A method for constructing interference knowledge base based on deep learning network

Abstract

The present invention relates to a method for constructing an interference knowledge base based on a deep learning network, which belongs to the field of interference identification technology and comprises the following steps: S1: analyzing the characteristics of interference signals in complex scenarios, and providing different interference suppression strategies for different types of interference signals in the surrounding environment of communication terminal equipment; S2: constructing an interference knowledge base offline training module: based on scenario analysis, using original interference data to establish an initial knowledge base and train a deep learning network to obtain an interference identification model with online identification capability; S3: constructing an interference knowledge base online learning module: matching a feature data set obtained after module identification of interference data collected in real time with an initial database, and if it is a new data set, updating the knowledge base online to obtain an interference knowledge base construction module with online learning capability.

Description

Technical Field

[0001] The present invention belongs to the technical field of interference identification and relates to a method for constructing an interference knowledge base based on a deep learning network. Background Art

[0002] The complex electromagnetic environment leads to different forms of interference in the actual environment, which seriously affects the effectiveness and reliability of information transmission of terminal devices during communication. Therefore, in order to improve the adaptive communication capabilities of terminal devices in known, partially known, and unknown electromagnetic environments, corresponding anti-interference technologies are urgently needed to meet the waveform design and anti-interference decision-making requirements of anti-interference in actual communication environments.

[0003] In anti-interference technology, decision trees, BP (Back propagation) neural networks, deep neural networks (DNN) and other algorithms are usually used to build interference knowledge bases. However, compared with decision trees and BP neural networks, DNN has the highest recognition accuracy under low jamming-to-noise ratio (JNR) conditions. Therefore, a large number of scholars have studied the application of interference classifiers based on DNN algorithms in wireless network scenarios. For example, the DNN-based radar interference signal recognition method and the DNN-based multi-node collaborative interference recognition algorithm can improve the interference recognition rate under low signal-to-noise ratio.

[0004] However, the above literature ignores the interference identification and knowledge base construction in complex electromagnetic environments and lacks consideration of the knowledge base application method. In fact, the interference knowledge base has an important impact on the identification of interference and the designation of interference suppression schemes. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a method for constructing an interference knowledge base based on a deep learning network, to improve the efficiency and accuracy of identifying various interferences in complex electromagnetic environments, to support terminal equipment in formulating appropriate interference suppression schemes, and to improve communication quality.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for constructing an interference knowledge base based on a deep learning network comprises the following steps:

[0008] S1: Analyze the characteristics of interference signals in complex scenarios, and provide different interference suppression strategies for different types of interference signals in the environment around the communication terminal equipment;

[0009] S2: Building an interference knowledge base offline training module: Based on scenario analysis, the original interference data is used to build an initial knowledge base and train a deep learning network to obtain an interference recognition model with online recognition capabilities;

[0010] S3: Construct an interference knowledge base online learning module: match the feature data set obtained after the interference data collected in real time is identified by the module with the initial database. If it is a new data set, update the knowledge base online to obtain an interference knowledge base construction module with online learning capabilities.

[0011] Furthermore, the interference signals in the surrounding environment of the communication terminal device in step S1 include the following types: interference center frequency, interference power, and interference period.

[0012] Further, step S2 constructs an interference knowledge base offline training module, specifically including the following steps:

[0013] S21: Based on scenario analysis, build a single-class and multi-class mixed interference raw data generation module;

[0014] S22: construct a feature extraction module based on the original data to extract a 3-domain 10-dimensional feature vector dataset;

[0015] S23: Construct an interference identification module based on deep learning network.

[0016] Further, the construction of the single-class and multi-class mixed interference raw data generation module in step S21 includes: according to the scene analysis, obtaining the distribution characteristics of different interference signals and using them to generate an initial data set, specifically including the following steps:

[0017] S211: Setting input parameters for generating each interference signal;

[0018] S212: Then, based on the known input parameters, by adding noise at different drying ratios, calling each interference signal generation function, and obtaining multiple groups of time domain and frequency domain data of each interference signal at different drying ratios and their corresponding label data;

[0019] S213: Save the acquired original signal data and corresponding label data.

