A radar target recognition classification method and system based on deep learning

Abstract

This invention relates to the field of radar signal technology, specifically to a radar target recognition and classification method and system based on deep learning. By extracting neural network depth features and normalized autocorrelation features in parallel from standardized echo samples within the same target observation window, the local amplitude-phase combination pattern and the cross-pulse time-domain correlation structure can maintain a common sample identifier and time range. Furthermore, the autocorrelation coefficient sequence is projected onto the same dimension as the depth features, allowing the two types of information to complement each other in the same fusion operation. This enhances the ability to represent target motion and scattering changes, reduces category boundary fluctuations that may occur when relying solely on local depth responses, and alleviates the problem of insufficient differentiation criteria when the local echo features of UAVs, birds, people, and vehicles are similar.

Description

Technical Field

[0001] This invention belongs to the field of radar signal technology, specifically relating to a radar target recognition and classification method and system based on deep learning. Background Technology

[0002] Radar can detect and identify targets based on the echoes formed by their reflections. The movement, size, and scattering locations of drones, birds, people, and vehicles vary, and these differences are reflected in the amplitude and phase variations and time-correlation structure of the trans-pulse echoes. In practical applications, target identification equipment typically performs windowing processing on continuous echoes within a target range cell, extracts features from each echo window, and determines the target category accordingly.

[0003] Using neural networks to extract features from radar echoes allows the learning of local waveform and amplitude-phase combination patterns from training samples. However, noise and ground clutter in the radar operating environment can alter local response patterns. When classification decisions primarily rely on depth features formed by a single path, such as in pure CNN / Transformer end-to-end radar target recognition models, local perturbations may propagate to the classification boundary, affecting the stability of the judgment. On the other hand, autocorrelation coefficients can describe the correlation of the same echo window at different time delays, offering unique advantages in characterizing the echo time-series structure. However, the autocorrelation coefficient sequence differs significantly from the neural network output in dimensionality and numerical scale. Directly concatenating the two makes it difficult to reasonably determine the actual contribution of the two types of representations to the classification result, affecting the reliability and interpretability of the classification.

[0004] To address the aforementioned problems, this invention proposes a radar target recognition and classification method and system based on deep learning. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, the present invention aims to provide a radar target recognition and classification method and system based on deep learning.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A radar target recognition and classification method based on deep learning, comprising: The radar echoes are acquired and then processed through windowing, mean removal, and energy normalization to form standardized echo samples with sample labels. The standardized echo samples are input into the deep feature extraction network to obtain the deep feature vector. The normalized complex autocorrelation coefficient sequence of the standardized echo samples is calculated on the preset time delay set. The correlation coefficient sequence is subjected to linear projection, normalization and activation transformation to obtain the correlation feature vector with the same dimension as the deep feature vector. Based on the sample identifier, pair the deep feature vector with the relevant feature vector, jointly generate the dimension-wise fusion weight and its complementary weight, and form the fusion feature vector; The class probabilities of each target category are generated based on the fused feature vectors, and the category corresponding to the highest probability is determined as the target category.

[0007] Preferably, the radar echo is acquired and subjected to windowing, mean removal, and energy normalization to form a standardized echo sample with sample identification, including: Complex sequences are extracted from radar echoes according to a preset window length. A sliding window is used to extract complex sequences between adjacent windows, and overlapping sampling points are set between adjacent windows to maintain time continuity. The mean of the complex sequence within the window is calculated, and the mean is subtracted from each sampling point to remove DC and static components, resulting in a mean-removed sequence. The mean-removed sequence is normalized by the average energy of the window to obtain standardized echo samples. The denominator during normalization includes a stability constant to prevent division by zero caused by a zero-energy window. When the average energy of the window reaches the preset effective threshold, the sample validity flag is set to valid. If it is lower than the effective threshold, the window is slid backward and subsequent pulses are included, and the window energy is recalculated until the effective threshold is met.

