A system and method for intelligent classification and recognition of multiple gunshot mixtures based on data distillation

Through data distillation and knowledge distillation technology, combined with FPGA and neural network acceleration module, the problems of low recognition accuracy of multi-gun sound mixed signal and high consumption of computing resources are solved, and an efficient and lightweight gun sound recognition system is realized, which is convenient for portable deployment.

CN114299993BActive Publication Date: 2024-12-03ZHONGBEI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111602267.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-24
Publication Date
2024-12-03
Estimated Expiration
2041-12-24

AI Technical Summary

Technical Problem

Existing gunshot recognition technology is difficult to handle multiple gunshot mixed signals, with reduced recognition accuracy, high computing resources and bulky equipment, making it difficult to achieve portable deployment.

Method used

Using a multi-gun sound hybrid intelligent classification recognition system based on data distillation, the mixed gun sound signal is separated into a single gun sound signal through data acquisition, separation, preprocessing and feature extraction modules, and the knowledge distillation method of teacher network and student network is reduced. Network parameters are combined with FPGA and neural network acceleration module for lightweight processing to realize gunshot category recognition.

Benefits of technology

It realizes simultaneous recognition of multiple gunshot mixed signals, improves recognition accuracy and real-timeness, reduces the demand for computing resources, and reduces the system size to the palm size, making it easy to deploy with portable deployment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114299993B_ABST
    Figure CN114299993B_ABST
Patent Text Reader

Abstract

The present invention relates to a multi-gunshot intelligent classification and recognition system and method based on data distillation, including a data acquisition module, a data separation module, a data preprocessing module, a feature extraction module, a neural network and acceleration module, and an output module; first, a plurality of mixed gunshot signals are collected by using the data acquisition module, and these mixed gunshot signals are transmitted to the data separation module, multiple gunshot data are decoupled and separated, and single gunshot data is output and input into the data preprocessing module, each single gunshot data is adaptively filtered to remove aliased noise signals, and the filtered multiple sound signals are input into the feature extraction module, and finally the classification structure is input into the output module. The present invention can separate multiple gunshot mixed signals into multiple single gunshot signals, and then identify each single gunshot signal separately, so as to achieve the function of simultaneous classification of multiple gunshots. The number of network parameters can be effectively reduced, and the volume of the entire system can be reduced to the size of a palm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of gunshot recognition, and in particular relates to a system and method for intelligent classification and recognition of multiple gunshot mixtures based on data distillation. Background Art

[0002] In conventional warfare, peacekeeping and counter-terrorism operations, snipers can efficiently kill enemy personnel and damage key equipment, causing huge casualties and psychological panic to the army, and their threats are increasing. In order to alleviate the threat posed by snipers, it is necessary to study an efficient anti-sniper detection system. Among them, sniper gunshot category recognition technology is one of the key technologies of the anti-sniper detection system.

[0003] With the development of artificial intelligence, gunshot recognition technology mainly draws on deep neural networks for sound source classification such as conventional urban noise classification, mixed sound event detection, and medical heart sound classification diagnosis. The following problems exist when applying the above deep neural networks to gunshot recognition:

[0004] When multiple gunshots are mixed and input into the system at the same time, only a single gunshot can be identified.

[0005] In current gunshot recognition tasks, a single gunshot is assumed as input. Although the recognition accuracy is constantly improving, in actual gunshot recognition tasks, most of them require multiple gunshots to be recognized. When a single gunshot recognition system is used for mixed recognition of multiple gunshots, not only will the recognition accuracy be reduced, but it can only recognize one type of gunshot in the mixed gunshot.

[0006] 2. The gunshot recognition network consumes a lot of computing power and time, making it difficult to achieve strong real-time performance and high accuracy.

[0007] As the accuracy of gunshot recognition increases, the network model is gradually deepened. Large models can meet the accuracy requirements, but they also require extreme computing power of computers and increase the time required for recognition.

[0008] 3. The terminals for deploying deep neural networks are bulky and difficult to be portable.

[0009] Deep neural networks are composed of a large number of neuron units, and powerful computing power is also required when loading model tests. Currently, almost all deep neural networks require graphics cards for acceleration. Large, high-power graphics cards cannot be portable and deployed at the edge. Summary of the invention

[0010] The present invention provides a multi-gunshot mixed intelligent classification and recognition system and method based on data distillation to solve the above problems.