[0020] Further, step S22 constructs a feature extraction module for extracting a 3-domain 10-dimensional feature vector data set based on the original data, including: extracting interference feature parameters from the time domain, frequency domain and transform domain to characterize the interference signal according to the original data sets of each interference signal under different JNRs generated; specifically including the following steps:

[0021] S221: obtaining original data of each generated interference signal, and setting input parameters for generating a variety of multi-domain interference features;

[0022] S222: obtaining multiple groups of interference feature parameter data and corresponding feature label data of each interference signal at different JNRs by adding noises of different JNRs;

[0023] S223: Save the obtained interference feature parameter data and corresponding feature label data.

[0024] Further, step S23 of constructing an interference identification module based on a deep learning network includes:

[0025] S231: Initialize the parameters required by the module;

[0026] S232: Import the data set into the data acquisition and batch processing submodule to obtain a training set and a test set;

[0027] S233: constructing a multi-layer perception network according to the initialized neural network parameters, etc.;

[0028] S234: input the training set into the DNN model training submodule to obtain a DNN model after the neural nodes of each layer are optimized;

[0029] S235: The trained model and test set are sent to the DNN model testing submodule to verify the recognition accuracy and generalization ability of the model;

[0030] S236: Encapsulate the evaluated model and compare it with all the evaluated models to select the model with the best performance.

[0031] Further, the step S3 specifically includes:

[0032] S31: digitally processing the interference data collected from the outside in real time to obtain a cleaned effective measured data set;

[0033] S32: import the data set into the interference detection module for detection, and ignore it if there is no interference; otherwise, give basic information of the interference;

[0034] S33: preprocessing the data set and inputting it into the feature extraction module to extract feature sets for each dimension;

[0035] S34: input the feature set into the DNN module for identification, and output the type parameter of the interference;

[0036] S35: Match the feature parameter information with the knowledge base. If the feature parameter information already exists in the knowledge base, it will be ignored. Otherwise, it will be updated to the knowledge base to realize the online learning and updating function of the knowledge base.

[0037] Further, the process of the interference detection module is:

[0038] Import the original data set into the detection module, first check the existence of the interference signal, then perform interference spectrum analysis based on the energy detection algorithm to obtain basic interference parameter information. The steps are as follows:

[0039] (1) The modulus of the spectrum of the data bit signal received in the DSP is first squared and sorted from small to large;

[0040] (2) Select the first 500 smallest spectrum values ​​to set the first interference detection threshold, with a threshold factor of c = 2.7, that is,

[0041] (3) All spectral line values ​​are compared with T1, and all spectral line values ​​less than T1 are used as reference spectral lines for the second detection threshold setting. The second detection threshold factor is 2.7, and the final threshold in The total number is M;

[0042] (4) All spectral line values ​​are compared with the final threshold T. If The frequency spectrum position outputs 1, indicating that interference is received, otherwise it outputs 0, indicating that there is no interference.

[0043] The beneficial effect of the present invention is that the method of the present invention effectively improves the recognition accuracy while ensuring the interference recognition efficiency, and meets the needs of interference recognition of terminal communication equipment in complex environments.

[0044] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0046] Figure 1 Constructing a block diagram for the knowledge base of the present invention;

[0047] Figure 2 Constructing a flow chart for the modules of the present invention;

[0048] Figure 3 This is a diagram showing the recognition effect of the present invention on interference signals under different JNRs. DETAILED DESCRIPTION

[0049] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0050] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0051] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0052] The present invention provides a technology for constructing an interference knowledge base based on a deep learning network. Figure 1-2 As shown, including:

[0053] S1: Analyze the characteristics of interference signals in complex scenarios. The interference signals in the environment around the communication terminal equipment have different types, interference center frequencies, interference powers, interference periods, etc., which correspond to different interference suppression strategies;

[0054] S2: Construct an interference knowledge base offline training module. Based on scenario analysis, use the original interference data to establish an initial knowledge base and train the deep learning network to obtain an interference recognition model with online recognition capabilities. Specifically, it includes:

[0055] S21: Based on scenario analysis, build a single-class and multi-class mixed interference raw data generation module;

[0056] S22: construct a feature extraction module based on the original data to extract a 3-domain 10-dimensional feature vector dataset;

[0057] S23: Construct an interference identification module based on deep learning network.

