Radar dynamic multi-feature data fusion target classification method

Abstract

The invention relates to a target classification method based on dynamic multi-feature data fusion, and belongs to the technical field of radar signal processing. Comprising the following steps: 1) acquiring echo data of a target by a radar, preprocessing the echo data, and constructing a training data set by using time-frequency information, distance period information and distance period data of a signal; 2) constructing a dynamic multi-feature fusion network target classification model which comprises a global feature extraction module, a local feature extraction module, a dynamic feature extraction module and a feature fusion module; 3) inputting a training data set to carry out iterative optimization training on the dynamic multi-feature fusion network to obtain an optimal network parameter; and 4) preprocessing real-time radar echo signals, inputting the real-time radar echo signals into the trained dynamic multi-feature fusion network for testing, and completing target classification. According to the method, the Doppler information and the dynamic distance amplitude information of the radar signal can be subjected to feature extraction at the same time, the target classification accuracy is improved, and the actual application requirement is met.

Description

Technical Field

[0001] The invention relates to a target classification method based on dynamic multi-feature data fusion, which is particularly suitable for radar target classification intelligent processing and belongs to the technical field of radar signal processing. Background Art

[0002] Radar target detection and classification are widely used in military and civilian fields. However, due to the influence of clutter or noise generated by complex environments and the diversity of target types, reliable and robust radar target detection and classification has always been one of the key technologies that need to be studied. At present, the difficulties in radar target detection and recognition mainly lie in background suppression, target high-resolution feature extraction, complex feature classification, etc. In recent years, small targets such as drones have been popular with the public due to their small size, low price, and easy control. They have been widely used in aerial photography, environmental monitoring, mobile communications and other fields, but there are certain difficulties in the recognition and classification of small targets.

[0003] Some methods based on FMCW radar have been applied to aerial target recognition, using radar cross-section features obtained by radar to provide useful information for aerial target classification. However, when two targets are of similar size or small, RCS cannot reliably distinguish between two targets. In addition, RCS is affected by the physical material, size, shape, etc. of the scatterer, which reduces the accuracy of target classification. Micro-Doppler is the Doppler frequency shift caused by tiny components other than the target body. It was first used by Chen et al. (VCChen, F. Li, S.-S. Ho, and H. Wechsler, "Micro-Doppler effect in radar: Phenomenon, model, and simulation study," IEEE Trans. Aerosp. Electron. Syst., vol. 42, no. 1, pp. 2–21, Jan. 2006.) for aerial target recognition. The rotation of the drone's rotor and the flapping of the bird's wings will generate Doppler in the radar echo. The micro-motion characteristics are closely related to the type and motion state, so micro-Doppler is an effective feature for the classification of drones and flying birds. When analyzing micro-Doppler, the frequency of the spectrogram is an important factor because the micro-motion of the target is repeatedly represented in the spectrogram. By performing time-frequency transformation on the micro-motion signal, the rotation frequency and blade length of the drone blades are estimated to distinguish the type of target. However, due to the influence of time-frequency resolution, there is aliasing in the drone rotor in actual measurement, and it is difficult to clearly observe the flicker of each blade, which affects the classification of the target. In order to further analyze the micro-motion characteristics, the CVD graph is introduced. The CVD is obtained by iterating the Fourier transform of each row of the spectrogram. As an important feature for classification, considering that MDS is a Doppler signal in the time domain and CVD is a Doppler signal in the frequency domain, the CVD method is also effective for the analysis of radar echo signals.

