

[0087] In step S3, feature fusion means that the two-dimensional graphs of the various features of the radar pulse sequence and the arrival time TOA are drawn in the same image with different colors, so that all the features of the intercepted radar pulse signal can be included in one image; the image of the original radar pulse sequence after feature extraction can only reflect some features of the radar signal, and cannot reflect all feature information in the sequence;

[0088] A feature fusion operation refers to merging all the single-dimensional feature graphs extracted from a segment of intercepted radar pulse sequence features into one graph in sequence;

[0089] Feature fusion specifically includes the following steps:

[0090] A1: First, convert the single-dimensional feature image after feature extraction into a grayscale image. The specific value of the pixel after grayscale processing is shown in formula (2):

[0091] Gray=0.2989×R+0.5870×G+0.1140×B (2)

[0092] Among them, the single-dimensional feature image is stored in 24-bit RGB mode; R, G, B are the brightness values ​​of the pixel in the red, green, and blue channels, with a value range of 0-255, all of which are integers, indicating the brightness of each channel; after grayscale conversion, the value range of the pixel is 0-255, indicating different grayscale levels, and the larger the value, the brighter it is;

[0093] A2: Perform color inversion on the grayscale processed image, convert black pixels into white pixels and white pixels into black pixels, and use them as the subsequent mask image. The color inversion formula is shown in formula (3):

[0094] t_pixel = 255 - s_pixel (3)

[0095] Among them, t_pixel is the pixel value after color inversion, and s_pixel is the pixel value before color inversion;

[0096] A3: Set the color used to represent the features in each single-dimensional feature map; the current feature fusion single-dimensional feature map is F-1, and create two images with the same size as the image after feature extraction, one with all pixels white and the other with all pixels of the color of the feature in F-1;

[0097] A4: Merge each single-dimensional feature map into one image in turn. Specifically, paste the feature color image in step A3 onto the newly created white image in step A3, and use a mask to make it transparent. The mask image is usually a grayscale image, in which each pixel value represents the transparency of the corresponding pixel. The value range is 0-255, where 0 means completely transparent and 255 means completely opaque.

[0098] Each single-dimensional feature image processed by step A1 and step A2 is used as a mask image in turn; in the processed single-dimensional feature image, the background pixel value is 0, and the feature pixel value is 255; in step A3, the pixel value of the background coordinate of the single-dimensional feature image in the feature color image will not be copied to the white image, and the pixel value of the feature coordinate of the single-dimensional feature image in the feature color image will be completely copied to the pixel of the corresponding coordinate in the white image; the formula of the weight α at each coordinate of the mask image is shown in formula (4):

[0099] α=mask[src_x,src_y] / 255 (4)

[0100] α is the weight value of the mask image, mask[src_x,src_y] is the pixel value of the mask image at the coordinate (src_x,src_y), src_pixel is the pixel value at the corresponding position of the characteristic color image to be pasted, and tar_pixel is the pixel value at the original corresponding position of the white image to be pasted;

[0101] The formula for the pixel value in the final image is shown in formula (5):

[0102] final_pixel=α*src_pixel+(1-α)*tar_pixel (5)

[0103] After the above operations, each feature map after the radar signal pulse sequence feature extraction of the intercepted time period is merged into one image in turn; the background color of the fused image is white, pixels of different colors represent different characteristics of the radar signal pulse sequence, the length of the data contains information about the number of coherently processed pulses, the multi-segment data contains information about the number of wave position residences, and the intervals in the data contain echo information; the obtained image is used as the input data of the neural network.

[0104] In this embodiment, the phased array radar tracking while searching (TWS) mode is taken as an example to obtain Figure 4 The input image is shown, where red is frequency (CF), green is amplitude (PA), yellow is pulse width (PW), and dark blue is first-order difference (PRI).