[0058] In this embodiment, if Figure 1 The construction of interference knowledge base is generally divided into two processes: offline training and online learning. Among them, offline training is divided into data preprocessing, signal processing, feature extraction and interference identification. Data preprocessing is used to generate initial interference signal data; signal processing is used to determine whether there is interference in the input signal; feature extraction is used to obtain feature parameters of different domains of interference signals; interference identification is used to initialize the interference knowledge base.

[0059] Optionally, an interference knowledge base offline training module is constructed. First, the characteristics of interference signals in complex scenarios are analyzed. The interference signals in the environment surrounding the communication terminal equipment have different types, interference center frequencies, interference powers, interference periods, etc. Specifically:

[0060] Secondly, analyze the interference existing in typical complex communication scenarios, taking the urban environment as an example: in the city, there are mainly natural interference, interference from the same-frequency civil communication system, adjacent channel interference, suppression interference, and deceptive interference. This technology mainly identifies suppression interference (which deteriorates the signal-to-interference-to-noise ratio of the received signal by suppressing the power of the communication signal, thereby causing the receiver to demodulate.). The corresponding types of suppression interference in this scenario mainly include multi-tone interference, linear sweep interference, partial band noise interference, and noise frequency modulation interference. The mathematical statistical model of the above interference is shown in Table 1 below.

[0061] Among them, P J,i represents the signal power of the i-th STJ; N represents the number of STJs, i.e. the number of interference frequencies; f J,i is the frequency of the i-th STJ; is the initial phase of the i-th STJ signal, which is uniformly distributed in [0,2π); P J represents the power of LFSJ; μ0 represents the frequency sweep rate; f0 is the starting frequency; is the initial phase and is uniformly distributed in [0,2π); P w is the total interference bilateral power, f J is the single-sided center frequency, W J The ratio of the total bandwidth W to the interference bandwidth W0 is called the interference factor, which is represented by ρ, and ρ=W0 / W, 0<ρ<1.

[0062] Table 1

[0063]

[0064] Optionally, based on scenario analysis, a single-class or multi-class mixed interference raw data generation module is constructed:

[0065] The module is developed in C++ language and runs on MATLAB compiler. It uses the function library in MATLAB to generate the original data of the interference signal. The input of this module is shown in Table 2. The output includes: the original data file of the interference signal (.mat format file) and the label file of the original data of the interference signal (.mat format file).

[0066] According to the scene analysis, the distribution characteristics of five different interference signals (multi-tone interference, linear frequency sweep interference, periodic pulse interference, partial frequency band interference, and noise frequency modulation interference) are obtained and used to generate the initial data set. First, the input parameters for generating each interference signal are set; then, based on the known input parameters, by adding noise at different drying ratios, each interference signal generation function is called to obtain multiple sets of time domain and frequency domain data of each interference signal at different drying ratios and its corresponding label data; finally, the obtained original signal data and the corresponding label data are saved in a .mat file.

