A method for identifying electromagnetic interference sources based on electromagnetic principle-optimized neural network

Abstract

The present invention discloses an electromagnetic interference source identification method based on an optimized neural network based on electromagnetic principles, comprising the following steps: S1. constructing a combined electromagnetic interference signal set under electromagnetic environments with different noise intensities, and performing time-frequency transformation; S2. performing multi-dimensional feature extraction on the constructed combined electromagnetic interference signal set, extracting typical feature indicators in the time domain, frequency domain and energy domain respectively, and preprocessing the feature indicator data; S3. randomly selecting the feature indicator data extracted in S2 to form a small sample set, performing feature selection on the processed data in the small sample set, reducing the feature dimension, preventing overfitting in the process of establishing the recognition model and effectively reducing the calculation cost of the model; S4. constructing an electromagnetic interference source identification model of an optimized neural network based on electromagnetic principles, and performing training based on a training set; S5. using the electromagnetic interference source identification model obtained by training to identify the electromagnetic interference source.

Description

Technical Field

[0001] The invention relates to the field of electromagnetism, and in particular to an electromagnetic interference source identification method based on an electromagnetic principle-based optimized neural network. Background Art

[0002] With the rapid development of electronic technology, the types of electronic and electrical equipment in the system are increasing. As electromagnetic interference sources, they constitute a complex electromagnetic environment. In order to analyze the electromagnetic compatibility of various electronic and electrical equipment in the system, the identification of electromagnetic interference sources is particularly important. The existing electromagnetic interference source identification technologies mainly include matching filtering method, time-frequency analysis method, and machine learning-based algorithms. Although these technologies can achieve a certain recognition accuracy in certain occasions, they also have problems such as difficulty in identifying large sample signals, high algorithm noise sensitivity, insufficient generalization ability, and low adaptability to unknown signals. In short, the existing electromagnetic interference source identification technology still has obvious deficiencies in dealing with complex electromagnetic environments and unknown signals. Improving the real-time, robustness and generalization ability of the identification method to better adapt to the complex and changing electromagnetic environment is a key issue to be studied.

[0003] In the work of electromagnetic interference source identification, neural networks have shown significant advantages. Neural networks not only perform well in processing complex and nonlinear signals, but also have good adaptability and generalization capabilities, and can maintain a high recognition accuracy rate when facing diverse and complex electromagnetic signals. However, the performance of neural networks is highly dependent on the selection of network parameters, including weights, biases, and network structures. At this time, it is necessary to establish a neural network model with optimized network parameters to effectively search for the global optimal solution of network parameters.

[0004] Currently, in the research on the identification of electromagnetic interference sources, there are cases of using optimized neural networks as identification models, but the corresponding optimization algorithms still have problems such as being unable to converge within a limited number of iterations. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an electromagnetic interference source identification method based on an electromagnetic principle-based optimized neural network. Background noise is added to the combined interference signal to simulate the complex electromagnetic environment in actual industrial engineering, and an electromagnetic interference source identification model based on the electromagnetic principle is used to identify the combined electromagnetic interference signal sample set.

[0006] The object of the present invention is achieved through the following technical solution: a method for identifying electromagnetic interference sources based on electromagnetic principle optimization neural network, comprising the following steps:

[0007] S1. Construct a combined electromagnetic interference signal set under electromagnetic environments with different interference intensities and perform time-frequency transformation;

[0008] S2. Perform multi-dimensional feature extraction on the constructed combined electromagnetic interference signal set, extract typical characteristic indicators in the time domain, frequency domain and energy domain respectively, and pre-process the characteristic indicator data;

[0009] S3. Randomly select the feature index data extracted in S2 to form a small sample set, perform feature selection on the processed data in the small sample set, reduce the feature dimension, prevent overfitting in the process of establishing the recognition model, and effectively reduce the calculation cost of the model;

[0010] S4. Construct an electromagnetic interference source identification model based on an optimized neural network based on electromagnetic principles, and train it based on the training set;

[0011] S5. Use the trained electromagnetic interference source identification model to identify the electromagnetic interference source.