[0105] The DenseNet neural network based on the attention mechanism in step S4 includes three main parts: DenseBlock, Transition layer and CBAM module. The construction method of the DenseNet neural network is as follows:

[0106] The existing DenseNet neural network connects all the previous layers as input and has the characteristic of feature reuse. However, the input of each layer in DenseBlock and the Transition layer will compress the channel and size of the feature map, which may lose some key information. In order to emphasize the channel and spatial features in the feature map and reduce the loss of key feature information, the present invention improves DenseNet by adding CBAM (channel and space attention mechanism) before the last layer of DenseBlock, so as to facilitate the next layer of DenseBlock to extract features, improve the accuracy of radar working mode recognition, and form a DenseNet network based on the attention mechanism. The specific architecture of the DenseNet neural network based on the attention mechanism in this embodiment is as follows: Figure 2 shown.

[0107] The operation process of the DenseNet neural network based on the attention mechanism is:

[0108] The image first passes through the convolution layer to obtain the feature image, which is input into the three-layer DenseBlock and Transition layer. DenseBlock has at least three layers. Too few layers will affect the performance of the network model and lead to unsatisfactory recognition accuracy. The calculation formula of the convolution layer is shown in formula (6):

[0109] a l =f(z l )=f(a l-1 *W l +b l ) (6)

[0110] Among them, W l and b l They represent the corresponding weights and biases in the lth convolutional layer, and f is the ReLU activation function, which is mainly used to solve the overfitting problem and the gradient disappearance phenomenon;

[0111] The pooling layer in the DenseBlock module is also called the downsampling layer. It is mainly responsible for controlling the network parameters and calculation amount, preventing overfitting, and improving the network model generalization ability and training efficiency. The pooling operation formula is shown in formula (7):

[0112] a l =f(down(a l-1 )+b l ) (7)

[0113] Among them, b l represents the bias of the lth pooling layer, and down(·) represents the pooling function.

[0114] Dense Block connects the output of all previous layers with the output of the previous layer, so there are a total of (L+1)*L / 2(1+2+...+L)*L=(L+1)*L / 2 connections; the Dense Block module connects all previous layers as input, where H(.) represents a nonlinear combination function, which is a combination operation. The feature transfer method is to directly concat the features of all previous layers and pass them to the next layer, instead of having an arrow pointing to all the subsequent layers in the previous layer, as shown in formula (7):

[0115] x l =H l ([x0,x0,...x l-1 ]) (8)

[0116] The nonlinear combination function in the above DenseBlock refers to the combination of BN+ReLU+3x3 Conv, although only k new feature maps are generated for each layer. There may be many layers in Dense Block. For example, when there are l layers, the input of the last layer can reach l*k layers, which will be very large. Therefore, before performing the 3×3 convolution, a 4*k 1×1 convolution kernel is used to reduce the number of input feature maps to 4*k. This design is particularly effective for DenseBlock. Therefore, the nonlinear combination function of the present invention also adopts the structure of BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv;

[0117] The BN layer in the Dense Block module is the abbreviation of BatchNormalization, which is a regularization method in deep learning that can accelerate the convergence speed of the network, suppress overfitting, and improve generalization ability. The BN layer needs to calculate a minibatch input feature (x i ) and then calculate the mean μ and variance δ of all elements in x i Subtract the mean and divide by the standard deviation, and finally use the learnable parameters γ and β to perform an affine transformation to get the final BN output y i The specific formulas are shown in (9), (10), (11), and (12):

[0118]

[0119]

[0120]

[0121]

[0122] The Transition module connects two adjacent DenseBlocks, and reduces the size of the feature map through the pooling layer to integrate the features obtained from the previous DenseBlock; that is, a Transition layer is used between each Dense Block for downsampling transition connection, including BN, ReLU, 1×1 convolution, and 2×2 average pooling operations. The number of channels is reduced through a 1×1 convolution layer, and the average pooling layer with a stride of 2 is used to halve the height and width.

[0123] Add an attention mechanism to the DenseNet neural network. Specifically, add a CBAM module after the three-layer Dense Block and Transition layers. The CBAM module includes a channel attention mechanism module and a spatial attention mechanism module. The CBAM module focuses on the channel and space of the feature map. The attention maps output by the model will be multiplied by the input feature map and then used as the output of the module.