[0067] Table 2

[0068]

[0069] Optionally, construct a feature extraction module that extracts a 3-domain 10-dimensional feature vector dataset based on the original data:

[0070] According to the original data sets of each interference signal under different JNR, 10 interference feature parameters, including time domain moment kurtosis coefficient (TDK), 3dB bandwidth factor (3dBBF), interference detection bandwidth factor (JDBF), frequency domain moment kurtosis coefficient (FDK), average spectrum flatness coefficient (ASFC), differential signal spectrum kurtosis coefficient (DSSK), quadratic spectrum bandwidth factor (SSBF), quadratic spectrum kurtosis coefficient (SSK), and quartic spectrum kurtosis coefficient (QSK), are extracted from the time domain, frequency domain, and transform domain (as shown in Table 3 below) to characterize the interference signal. First, the original data of each interference signal is obtained, and the input parameters for generating 10 multi-domain interference features are set; secondly, by adding noise of different JNR, multiple groups of interference feature parameter data and corresponding feature label data of each interference signal under different JNR are obtained; finally, the obtained interference feature parameter data and corresponding feature label data are saved in a .mat file.

[0071] Among them, r re and r im are the real and imaginary parts of the input signal r(n), respectively; μ re and μ im They are r re and r im The mean of re and σim They are r re and r im The standard deviation of P, P d , P s and P q They are r(n), d(n), r 2 (n) and r 4 The power spectral density of (n), where d(n) = r(n+m)×r(n) * .P p (k) is the impulse part of P, YesP p The mean of (k). P , μ d , μ s and μ q They are P and P d , P s and P q The mean of P , σ d , σ s and σ q They are P and P d , P s and P q The standard deviation of W c is the bandwidth of interference identification analysis; W j is the bandwidth determined by interference detection; W 3dB , and They are r(n), r 2 (n) and r 4 (n) 3dB interference bandwidth.

[0072] Table 3

[0073]

[0074]

[0075] Optionally, build an interference identification module based on a deep learning network:

[0076] The module is developed in Python and runs on the PyCharm compiler. It relies on the pytorch learning framework to build the DNN network model and uses CUDA to accelerate the GPU training process. The main libraries it relies on are numpy, h5py, and torch. The data connection between this module and the previous module is based on the File function of the library h5py to achieve data access interaction. The input of this module is feature data and label data (.mat file), and the output includes: the model of the evaluation index (.pkl format file), the classification results of the interference signal and its unique parameters (.csv format file).

[0077] The input data of the module as a whole is the eigenvalue data (.mat file) processed by matlab and its corresponding interference signal type label (.mat file). First, initialize the parameters required by the module (such as learning rate, batch size, neural network parameters, etc.); then, import the data set (.mat file) into the data acquisition and batch processing submodule to obtain the training set and test set; secondly, build a multi-layer perception network (i.e., DNN) based on the initialized neural network parameters; thirdly, input the training set into the DNN model training submodule to obtain the DNN model with optimized neural nodes in each layer; then, send the trained model and test set to the DNN model testing submodule to verify the recognition accuracy and generalization ability of the model; finally, encapsulate the evaluated model (.pkl file) and compare it with all evaluated models to select the best performance.

[0078] The activation function of the hidden layer of the model uses Leaky-ReLU, and when the input x < 0, a very small gradient γ is maintained. In this way, when the output value of the neuron is negative, there can be a non-zero gradient to update the parameters to avoid being unable to be activated forever. The definition of Leaky-ReLU is as follows:

[0079]

[0080] Among them, γ is an auxiliary variable.

[0081] The output layer activation function uses Softmax, which is used for multi-class classification problems. In multi-class classification problems, if there are more than two class labels, class membership is required. It is defined as follows:

[0082]

[0083] The model uses a fully connected network, and its forward propagation process can be To describe, suppose the (l-1)th layer contains N l-1 neurons, where the activation value of the kth neuron can be expressed as Assume that the lth layer contains N lneurons, then after propagation from the (l-1)th layer to the lth layer, the activation value of the jth neuron is:

[0084]

[0085] Where j = 1, 2, ..., N l , is the weighted input of the jth neuron in the lth layer, is the weight from the kth neuron in the (l-1)th layer to the jth neuron in the lth layer, is the bias of the jth neuron in the lth layer, and σ(x) is the activation function.