[0004] In recent years, many DNN-based radar automatic target recognition methods have been developed. Deep learning-based methods extract high-dimensional features through deep neural networks, avoiding the limitations of manual feature extraction. Xu et al. (Xu B, Chen B, Wan J, et al. Target-Aware Recurrent Attentional Network for Radar HRRP Target Recognition. Signal Processing 2019; 155: 268-80.) proposed a target-aware recurrent attention network (TARAN), which uses the temporal dependency between range units to achieve planar target recognition, and proposed a reused long short-term memory (RLSTM) network to extract the spatial features of gestures. For three-dimensional data analysis, 3D-CNN and hybrid complex architecture 3D-CNN-LSTM were proposed, and experiments were conducted on radar data collected by ultra-wideband radar (UWB). Rahman et al. used the GoogleNet framework to train the classification of drones and birds based on the micro-Doppler features of drones and flying birds. Kim et al. (Kim BK, Kang HS, Park S O. Drone classification using convolutional neural networks with merged Doppler images [J]. IEEE Geoscience and Remote Sensing Letters, 2016, 14 (1): 38-42.) merged micro-Doppler features and CVD features into a new image, and improved the classification accuracy through CNN structure classification. Chen et al. (Chen X, Zhang H, Song J, et al. Micro-motion classification of flying bird and rotor drones via data augmentation and modified multi-scale cnn [J]. Remote Sensing, 2022, 14 (5): 1107.) used FMCW radar to collect data of drones and flying birds, fused the range periodicity map and micro-Doppler features, and used multi-channel DCNN to classify targets to improve the classification accuracy of drones and flying birds. It can be seen that using deep learning to extract target micro-Doppler features is very helpful for target classification and recognition. However, extracting features from a spectrum has certain limitations and is easily affected by complex environments, which reduces the accuracy of target classification. Summary of the invention

[0005] In view of the shortcomings of the above-mentioned prior art, the present invention provides a radar dynamic multi-feature data fusion target classification method, which can simultaneously extract features of the Doppler information and dynamic range amplitude information of the radar signal, improve the accuracy of target classification, and meet practical application needs.

[0006] The present invention provides a radar dynamic multi-feature data fusion target classification method, which is special in that it comprises the following steps:

[0007] Step 1: The radar collects the echo data of the target and preprocesses the echo data, and constructs a training data set using the signal's time-frequency information, range-period information, and range-period data;

[0008] The echo data preprocessing method described in step 1 is: collecting radar echo data under various observation conditions and areas, preprocessing the data, including removing negative frequencies and DC components of the signal, performing MTI processing on the signal, and effectively filtering out stationary clutter on the ground and in the air. After the above data preprocessing steps, a range period diagram of the target is drawn, the radar signal time series is intercepted according to the set sample observation time, and each signal sample sequence is subjected to time-frequency analysis. The time-frequency analysis result is Fourier transformed along the time dimension to obtain a CFD spectrum, and the time-frequency spectrum diagram and the CFD spectrum diagram are spliced ​​into a TCD spectrum diagram;

[0009] Preferably, the specific steps of the echo data preprocessing method described in step 1 are:

[0010] First, the radar echo signal is processed by removing negative frequency modulation, DC signal and MTI, and then the distance period information of the target is obtained. The distance unit information of the target is extracted and the echo is converted into a time-frequency spectrum using Short-time Fourier Transform (STFT). Assuming s(t) is the extracted target echo signal, the short-time Fourier transform formula can be expressed as:

[0011]

[0012] Among them, g(·) represents the sliding window, taking the Gaussian window function, t represents the time dimension, and ω represents the frequency dimension. By observing the range Doppler of the UAV, the distance unit where the target is located is selected, and the signal of the distance unit where the target is located is extracted to perform STFT to obtain the time-frequency diagram of the target signal. By performing short-time Fourier transform on the preprocessed echo data, the data matrix STFT is obtained, and the time-frequency transform result is FFT along the time dimension to obtain the CFD spectrum, that is,

[0013] CFD=F t {STFT} (2)

[0014] F tIndicates that FFT is performed on the time-frequency diagram along the time dimension;

[0015] The time-frequency spectrum and CFD spectrum are horizontally spliced ​​into a TCD spectrum, expressed as

[0016] y tcd =[x md ,x cfd ] (3)

[0017] where y tcd represents the spliced ​​TCD spectrum matrix, x md represents the micro-Doppler spectrum matrix, x cfd Represents the CFD spectrum matrix.