[0011] In order to solve the above technical problems, the present invention provides a multi-gunshot mixed intelligent classification and recognition system based on data distillation, which is characterized by: comprising a data acquisition module, a data separation module, a data preprocessing module, a feature extraction module, a neural network and acceleration module, and an output module;

[0012] The data acquisition module collects multiple mixed gunshot signals and transmits these mixed gunshot signals to the data separation module, decouples and separates the multiple gunshot data, outputs single gunshot data, and inputs it to the data preprocessing module, adaptively filters each single gunshot data to remove aliased noise signals, and inputs the multiple sound signals after filtering to the feature extraction module. The feature extraction module is used to perform Mel mapping on each sound signal to make the one-dimensional sound signal become a two-dimensional image signal, and input these image signals into the neural network classification and acceleration module to classify the above features and accelerate the neural network operation, and finally input the classification structure into the output module to obtain the gunshot category;

[0013] The neural network and acceleration module include a teacher network, a student network and an acceleration module;

[0014] The two-dimensional image is input into the teacher network, which learns an accurate probability distribution and guides the student network to obtain a probability distribution similar to that of the teacher network. Finally, the student network extracts features from the two-dimensional image to obtain the classification of gunshot features, and accelerates the calculation of the student network through the acceleration module.

[0015] The data acquisition module collects the gunshot information collected by the sound field sensor and converts it into a digital signal output after each segment lasts for a certain period of time.

[0016] The data separation module decomposes the digital signal into a series of intrinsic mode components through the EEMD algorithm and inputs them into the FastICA algorithm to obtain multiple single-gun sound separation signals.

[0017] The data preprocessing module performs filtering operations on the multiple single-gunshot sound signals output by the data separation module through an adaptive filter, and outputs multiple pure sound signals of the single-gunshot sounds.

[0018] The feature extraction module transformation method is as follows:

[0019] 1) Pre-emphasis: Filter the gunshot signal through a high-pass filter;

[0020] 2) Framing: After pre-emphasis, the signal is divided into multiple short-time frames;

[0021] 3) Windowing: After dividing the signal into frames, each frame is multiplied by a window function;

[0022] 4) Calculate the signal energy spectrum: Perform a fast Fourier transform operation on each frame of the signal to obtain the frequency spectrum and time-phase spectrogram of the gunshot signal, and then take the square of the modulus of the frequency spectrum to obtain the power spectrum of the gunshot signal;

[0023] 5) Mel filtering: extract the frequency bands of the power spectrum through a set of triangular filters with equal Mel scales to obtain a Mel spectrogram;

[0024] 6) Take the logarithm of the obtained Mel-spectrogram to obtain a logarithmic Mel-spectrogram.

[0025] The teacher network consists of 10 two-dimensional residual modules and a global average pooling module. The two-dimensional residual module is used to increase the network depth. After the logarithmic Mel-spectrogram is extracted by ten residual modules, it is input into the global average pooling module for classification, and finally the probability distribution of gunshot categories is obtained.

[0026] In each two-dimensional residual network module, one path of the input signal is first processed by the batch normalization layer and Relu activation, and then a two-dimensional convolution operation is performed. The above batch normalization layer, Relu activation, and two-dimensional convolution operation are repeated once before output. The other path is downsampled through the average pooling layer and then concatenated with the output data of the first path on the feature channel to obtain the output of the two-dimensional residual module.

[0027] The student network consists of three depthwise separable convolution modules and one global average pooling module. The input features are processed in two ways. One way is activated by BN and Relu, and then processed by the depthwise separable convolution module. Finally, it is superimposed with the other way of features to form a residual structure. This structure is repeated twice and finally classified by the global average pooling module to obtain the probability distribution of gunshot categories.