[0086] The optimizer of the model is set to Adam; learning rate (controls the update rate of weights): 0.01, minimum learning rate: 10^(-7); exponential decay rate of first-order moment estimate: 0.99; exponential decay rate of second-order moment estimate: 0.9;

[0087] In order to use the data set to train the neural network to search for the optimal weights and biases, it is necessary to define a cost function to evaluate the difference between the current network output and the expected network output. Therefore, the cross entropy loss function is selected as the cost function of the model:

[0088]

[0089] Among them, N t is the number of samples input to the network, p j (n) represents the probability that the nth sample is expected to belong to the jth category, y j (n) represents the probability that the network actually predicts that the nth sample belongs to the jth category.

[0090] The above are the main components of the DNN model.

[0091] S3: Construct an interference knowledge base online learning module, match the feature data set obtained after the real-time interference data is identified by the module with the initial database, and if it is a new data set, update the knowledge base online, that is, obtain an interference knowledge base construction module with online learning capabilities. Online learning is for the interaction between the outside world and the knowledge base, including the output of unknown (not in the interference knowledge base) interference signal parameter information and the interference knowledge base update function.

[0092] The interference data collected from the outside world in real time is digitally processed to obtain a cleaned effective measured data set. The data set is imported into the interference detection module for detection. If there is no interference, it is ignored; otherwise, basic information of the interference is given, such as interference center frequency, interference power, interference bandwidth, etc.; secondly, the data set is preprocessed and input into the feature extraction module to extract the corresponding 10-dimensional feature set; then, the feature set is input into the DNN module for identification, and the type of interference and other parameters are output; then, its feature parameter information is matched with the knowledge base. If it already exists in the knowledge base, it is ignored. Otherwise, it is updated to the knowledge base to realize the online learning and update function of the knowledge base.

[0093] The process of building the interference detection module is as follows: import the original data set into the detection module, first check the existence of the interference signal, then perform interference spectrum analysis based on the energy detection algorithm to obtain basic interference parameter information. The interference detection algorithm uses the forward continuous mean elimination algorithm, which is superimposed twice. The steps are as follows:

[0094] (1) The modulus of the spectrum of the data bit signal received in the DSP is first squared and sorted from small to large;

[0095] (2) Select the first 500 smallest spectrum values ​​to set the first interference detection threshold, with a threshold factor of c = 2.7, that is,

[0096] (3) All spectral line values ​​are compared with T1. All spectral line values ​​less than T1 are used as reference spectral lines for the second detection threshold setting. The second detection threshold factor is still 2.7. The final threshold in The total number is M;

[0097] (4) All spectral line values ​​are compared with the final threshold T. If The frequency spectrum position outputs 1, indicating that interference is received, otherwise it outputs 0, indicating that there is no interference.