[0018] Preferably, the method for constructing the training data set in step 1 is:

[0019] The training data set contains the range period graph, time-frequency spectrum graph, range period matrix and corresponding labels of multiple target signals. A training sample is constructed according to the target location information. The distance unit where the target echo is located in the radar signal at each moment is determined, and the distance unit data is taken out. The data is subjected to time-frequency analysis to obtain the time-frequency spectrum graph. The result of the time-frequency analysis is subjected to Fourier transform along the time dimension to obtain the CFD spectrum graph. The two spectra are horizontally spliced ​​into one to construct the second training sample. In order to obtain the dynamic characteristics of the data, a multi-frame range period data cube is constructed and stored in a matrix:

[0020]

[0021] in Represents multi-frame periodic cube data, FFT represents Fourier transform along the time dimension, s IF (t m ,t s ) represents the intermediate frequency signal, t m represents the fast time of the time index within the mth chirp, t s represents the slow time between adjacent chirps;

[0022] Step 2: Construct a dynamic multi-feature fusion network target classification model, including a global feature extraction module, a local feature extraction module, a dynamic feature extraction module, and a feature fusion module;

[0023] Preferably, the dynamic multi-feature fusion network in step 2 includes an input layer, channel 1, channel 2 and channel 3, the input layer extracts preliminary features by inputting three data sets, channel 1 connects multiple local feature extraction modules to extract local features in the target spectrogram, channel 2 is a ResNet3D network model to extract dynamic distance information of the target, and channel 3 is composed of multiple global feature extraction modules to extract global features in the target signal spectrogram; the feature fusion network further extracts local features and global features through spatial attention and channel attention, and finally adds the features extracted by the three channels, and the number of nodes in the output layer of the feature fusion network is 6, indicating six categories; assuming that the input x1, x2, and x3 respectively represent the input time-frequency spectrum, the distance periodogram and the distance period data are respectively input into the network modules of the three channels, flattened in the channel direction, and then the number of channels of each pixel is linearly transformed, and finally the features extracted by the three channels are input into the feature fusion module;

[0024] Step 3: Input the training data set to iteratively optimize the dynamic multi-feature fusion network to obtain the optimal network parameters;

[0025] Preferably, the specific method of step 3 is:

[0026] Input the training data set constructed in step 1 into the target classification model established in step 2, and train the model using the improved gradient descent method AdamW. After continuous iterative optimization training, the target classification model training is completed and the network parameters are optimized when the labels detected by the network are close to the set labels.

[0027]

[0028] m t ←β1m t-1 (1-β1)g t (6)

[0029] v t ←β2v t-1 +(1-β2)g 2 t (7)

[0030]

[0031] Where t represents the number of current iterations and η represents the learning rate.

[0032] Step 4: Preprocess the real-time radar echo signal and input it into the trained dynamic multi-feature fusion network for testing to complete the target classification.

[0033] Preferably, the specific method of step 4 is:

[0034] The target echo data generated in real time is preprocessed, and the time-frequency information and distance information obtained after processing are made into time-frequency spectrum, distance periodicity diagram and distance data tensor in the form of input data set, and input into the optimal target detection model trained based on a large amount of high-quality data. After feature extraction and feature fusion, the output value is obtained to realize the classification of the target;

[0035] The output vector z of the neural network is [z1,z2,....,z6], where z i is the score of the model on the i-th category, and the Softmax function calculates the probability p of the i-th category i The formula is as follows:

[0036]

[0037] where z i is the score of the i-th category, p i is the predicted probability of the ith class.

