A Deep Learning-Based Micro-Doppler Feature Target Classification Method

Abstract

This invention relates to a deep learning-based micro-Doppler feature target classification method, belonging to the technical field of radar target recognition. The implementation steps include: (1) establishing a target micro-Doppler dataset; (2) constructing single-frame and multi-frame micro-Doppler target recognition networks respectively; (3) training the single-frame and multi-frame micro-Doppler feature recognition networks respectively; and (4) target micro-Doppler classification. This invention solves the problem of low classification accuracy caused by traditional features and classifiers. It can effectively realize radar target classification based on micro-Doppler features and can be used for radar target recognition.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning and radar target recognition, and further relates to a micro-Doppler feature classification method based on deep learning. This invention can effectively achieve radar target classification based on micro-Doppler features and can be used for radar target recognition. Background Technology

[0002] Radar Automatic Target Recognition (RATR) technology is essentially an application of pattern recognition and machine learning theories in the radar field. It primarily utilizes theoretical knowledge of pattern recognition and machine learning, combined with the characteristics of radar signals themselves, to extract feature information describing the characteristics of a given radar target echo. Then, it uses various classification methods to determine the target's category or type. Many objects in nature, whether man-made or natural, mechanical or animal-made, contain "micro-moving parts" within their structures. These micro-movements are typically generated by the movement of components such as aircraft rotors, rotating antennas on ships, vehicle running parts, the limbs of a person walking or running, the wings of a bird in flight, and vibrations on the surface of moving objects. These micro-moving parts all share similar forms of motion, namely rotation or vibration. When a target moves, its primary form of motion—translation—does Doppler modulate the radar signal. Simultaneously, the motion generated by these micro-moving parts on the target also does Doppler modulate the radar signal. Doppler modulation generated by micro-moving components is typically time-varying, and its Doppler frequencies often surround the Doppler frequencies generated by the translational motion of the target. These Doppler modulation phenomena caused by micro-moving components on the target are called the micro-Doppler effect, and the corresponding Doppler frequencies are called micro-Doppler frequencies, used to distinguish them from the Doppler frequencies generated by the translational motion of the target. Micro-Doppler modulation can be considered a unique characteristic of the target. Research on the micro-Doppler effect can provide more information about radar targets and offer a new approach to existing radar target identification methods.

[0003] With the rise of artificial intelligence, classification and recognition methods based on big data and deep learning are becoming more and more numerous and have better performance. Compared with traditional methods that use features such as waveform entropy and second-order central moments, and use support vector machines and decision trees as classifiers to achieve micro-Doppler classification, deep learning-based methods can achieve higher classification accuracy by mining higher-dimensional features, while having good generalization ability, thus realizing target classification and recognition based on micro-Doppler features. Summary of the Invention

[0004] Technical problems to be solved

[0005] To overcome the shortcomings of existing technologies, this invention provides a deep learning-based micro-Doppler feature target classification method to improve the classification accuracy of target recognition methods based on micro-Doppler features. The proposed micro-Doppler target classification method is based on a deep learning framework and classifies radar echo time-frequency maps corresponding to different targets.

[0006] Technical solution

[0007] A deep learning-based method for classifying and recognizing micro-Doppler feature targets, characterized by the following steps:

[0008] Step 1: Establish the target micro-Doppler dataset

[0009] Step 1a: Data preprocessing. Raw echo data is acquired through radar equipment and pulse compression is performed to obtain each frame of data containing the target.

[0010] Step 1b: Use the data from each frame to obtain single-frame time-frequency diagrams and multi-frame time-frequency diagrams, and set the corresponding category labels;

[0011] Step 1c: Divide the single-frame and multi-frame time-frequency maps into training set, validation set and test set in a ratio of 7:1:2, respectively, and use them to train and test the single-frame and multi-frame deep learning models.

[0012] Step 2: Construct single-frame and multi-frame micro-Doppler target recognition networks respectively. The single-frame and multi-frame micro-Doppler target recognition networks are two independent networks. The single-frame network is used to identify targets when a sufficient number of consecutive frames cannot be obtained.