[0008] Preferably, the standardized echo samples are input into the depth feature extraction network to obtain a depth feature vector, including: The real and imaginary parts of each complex sampling point in the standardized echo sample are arranged into two input channels to form a dual-channel sequence. The dual-channel sequence is then processed by one-dimensional convolution, nonlinear transformation and downsampling in sequence. The one-dimensional convolution kernel slides along the slow time direction to extract the local amplitude and phase changes between adjacent pulses. The downsampling operation expands the time sensing range step by step. Global average pooling is performed on the time dimension and then transformed by a fully connected layer to obtain a deep feature vector of a preset dimension.

[0009] Preferably, the normalized complex autocorrelation coefficient sequence of standardized echo samples is calculated on a preset time delay set, and the correlation coefficient sequence is transformed into a correlation feature vector with the same dimension as the depth feature vector, including: Set a time delay set consisting of the minimum positive time delay to the maximum time delay, where the maximum time delay is less than the window length; for each time delay, calculate the normalized complex autocorrelation coefficient of the corresponding sequence segment under that time delay, where the denominator includes the stability constant during normalization; Following the order of the time delay set, the real part, imaginary part, and magnitude of each normalized complex autocorrelation coefficient are extracted sequentially to form a correlation coefficient vector. The correlation coefficient vector is then subjected to linear projection, normalization, and activation transformation to form a correlation feature vector with the same dimension as the deep feature vector.

[0010] Preferably, the sample identifiers in the depth feature records and related feature records are checked, and only depth feature vectors and related feature vectors with the same identifiers are paired; The paired deep feature vectors are concatenated with the relevant feature vectors in a dimensional manner, and then gating transformation is performed to generate dimensional fusion weights. The values ​​of each dimension of the fusion weights are between zero and one. The fused feature vector is formed by the following formula: the product of the first weight and the depth feature vector in each dimension plus the product of the complementary weight of the first weight and the relevant feature vector in each dimension, where the first weight is the fused weight and the complementary weight is one minus the fused weight.

[0011] Preferably, the class probabilities of each target category are generated based on the fused feature vector, and the category corresponding to the highest probability is determined as the target category, including: A linear transformation is performed on the fused feature vector to obtain the classification score of each target category; the classification score is normalized using the Softmax function to obtain the category probability of each target category; the category with the highest probability is selected as the target category, and the complete probability vector and sample identifier are output simultaneously.

[0012] This invention also provides a radar target recognition and classification system based on deep learning, comprising: The echo processing unit is used to acquire radar echoes and perform windowing, mean removal, and energy normalization to form standardized echo samples with sample labels. The depth feature extraction unit is used to input standardized echo samples into the depth feature extraction network to obtain depth feature vectors; The autocorrelation feature calculation unit is used to calculate the normalized complex autocorrelation coefficient sequence of standardized echo samples on a preset time delay set, and transform the correlation coefficient sequence into a correlation feature vector with the same dimension as the depth feature vector. The feature fusion unit is used to pair deep feature vectors and related feature vectors according to sample identifiers, jointly generate dimension-wise fusion weights and their complementary weights, and form a fusion feature vector. The classification unit is used to generate the class probability of each target category based on the fused feature vector, and the category corresponding to the highest probability is determined as the target category.

[0013] Preferably, the echo processing unit extracts a complex sequence from the radar echo according to a preset window length, calculates the mean of the complex sequence within the window and subtracts the mean from each sampling point, and normalizes the mean-free sequence according to the average energy of the window; the denominator includes a stability constant during normalization; the echo processing unit sets the sample validity flag to valid when the average energy of the window reaches a preset effective threshold, and continues to accumulate subsequent pulses until the effective threshold is met when the average energy of the window is lower than the effective threshold; The autocorrelation feature calculation unit extracts the real part, imaginary part, and magnitude of each normalized complex autocorrelation coefficient in the order of the time delay set to form a correlation coefficient vector. The correlation coefficient vector is then transformed into a correlation feature vector through linear projection, normalization, and activation transformation.

[0014] Preferably, the deep feature extraction unit arranges the real and imaginary parts of each complex sampling point in the standardized echo sample into two input channels to form a dual-channel sequence. The dual-channel sequence is then processed by one-dimensional convolution, nonlinear transformation and downsampling in sequence, and then global average pooling is performed on the time dimension and a fully connected transformation is performed to obtain a deep feature vector of a preset dimension.

