A radar signal classification method under low signal-to-noise ratio

Abstract

The present invention discloses a radar signal classification method under low signal-to-noise ratio (SNR). The method comprises the following steps: first, the radar signal is converted into a time-frequency image using time-frequency transformation to obtain the time-frequency domain features of the radar signal under different SNRs, and the time-frequency image is scaled using bilinear interpolation. Then, the first convolutional layer of a convolutional spiking neural network is used to extract spatial features from the radar signal time-frequency image and adaptively encode it. The encoded feature information is pulsed using leaky integral spiking neurons to obtain the signal's time series features. The pulsed information is then accumulated within an optimized time step, and information is alternately transferred between the convolutional layer and the spiking neuron layer to complete the fusion of spatiotemporal features. Finally, the fused features are used to complete signal classification, and the gradient factor is optimized. The radar signal is classified using the parameter-updated convolutional spiking neural network. The present invention improves the recognition speed and accuracy of radar signal classification under low SNR.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and in particular to a radar signal classification method under low signal-to-noise ratio. Background Art

[0002] Modern electronic warfare has rapidly developed with the rise of electronic information technology. The electromagnetic spectrum is diverse and complex, and radar signal waveforms are characterized by agility and high interference. Identifying key radar signals within complex electromagnetic environments can gain an advantage in modern electronic warfare.

[0003] Currently, mainstream radar signal classification methods can be broadly categorized into three types: traditional probability and statistics-based classification methods, feature-based classification methods based on machine learning, and learning models based on deep neural networks. Signal features used for classification can be broadly categorized into inter-pulse features and intra-pulse features.

[0004] Due to the surge in radar signal types and frequency agility, traditional methods based on pulse-to-pulse features have significantly reduced their classification capabilities. Deep learning, however, has been widely used in radar signal classification due to its powerful learning, generalization, and robustness capabilities. While radar signal classification technology using deep learning has made significant progress in recent years, the complexity of the electromagnetic environment means that radar signal features at low signal-to-noise ratios are significantly affected by noise, making it difficult to extract effective features. This results in low radar signal classification accuracy at low signal-to-noise ratios. Therefore, overcoming noise interference and efficiently identifying different radar signals is a key issue that needs to be addressed in radar signal classification. Summary of the Invention

[0005] The object of the present invention is to provide a radar signal classification method under low signal-to-noise ratio, thereby reducing noise interference and efficiently identifying various radar signals.

[0006] The technical solution for achieving the purpose of the present invention is: a radar signal classification method under low signal-to-noise ratio, comprising the following steps:

[0007] Step 1: Use Choi-Williams distribution to perform time-frequency transformation on the radar signal, convert the radar signal into a radar signal time-frequency graph, and obtain the time-frequency characteristics of the radar signal under different signal-to-noise ratios;

[0008] Step 2: Use bi-trilinear interpolation to scale the radar signal time-frequency graph to the input size of the convolutional spike neural network, and use random flipping and random permutation to expand the features of the signal time-frequency graph;

[0009] Step 3: Use the first convolutional layer of the convolutional spike neural network to extract spatial features from the radar signal time-frequency graph and perform adaptive coding;

[0010] Step 4: The encoded radar signal spatial feature information is fed into the leaky integral-and-release pulse neuron of the convolutional spike neural network and converted into a pulse sequence to obtain the time series feature information of the radar signal;

[0011] Step 5: By comparing the effects of different time steps on network stability and inference delay, the optimal time step of the network is selected. The characteristic values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spiking neurons to stimulate them to release pulses, and thus complete the forward transmission of the characteristic information;

[0012] Step 6: The convolutional layer and the leaky integrated release pulse neuron layer in the convolutional spike neural network alternately transmit the spatiotemporal information of the radar signal to complete the spatiotemporal feature fusion of the radar signal;

[0013] Step 7: The radar signal after the spatiotemporal feature fusion is passed through the fully connected layer of the convolutional spiking neural network composed of integral spiking neurons to classify the radar signal;

[0014] Step 8. When the convolutional spiking neural network backpropagates the error through the chain rule, the gradient is used to replace the backpropagation error, the gradient factor is optimized, the convolutional spiking neural network parameters are updated, the convergence and stability of the convolutional spiking neural network are enhanced, and the convolutional spiking neural network with updated parameters is used to complete the classification of the radar signal.