[0038] The radar dynamic multi-feature data fusion target classification method of the present invention has the following beneficial effects:

[0039] (1) Breaking through the limitations of traditional classification methods, it can realize intelligent classification of radar targets in complex environments;

[0040] (2) The present invention overcomes the shortcomings of single feature neural network feature extraction, and simultaneously extracts and fuses the time-frequency information, distance information, and dynamic information in the radar echo signal to improve the target classification performance;

[0041] (3) The present invention enhances the feature extraction capability of the global feature extraction module by utilizing the attention mechanism, uses channel attention and spatial attention in the feature fusion module to further extract local features and local features, and then fuses these features to improve the accuracy of target classification and meet practical application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flowchart of the target classification method;

[0043] Figure 2 It is a dynamic multi-feature fusion network framework diagram;

[0044] Figure 3 It is the structure diagram of the global feature extraction module;

[0045] Figure 4 It is the structural diagram of the local feature extraction module;

[0046] Figure 5 It is the structural diagram of the dynamic feature extraction module;

[0047] Figure 6 It is the structure diagram of the feature fusion module. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] A radar dynamic multi-feature data fusion target classification method of this embodiment includes the following steps:

[0050] 1. Radar signal preprocessing and data set construction

[0051] Radar signal preprocessing:

[0052] First, in order to eliminate the negative frequency signal in the radar echo data, the data was preliminarily processed. A high-pass filter was applied to effectively remove the DC component in the signal. This step is to enhance the dynamic range of the signal and ensure the accurate extraction of high-frequency components. The moving target indication (MTI) technology was used to effectively filter out the stationary clutter on the ground and in the air, thereby significantly improving the detection capability of the target. MTI can effectively suppress the stationary or slow-moving background noise by comparing the phase changes between adjacent pulses, making the target signal more prominent. After the above data preprocessing steps, the range period diagram of the target was drawn to clarify the range unit where the target is located. By analyzing the range period diagram, the echo data of the target can be accurately extracted. In order to further extract the micro-motion characteristics of the target, short-time Fourier transform (STFT) is used for time-frequency analysis. After the radar signal is preprocessed to remove jitter interference, the range unit where the target is located is found, and the echo signal of the range unit where the target is located is extracted for STFT. The signal in the time window is FFTed by sliding the window to obtain the time-frequency spectrum:

[0053]

[0054] Among them, g(·) represents a sliding window, and this paper uses a Gaussian window function, t represents the time dimension, and w represents the frequency dimension. By observing the range Doppler of the drone, the distance unit where the target is located is selected, and the signal of the distance unit where the target is located is extracted and STFT is performed to obtain the time-frequency diagram of the target signal.

[0055] The data matrix STFT is obtained by performing short-time Fourier transform on the preprocessed echo data, and the CFD spectrum is obtained by performing FFT on the time-frequency diagram along the time axis.

[0056] CFD=F t {STFT} (2)

[0057] F tIndicates that FFT is performed on the time-frequency diagram along the time axis.

[0058] The time-frequency spectrum and CFD spectrum are horizontally spliced ​​into a TCD spectrum, expressed as

[0059] y tcd =[x md ,x cfd ] (3)

[0060] where y tcd represents the spliced ​​TCD spectrum matrix, x md represents the micro-Doppler spectrum matrix, x cfd Represents the CFD spectrum matrix.

[0061] In order to obtain the dynamic characteristics of the data, a multi-frame distance period data cube is constructed, and the dynamic characteristics between multiple frames are extracted through a three-dimensional network to improve the classification accuracy of the target. After preprocessing the collected original echo, the distance period data R is obtained by fast Fourier transforming the intermediate frequency signal along the fast time index. It is an N×M covariance matrix, where N represents the distance sampling unit and M represents the number of cycles. Due to the displacement of the distance caused by the movement of the target, in order to extract the dynamic characteristics of the target, L frames of the distance period sequence can be obtained within 1 total imaging time, and then the multi-frame distance period sequence plane is superimposed along the slow time axis to form the distance period sequence tensor data. Where N represents the distance sampling unit, M represents the number of cycles, and L represents the number of frames.

[0062]

[0063] Since each sample is input into the network for processing multiple times during the training process, in order to reduce the number of time-frequency transformation operations, the data is first preprocessed and the preprocessed data is used as the data set.