[0015] Preferably, the feature fusion unit checks the sample identifiers in the deep feature records and related feature records, and only pairs deep feature vectors and related feature vectors with the same identifier; The paired deep feature vectors are concatenated with the relevant feature vectors in a dimensional manner, and then gating transformation is performed to generate dimensional fusion weights. The values ​​of each dimension of the fusion weights are between zero and one. The fused feature vector is the product of the first weight and the depth feature vector in one dimension, plus the product of the complementary weight of the first weight and the relevant feature vector in one dimension. The first weight is the fused weight, and the complementary weight is one minus the fused weight. The classification unit performs a linear transformation on the fused feature vector to obtain the classification score of each target category. The classification score is then normalized using the Softmax function to obtain the category probability of each target category, and the category with the highest probability is selected as the target category.

[0016] Beneficial effects 1. This invention extracts neural network depth features and normalized autocorrelation features in parallel from standardized echo samples of the same target observation window, enabling local amplitude-phase combination patterns and cross-pulse temporal correlation structures to maintain common sample identification and time range. On this basis, the autocorrelation coefficient sequence is projected onto the same dimension as the depth features, so that the two types of information can complement each other in the same fusion operation, thereby enhancing the ability to represent target motion and scattering changes, reducing the category boundary fluctuations that may occur when relying solely on local depth responses, and alleviating the problem of insufficient differentiation criteria when the local echo features of UAVs, birds, people and vehicles are similar.

[0017] 2. This invention applies mean removal, energy normalization, and normalized complex autocorrelation calculation to the echo window processing, and sets a stabilizing term in the autocorrelation denominator, effectively constraining the impact of overall amplitude variations and low-energy perturbations on temporal correlation features. The deep path preserves learnable local patterns, while the autocorrelation path provides a deterministic structure across time delays. The two paths, weighted sequentially, form a fused feature, thereby improving the stability of the representation under noise and clutter perturbations and reducing the possibility of a single feature component being amplified abnormally and directly dominating the classification result. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the invention and are not intended to limit the scope of protection of the invention.

[0020] Example 1 Please see Figure 1 As shown, this embodiment provides a radar target recognition and classification method based on deep learning, which includes the following steps: S1. Acquire radar echoes and perform windowing, mean removal, and energy normalization to form standardized echo samples with sample labels; The echo processing unit receives the complex baseband sampling sequence output by the radar receiver, and then reads the slow-time echo corresponding to the continuous pulses based on the range gate or target range unit given by the target detection result. For each range gate, the complex sequence is truncated according to a sliding window of length N; overlapping sampling points may exist between adjacent windows to maintain the temporal continuity of continuously observed objects at the window boundaries. Preferably, the window length N can be set to 256.

[0021] Let the original complex echo within a window be , where This is the index of the sampling point within the window, with a value ranging from 0 to... , This represents the number of sampling points contained in the window. The complex mean of the window is calculated, and this mean is subtracted from each sampling point to remove DC and static components, resulting in the mean-removed sequence. :

[0022] In the formula, The first one after removing the mean A complex number of echoes, For the first One original complex echo, For summation index The corresponding original complex echo, From 0 to The summation index; This indicates all the original complex echoes within the window. Summate the results and divide the sum by the number of sampling points in the window. The result is the window's complex mean; subtraction means subtracting the window's complex mean point by point in the complex domain to remove DC and static components. The mean-removed sequence is then normalized according to the window's average energy.

[0023] In the formula, For the standardized first A complex number of echoes, For summation index The corresponding mean is used to remove the echo, the absolute value sign represents the complex modulus, and the square represents the square of the modulus; This represents the average energy of the mean-free sequence within the window. This indicates the relationship between the average energy and the stability constant. Taking the square root of the sum, we get the sum of the square roots and the sum of the square roots. A normalized denominator with the same magnitude dimensions, thus resulting in a normalized sequence. For dimensionless standardized complex echoes, The stability constant is greater than zero. Preferably, the stability constant is... It can be set to 0.000001. When the average energy of the window reaches the valid sampling threshold, the sample validity flag is set to valid; if it is lower than the threshold, the echo processing unit continues to accumulate subsequent pulses until a window that meets the valid threshold is formed.