[0028] A method for intelligent classification and recognition of multiple gunshot mixtures, characterized in that a neural network adopts a knowledge distillation method, and the specific steps are as follows:

[0029] S1. Train a deep neural network model as the teacher network, and use softmax to soften the features extracted by the deep neural network. The softened softmax formula is as follows to generate softened labels, and then generate the teacher network loss value;

[0030]

[0031] T is the softening time, z i is the input matrix of the i-th gunshot signal, q i is the mapping result of the i-th gunshot signal, and j is one of all gunshot signals;

[0032] S2. Create a student network, soften the final result through softmax, output the actual label represented by a one-hot vector as the softened prediction value, and generate the student network loss value, which is consistent with the teacher network loss value formula;

[0033] S3, weighted sum of the teacher network loss value and the student network loss value to generate a total loss value;

[0034] S4, the total loss value is used to iteratively optimize the student network until it is corrected to the best;

[0035] S5. Change T in the student network loss value to 1, and the trained student network model is obtained. The corresponding prediction label, i.e., the classification of gunshot features, can be obtained by inputting data.

[0036] Beneficial effects:

[0037] 1. The present invention integrates a gunshot separation algorithm, which can separate multiple gunshot mixed signals into multiple single gunshot signals, and then identify each single gunshot signal separately, thus achieving the function of simultaneous classification of multiple gunshot mixtures.

[0038] 2. The present invention adopts a knowledge distillation network structure in which a large network drives a small network to learn, which can effectively reduce the number of network parameters, make the model lightweight, and effectively improve the recognition rate and real-time performance.

[0039] 3. The present invention integrates the entire recognition system into the FPGA, and uses the neural network accelerator to accelerate the network optimization. The gunshot preprocessing is performed in the ARM core integrated in the FPGA. In the neural network accelerator optimization network, the two operations are executed in parallel in a multi-threaded manner. Since GPU acceleration is abandoned and DPU in neural network acceleration is used for acceleration, power consumption is reduced and the size of the entire system is reduced to the size of a palm. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Overall design of gunshot recognition system

[0041] Figure 2 Schematic diagram of knowledge distillation network training

[0042] Figure 3 Teacher network structure diagram

[0043] Figure 4 Schematic diagram of the 2D residual network module structure

[0044] Figure 5 Student network diagram

[0045] Figure 6 DPU flow chart. DETAILED DESCRIPTION

[0046] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below.

[0047] The invention proposes a multi-gunshot mixed intelligent classification and recognition system based on data distillation, which includes a data acquisition module, a data separation module, a data preprocessing module, a feature extraction module, a neural network classification and acceleration module, and an output module. Figure 1 shown.

[0048] First, use the data acquisition module to collect multiple mixed gunshot signals, and pass these mixed gunshot signals into the data separation module, decouple and separate the multiple gunshot data, output single gunshot data, and input it into the data preprocessing module, adaptively filter each single gunshot data to remove aliased noise signals, and input the multiple sound signals after filtering into the feature extraction module. In this module, each gunshot data will be Mel-mapped to make the one-dimensional sound signal into a two-dimensional image signal, and these image signals will be input into the neural network classification and acceleration module to classify the above features and accelerate the neural network operation, and finally input the classification structure into the output module.

[0049] Data acquisition module: Use the PL end of FPGA to control the multi-channel AD7606 chip to collect gunfire information collected by the sound field sensor at a sampling frequency of 44.1KHz. Each segment lasts for 5S and is converted into digital signal output.

[0050] Data separation module: The EEMD-Fast ICA model method is used to separate the mixed gunshot signal into multiple single gunshot signals. The specific steps are as follows: 1. Decompose it into a series of intrinsic mode components through EEMD (ensemble average empirical mode decomposition algorithm). 2. Input the intrinsic mode components obtained by EEMD decomposition into the Fast ICA algorithm to obtain multiple single gunshot separation signals.

[0051] Data preprocessing module: An adaptive filter is designed using FPGA to filter the multiple single gunshot signals output by the data separation module, reduce the ambient noise in the mixed gunshots, and output multiple pure sound signals of single gunshots.

[0052] Feature extraction module: Perform Mel transform on the one-dimensional sound signal output by the data preprocessing module and map it into a logarithmic Mel spectrum. The specific steps are as follows:

[0053] 1) Pre-emphasis: The gunshot signal is passed through a high-pass filter to balance the spectrum and improve the signal to avoid numerical problems during the Fourier transform operation.