[0064] 2. Dataset Construction

[0065] After preprocessing the data, we get the target range periodogram, time-frequency spectrum and TCD spectrum, which are then constructed into three data sets. The training data set contains the range periodogram, time-frequency spectrum and range period matrix of multiple signals and corresponding labels. A training sample is constructed based on the target location information. The distance unit where the target echo is located in the radar signal at each moment is determined, and the distance unit data is taken out. The data is analyzed for time and frequency to obtain the time-frequency spectrum. The result of the time-frequency analysis is Fourier transformed along the time dimension to obtain the CFD spectrum. The two spectra are horizontally spliced ​​into one to construct the second training sample. In order to obtain the dynamic characteristics of the data, a multi-frame range period data cube is constructed and stored in a matrix.

[0066] A dynamic multi-feature fusion convolutional neural network classification method includes the following steps:

[0067] Construct a dynamic multi-feature fusion network.

[0068] Compare with Figure 2 ,The structure of the constructed classification network model is divided into three parts: global ,feature extraction module, local feature extraction module, dynamic feature ,fusion module, and softmax classification.

[0069] (1) Global feature extraction module

[0070] In the global feature extraction module channel, firstly, Patch Partition is used to divide the input (H, W, 3) image into (4, 4) small blocks. The size of the divided image is (H / 4, W / 4, 48) dimensions. Linear Embedding maps the divided image to 96 dimensions. Figure 3 As shown in the figure, Windows Multi-Head Self-Attention (W-MSA) is introduced in the global feature extraction branch. W-MSA is a module in Swin-transformer. The computational complexity of extracting feature maps grows linearly with the size of the feature maps, and it utilizes the prior knowledge of image locality, but lacks information interaction between windows. Therefore, the next module introduces SW-MSA, which moves the window toward the lower right corner to achieve the interaction of pixels in different windows, better obtain the context information, and extract the correlation information between each window.

[0071] Ω(MSA)=4hwC 2 +2(hw) 2 C (12)

[0072] Ω(W-MSA)=4hwC 2 +2M 2 hwC (13)

[0073] Where h represents the height of the feature map, w represents the width of the feature map, C represents the depth of the feature map, and M represents the size of each window. For each stage, by incorporating the patch into the global feature block, the feature map enters the W-MSA through the LayerNorm layer and then passes through the linear layer with the GELU activation function, as shown in (14).

[0074] G i =f(SW-MSA(LN(f(W-MSA(LN(G i-1 ))))+G i-1 ))+f(W-MSA(LN(G i-1 )))+G i-1 (14) (2)

[0075] Among them G i represents the output of the global feature, f is the convolution operation with a convolution kernel size of 1×1, LN represents the LayerNorm operation, and finally, the global features in the extracted time-spectrum graph are input into the feature fusion module.

[0076] (2) Local feature extraction module

[0077] For the local feature extraction module, we hope that the network can only extract local features. By borrowing the LN and GELU activation functions in Transform, we can obtain good performance in different scenarios. Specifically, the local feature extraction module. The local features extracted from the distance period map are input into the feature fusion module. This process is shown in (15).

[0078] L i =f(f(LN(f 7×7 (L i-1 ))))+L i-1 (15)

[0079] Where L i is the local feature extracted, f represents the convolution operation with a convolution kernel size of 1×1, LN represents the LayerNorm module, and f 7×7 Indicates a convolution operation with a convolution kernel of 7×7.

[0080] (3) Dynamic feature extraction module

[0081] ResNet3D has a specific structure similar to ResNet, but its convolution is extended to three-dimensional convolution to process tensor data. Specifically, the input data is confined to (batch-size, channels, depth, height, width), and through a series of 3D convolution layers, 3D pooling layers, and residual connections, a feature representation for classification or regression is finally obtained. The input of the network is a 16×513×64 matrix, which divides the continuous distance period data into 16 frames, which enables the network to learn rich spatial features, which is crucial for accurately recognizing complex data. The three-dimensional convolution kernel slides on the distance period sequence tensor data, which can capture the long-term dependencies and related features between different numbers of frames in the sequence, and performs convolution operations on the data set to obtain a three-dimensional feature map. The structure of ResNet3D includes 4 3×3×3 convolutions with 64 channels, 4 3×3×3 convolutions with 128 channels, 4 3×3×3 convolutions with 256 channels, and 4 3×3×3 convolutions with 512 channels, two of which have a residual connection.