[0012] The beneficial effects of the present invention are as follows: the present invention optimizes the parameters of the electromagnetic interference identification model by optimizing the algorithm based on the electromagnetic principle, so that the accuracy and stability of the electromagnetic interference source identification model are improved. The present invention also establishes a new set of emission signals of electromagnetic interference sources, provides identification samples for the identification model through data processing algorithms and feature extraction algorithms, and more intuitively reflects the characteristics of the time-frequency energy domain of electromagnetic interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the method of the present invention;

[0014] Figure 2 It is the time domain waveform diagram of the sample signal when the power ratio of the interference signal to the background noise is 3dB;

[0015] Figure 3 It is the two-dimensional time-frequency diagram of the sample signal when the power ratio of the interference signal and the background noise is 3dB;

[0016] Figure 4 It is a schematic diagram of the topological structure of the BP neural network;

[0017] Figure 5 A flow chart is established for the electromagnetic interference source identification algorithm model based on electromagnetic optimization neural network. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0019] Maxwell derived optimization is an optimization algorithm based on electromagnetic principles. It shows excellent performance in solving parameter optimization problems. Therefore, it can be used as an algorithm for optimizing neural network parameters to build an electromagnetic interference source identification model. The present invention establishes an electromagnetic interference source identification technology based on electromagnetic principle optimization neural network. The technology is used for the identification process of electromagnetic interference emission sources. Figure 1 As shown, a method for identifying electromagnetic interference sources based on an electromagnetic principle-based neural network optimization comprises the following steps:

[0020] S1. Construct a combined electromagnetic interference signal set under electromagnetic environments with different noise intensities and perform time-frequency transformation;

[0021] The present invention first constructs a basic electromagnetic interference signal set, takes four basic electromagnetic emission signals as individuals of the combined signal, and numbers the four types of signals 1-4 in the order of pulse signal, mismatch signal, analog signal and digital signal, extracts any one, two, three and all four types of signals from the four types of signals numbered 1-4, and performs random convex combination on the extracted signals to finally form a combined signal, thereby finally constructing a basic signal set of the combined electromagnetic interference signal.

[0022] Extract any one, two, three or all four types of signals from the four basic signals numbered 1-4, and perform random convex combination on the extracted single and multiple signals to form a combined signal. The specific method of constructing a combined signal through random convex combination is as follows:

[0023] Four basic electromagnetic emission signals are used as individuals of the combined signal, and the four basic signals are numbered 1-4 in the order of pulse signal, mismatch signal, analog signal and digital signal;

[0024] Extract any one, two, three or all four types of signals from the four basic signals numbered 1-4, and perform random convex combination on the extracted single and multiple signals to form a combined signal. The specific method of constructing a combined signal through random convex combination is as follows:

[0025] First, the basic signals of categories 1 to 4 are represented as {h i (t)}, i = 1, 2, 3, 4, there is a real number u i Satisfy the condition 0.2≤u i ≤0.8 and u i Randomly select values ​​under this condition, where n is the number of signal categories of the sample vector, and the sample vector {w m (t)}, m = 1, 2, ..., 15 is a combination signal. There are 15 types of combination signals. The expression of the basic combination signal is as follows:

[0026]

[0027] Where p, q = 1, 2, 3, 4, and the number of categories of basic signals contained in the combined signal is n = q-p+1. The 15 combined signals constructed are numbered 1-15 respectively, and the numbers correspond to the sample label values ​​of each category of sample signals.

[0028] Taking the combined signal of the sample category containing all four basic signals as an example, at this time p = 1, q = 4, m = 15, the basic combined signal expression constructed by random convex combination is as follows:

[0029] w 15 (t)=u1h1(t)+u2h2(t)+u3h3(t)+u4h4(t).

[0030] The basic interference signal set is used to generate the signal, and Gaussian white noise is used as the environmental background noise ε during signal generation. m , construct a combined electromagnetic interference signal x containing background noise m (t), m=1,2,…,15, its expression is as follows:

[0031] x m (t) = w m (t)+ε m

[0032] Let the basic combination signal w m The effective power of (t) is P n , environmental background noise ε m The effective power is P s , the power ratio of the interference signal to the background noise is S, and the interference signals when S is 3dB, 6dB and 10dB are selected to form the electromagnetic interference signal sample set of the model. The calculation formula is as follows:

[0033] S=10lg(P s / P n ).

[0034] It is worth noting that for the basic recognition sample, the value of S is infinite. Figure 2 Taking S as 3dB as an example, the time domain waveform of the electromagnetic interference sample signal set is drawn. Figure 3 The two-dimensional time-frequency diagram of the electromagnetic interference sample signal set is drawn when S is also 3dB. The construction of the electromagnetic interference signal sample set provides input signals for the subsequent optimization of the neural network model based on electromagnetic principles.

[0035] S2. Perform multi-dimensional feature extraction on the combined electromagnetic interference signal in the constructed signal set, extract typical characteristic indicators in the time domain, frequency domain and energy domain respectively, and pre-process the characteristic indicator data;

[0036] The invention extracts multi-domain features of electromagnetic interference. The time domain feature extraction process adopts the time domain analysis method to extract the peak feature, trend feature, margin feature and pulse feature in the time domain, wherein the time domain peak feature includes the main peak feature and the peak factor feature; the frequency domain feature extraction process adopts the fast Fourier transform and wavelet transform method to extract the trend feature, peak feature, center of gravity frequency feature, mean square frequency feature and frequency variance feature in the frequency domain, wherein the frequency domain peak feature includes the main peak feature and the secondary peak feature, and the frequency domain trend feature includes the main trend feature and the trend position feature; the energy domain feature extraction process adopts the entropy analysis method to extract the power spectrum entropy feature and the energy entropy feature in the energy domain. The extracted l features are denoted as X j ,j=1,2,…,l, the total feature number of the electromagnetic interference signal set in the present invention is l=14.