[0124] Reference Figure 2 , the output of the CBAM module is shown in equations (13) and (14):

[0125]

[0126]

[0127] Among them, F is the input feature map before the attention mechanism module, F' is the output of the channel attention mechanism module, and M c is to extract channel features, F" is the output of the spatial attention mechanism module, M s To extract spatial features.

[0128] When the input feature map F passes through the CBAM module, it first passes through the channel attention mechanism module, and the output M c (F) is multiplied by the previous input feature map F to obtain the feature map F', which is then sent to the spatial attention mechanism module. The feature map M output by the spatial attention module s (F') is then multiplied by F' to get the final output F" of the attention mechanism module.

[0129] In the attention mechanism, the channel attention mechanism first performs a squeeze operation to compress the spatial dimension, that is, each two-dimensional feature map becomes a real number, and the number of feature channels remains unchanged; then the excitation operation is performed to generate weights for each feature channel through parameters ω, ω. ω is learned to explicitly model the correlation between feature channels. After obtaining the weight of each feature channel, the weight is applied to each original feature channel. The specific calculation of the channel attention mechanism module is shown in formula (15):

[0130] M c (F)=δ(MLP(AvgPool(F))+MLP(MaxPool(F))) (15)

[0131] The spatial attention mechanism is to find the most important part of the network for processing. First, the channel itself is reduced in dimension to obtain the maximum pooling and mean pooling results respectively, and then spliced ​​into a feature map. Then a convolutional layer is used for learning to obtain the result of the spatial attention mechanism. The specific calculation formula of the spatial attention module is shown in formula (16):

[0132]

[0133] Add the feature map F" after the channel attention mechanism and the spatial attention mechanism to the original input F to obtain the final feature map and output it.

[0134] After passing through the CBAM module, it passes through a Dense Block and is sent to the Softmax layer to obtain the output result:

[0135] The feature map after the CBAM module passes through a Dense Block again and is sent to the pooling layer and the Softmax layer for classification. The Softmax calculation formula is shown in formula (17):

[0136]

[0137] Among them, i represents the i-th neuron, so the Softmax layer has a total of N neurons, so the total number of classifications of the network is N.

[0138] Based on the above content, in order to verify the effectiveness and practical effect of the method of the present invention, the method of the present invention is compared with the existing radar working mode recognition method based on BP neural network in this embodiment, and the characteristic parameters of the working mode are as follows: Figure 5 As shown, Figure 5 The characteristic parameters of the six phased array radar working modes extracted from the radar signal include frequency (RF), pulse width (PW), first-order difference (PRI), and amplitude (PA). After training the training sets with the same number and parameters, the test results are as follows: Figure 3 As shown, it can be seen that the accuracy of the present invention has been significantly improved, proving the superiority of working mode recognition based on attention mechanism and image fusion.

Claims

1. A phased array radar working mode recognition method based on attention mechanism and feature fusion, characterized in that: The steps include: S1: Based on the acquired radar signal pulse sequence data, the radar signal pulse sequence is feature enhanced, and the part of the radar signal pulse sequence with a large TOA difference between the front and rear arrival times is scaled; S2: extracting features from the radar signal pulse sequence and generating images corresponding to each characteristic parameter of the processed radar pulse sequence; S3: performing feature fusion on the radar signal pulse sequence, fusing the images drawn by the characteristic parameters in step S2 into one image to obtain a radar signal fusion image; S4: The radar signal fusion image is input into the DenseNet neural network based on the attention mechanism. After passing through multiple layers of the network, the classification result is obtained through the Softmax classifier output.

2. According to claim 1, a phased array radar working mode recognition method based on attention mechanism and feature fusion is characterized in that: In step S1, feature enhancement is performed on the received multiple long-duration radar signal pulse sequences, specifically: In the input radar signal pulse sequence, multiple fixed-time data are randomly intercepted and scaled for preprocessing. The arrival time TOA processing formula is shown in the following formula (1): TOA(i)=TOA(i)-coef*diff (1) Wherein, coef is the scaling factor, diff is the current excessive arrival time TOA difference, and TOA(i) is the arrival time of a pulse sequence after two pulses with excessive first-order difference.