[0024] The echo processing unit writes the sample identifier, sampling time range, distance gate, normalized complex sequence, and validity identifier into the echo sample record. The same normalized complex sequence in this record is output to the depth feature extraction unit and the autocorrelation feature calculation unit respectively, thereby ensuring that the subsequent depth feature extraction and autocorrelation feature calculation stages can receive the exact same observation objects, avoiding the loss of correspondence between the two types of features due to different windows or distance gates; the outputs of the two processing branches will be merged in the subsequent feature fusion operation.

[0025] S2. Input the standardized echo samples into the depth feature extraction network to obtain depth feature vectors with sample labels; The deep feature extraction unit arranges the real and imaginary parts of each complex sampling point into two input channels, resulting in a dual-channel sequence of length N. This sequence is then processed sequentially through one-dimensional convolution, nonlinear transformation, and downsampling. Specifically, the one-dimensional convolution kernel slides along the slow-time direction to extract local amplitude-phase changes between adjacent pulses; the downsampling operation progressively expands the time-sensing range; finally, global average pooling is performed on the time dimension, followed by a fully connected transformation to obtain a D-dimensional deep feature vector. .

[0026] Furthermore, the deep feature extraction unit uses three sets of one-dimensional convolutions for processing. After each set of convolutions, normalization, activation, and max pooling operations are set, with the number of convolution channels increasing sequentially. Global average pooling merges the responses at different time positions into a vector of fixed length.

[0027] Furthermore, the depth feature dimension D can be set to 128. During the training phase, the convolutional kernels, normalization parameters, and fully connected parameters are updated based on the class loss; during the recognition phase, the trained parameters are used to perform forward computation on the echo to be recognized. The depth feature extraction stage outputs a depth feature record with sample labels, where... This will serve as the first input for subsequent feature fusion. Deep features retain the learned local amplitude-phase combination patterns, which, together with the cross-temporal correlation structure explicitly formed in the autocorrelation feature calculation stage, constrain subsequent classification.

[0028] The deep feature extraction network adopts the following complete architecture: The input layer arranges the real and imaginary parts of each complex sampling point in the normalized echo samples into two independent input channels, forming a dual-channel sequence of length N as the initial input to the network. The main body of the network uses three sets of one-dimensional convolutions for progressive processing. After each set of convolutions, normalization, nonlinear transformation, and max pooling operations are applied sequentially. The one-dimensional convolution kernel slides along the slow-time direction to extract local amplitude and phase changes between adjacent pulses. The number of convolution channels increases progressively with the advancement of the three sets of convolutional blocks, and each max pooling operation progressively expands the network's temporal perception range. After passing through the three sets of convolutional blocks... After processing, global average pooling is performed on the current time dimension to merge the response features at different time positions into a fixed-length feature vector. The fixed-length feature vector obtained after global average pooling is then mapped to a depth feature vector of a preset dimension through a fully connected transformation, with the preset dimension D preferably set to 128. During the training phase, the kernel parameters, normalization parameters, and mapping parameters of the fully connected layer in the above three sets of convolutional blocks are updated and optimized according to the category loss. During the recognition phase, the fixed parameters that have been trained and saved are used to perform forward computation on the echo samples to be recognized.

[0029] The specific parameters for the three sets of one-dimensional convolutions and their subsequent operations are set as follows: The first group of one-dimensional convolutions has a kernel size of 7, a stride of 1, and 32 output channels. Zero padding is used to keep the time length unchanged. After the first group of convolutions, a batch normalization layer, a ReLU activation function layer, and a max pooling layer are connected in sequence. The pooling window size of the max pooling layer is 2, and the stride is 2, which reduces the time length to half of the original.

[0030] The second group of one-dimensional convolutions has a kernel size of 5, a stride of 1, 32 input channels, and 64 output channels. Zero padding is used to keep the time length unchanged. After the second group of convolutions, a batch normalization layer, a ReLU activation function layer, and a max pooling layer are connected in sequence. The pooling window size of the max pooling layer is 2, and the stride is 2, which reduces the time length to half of the original.