[0013] Step 2a: The structure of the single-frame micro-Doppler target recognition network is as follows: input part, four dilated convolutional residual modules, and output part; the input part consists of an input layer and a convolutional layer; the input layer normalizes the input data so that the data range is [0,1], and the convolutional kernel size is 1×3; the structure of the dilated convolutional residual module is: 1×3 convolutional layer, batch normalization layer, 1x3 convolutional layer, batch normalization layer, activation layer, and Add layer. In this module, all convolutional layers are dilated convolutions. The convolutional layers in the four dilated convolutional residual modules all use dilated convolutions, and the dilation coefficients are 1, 2, 4, and 8 respectively, which aims to increase the receptive field of the network without increasing the network parameters. The number of convolutional kernels in all convolutional layers is 64; the output part consists of a convolutional layer, an activation layer, a fully connected layer, and a Softmax activation layer. The number of convolutional kernels in the convolutional layers is 64, and the Softmax activation layer uses the Softmax activation function to calculate the probability that the input time-frequency image is classified into each class;

[0014] Step 2b: The structure of the multi-frame micro-Doppler target recognition network is as follows: input part, three multi-scale residual modules, channel attention module, and output part; the input part consists of an input layer, a convolutional layer, a convolutional layer, an activation layer, a convolutional layer, and an activation layer; the input layer normalizes the input two-dimensional data to make the data range [0,1], the first convolutional layer has a kernel size of 3×3, the second convolutional layer has a kernel size of 512×1, the third convolutional layer has a kernel size of 1×3, and the activation layer uses Leakly... The ReLU activation function is used. The structure of the multi-scale residual module is as follows: 1×3 convolutional layer, 1×3 convolutional layer, 1×1 convolutional layer, 1×3 convolutional layer, 1×5 convolutional layer, Concat layer, 1x3 convolutional layer, batch normalization layer, activation layer, and Concat layer. This module utilizes convolutional layers with multiple kernels of different sizes to learn features at different scales, thereby improving the accuracy of the network in learning features. The number of convolutional kernels in the three multi-scale residual modules are 32, 64, and 128, respectively. The channel attention module has the following structure: convolutional layer, activation layer, global pooling layer, fully connected layer, activation layer, fully connected layer, activation layer, Sigmoid activation layer, M... The `ultiple` layer first compresses the spatial dimension of the features to 1 through pooling, converting each channel into a real number to calculate the statistical information corresponding to each channel. Two fully connected layers calculate the statistical information of each channel, learn the correlation between feature channels, and obtain the potential nonlinear relationship between channels. Finally, a Sigmoid activation layer is used to obtain the channel dimension weight coefficients, which are then used to perform a weighted operation on the features, implementing a channel attention mechanism. The output part consists of convolutional layers, activation layers, fully connected layers, and a Softmax activation layer. The Softmax activation layer uses the Softmax activation function to calculate the probability that the input time-frequency image is classified into each class.

[0015] Step 3: Train the network

[0016] Step 3a: Input the data from the single-frame and multi-frame training sets into the single-frame and multi-frame networks respectively, and the two networks output the predicted classification labels of the networks respectively;

[0017] Step 3b: Calculate the loss between the predicted target class label and the true target class label using the cross-entropy loss function, and then iteratively update the network parameters using the backpropagation algorithm until the preset number of training iterations is reached. At the same time, during the training process, test the network classification accuracy using the validation dataset, and calculate the network loss on the validation data using the cross-entropy loss function. Save the network parameters that minimize the loss function on the validation set to obtain the trained network.

[0018] Step 4: Target Doppler Classification

[0019] Step 4a: After the input single-frame and multi-frame time-frequency maps are processed by the single-frame network and multi-frame network modules respectively, the probability of the target to be identified being classified into each category is calculated by the Softmax layer, and the category corresponding to the highest probability is selected as the recognition result.

[0020] The ratio of the training set, validation set, and test set mentioned in step 1c is 7:1:2.

[0021] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.

[0022] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0023] Beneficial effects

[0024] The present invention provides a deep learning-based micro-Doppler feature target classification method, which has the following advantages compared with the prior art:

[0025] First, because the present invention constructs a multi-scale channel attention network, which includes a multi-scale feature extraction module and a channel attention module, the present invention can make full use of the target echo signal and improve the accuracy of target recognition.