[0015] Compared with the existing technology, the present invention has the following significant advantages: (1) the convolutional pulse neural network is used for radar signal classification, the network model is small, easy to train, and the radar signal can be classified more efficiently; (2) the spatial and temporal characteristics of the radar signal are integrated to reduce the interference of noise, highlight the differences between different signal characteristics, and improve the accuracy of radar signal classification; (3) the training method of gradient substitution is used to accelerate the network convergence speed and improve the recognition speed of radar signals with low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The figure is a flow chart of a radar signal classification method under low signal-to-noise ratio according to the present invention.

[0017] Figure 2 Schematic diagram of the structure of the convolutional pulse neural network constructed in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Combine Figure 1The present invention is a radar signal classification method under low signal-to-noise ratio. First, the radar signal is converted into a time-frequency diagram using the Choi-Williams distribution, and the bilinear interpolation method is used to scale it to the size required by the convolutional spike neural network input. The characteristics of the signal time-frequency diagram are expanded by random flipping and random arrangement. Then, the first convolution layer of the convolutional spike neural network is used to extract the spatial features of the radar signal time-frequency diagram and adaptively encode it through training. The encoded feature information is then sent to the leakage integral release spike neuron to convert it into a spike sequence, thereby obtaining the signal characteristics in the time series. By comparing the effects of different time steps on network stability and inference delay, the optimal time step of the network is selected. The feature values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spike neuron to stimulate it to release pulses, and then complete the forward transmission of the feature information. Gradient is used to replace the back-propagation error, optimize the gradient factor, update the network parameters, enhance the convergence and stability of the network, and finally use the convolutional spike neural network with updated parameters to complete the recognition of the radar signal. Specifically, the following steps are included:

[0020] Step 1: Use Choi-Williams distribution to perform time-frequency transformation on the radar signal, convert the radar signal into a radar signal time-frequency graph, and obtain the time-frequency characteristics of the radar signal under different signal-to-noise ratios;

[0021] Step 2: Use bi-trilinear interpolation to scale the radar signal time-frequency graph to the input size of the convolutional spike neural network, and use random flipping and random permutation to expand the features of the signal time-frequency graph;

[0022] Step 3: Use the first convolutional layer of the convolutional spike neural network to extract spatial features from the radar signal time-frequency graph and perform adaptive coding;

[0023] Step 4: The encoded radar signal spatial feature information is fed into the leaky integral-and-release pulse neuron of the convolutional spike neural network and converted into a pulse sequence to obtain the time series feature information of the radar signal;

[0024] Step 5: By comparing the effects of different time steps on network stability and inference delay, the optimal time step of the network is selected. The characteristic values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spiking neurons to stimulate them to release pulses, and thus complete the forward transmission of the characteristic information;

[0025] Step 6: The convolutional layer and the leaky integrated release pulse neuron layer in the convolutional spike neural network alternately transmit the spatiotemporal information of the radar signal to complete the spatiotemporal feature fusion of the radar signal;

[0026] Step 7: The radar signal after the spatiotemporal feature fusion is passed through the fully connected layer of the convolutional spiking neural network composed of integral spiking neurons to classify the radar signal;

[0027] Step 8. When the convolutional spiking neural network backpropagates the error through the chain rule, the gradient is used to replace the backpropagation error, the gradient factor is optimized, the convolutional spiking neural network parameters are updated, the convergence and stability of the convolutional spiking neural network are enhanced, and the convolutional spiking neural network with updated parameters is used to complete the classification of the radar signal.

[0028] As a specific implementation method, in step 1, the radar signal is transformed into a radar signal time-frequency diagram using the Choi-Williams distribution to obtain the time-frequency characteristics of the radar signal under different signal-to-noise ratios, as follows:

[0029] Different Gaussian white noises are added to different radar signals, and the Choi-Williams transform is used to convert the radar signals into time-frequency diagrams of radar signals with different noise interferences.

[0030] As a specific implementation, step 2 uses bi-trilinear interpolation to scale the radar signal time-frequency graph to the input size of the convolutional spike neural network, and uses random flipping and random permutation to expand the features of the signal time-frequency graph, as follows:

[0031] Step 2.1: Use bi-trilinear interpolation to scale the radar signal time-frequency graph to the convolutional spike neural network input size.

[0032] Step 2.2: Use random flipping and random arrangement to expand the features of the time-frequency graph to enhance generalization.