[0098] In this embodiment, the proposed interference knowledge base construction technology based on deep learning network is compared with the traditional decision tree for multi-class interference recognition effect. Figure 3 It can be seen that the interference identification using the DNN algorithm can achieve a higher accuracy rate than the traditional decision tree algorithm, reaching 93.1%, an increase of 8.2%. By looking at the interference identification rate of different interference signals under different JNRs, it can be seen that the knowledge base construction technology based on the DNN algorithm can better distinguish the types of different interference signals, verifying its effectiveness and reliability.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for constructing an interference knowledge base based on a deep learning network, characterized in that: The following steps are involved: S1: Analyze the characteristics of interference signals in complex scenarios, and provide different interference suppression strategies for different types of interference signals in the environment around the communication terminal equipment; S2: Building an interference knowledge base offline training module: Based on scenario analysis, the original interference data is used to build an initial knowledge base and train a deep learning network to obtain an interference recognition model with online recognition capabilities; S3: Constructing an online learning module for the interference knowledge base: matching the feature data set obtained after the interference data collected in real time is identified by the module with the initial database. If it is a new data set, the knowledge base is updated online, thus obtaining an interference knowledge base construction module with online learning capabilities; The step S3 specifically includes: S31: digitally processing the interference data collected from the outside in real time to obtain a cleaned effective measured data set; S32: import the data set into the interference detection module for detection, and ignore it if there is no interference; otherwise, give basic information of the interference; S33: preprocessing the data set and inputting it into the feature extraction module to extract feature sets for each dimension; S34: input the feature set into the DNN module for identification, and output the type parameter of the interference; S35: Match the feature parameter information with the knowledge base. If the feature parameter information already exists in the knowledge base, it is ignored. Otherwise, it is updated to the knowledge base to realize the online learning and updating function of the knowledge base. The process of the interference detection module is as follows: Import the original data set into the detection module, first check the existence of the interference signal, then perform interference spectrum analysis based on the energy detection algorithm to obtain basic interference parameter information. The steps are as follows: (1) The modulus of the spectrum of the data bit signal received in the DSP is first squared and sorted from small to large; (2) Select the first 500 smallest spectrum values ​​to set the first interference detection threshold, with a threshold factor of c = 2.7, that is, (3) All spectrum values ​​are compared with T1, and all spectrum values ​​less than T1 are used as the reference spectrum line for the second detection threshold setting. The second detection threshold factor is 2.7, and the final threshold in The total number is M; (4) All spectrum values ​​are compared with the final threshold T. If The frequency spectrum position outputs 1, indicating that interference is received, otherwise it outputs 0, indicating that there is no interference.

2. The method for constructing an interference knowledge base based on a deep learning network according to claim 1, characterized in that: The interference signals in the environment surrounding the communication terminal device in step S1 include the following types: interference center frequency, interference power, and interference period.

3. The method for constructing an interference knowledge base based on a deep learning network according to claim 2, characterized in that: Step S2 constructs an interference knowledge base offline training module, specifically including the following steps: S21: Based on scenario analysis, build a single-class and multi-class mixed interference raw data generation module; S22: construct a feature extraction module based on the original data to extract a 3-domain 10-dimensional feature vector dataset; S23: Construct an interference identification module based on deep learning network.

4. The method for constructing an interference knowledge base based on a deep learning network according to claim 3, characterized in that: The construction of the single-class and multi-class mixed interference raw data generation module in step S21 includes: according to the scene analysis, obtaining the distribution characteristics of different interference signals and using them to generate an initial data set, specifically including the following steps: S211: Setting input parameters for generating each interference signal; S212: Then, based on the known input parameters, by adding noise at different drying ratios, calling each interference signal generation function, and obtaining multiple groups of time domain and frequency domain data of each interference signal at different drying ratios and their corresponding label data; S213: Save the acquired original signal data and corresponding label data.

5. The method for constructing an interference knowledge base based on a deep learning network according to claim 3, characterized in that: The step S22 constructs a feature extraction module based on the original data to extract a 3-domain 10-dimensional feature vector data set, including: extracting interference feature parameters from the time domain, frequency domain and transform domain to characterize the interference signal according to the original data sets of each interference signal under different JNRs generated; specifically including the following steps: S221: obtaining original data of each generated interference signal, and setting input parameters for generating a variety of multi-domain interference features; S222: obtaining multiple groups of interference feature parameter data and corresponding feature label data of each interference signal at different JNRs by adding noises of different JNRs; S223: Save the obtained interference feature parameter data and corresponding feature label data.

6. The method for constructing an interference knowledge base based on a deep learning network according to claim 3, characterized in that: The construction of the interference identification module based on the deep learning network in step S23 includes: S231: Initialize the parameters required by the module; S232: Import the data set into the data acquisition and batch processing submodule to obtain a training set and a test set; S233: constructing a multi-layer perception network according to the initialized neural network parameters, etc.; S234: input the training set into the DNN model training submodule to obtain a DNN model after the neural nodes of each layer are optimized; S235: The trained model and test set are sent to the DNN model testing submodule to verify the recognition accuracy and generalization ability of the model; S236: Encapsulate the evaluated model and compare it with all the evaluated models to select the model with the best performance.

Images