[0054] 2) Framing: After pre-emphasis, the signal is divided into short time frames of 50ms in length, and each frame shift length is 1024 points. The gunshot signal is non-stationary, and it is meaningless to perform Fourier transform on the entire signal because the frequency contour of the signal is lost over time.

[0055] 3) Windowing: After dividing the signal into frames, each frame is multiplied by a window function, such as a Hamming window, to increase the continuity of the left and right ends of the frame.

[0056] 4) Calculate the signal energy spectrum: Perform a 2048-point fast Fourier transform operation on each frame of the signal to obtain the frequency spectrum, collect the vector angles, and obtain a phase spectrum with a dimension size of [64, 216], where 64 is the phase scale and 216 is the time dimension. Then take the square of the modulus to obtain the spectral line energy of the gunshot signal.

[0057] 5) Mel filtering: The Mel filter bank is calculated and the power spectrum is passed through a set of 64 filters with equal high Mel scale triangular filters to extract the frequency bands.

[0058] 6) Take the logarithm of the obtained Mel spectrogram to obtain an approximate homomorphic transformation effect, and the output dimension size is (64, 216) logarithmic Mel spectrogram, where 64 is the Mel scale and 216 is the time dimension.

[0059] The neural network and acceleration module include a teacher network, a student network and an acceleration module; a two-dimensional image is input into the teacher network, which learns an accurate probability distribution and guides the student network to obtain a probability distribution similar to that of the teacher network. Finally, the student network extracts features from the two-dimensional image to obtain a classification of gunshot features, and the acceleration module accelerates the calculation of the student network.

[0060] The neural network in this design adopts the knowledge distillation method to optimize the neural network structure to the greatest extent and reduce the number of network parameters. Model distillation is to transfer the "knowledge" of the promotion ability of the trained complex model to a network with a simpler structure. A large-scale network with a large number of parameters is used as the teacher network, and a small-scale network with a small number of parameters is used as the student network. By training the large-scale teacher network, a more accurate probability distribution is obtained, thereby guiding the student network to iterate and obtain a probability distribution similar to that of the teacher network, such as Figure 2 , the specific steps are as follows:

[0061] 1. Train a deep neural network as the teacher network. The model has high accuracy after training with a large amount of data, but the model is bulky. The features extracted by this deep neural network are softened by the final result of softmax. The softened softmax formula is as follows to generate softened labels, and then generate the teacher network loss value;

[0062]

[0063] T is the softening time, z i is the input matrix of the i-th gunshot signal, q i is the mapping result of the i-th gunshot signal, j is one of all gunshot signals. It can be obtained from the formula that the softening process is to divide the softmax input by time T. T can be taken at will. The higher T is, the smoother the output probability distribution of softmax is, and the greater the entropy of its distribution is. The information carried by the negative label will be relatively amplified, and the model training will pay more attention to the negative label.

[0064] Create a lightweight neural network with normal foreimage propagation, soften the final result after softmax, output the actual label with a one-hot vector, that is, the softened prediction value, and generate the student network loss value;

[0065] The teacher network loss value and the student network loss value are weighted and summed to generate the total loss value;

[0066] The total loss value is used to iteratively optimize the lightweight network until it is corrected to the best;

[0067] By changing T to 1, we can obtain a trained lightweight network model. By inputting data, we can get the corresponding prediction label, which can be deployed to the edge.

[0068] Teacher network: The teacher network consists of 10 two-dimensional residual modules and a global average pooling module. The network structure is as follows: Figure 3 As shown in Figure 1, the two-dimensional residual module can effectively increase the network depth and prevent the gradient from vanishing or exploding. After the logarithmic Mel spectrogram is extracted through ten residual modules, it is input into the global average pooling module for classification, and finally the probability distribution of gunshot categories is obtained.

[0069] Each two-dimensional residual network module structure is as follows Figure 4 As shown in the figure, one path of the input signal first undergoes batch normalization layer processing (BN) and Relu activation function, and then performs a two-dimensional convolution operation with a convolution kernel size of 3*3, and then repeats the above batch normalization layer, Relu activation function, and two-dimensional convolution operation before output. The other path is downsampled through the average pooling layer and then concatenated with the output data of the first path on the feature channel to obtain the output of the two-dimensional residual module.