3. According to claim 1, a phased array radar working mode recognition method based on attention mechanism and feature fusion is characterized in that: The feature extraction in step S2 refers to plotting the radar signal pulse sequence into an image with the arrival time TOA as the horizontal axis and the various features of the sequence as the vertical axis. The generated image is in RGB mode, with a white background and black features.

4. According to claim 1, a phased array radar working mode recognition method based on attention mechanism and feature fusion is characterized in that: The feature fusion in step S3 refers to drawing the two-dimensional graphs of the various features of the radar pulse sequence and the arrival time TOA in the same image with different colors, so that all the features of the intercepted radar pulse signal can be included in one image; A feature fusion operation refers to fusing all the single-dimensional feature maps extracted from a segment of intercepted radar pulse sequence features into one map in sequence.

5. According to claim 4, a method for phased array radar working mode recognition based on attention mechanism and feature fusion is characterized in that: The feature fusion in step S3 specifically includes the following steps: A1: First, convert the single-dimensional feature image after feature extraction into a grayscale image. The specific value of the pixel after grayscale processing is shown in formula (2): Gray=0.2989×R+0.5870×G+0.1140×B (2) Among them, the single-dimensional feature image is stored in 24-bit RGB mode; R, G, B are the brightness values ​​of the pixel in the red, green, and blue channels, with a value range of 0-255, all of which are integers, indicating the brightness of each channel; after grayscale conversion, the value range of the pixel is 0-255, indicating different grayscale levels, and the larger the value, the brighter it is; A2: Perform color inversion on the grayscale processed image, convert black pixels into white pixels and white pixels into black pixels, and use them as the subsequent mask image. The color inversion formula is shown in formula (3): t_pixel = 255 - s_pixel (3) Among them, t_pixel is the pixel value after color inversion, and s_pixel is the pixel value before color inversion; A3: Set the color used to represent the features in each single-dimensional feature map; the current feature fusion single-dimensional feature map is F-1, and create two images with the same size as the image after feature extraction, one with all pixels white and the other with all pixels of the color of the feature in F-1; A4: Merge each single-dimensional feature map into one image in turn. Specifically, paste the feature color image in step A3 onto the newly created white image in step A3, and use a mask to make it transparent. The mask image is a grayscale image, in which each pixel value represents the transparency of the corresponding pixel. The value range is 0-255, where 0 means completely transparent and 255 means completely opaque. Each single-dimensional feature image processed by step A1 and step A2 is used as a mask image in turn; in the processed single-dimensional feature image, the background pixel value is 0, and the feature pixel value is 255; in step A3, the pixel value of the background coordinate of the single-dimensional feature image in the feature color image will not be copied to the white image, and the pixel value of the feature coordinate of the single-dimensional feature image in the feature color image will be completely copied to the pixel of the corresponding coordinate in the white image; the formula of the weight α at each coordinate of the mask image is shown in formula (4): α=mask[src_x,src_y] / 255 (4) α is the weight value of the mask image, mask[src_x,src_y] is the pixel value of the mask image at the coordinate (src_x,src_y), src_pixel is the pixel value at the corresponding position of the characteristic color image to be pasted, and tar_pixel is the pixel value at the original corresponding position of the white image to be pasted; The formula for the pixel value in the final image is shown in formula (5): final_pixel=α*src_pixel+(1-α)*tar_pixel (5) After the above operations, each feature map after the radar signal pulse sequence feature extraction of the intercepted time period is merged into one image in turn; the background color of the fused image is white, pixels of different colors represent different characteristics of the radar signal pulse sequence, the length of the data contains information about the number of coherently processed pulses, the multi-segment data contains information about the number of wave position residences, and the intervals in the data contain echo information; the obtained image is used as the input data of the neural network.