[0033] As a specific implementation, step 3 uses the first convolutional layer of the convolutional spike neural network to extract spatial features from the radar signal time-frequency diagram and perform adaptive encoding, as follows:

[0034] Step 3.1, the first convolutional layer of the convolutional spike neural network extracts spatial features from the radar signal time-frequency graph through convolution operation;

[0035] Step 3.2: The first convolutional layer adaptively encodes the spatial features of the radar signal time-frequency graph through training.

[0036] As a specific implementation method, in step 4, the encoded radar signal spatial feature information is fed into the leaky integral-and-release pulse neuron of the convolutional spike neural network and converted into a pulse sequence to obtain the time series feature information of the radar signal, as follows:

[0037] Step 4.1: Send the encoded radar signal spatial eigenvalues ​​to the leakage integration and release pulse neuron for pulse conversion. The pulse conversion first undergoes a charging process, as shown in formula (1):

[0038] τ n (U[t]-U[t-1])=-(U[t-1]-U reset )+X[t] (1)

[0039] Where τ n is the time constant, U reset is the reset voltage, X(t) is the input characteristic value, where U[t] is the membrane potential at the current moment, and U[t-1] is the membrane potential at the end of the previous moment;

[0040] The accumulation of input characteristic values ​​is manifested as an increase in membrane potential. Within each accumulation time step, if the membrane potential threshold is reached, the leakage integral pulse neuron releases a pulse, completing the pulse conversion of radar signal characteristic information, as shown in formula (2):

[0041] P(t)=Θ(U tmp -U threshold ) (2)

[0042] Among them U threshold is the threshold voltage for the activation of the spiking neuron, P(t) is the output pulse signal, and Θ(x) is the step function;

[0043] Step 4.2: The pulsed radar signal spatial characteristic information is a pulse train with the radar signal time information. The leakage integrating and releasing neurons accumulate the pulse train carrying the radar signal time information, which is manifested as an increase in membrane potential, as shown in formula (3):

[0044] U t+1 =U t (1-P t-1 )+X(t) (3)

[0045] where p t-1 The threshold is U threshold The step function represents the pulses released by the leaky integrate-and-send neuron.

[0046] As a specific implementation method, step 5 compares the effects of different time steps on network stability and inference delay to select the optimal time step of the network. The characteristic values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spiking neuron to stimulate it to release pulses, and thus complete the forward transmission of the characteristic information. The details are as follows:

[0047] By comparing the effects of different time steps on network stability and inference delay, the optimal time step is optimized. Then, the time series features of the pulsed radar signal are continuously accumulated until the given optimal time step ends. The membrane potential of the spiking neuron increases. When the membrane potential reaches the threshold, the spiking neuron releases a pulse, completing the forward transmission of the feature information.

[0048] As a specific implementation, the convolutional layer and the leaky integrated discharge pulse neuron layer in the convolutional spike neural network described in step 6 alternately transmit the spatiotemporal information of the radar signal to complete the spatiotemporal feature fusion of the radar signal, as follows:

[0049] The first convolutional layer of the convolutional spiking neural network and the leaky integral release pulse neuron layer alternately transmit information to complete the spatiotemporal feature fusion of the radar signal and enhance the learning ability of the convolutional spiking neural network for different radar signal features.

[0050] As a specific implementation, in step 7, the radar signal after the spatiotemporal feature fusion is classified by passing it through a fully connected layer of a convolutional spiking neural network composed of integral spiking neurons, as follows:

[0051] Step 7.1: Pass the feature-fused radar signal through the fully connected layer of the convolutional spiking neural network composed of integral-spiking neurons. The pulse frequency output by the integral-spiking neurons is as follows:

[0052]

[0053] Where k represents the kth neuron, and the maximum value is the number of categories of radar signals; It represents the cumulative number of pulses emitted by the k-th neuron within the total simulation step length;

[0054] Step 7.2: Among all neurons, the neuron number with the largest output pulse frequency is the predicted category of the input signal.

[0055] As a specific implementation, in step 8, when the convolutional spiking neural network backpropagates errors through the chain rule, gradients are used to replace backpropagated errors, the gradient factors are optimized, the convolutional spiking neural network parameters are updated, the convergence and stability of the convolutional spiking neural network are enhanced, and the convolutional spiking neural network with updated parameters is used to complete the classification of the radar signal, as follows:

[0056] Step 8.1: When the convolutional spiking neural network backpropagates the error through the chain rule, a differentiable function that approximates the step function is used to replace the gradient, complete the backpropagation of the error, and update the convolutional spiking neural network parameters:

[0057]

[0058] Where α is the gradient factor and x is the independent variable;

[0059] Step 8.2. Compare the performance of the convolutional spiking neural network under different gradient factors α, optimize the gradient factor that affects the gradient size in the substitution function, and increase the convergence of the convolutional spiking neural network.