[0026] Secondly, since the present invention directly takes the target video image as input at the input end and directly outputs the category of the target to be identified at the output end, it overcomes the problem of existing technologies that require manual design and selection of features, resulting in large computational load and poor real-time performance of the classification model. This invention improves the real-time performance of target recognition. Attached Figure Description

[0027] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0028] Figure 1 This is a flowchart of the present invention;

[0029] Figure 2 This is a single-frame network structure diagram of the present invention;

[0030] Figure 3 Diagram of the structure of the dilated convolution residual module;

[0031] Figure 4 This is a diagram of the multi-frame network structure of the present invention;

[0032] Figure 5This is a structural diagram of the multi-scale residual module in the network of the present invention;

[0033] Figure 6 This is a structural diagram of the channel attention module in the network of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0035] Step 1: Dataset Construction: Using a certain type of radar, actual echo signals from three types of aircraft are obtained to construct an aircraft target dataset. A ground target dataset is constructed using the 933A and 953D radars. Data preprocessing involves demodulating the narrowband radar echoes and then obtaining high resolution through pulse compression. Short-time Fourier transform is used to obtain single-frame time-frequency signals from the demodulated and pulse-compressed radar echo signals, thus generating a single-frame time-frequency map. The number of frames is set to 20, and the number of overlapping frames is 10. The single-frame time-frequency signals are combined to obtain multi-frame time-frequency maps, and corresponding category labels are assigned to the obtained single-frame and multi-frame time-frequency maps. The single-frame and multi-frame time-frequency map data are divided into training, validation, and test sets in a ratio of 7:1:2 for training and testing the deep learning model.

[0036] Step 2, as follows Figure 2 As shown, the structure of the single-frame micro-Doppler target recognition network is as follows: input part, four dilated convolutional residual modules, and output part; the input part consists of an input layer and a convolutional layer; the input layer normalizes the input data so that the data range is [0,1], and the convolutional kernel size is 1×3; as shown... Figure 3 As shown, the structure of the dilated convolution residual module is as follows: 1×3 convolutional layer, batch normalization layer, 1×3 convolutional layer, batch normalization layer, activation layer, and Add layer. All convolutional layers in this module are dilated convolutions. The convolutional layers in all four dilated convolution residual modules use dilated convolutions, with dilation coefficients of 1, 2, 4, and 8 respectively, aiming to increase the receptive field of the network without increasing network parameters. All convolutional layers have 64 kernels. The output section consists of convolutional layers, activation layers, fully connected layers, and Softmax activation layers. The convolutional layers have 64 kernels, and the Softmax activation layers use the Softmax activation function to calculate the probability that the input time-frequency image is classified into each class.

[0037] like Figure 4As shown, the structure of the multi-frame micro-Doppler recognition network is as follows: input part, three multi-scale residual modules, channel attention module, and output part. The input part consists of an input layer, a convolutional layer, a convolutional layer, an activation layer, a convolutional layer, and an activation layer. The input layer normalizes the input two-dimensional data to make the data range [0,1]. The first convolutional layer has a kernel size of 3×3, the second convolutional layer has a kernel size of 512×1, and the third convolutional layer has a kernel size of 1×3. The activation layer uses the Leaky ReLU activation function. Figure 5 As shown, the structure of the multi-scale residual module is: a 1×3 convolutional layer, a 1×3 convolutional layer, a 1×1 convolutional layer, a 1×3 convolutional layer, a 1×5 convolutional layer, a concat layer, a 1x3 convolutional layer, a batch normalization layer, an activation layer, and another concat layer. This module utilizes convolutional layers with multiple kernels of different sizes to learn features at different scales, thereby improving the accuracy of the network in learning features; for example... Figure 6 As shown, the channel attention module has the following structure: convolutional layer, activation layer, global pooling layer, fully connected layer, activation layer, fully connected layer, activation layer, sigmoid activation layer, and multiple layer. This module first compresses the spatial dimension of the features to 1 through pooling, converting each channel into a real number to calculate the statistical information corresponding to each channel. Two fully connected layers are used to calculate the statistical information of each channel, learn the correlation between feature channels, and obtain the potential nonlinear relationship between channels. Finally, the sigmoid activation layer is used to obtain the channel dimension weight coefficients, which are then used to perform a weighted operation on the features to implement the channel attention mechanism. The output part consists of a convolutional layer, activation layer, fully connected layer, and softmax activation layer. The softmax activation layer uses the softmax activation function to calculate the probability that the input time-frequency image is classified into each class.