6. According to claim 1, a phased array radar working mode recognition method based on attention mechanism and feature fusion is characterized in that: The DenseNet neural network based on the attention mechanism in step S4 includes three main parts: Dense Block, Transition layer and CBAM module. The construction method of the DenseNet neural network is: improve DenseNet, add CBAM before the last layer of Dense Block, and form a DenseNet network based on the attention mechanism.

7. The method for phased array radar working mode recognition based on attention mechanism and feature fusion according to claim 6 is characterized in that: The operation process of the DenseNet neural network based on the attention mechanism in step S4 is: The image first passes through the convolution layer to obtain the feature image, which is then input into the three-layer Dense Block and Transition layers. The calculation formula of the convolution layer is shown in formula (6): a l =f(z l )=f(a l-1 *W l +b l ) (6) Among them, W l and b l They represent the corresponding weights and biases in the lth convolutional layer, and f is the ReLU activation function; Dense Block connects the output of all previous layers with the output of the previous layer, so there are a total of (L+1)*L / 2(1+2+...+L)*L=(L+1)*L / 2 connections; Dense Block module connects all previous layers as input, where H(.) represents a nonlinear combination function, which is a combination operation. The feature transfer method is to directly concat the features of all previous layers and pass them to the next layer, as shown in formula (7): x l =H l ([x0,x0,...x l-1 ]) (7) By using 1*1 convolution before the 3*3 convolution layer to reduce the number of feature maps, the structure of Dense Block is BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv; A Transition layer is used between each Dense Block for downsampling transition connection, including BN, ReLU, 1×1 convolution, and 2×2 average pooling operations. The number of channels is reduced by a 1×1 convolution layer, and the average pooling layer with a stride of 2 is used to halve the height and width.

8. The method for phased array radar working mode recognition based on attention mechanism and feature fusion according to claim 7 is characterized in that: In step S4, an attention mechanism is added to the DenseNet neural network, specifically, a CBAM module is added after the three-layer Dense Block and Transition layers. The CBAM module includes a channel attention mechanism module and a spatial attention mechanism module. The CBAM module pays attention to the channel and space of the feature map. The attention maps output by the model will be multiplied by the input feature map and then used as the output of the module; The output of the CBAM module is shown in equations (8) and (9): Among them, F is the input feature map before the attention mechanism module, F' is the output of the channel attention mechanism module, and M c is to extract channel features, F" is the output of the spatial attention mechanism module, M s To extract spatial features.

9. The method for phased array radar working mode recognition based on attention mechanism and feature fusion according to claim 8, characterized in that: In the attention mechanism, The channel attention mechanism first performs a squeeze operation to compress the spatial dimension, that is, each two-dimensional feature map becomes a real number, and the number of feature channels remains unchanged; Then comes the Excitation operation, which generates weights for each feature channel through parameters ω, ω. ω is learned to explicitly model the correlation between feature channels. After obtaining the weight of each feature channel, the weight is applied to each original feature channel. The specific calculation of the channel attention mechanism module is shown in formula (10): M c (F)=δ(MLP(AvgPool(F))+MLP(MaxPool(F))) (10) The spatial attention mechanism is to find the most important part of the network for processing. First, the channel itself is reduced in dimension to obtain the maximum pooling and mean pooling results respectively, and then spliced ​​into a feature map. Then a convolutional layer is used for learning to obtain the result of the spatial attention mechanism. The specific calculation formula of the spatial attention module is shown in formula (11): Add the feature map F" after the channel attention mechanism and the spatial attention mechanism to the original input F to obtain the final feature map and output it.

10. The method for phased array radar working mode recognition based on attention mechanism and feature fusion according to claim 9, characterized in that: After passing through the CBAM module in step S4, it passes through a Dense Block and is sent to the Softmax layer to obtain the output result: The feature map after the CBAM module passes through a Dense Block again and is sent to the pooling layer and the Softmax layer for classification. The Softmax calculation formula is shown in formula (12): Among them, i represents the i-th neuron, so the Softmax layer has a total of N neurons, so the total number of classifications of the network is N.

Citation Information

Cited By

  • Enterprise energy supply and demand dynamic balance management and control method and system

    CN122089022A