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] Example

[0062] The radar signal classification method under low signal-to-noise ratio of the present invention comprises the following specific steps:

[0063] Step 1: Use Choi-Williams distribution to perform time-frequency transformation on the radar signal to obtain the time-frequency characteristics of the signal under different signal-to-noise ratios;

[0064] Step 2: Use convolutional spike neural networks to extract spatial features, pulse feature information, accumulate temporal feature information, and fuse spatiotemporal features from time-frequency images to enhance the characteristics of different signals, optimize the feature accumulation time step, and improve the network's noise generalization ability;

[0065] Step 3: Use the accumulated features to classify radar signals, use the pulse frequency output by the integrated pulse neuron as the classification criterion, use the gradient to replace the back propagation error, optimize the gradient factor, update the network parameters, and enhance the convergence and stability of the network.

[0066] Furthermore, the radar signal is subjected to time-frequency transformation using Choi-Williams distribution as described in step 1. The specific steps are as follows:

[0067] Step 1.1: Add different Gaussian white noises to different radar signals and use Choi-Williams distribution to convert the signals into time-frequency images with different noise interferences.

[0068] Step 1.2: Use bi-trilinear interpolation to scale the time-frequency image to the 32 pixel × 32 pixel size required for the convolutional spike neural network input.

[0069] Furthermore, the convolutional spike neural network described in step 2 is used to perform spatial feature extraction, feature information pulsing, temporal feature information accumulation, and spatiotemporal feature fusion on the time-frequency image to enhance the features of different signals, optimize the feature accumulation time step, and improve the network noise generalization ability. The specific steps are as follows:

[0070] Step 2.1: The first convolutional layer extracts the spatial features of the time-frequency image and adaptively encodes the extracted eigenvalues. The encoded eigenvalues ​​are fed into the leaky integral-spark neuron for pulsing. The pulsing first undergoes a charging process, as shown in the following formula:

[0071] τ n (U[t]-U[t-1])=-(U[t-1]-U reset )+X[t]

[0072] Where τ n is the time constant, U reset is the reset voltage, X(t) is the input eigenvalue, where U[t] is the membrane potential at the current moment, and U[t-1] is the membrane potential at the end of the previous moment. The accumulation of input eigenvalues ​​is manifested as an increase in membrane potential. Within each accumulation time step, if the membrane potential threshold is reached, the leakage current integration and spiking neuron will release a pulse, completing the pulsing of characteristic information, as shown in the following formula:

[0073] P(t)=Θ(U tmp -U threshold )

[0074] Among them U threshold is the threshold voltage for neuron activation, P(t) is the output pulse signal, and Θ(x) is the step function.

[0075] Step 2.2: The characteristic information of the pulse is a pulse train with time information. The leakage current integrating and releasing neurons accumulates the pulse train carrying time information and manifests as an increase in membrane potential as follows:

[0076] U t+1 =U t (1-P t-1 )+X(t)

[0077] where p t-1 The threshold is U threshold The step function represents the pulses released by the leaky integrate-and-send neuron.

[0078] Step 2.3: The convolutional layer and the pulse neuron layer alternately transmit feature information to complete the spatiotemporal information fusion.

[0079] Step 2.4: Optimize the feature accumulation time step to enhance network convergence and reduce inference latency.

[0080] Furthermore, the radar signal classification is performed using the accumulated features described in step 3, with the pulse frequency of the integrated pulse neuron output as the classification criterion. The gradient is used to replace the back-propagation error, the gradient factor is optimized, the network parameters are updated, and the convergence and stability of the network are enhanced. The specific steps are as follows:

[0081] Step 3.1: The radar signal after feature fusion is classified by a fully connected layer composed of integral and spiking neurons. The pulse frequency output by the integral and spiking neurons is as follows:

[0082]

[0083] Where k represents the kth neuron, and the maximum value is the number of categories of radar signals. It represents the cumulative number of pulses emitted by the kth neuron within the total simulation step. Among all neurons, the number of the neuron with the highest output pulse frequency is the predicted category of the input signal.

[0084] Step 3.2: When the convolutional spike neural network backpropagates the error through the chain rule, a differentiable function that approximates the step function is used to replace the gradient to complete the backpropagation of the error and update the network parameters.