[0038] Step 3, train the network. Input the data in the training set into the network, and output the network's predicted classification labels. At the same time, during the training process, use the validation dataset to test the network's classification accuracy, and use the cross-entropy loss function to calculate the network's loss on the validation data. Use the cross-entropy loss function to calculate the loss between the predicted target class label and the target true class label, and then use the backpropagation algorithm to iteratively update the network parameters until the preset number of training iterations is reached. Save the network parameters with the minimum loss function on the validation set, and you will get the trained network.

[0039] Step 4, Target Classification. After the input time-frequency image passes through each network module, the Softmax layer calculates the probability of the target being classified into each category, and selects the category with the highest probability as the recognition result.

[0040] The effects of this invention can be further illustrated by the following simulation experiments.

[0041] 1. Simulation conditions:

[0042] The simulation experiment of this invention was conducted using MATLAB 2021b software on a CPU consisting of an Intel(R) Core i7-9700 3.0GHz processor, 16GB of memory, and a Windows 7 operating system.

[0043] 2. Simulation content:

[0044] The data used in the experiments of this invention includes three types of aircraft and ground target data. The three types of aircraft include helicopters, propeller aircraft, and jet aircraft. The ground target data includes five types: single person, small team, tracked armored vehicle, wheeled armored vehicle, and wheeled civilian vehicle.

[0045] The data consists of three types of aircraft: a single-frame dataset with a total of 11,060 samples (1,136 jet aircraft, 4,588 propeller aircraft, and 5,366 helicopters) and a multi-frame dataset with a total of 3,566 samples (359 jet aircraft, 1,472 propeller aircraft, and 1,735 helicopters). All categories are divided into training, validation, and test sets in a 7:1:2 ratio. Due to the limited number of jet aircraft samples, this process was repeated to ensure that the number of jet aircraft samples was roughly equal to that of the other categories, thus preventing data imbalance and potential decreases in classification accuracy. The training set is used for training, while the validation set is used only for testing to prevent data leakage that could increase classification accuracy.

[0046] Ground target data: A total of 24,573 samples were collected in single-frame data, including 4,279 individuals, 8,704 squads, 2,352 tracked armored vehicles, 5,143 wheeled civilian vehicles, and 4,095 wheeled armored vehicles; a total of 2,699 samples were collected in multi-frame data, including 421 individuals, 856 squads, 513 tracked armored vehicles, 501 wheeled civilian vehicles, and 408 wheeled armored vehicles. All categories were divided into training, validation, and test sets in a 7:1:2 ratio. During training, the training set was used for training, and only the validation set was used to test the network performance to avoid data leakage that could lead to increased classification accuracy.

[0047] Tables 1 and 2 show the classification results of the multi-frame method on the two types of data, respectively. Tables 3 and 4 show the classification results of the single-frame method on the two types of data, respectively.

[0048] Table 1. Obfuscation matrix of the multi-frame method on ground target data.

[0049]

[0050] Table 2. Confusion matrix of the multi-frame method on three types of aircraft data.

[0051]

[0052] Table 3. Confusion matrix of the single-frame method on ground target data

[0053]

[0054] Table 4. Confusion matrix of the single-frame method on three types of aircraft data.

[0055]

[0056]

[0057] The simulation experiments above show that the radar micro-Doppler target recognition method based on deep learning proposed in this invention uses deep learning to achieve micro-Doppler target recognition, replacing the traditional method of manually designing and extracting features and then retraining the classifier, and effectively improving the recognition accuracy. Therefore, the method proposed in this invention has better performance for radar target recognition based on micro-Doppler and has important practical significance.