[0031] The third group of one-dimensional convolutions has a kernel size of 3, a stride of 1, 64 input channels, and 128 output channels. Zero padding is used to keep the time length unchanged. After the third group of convolutions, a batch normalization layer, a ReLU activation function layer, and a max pooling layer are connected in sequence. The pooling window size of the max pooling layer is 2, and the stride is 2, which reduces the time length to one-eighth of the original.

[0032] After processing the three sets of convolutional blocks, global average pooling is performed on the current time dimension to merge the response features at different time positions into a fixed-length feature vector with a length of 128. Finally, a fully connected layer maps the fixed-length feature vector into a depth feature vector of a preset dimension D. When D is 128, the input dimension of the fully connected layer is 128 and the output dimension is 128.

[0033] S3. Calculate the normalized complex autocorrelation coefficient sequence of the standardized echo samples on the preset time delay set, and transform it into a correlation feature vector with the same dimension as the depth feature vector. The autocorrelation feature calculation unit is set with a time delay set consisting of 1 to K, where K is less than N. The maximum number of time delays K can be set to 32. For each time delay... Calculate the normalized complex autocorrelation coefficient using the following formula. :

[0034] In the formula, For delay The corresponding normalized multiple autocorrelation coefficient, The delay index is a positive integer, ranging from 1 to K, where K is the maximum delay value. for The complex conjugate, For relative Postponed The normalized echoes from each sampling point, the two summation terms give the energy of the corresponding sequence segment, and the multiplication represents the multiplication of the square roots of the two energies. To maintain a stable constant with a positive denominator, normalization was applied to effectively constrain the direct impact of different window scales on the correlation coefficient.

[0035] The autocorrelation feature calculation unit extracts each feature sequentially in ascending order of time delay. The real part, imaginary part, and magnitude of the correlation coefficient vector are used to form a correlation coefficient vector of length 3K. The real part retains the in-phase correlation component, the imaginary part retains the orthogonal component caused by phase changes, and the magnitude describes the correlation strength. This correlation coefficient vector is then subjected to linear projection, normalization, and activation transformation to form a D-dimensional correlation feature vector with the same dimension as the depth feature vector. .

[0036] Furthermore, the relevant feature dimension D is consistent with the deep feature dimension, and can be set to 128. The relevant feature calculation stage outputs relevant feature records with the same sample identifier, and... This serves as the second input for subsequent feature fusion. The linear projection parameters are updated during the training phase based on the classification loss, while the autocorrelation coefficient itself is calculated according to the aforementioned rules and does not participate in the training.

[0037] S4. Based on the sample identifier, pair the deep feature vector with the relevant feature vector, and jointly generate the dimension-wise fusion weight and its complementary weight to form the fusion feature vector; The feature fusion unit checks the sample identifiers in the two feature records to ensure that only samples with the same identifier are included. and This forms a fused input. After concatenating the two vectors in a fixed order, a gated transformation is performed to obtain the D-dimensional fused weights. :

[0038] In the formula, This represents the D-dimensional deep feature vector. Related feature vectors in D dimensions The resulting 2D connection vector after dimensional concatenation; Let D×2D be the gated weight matrix. Let D be the gated bias vector. This represents the element-wise Sigmoid activation function. Preferably, the gate weight matrix is ​​initialized using the Xavier uniform initialization method, and the gate bias vector is initialized to a zero vector, so that the values ​​of each dimension of the fusion weight are around 0.5 in the initial stage of training, thereby providing a balanced initial contribution to deep features and related features.

[0039] Subsequently, the fusion feature is formed according to the following formula. :

[0040] In the formula, For D-dimensional fused feature vectors, This indicates element-wise multiplication. To and The weights are complementary in each dimension. The first term retains the contribution of the depth features in the corresponding dimension, and the second term retains the contribution of the autocorrelation features in the corresponding dimension. Addition represents the sum of the two terms in each dimension. The gating parameters are updated together with the classifier during the training phase, allowing the fusion weights to be adaptively determined based on the two types of features of the current sample. The feature fusion stage outputs a fusion feature record with sample identification, which serves as the direct input for the subsequent classification stage. Since both the depth features and the correlation features originate from the same normalization sequence in the echo normalization process, the resulting weights can accurately reflect the correspondence between the two types of representations within the same observation window.