[0082] (4) Feature Fusion Module

[0083] For the feature fusion module, the global features are provided to the CBAM module, which uses the interdependence between channel maps to improve the feature representation of specific semantics. The local features are input into the spatial attention (SA) mechanism to enhance local details and suppress irrelevant areas. Finally, the results produced by each attention and fusion path are feature fused and connected to a residual inverted MLP. To a certain extent, it prevents problems such as gradient disappearance, explosion, and network degradation, and effectively captures global and local feature information at all levels.

[0084] CA(x)=σ(MLP(AvgPool(x))+MLP(MaxPool(x))) (16)

[0085] SA(x)=σ(f 7×7 (Concat[AvgPool(x),MaxPool(x)])) (17)

[0086] Among them, σ represents the Sigmoid function, f 1×1 Indicates a convolution operation with a convolution kernel of 1×1, f 7×7 The convolution operation with a convolution kernel of 7×7 is represented by the feature fusion module output as follows:

[0087]

[0088] in represents element-wise multiplication, represents the data after channel attention, represents the data after spatial attention, Represents the data output after feature fusion.

[0089] The specific steps of step 3 are:

[0090] The AdamW gradient descent method is used for iterative optimization training. Given the initial learning rate parameter α, momentum factors β1, β2, and optimization parameters θ t , the first instantaneous vector m t , the second instantaneous vector v t , progress multiplier η t

[0091]

[0092] m t ←β1m t-1 +(1-β1)g t (6)

[0093] v t ←β2v t-1 +(1-β2)g 2t (7)

[0094]

[0095] Where t represents the number of current iterations, and η represents the learning rate;

[0096] Finally, after the average pooling layer and the fully connected layer, the probability p of the i-th class is calculated by the Softmax function. i The formula is as follows:

[0097]

[0098] where z i is the score of the i-th category, p i is the predicted probability of the ith class.

Claims

1. A radar dynamic multi-feature data fusion target classification method, characterized in that The following steps are involved: Step 1: The radar collects the echo data of the target and preprocesses the echo data, and constructs a training data set using the signal's time-frequency information, range-period information, and range-period data; Step 2: Construct a dynamic multi-feature fusion network target classification model, including a global feature extraction module, a local feature extraction module, a dynamic feature extraction module, and a feature fusion module; Step 3: Input the training data set to iteratively optimize the dynamic multi-feature fusion network to obtain the optimal network parameters; Step 4: Preprocess the real-time radar echo signal and input it into the trained dynamic multi-feature fusion network for testing to complete the target classification.

2. A radar dynamic multi-feature data fusion target classification method according to claim 1, characterized in that: The echo data preprocessing method described in step 1 is: collecting radar echo data under various observation conditions and areas, preprocessing the data, including removing negative frequencies and DC components of the signal, performing MTI processing on the signal, and effectively filtering out stationary clutter on the ground and in the air. After the above data preprocessing steps, a range period diagram of the target is drawn, the radar signal time series is intercepted according to the set sample observation time, time-frequency analysis is performed on each signal sample sequence, the time-frequency analysis results are Fourier transformed along the time dimension to obtain a CFD spectrum, and the time-frequency spectrum diagram and the CFD spectrum are spliced ​​into a TCD spectrum.