[0070] Student network: The student network consists of three depthwise separable convolution modules and a global average pooling module. Depthwise separable convolution can achieve fewer parameters and less computation, making it more suitable for deployment on the edge. Unlike standard convolutional networks, depthwise separable convolution splits the convolution kernel into a single channel form. Without changing the depth of the input feature image, convolution operations are performed on each channel, thus obtaining an output feature map with the same number of channels as the input feature map. Figure 5 As shown in the figure, the input features are processed in two ways, one way is through BN and Relu activation function, and then through the depth separable convolution module, and finally superimposed with the other way of features to form a residual structure. This structure is repeated twice and finally classified through the global average pooling module.

[0071] A method for intelligent classification and recognition of multiple gunshot mixtures, characterized in that a neural network adopts a knowledge distillation method, and the specific steps are as follows:

[0072] S1. Train a deep neural network model as the teacher network, and use softmax to soften the features extracted by the deep neural network. The softened softmax formula is as follows to generate softened labels, and then generate the teacher network loss value;

[0073]

[0074] T is the softening time, z i is the input matrix of the i-th gunshot signal, q i is the mapping result of the i-th gunshot signal, and j is one of all gunshot signals;

[0075] S2. Create a student network, soften the final result through softmax, output the actual label represented by a one-hot vector as the softened prediction value, and generate the student network loss value, which is consistent with the teacher network loss value formula;

[0076] S3, weighted sum of the teacher network loss value and the student network loss value to generate a total loss value;

[0077] S4, the total loss value is used to iteratively optimize the student network until it is corrected to the best;

[0078] S5. Change T in the student network loss value to 1, and the trained student network model is obtained. The corresponding prediction label, i.e., the classification of gunshot features, can be obtained by inputting data.

[0079] The present invention uses Keras as a framework and Tensorflow as a backend to build a deep learning environment. Two 1080Ti GPUs are used to accelerate the image processing process in combination with the parallel computing architecture CUDA. The loss function of the overall network uses cross entropy, the optimizer uses SGD, and 8000 groups of gunshot data output by the data acquisition module are collected as training samples and input into the above-mentioned deep learning network. Iterate 500 rounds, take 128 batches of samples for training at a time during the training process, and the learning rate is 0.001. The corresponding deep learning network parameters are obtained.

[0080] Network acceleration module: FPGA uses DPU process as follows Figure 6 As shown:

[0081] 1. Model quantization and compression: The lightweight student model trained and saved in S5.2 is quantized and compressed, that is, floating-point numbers are converted to fixed-point numbers. The specific steps are: first use the Netro tool to visualize the network structure, then quantize the compressed model and calibrate it using the test set.

[0082] 2. DPU compilation: Deploy the network model obtained in 1 on the DPU platform, compile the DPU kernel, and generate the kernel file .elf.

[0083] 3. Mixed compilation: Cross-compile the kernel file obtained in 2 to get the final dynamic link library.

[0084] 4. Run the network: Deploy the entire network program file and dynamic link library to the FPGA end, run the network, and realize the network acceleration function.

[0085] The student network is deployed in the FPGA, and 2000 sets of gunshot data output by the data acquisition module are collected. After processing in steps S1, S2, S3 and S4, they are input into the student network for testing. After testing, the network recognition accuracy can reach 86%.

[0086] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-gunshot mixed intelligent classification and recognition system based on data distillation, characterized by: It includes data acquisition module, data separation module, data preprocessing module, feature extraction module, neural network and acceleration module, and output module; The data acquisition module collects multiple mixed gunshot signals and transmits these mixed gunshot signals to the data separation module, decouples and separates the multiple gunshot data, outputs single gunshot data, and inputs it to the data preprocessing module, adaptively filters each single gunshot data to remove aliased noise signals, and inputs the multiple sound signals after filtering to the feature extraction module. The feature extraction module is used to perform Mel mapping on each sound signal to make the one-dimensional sound signal become a two-dimensional image signal, and input these image signals into the neural network classification and acceleration module to classify the above features and accelerate the neural network operation, and finally input the classification structure into the output module to obtain the gunshot category; The neural network and acceleration module include a teacher network, a student network and an acceleration module; The teacher network is composed of 10 two-dimensional residual modules and a global average pooling module. The two-dimensional residual module is used to increase the network depth. After the logarithmic Mel spectrogram is extracted by ten residual modules, it is input into the global average pooling module for classification, and finally the probability distribution of gunshot categories is obtained; The student network consists of three depthwise separable convolution modules and one global average pooling module. The input features are processed in two ways. One way is activated by BN and Relu, and then processed by the depthwise separable convolution module. Finally, it is superimposed with the other way of features to form a residual structure. This structure is repeated twice and finally classified by the global average pooling module to obtain the probability distribution of gunshot categories. The two-dimensional image is input into the teacher network, which learns an accurate probability distribution and guides the student network to obtain a probability distribution similar to that of the teacher network. Finally, the student network extracts features from the two-dimensional image to obtain the classification of gunshot features, and accelerates the calculation of the student network through the acceleration module.