[0041] S5. Generate the category probabilities of drones, birds, people and vehicles based on the fused feature vectors, and determine the category corresponding to the highest probability as the target category; Classification unit for fusion features A linear transformation is performed to obtain four category scores, which are then converted into category probabilities using the Softmax function. Class score Sum of probabilities They are respectively:

[0042]

[0043] In the formula, For category indexing, The total number of categories is 4 in this embodiment. For the first The class classification weight row vector, For the first The classification bias of a class is represented by vector multiplication, which indicates the inner product of the classification weights and the fused features. For the first Class classification score, For the first Normalized probability of class, For probability-normalized summation index, The function is exponential. The four category indices correspond to drones, birds, people, and vehicles, respectively. The classification unit selects the category with the highest probability as the target category and outputs the complete probability vector, sample identifier, and sampling time range, so that the classification result can still be associated with the original echo window.

[0044] During the training phase, the Adam optimizer was used with a learning rate of 0.0003, a batch size of 128, and 300 iterations. A class label vector was set for each echo sample. And adopt cross-entropy loss As an optimization objective:

[0045] In the formula, For the classification loss of a single sample, For the first For the class's label component, the component corresponding to the actual category is set to 1, and the other components are set to 0. The natural logarithm is used; the negative sign reduces the loss as the true class probability increases, and the summation covers all values. There are several categories. The category loss updates the classification parameters, gating parameters, related feature projection parameters, and deep feature extraction parameters sequentially. It's worth noting that the autocorrelation coefficient remains constant, giving the cross-temporal correlation structure a definite physical meaning; the trainable parameters, based on this, adjust the classification boundary together with the deep features. After training, the model parameters are saved for use in the recognition phase. Parameter updates stop when training reaches the set termination conditions (e.g., validation set accuracy no longer improves or the preset number of iterations is reached).

[0046] Example 2 Please see Figure 2 As shown, this embodiment provides a radar target recognition and classification system based on deep learning, including the following modules: The system comprises an echo processing unit 100, a deep feature extraction unit 200, an autocorrelation feature calculation unit 300, a feature fusion unit 400, and a classification unit 500. Each unit can be implemented by a processor executing programs from memory, or by a programmable logic device working in conjunction with the processor. Data within the system is associated according to sample identifiers, and the deep feature extraction path and the autocorrelation feature calculation path converge at the feature fusion unit 400.

[0047] The echo processing unit 100 has a radar echo input terminal, which is connected to the input terminals of the depth feature extraction unit 200 and the autocorrelation feature calculation unit 300, respectively. This unit reads complex baseband echoes from the target range gate, performs windowing, mean removal, energy normalization, and validity determination, generating an echo sample record containing sample identifiers, time ranges, range gate identifiers, and a standardized complex sequence. Specifically, the echo processing unit 100 generates valid standardized echo samples when the average energy of the echo window reaches the valid sampling threshold; when the average energy of the echo window is below the threshold, it continues to accumulate subsequent pulses and reform the echo window. The same record generated is then read by both the depth feature extraction unit 200 and the autocorrelation feature calculation unit 300.

[0048] The deep feature extraction unit 200 reads the echo sample record, converts the real and imaginary parts of the complex sequence into a dual-channel time series, and outputs a D-dimensional deep feature record after one-dimensional convolution, nonlinear transformation, downsampling, global average pooling, and fully connected transformation. Simultaneously, the autocorrelation feature calculation unit 300 reads the same echo sample record, calculates the normalized complex autocorrelation coefficient on a preset time delay set, arranges the real, imaginary, and modulus values ​​of the coefficients into a correlation coefficient vector, and projects it to form a D-dimensional correlation feature record. Both records retain the sample identifier generated by the echo processing unit 100 and are output to the feature fusion unit 400 respectively.