[0085] Step 3.3: Optimize the gradient factor that affects the gradient size in the substitution function to enhance the convergence of the network.

[0086] Combine Figures 1 and 2 The data used in this embodiment is the radar signal time-frequency diagram generated by MATLAB simulation. The time-frequency diagram includes 2PSK, 2FSK, 4FSK, CW, LFM and NLFM signals from 0dB to -15dB, and the signal-to-noise ratio interval is 3dB.

[0087] Firstly, the Choi-Williams distribution is used to convert the radar signal into a radar signal time-frequency graph. Then, bi-trilinear interpolation is used to scale the radar signal time-frequency graph to 32 pixels × 32 pixels. Random flipping and random arrangement are used to expand the sample features and enhance generalization.

[0088] Then the radar signal time-frequency image is sent to Figure 2 The convolutional spike neural network (SCNN) shown in the figure extracts spatial features, accumulates temporal features, and fuses spatiotemporal information of radar signals. The optimized time step is T = 10.

[0089] Finally, the fused features are used to classify radar signals, and the inverse tangent function is used for gradient substitution, with the optimized gradient factor α = 2. Finally, a convolutional spike neural network with T = 10 and α = 2 is used to complete radar signal classification under low signal-to-noise ratio.

[0090] Table 1 Comparison of classification results between the method of the present invention and the traditional convolutional neural network method

[0091]

[0092] The classification results of the traditional convolutional neural network method and the classification results of the present invention are shown in Table 1. Compared with the traditional convolutional neural network classification method, the present invention has a smaller network model and significantly improves the classification accuracy under low signal-to-noise ratio.

Claims

1. A radar signal classification method under low signal-to-noise ratio, characterized in that: The following steps are involved: Step 1: Use Choi-Williams distribution to perform time-frequency transformation on the radar signal, convert the radar signal into a radar signal time-frequency graph, and obtain the time-frequency characteristics of the radar signal under different signal-to-noise ratios; Step 2: Use bi-trilinear interpolation to scale the radar signal time-frequency graph to the input size of the convolutional spike neural network, and use random flipping and random permutation to expand the features of the signal time-frequency graph; Step 3: Use the first convolutional layer of the convolutional spike neural network to extract spatial features from the radar signal time-frequency graph and perform adaptive coding; Step 4: The encoded radar signal spatial feature information is fed into the leaky integral-and-release pulse neuron of the convolutional spike neural network and converted into a pulse sequence to obtain the time series feature information of the radar signal; Step 5: By comparing the effects of different time steps on network stability and inference delay, the optimal time step of the network is selected. The characteristic values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spiking neurons to stimulate them to release pulses, and thus complete the forward transmission of the characteristic information; Step 6: The convolutional layer and the leaky integrated release pulse neuron layer in the convolutional spike neural network alternately transmit the spatiotemporal information of the radar signal to complete the spatiotemporal feature fusion of the radar signal; Step 7: The radar signal after the spatiotemporal feature fusion is passed through the fully connected layer of the convolutional spiking neural network composed of integral spiking neurons to classify the radar signal; Step 8. When the convolutional spiking neural network backpropagates the error through the chain rule, the gradient is used to replace the backpropagation error, the gradient factor is optimized, the convolutional spiking neural network parameters are updated, the convergence and stability of the convolutional spiking neural network are enhanced, and the convolutional spiking neural network with updated parameters is used to complete the classification of the radar signal.

2. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 1, the radar signal is transformed into a radar signal time-frequency diagram using the Choi-Williams distribution, and the time-frequency characteristics of the radar signal under different signal-to-noise ratios are obtained, as follows: Different Gaussian white noises are added to different radar signals, and the Choi-Williams transform is used to convert the radar signals into time-frequency diagrams of radar signals with different noise interferences.

3. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 2, the radar signal time-frequency graph is scaled to the input size of the convolutional spike neural network using bilinear interpolation, and the features of the signal time-frequency graph are expanded using random flipping and random permutation, as follows: Step 2.1: Use bi-trilinear interpolation to scale the radar signal time-frequency graph to the convolutional spike neural network input size. Step 2.2: Use random flipping and random arrangement to expand the features of the time-frequency graph to enhance generalization.

4. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 3, the first convolutional layer of the convolutional spike neural network is used to extract spatial features from the radar signal time-frequency diagram and perform adaptive coding, as follows: Step 3.1, the first convolutional layer of the convolutional spike neural network extracts spatial features from the radar signal time-frequency graph through convolution operation; Step 3.2: The first convolutional layer adaptively encodes the spatial features of the radar signal time-frequency graph through training.

5. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 4, the encoded radar signal spatial feature information is fed into the leaky integral-and-release pulse neuron of the convolutional pulse neural network and converted into a pulse sequence to obtain the time series feature information of the radar signal, as follows: Step 4.1: The encoded radar signal spatial eigenvalues ​​are fed into the leakage integration and release pulse neuron for pulse conversion. The pulse conversion first undergoes a charging process, as shown in formula (1): τ n (U[t]-U[t-1])=-(U[t-1]-U reset )+X[t] (1) Where τ n is the time constant, U reset is the reset voltage, X(t) is the input characteristic value, where U[t] is the membrane potential at the current moment, and U[t-1] is the membrane potential at the end of the previous moment; The accumulation of input characteristic values ​​is manifested as an increase in membrane potential. Within each accumulation time step, if the membrane potential threshold is reached, the leakage integral pulse neuron releases a pulse, completing the pulse conversion of radar signal characteristic information, as shown in formula (2): P(t)=Θ(U tmp -U threshold ) (2) Among them U threshold is the threshold voltage for the activation of the spiking neuron, P(t) is the output pulse signal, and Θ(x) is the step function; Step 4.2: The pulsed radar signal spatial characteristic information is a pulse train with the radar signal time information. The leakage integrating and releasing neurons accumulate the pulse train carrying the radar signal time information, which is manifested as an increase in membrane potential, as shown in formula (3): U t+1 =U t (1-P t-1 )+X(t) (3) where p t-1 The threshold is U threshold The step function represents the pulses released by the leaky integrate-and-send neuron.

6. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 5, by comparing the effects of different time steps on network stability and inference delay, the optimal time step of the network is selected. The characteristic values ​​are continuously accumulated within the optimal time step to accumulate the pulsed time series features, increase the membrane potential of the spiking neurons to stimulate them to release pulses, and then complete the forward transmission of the characteristic information. The details are as follows: By comparing the effects of different time steps on network stability and inference delay, the optimal time step is optimized. Then, the time series features of the pulsed radar signal are continuously accumulated until the given optimal time step ends. The membrane potential of the spiking neuron increases. When the membrane potential reaches the threshold, the spiking neuron releases a pulse, completing the forward transmission of the feature information.

7. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In the convolutional spike neural network described in step 6, the convolutional layer and the leaky integrated discharge pulse neuron layer alternately transmit the spatiotemporal information of the radar signal to complete the spatiotemporal feature fusion of the radar signal, as follows: The first convolutional layer of the convolutional spiking neural network and the leaky integral release pulse neuron layer alternately transmit information to complete the spatiotemporal feature fusion of the radar signal and enhance the learning ability of the convolutional spiking neural network for different radar signal features.

8. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 7, the radar signal after the spatiotemporal feature fusion is classified by passing it through a fully connected layer of a convolutional spiking neural network composed of integral spiking neurons, as follows: Step 7.1: Pass the feature-fused radar signal through the fully connected layer of the convolutional spiking neural network composed of integral-spiking neurons. The pulse frequency output by the integral-spiking neurons is as follows: Where k represents the kth neuron, and the maximum value is the number of categories of radar signals; It represents the cumulative number of pulses emitted by the k-th neuron within the total simulation step length; Step 7.2: Among all neurons, the neuron number with the largest output pulse frequency is the predicted category of the input signal.

9. The radar signal classification method under low signal-to-noise ratio according to claim 1, characterized in that: In step 8, when the convolutional spiking neural network backpropagates errors through the chain rule, the gradient is used to replace the backpropagation error, the gradient factor is optimized, the convolutional spiking neural network parameters are updated, the convergence and stability of the convolutional spiking neural network are enhanced, and the convolutional spiking neural network with updated parameters is used to complete the classification of the radar signal, as follows: Step 8.1: When the convolutional spiking neural network backpropagates the error through the chain rule, a differentiable function that approximates the step function is used to replace the gradient, complete the backpropagation of the error, and update the convolutional spiking neural network parameters: Where α is the gradient factor and x is the independent variable; Step 8.

2. Compare the performance of the convolutional spiking neural network under different gradient factors α, optimize the gradient factor that affects the gradient size in the substitution function, and increase the convergence of the convolutional spiking neural network.

Images