2. According to claim 1, a multi-gunshot mixed intelligent classification and recognition system based on data distillation is characterized by: The data acquisition module collects the gunshot information collected by the sound field sensor and converts it into a digital signal output after each segment lasts for a certain period of time.

3. The multi-gunshot mixed intelligent classification and recognition system based on data distillation according to claim 1 is characterized by: The data separation module decomposes the digital signal into a series of intrinsic mode components through the EEMD algorithm and inputs them into the Fast ICA algorithm to obtain multiple single-gun sound separation signals.

4. The multi-gunshot mixed intelligent classification and recognition system based on data distillation according to claim 1 is characterized by: The data preprocessing module performs filtering operations on the multiple single-gunshot sound signals output by the data separation module through an adaptive filter, and outputs multiple pure sound signals of the single-gunshot sounds.

5. The multi-gunshot mixed intelligent classification and recognition system based on data distillation according to claim 1 is characterized by: The feature extraction module transformation method is as follows: 1) Pre-emphasis: Filter the gunshot signal through a high-pass filter; 2) Framing: After pre-emphasis, the signal is divided into multiple short-time frames; 3) Windowing: After dividing the signal into frames, each frame is multiplied by a window function; 4) Calculate the signal energy spectrum: Perform a fast Fourier transform operation on each frame of the signal to obtain the frequency spectrum and time-phase spectrum of the gunshot signal, and then take the square of the modulus of the spectrum to obtain the power spectrum of the gunshot signal; 5) Mel filtering: A set of triangular filters with equal Mel scales are used to extract the frequency bands of the power spectrum to obtain a Mel spectrogram; 6) Take the logarithm of the obtained Mel-spectrogram to obtain a logarithmic Mel-spectrogram.

6. The multi-gunshot mixed intelligent classification and recognition system based on data distillation according to claim 1 is characterized by: In each two-dimensional residual network module, one path of the input signal is first processed by the batch normalization layer and Relu activation, and then a two-dimensional convolution operation is performed. The above batch normalization layer, Relu activation, and two-dimensional convolution operation are repeated once before output. The other path is downsampled through the average pooling layer and then concatenated with the output data of the first path on the feature channel to obtain the output of the two-dimensional residual module.

7. A method for multi-gunshot mixed intelligent classification and recognition based on a multi-gunshot mixed intelligent classification and recognition system based on data distillation according to any one of claims 1 to 6, characterized in that: The neural network uses the knowledge distillation method. The specific steps are as follows: S1. Train a deep neural network model as the teacher network, and use softmax to soften the features extracted by the deep neural network. The softened softmax formula is as follows to generate softened labels, and then generate the teacher network loss value; (1) T is the softening time, z i is the input matrix of the i-th gunshot signal, q i is the mapping result of the i-th gunshot signal, and j is one of all gunshot signals; S2. Create a student network, soften the final result through softmax, output the actual label represented by a one-hot vector as the softened prediction value, and generate the student network loss value, which is consistent with the teacher network loss value formula; S3, weighted sum of the teacher network loss value and the student network loss value to generate a total loss value; S4, the total loss value is used to iteratively optimize the student network until it is corrected to the best; S5. Change T in the student network loss value to 1, and the trained student network model is obtained. The corresponding prediction label, i.e., the classification of gunshot features, can be obtained by inputting data.

8. The multi-gunshot mixed intelligent classification and recognition system based on data distillation according to claim 7 is characterized by: The student network is deployed to the FPGA.