[0049] The two inputs of the feature fusion unit 400 are connected to the outputs of the deep feature extraction unit 200 and the autocorrelation feature calculation unit 300, respectively. This unit verifies the sample identifiers of the two feature records, and only connects the deep features and related features with the same identifier, calculates the dimension-wise fusion weights, and forms a fused feature record. This record is then output to the classification unit 500. The classification unit 500 reads the fused feature record, calculates the classification scores and probabilities for four categories: drones, birds, people, and vehicles, and generates a classification result containing the sample identifier, category probability vector, and target category.

[0050] During training, the classification unit 500 also receives the category labels corresponding to the echo samples and calculates the cross-entropy loss. The parameter update path connects sequentially from the classification unit 500 to the feature fusion unit 400, the projection part of the autocorrelation feature calculation unit 300, and the depth feature extraction unit 200. During recognition, each unit reads the trained parameters, and the classification unit 500 provides the classification results to the radar target display, tracking, or alarm device. It should be noted that the display, tracking, or alarm device does not alter the feature formation and classification relationships within this system.

[0051] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A radar target recognition and classification method based on deep learning, characterized in that, include: The radar echoes are acquired and then processed through windowing, mean removal, and energy normalization to form standardized echo samples with sample labels. The standardized echo samples are input into the deep feature extraction network to obtain the deep feature vector. The normalized complex autocorrelation coefficient sequence of the standardized echo samples is calculated on the preset time delay set. The correlation coefficient sequence is subjected to linear projection, normalization and activation transformation to obtain the correlation feature vector with the same dimension as the deep feature vector. Based on the sample identifier, pair the deep feature vector with the relevant feature vector, jointly generate the dimension-wise fusion weight and its complementary weight, and form the fusion feature vector; The class probabilities of each target category are generated based on the fused feature vectors, and the category corresponding to the highest probability is determined as the target category.

2. The radar target recognition and classification method based on deep learning according to claim 1, characterized in that, The radar echoes are acquired and processed through windowing, mean removal, and energy normalization to form standardized echo samples with sample labels, including: Complex sequences are extracted from radar echoes according to a preset window length. A sliding window is used to extract complex sequences between adjacent windows, and overlapping sampling points are set between adjacent windows to maintain time continuity. The mean of the complex sequence within the window is calculated, and the mean is subtracted from each sampling point to remove DC and static components, resulting in a mean-removed sequence. The mean-removed sequence is normalized by the average energy of the window to obtain standardized echo samples. The denominator during normalization includes a stability constant to prevent division by zero caused by a zero-energy window. When the average energy of the window reaches the preset effective threshold, the sample validity flag is set to valid. If it is lower than the effective threshold, the window is slid backward and subsequent pulses are included, and the window energy is recalculated until the effective threshold is met.

3. The radar target recognition and classification method based on deep learning according to claim 1, characterized in that, The standardized echo samples are input into the depth feature extraction network to obtain a depth feature vector, including: The real and imaginary parts of each complex sampling point in the standardized echo sample are arranged into two input channels to form a dual-channel sequence. The dual-channel sequence is then processed by one-dimensional convolution, nonlinear transformation and downsampling in sequence. The one-dimensional convolution kernel slides along the slow time direction to extract the local amplitude and phase changes between adjacent pulses. The downsampling operation expands the time sensing range step by step. Global average pooling is performed on the time dimension and then transformed by a fully connected layer to obtain a deep feature vector of a preset dimension.

4. The radar target recognition and classification method based on deep learning according to claim 1, characterized in that, Calculate the normalized complex autocorrelation coefficient sequence of standardized echo samples on a preset time delay set, and transform the correlation coefficient sequence into a correlation feature vector with the same dimension as the depth feature vector, including: Set a time delay set consisting of the minimum positive time delay to the maximum time delay, where the maximum time delay is less than the window length; for each time delay, calculate the normalized complex autocorrelation coefficient of the corresponding sequence segment under that time delay, where the denominator includes the stability constant during normalization; Following the order of the time delay set, the real part, imaginary part, and magnitude of each normalized complex autocorrelation coefficient are extracted sequentially to form a correlation coefficient vector. The correlation coefficient vector is then subjected to linear projection, normalization, and activation transformation to form a correlation feature vector with the same dimension as the deep feature vector.