3. A radar dynamic multi-feature data fusion target classification method according to claim 2, characterized in that: The specific steps of the echo data preprocessing method described in step 1 are: First, the radar echo signal is processed by removing negative frequency modulation, DC signal and MTI, and then the distance period information of the target is obtained. The distance unit information of the target is extracted and the echo is converted into a time-frequency spectrum using Fourier transform STFT. Assume that s(t) is the extracted target echo signal, where the short-time Fourier transform formula is expressed as: Among them, g(·) represents the sliding window, taking the Gaussian window function, t represents the time dimension, and ω represents the frequency dimension. By observing the range Doppler of the UAV, the distance unit where the target is located is selected, and the signal of the distance unit where the target is located is extracted to perform STFT to obtain the time-frequency diagram of the target signal. By performing short-time Fourier transform on the preprocessed echo data, the data matrix STFT is obtained, and the time-frequency transform result is FFT along the time dimension to obtain the CFD spectrum, that is, CFD=F t {STFT} (2) F t Indicates that FFT is performed on the time-frequency diagram along the time dimension; The time-frequency spectrum and CFD spectrum are horizontally spliced ​​into a TCD spectrum, expressed as y tcd =[x md ,x cfd ] (3) where y tcd represents the spliced ​​TCD spectrum matrix, x md represents the micro-Doppler spectrum matrix, x cfd Represents the CFD spectrum matrix.

4. A radar dynamic multi-feature data fusion target classification method according to claim 1, characterized in that: The method for constructing the training data set described in step 1 is: The training data set contains the range period graph, time-frequency spectrum graph, range period matrix and corresponding labels of multiple target signals. A training sample is constructed according to the target location information. The distance unit where the target echo is located in the radar signal at each moment is determined, and the distance unit data is taken out. The data is subjected to time-frequency analysis to obtain the time-frequency spectrum graph. The result of the time-frequency analysis is subjected to Fourier transform along the time dimension to obtain the CFD spectrum graph. The two spectra are horizontally spliced ​​into one to construct the second training sample. In order to obtain the dynamic characteristics of the data, a multi-frame range period data cube is constructed and stored in a matrix: in Represents multi-frame periodic cube data, FFT represents Fourier transform along the time dimension, s IF (t m ,t s ) represents the intermediate frequency signal, t m represents the fast time of the time index within the mth chirp, t s Indicates the slow time between adjacent chirps.

5. A radar dynamic multi-feature data fusion target classification method according to claim 1, characterized in that: The dynamic multi-feature fusion network in step 2 includes an input layer, channel 1, channel 2 and channel 3. The input layer extracts preliminary features by inputting three data sets. Channel 1 connects multiple local feature extraction modules to extract local features in the target spectrogram. Channel 2 is a ResNet3D network model to extract dynamic distance information of the target. Channel 3 is composed of multiple global feature extraction modules to extract global features in the target signal spectrogram. feature The fusion network will further extract local features and global features through spatial attention and channel attention, and finally add the features extracted from the three channels. The number of nodes in the output layer of the feature fusion network is 6, representing six categories. Assuming that the input x1, x2, and x3 represent the input time-frequency spectrum, the distance periodogram and distance period data are respectively input into the network modules of the three channels, flattened in the channel direction, and then the number of channels for each pixel is linearly transformed. Finally, the features extracted from the three channels are input into the feature fusion module.

6. A radar dynamic multi-feature data fusion target classification method according to claim 1, characterized in that: The specific method of step 3 is: Input the training data set constructed in step 1 into the target classification model established in step 2, and train the model using the improved gradient descent method AdamW. After continuous iterative optimization training, the target classification model training is completed and the network parameters are optimized when the labels detected by the network are close to the set labels. m t ←β1m t-1 +(1-β1)g t (6) v t ←β2v t-1 +(1-β2)g 2 t (7) Where t represents the number of current iterations and η represents the learning rate.

7. A radar dynamic multi-feature data fusion target classification method according to claim 1, characterized in that: The specific method of step 4 is: The target echo data generated in real time is preprocessed, and the time-frequency information and distance information obtained after processing are made into time-frequency spectrum, distance periodicity diagram and distance data tensor in the form of input data set, and input into the optimal target detection model trained based on a large amount of high-quality data. After feature extraction and feature fusion, the output value is obtained to realize the classification of the target; The output vector z of the neural network is [z1,z2,....,z6], where z i is the score of the model on the i-th category, and the Softmax function calculates the probability p of the i-th category i The formula is as follows: where z i is the score of the i-th category, p i is the predicted probability of the ith class.

Images