[0058] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for classifying and recognizing micro-Doppler feature targets, characterized in that... The steps are as follows: Step 1: Establish the target micro-Doppler dataset Step 1a: Data preprocessing. Raw echo data is acquired through radar equipment and pulse compression is performed to obtain each frame of data containing the target. Step 1b: Use the data from each frame to obtain single-frame time-frequency diagrams and multi-frame time-frequency diagrams, and set the corresponding category labels; Step 1c: Divide the single-frame and multi-frame time-frequency maps into training sets, validation sets, and test sets, respectively, for training and testing deep learning models; Step 2: Construct single-frame and multi-frame micro-Doppler target recognition networks respectively. Step 2a: The structure of the single-frame micro-Doppler target recognition network is as follows: input part, four dilated convolutional residual modules, and output part; the input part consists of an input layer and a convolutional layer; the input layer normalizes the input data so that the data range is [0,1], and the convolutional kernel size is 1×3; the structure of the dilated convolutional residual module is: 1×3 convolutional layer, batch normalization layer, 1x3 convolutional layer, batch normalization layer, activation layer, and Add layer. In this module, all convolutional layers are dilated convolutions. The convolutional layers in the four dilated convolutional residual modules all use dilated convolutions, and the dilation coefficients are 1, 2, 4, and 8 respectively, which aims to increase the receptive field of the network without increasing the network parameters. The number of convolutional kernels in all convolutional layers is 64; the output part consists of a convolutional layer, an activation layer, a fully connected layer, and a Softmax activation layer. The number of convolutional kernels in the convolutional layers is 64, and the Softmax activation layer uses the Softmax activation function to calculate the probability that the input time-frequency image is classified into each class; Step 2b: The structure of the multi-frame micro-Doppler recognition network is as follows: input part, three multi-scale residual modules, channel attention module, and output part; the input part consists of an input layer, a convolutional layer, a convolutional layer, an activation layer, a convolutional layer, and an activation layer; the input layer normalizes the input two-dimensional data to make the data range [0,1], the first convolutional layer has a kernel size of 3×3, the second convolutional layer has a kernel size of 512×1, the third convolutional layer has a kernel size of 1×3, and the activation layer uses Leakly... The ReLU activation function is used. The structure of the multi-scale residual module is as follows: 1×3 convolutional layer, 1×3 convolutional layer, 1×1 convolutional layer, 1×3 convolutional layer, 1×5 convolutional layer, Concat layer, 1x3 convolutional layer, batch normalization layer, activation layer, and Concat layer. This module utilizes convolutional layers with multiple kernels of different sizes to learn features at different scales, thereby improving the accuracy of the network in learning features. The number of convolutional kernels in the three multi-scale residual modules are 32, 64, and 128, respectively. The channel attention module has the following structure: convolutional layer, activation layer, global pooling layer, fully connected layer, activation layer, fully connected layer, activation layer, Sigmoid activation layer, and Multiple layer. The channel attention module first compresses the spatial dimension of the features to 1 through pooling, converting each channel into a real number to calculate the statistical information corresponding to each channel. Two fully connected layers calculate the statistical information of each channel, learn the correlation between feature channels, and obtain the potential nonlinear relationships between channels. Finally, a sigmoid activation layer is used to obtain the channel dimension weight coefficients, which are then used to weight the features, implementing the channel attention mechanism. The output part consists of convolutional layers, activation layers, fully connected layers, and a softmax activation layer. The softmax activation layer uses the softmax activation function to calculate the probability that the input time-frequency image is classified into each class. Step 3: Train the network Step 3a: Input the data from the single-frame and multi-frame training sets into the single-frame and multi-frame networks respectively, and the two networks output the predicted classification labels of the network respectively; Step 3b: Calculate the loss between the predicted target class label and the true target class label using the cross-entropy loss function, and then iteratively update the network parameters using the backpropagation algorithm until the preset number of training iterations is reached. At the same time, during the training process, test the network classification accuracy using the validation dataset, and calculate the network loss on the validation data using the cross-entropy loss function. Save the network parameters that minimize the loss function on the validation set to obtain the trained network. Step 4: Target Doppler Classification Step 4a: After the input single-frame and multi-frame time-frequency maps are processed by the single-frame network and multi-frame network modules respectively, the probability of the target to be identified being classified into each category is calculated by the Softmax layer, and the category corresponding to the highest probability is selected as the recognition result.

2. The deep learning-based micro-Doppler feature target classification and recognition method according to claim 1, characterized in that: The ratio of the training set, validation set, and test set mentioned in step 1c is 7:1:

2.

3. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of claim 1.

4. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method of claim 1.

Images