5. The radar target recognition and classification method based on deep learning according to claim 1, characterized in that, Verify the sample identifiers in the deep feature records and related feature records, and only pair deep feature vectors and related feature vectors with the same identifier; The paired deep feature vectors are concatenated with the relevant feature vectors in a dimensional manner, and then gating transformation is performed to generate dimensional fusion weights. The values ​​of each dimension of the fusion weights are between zero and one. The fused feature vector is formed by the following formula: the product of the first weight and the depth feature vector in each dimension plus the product of the complementary weight of the first weight and the relevant feature vector in each dimension, where the first weight is the fused weight and the complementary weight is one minus the fused weight.

6. The radar target recognition and classification method based on deep learning according to claim 1, characterized in that, Based on the fused feature vector, the class probabilities of each target category are generated, and the category corresponding to the highest probability is determined as the target category, including: A linear transformation is performed on the fused feature vector to obtain the classification score of each target category; the classification score is normalized using the Softmax function to obtain the category probability of each target category; the category with the highest probability is selected as the target category, and the complete probability vector and sample identifier are output simultaneously.

7. A radar target recognition and classification system based on deep learning, characterized in that, include: The echo processing unit is used to acquire radar echoes and perform windowing, mean removal, and energy normalization to form standardized echo samples with sample labels. The depth feature extraction unit is used to input standardized echo samples into the depth feature extraction network to obtain depth feature vectors; The autocorrelation feature calculation unit is used to calculate the normalized complex autocorrelation coefficient sequence of standardized echo samples on a preset time delay set, and transform the correlation coefficient sequence into a correlation feature vector with the same dimension as the depth feature vector. The feature fusion unit is used to pair deep feature vectors and related feature vectors according to sample identifiers, jointly generate dimension-wise fusion weights and their complementary weights, and form a fusion feature vector. The classification unit is used to generate the class probability of each target category based on the fused feature vector, and the category corresponding to the highest probability is determined as the target category.

8. The radar target recognition and classification system based on deep learning according to claim 7, characterized in that, The echo processing unit extracts a complex sequence from the radar echo according to a preset window length, calculates the mean of the complex sequence within the window, subtracts the mean from each sampling point, and normalizes the mean-free sequence according to the average energy of the window; the denominator during normalization includes a stability constant; when the average energy of the window reaches a preset effective threshold, the echo processing unit sets the sample validity flag to valid, and continues to accumulate subsequent pulses until the effective threshold is met when the average energy of the window is lower than the effective threshold; The autocorrelation feature calculation unit extracts the real part, imaginary part, and magnitude of each normalized complex autocorrelation coefficient in the order of the time delay set to form a correlation coefficient vector. The correlation coefficient vector is then transformed into a correlation feature vector through linear projection, normalization, and activation transformation.

9. A radar target recognition and classification system based on deep learning according to claim 7, characterized in that, The deep feature extraction unit arranges the real and imaginary parts of each complex sampling point in the standardized echo sample into two input channels to form a dual-channel sequence. The dual-channel sequence is then processed by one-dimensional convolution, nonlinear transformation and downsampling in sequence. Finally, global average pooling is performed on the time dimension and a fully connected transformation is performed to obtain a deep feature vector of a preset dimension.

10. A radar target recognition and classification system based on deep learning according to claim 7, characterized in that, The feature fusion unit checks the sample identifiers in the deep feature records and related feature records, and only pairs deep feature vectors and related feature vectors with the same identifier; The paired deep feature vectors are concatenated with the relevant feature vectors in a dimensional manner, and then gating transformation is performed to generate dimensional fusion weights. The values ​​of each dimension of the fusion weights are between zero and one. The fused feature vector is the product of the first weight and the depth feature vector in one dimension, plus the product of the complementary weight of the first weight and the relevant feature vector in one dimension. The first weight is the fused weight, and the complementary weight is one minus the fused weight. The classification unit performs a linear transformation on the fused feature vector to obtain the classification score of each target category. The classification score is then normalized using the Softmax function to obtain the category probability of each target category, and the category with the highest probability is selected as the target category.

Images