[0037] The present invention uses the standard score detection method to preprocess the extracted feature index data, judges the outliers based on the deviation between the data points and the mean, and removes the outliers, effectively improving the accuracy and robustness of the model training process, and providing optimized data for the subsequent feature selection process. In the present invention, each type of electromagnetic interference combined signal in the combined electromagnetic interference signal set has 500 samples, and the signal set contains a total of 7500 samples.

[0038] S3. Randomly select the feature index data extracted in S2 to form a small sample set, perform feature selection on the processed data in the small sample set, reduce the feature dimension, prevent overfitting in the process of establishing the recognition model, and effectively reduce the calculation cost of the model;

[0039] In the step S3, the sample data of the electromagnetic interference source identification technology based on the electromagnetic principle is processed, wherein the characteristic value contained in each sample in the sample data is the high-dimensional characteristic value extracted in S2, and the sample label value contained is the sample number specified in S101; it is necessary to perform feature dimensionality reduction on the sample data, select random forest as the model for feature dimensionality reduction, and perform dimensionality reduction on the sample features: first, it is necessary to extract some samples from the sample features, and then use the extracted samples as a small sample set to pre-train the network model to eliminate redundant features and prevent overfitting of the optimized network model during the training process.

[0040] The feature selection algorithm of the present invention is based on the principle of random forest model. Random forest improves the accuracy and stability of the model by combining the prediction results of multiple decision trees, and reduces overfitting by introducing sample randomness and feature randomness. In the feature selection of the present invention, random forest uses two main methods to evaluate feature importance: Gini importance I MDI and permutation importance IMDA Gini importance is to calculate the l features X in the electromagnetic interference signal sample set j ,j=1,2,…,lThe sum of purity reductions caused by splitting nodes in all trees, I MDI The calculation formula is as follows:

[0041]

[0042] Where N t is the total number of decision trees in the random forest, N t,j is the feature X used in the t-th tree j The set of all nodes to be split. ΔG n is the feature X at node n j The purity reduction when splitting.

[0043] The permutation importance is achieved by disrupting the l features X in the electromagnetic interference signal sample set. j ,j=1,2,…,l, the eigenvalue, the importance of the feature is measured by observing the change in model performance, I MDA The calculation formula is as follows:

[0044]

[0045] Among them A t is the classification accuracy or error rate of the t-th tree on the benchmark test set, is the feature X used in the t-th tree j Classification accuracy or error rate on the shuffled test set.

[0046] The present invention combines the above methods to evaluate the importance of each feature in the high-dimensional features, and selects features with higher importance scores to build a simplified and effective model. The specific feature selection process using the random forest model is as follows:

[0047] (1) Prepare the combined electromagnetic interference signal set in S2 after data preprocessing, randomly select 750 samples from the 7500 samples of the combined electromagnetic interference signal set, and use it as the small sample set X for feature selection j ,j=1,2,…,l.

[0048] (2) Construct a feature selection model based on the random forest algorithm. MDA and I MDI As a measurement standard, the importance of each feature of a small sample set with a feature number of l is calculated, and an importance bar chart is drawn to compare the importance of each feature, and the importance of each feature is evaluated and ranked.

[0049] (3) Use small sample data to pre-train the electromagnetic interference source recognition model and obtain the recognition accuracy P for the small sample set of electromagnetic interference signals with a feature number of l. m .

[0050] (4) Setting an importance threshold, removing features whose importance is lower than the threshold from the small sample set, and using the small samples containing the remaining features as pre-training data for the electromagnetic interference source identification model.

[0051] (5) Use the pre-training data for training, and judge whether the importance threshold set in (4) is reasonable by the recognition of the small sample set. If the recognition accuracy is much lower than P m Then adjust the importance threshold and repeat (4) until the recognition accuracy reaches P m The loop ends.

[0052] (6) K features that achieve the ideal recognition effect are retained, and the feature selection process ends. This process achieves feature compression, compressing the number of features from l to k, eliminating redundant features while preventing the occurrence of overfitting, and effectively reducing the computational cost of the model.

[0053] The present invention completes the feature selection process by pre-training the extracted small samples, and uses the feature extraction model based on random forest to compress l=14-dimensional features to k=8-dimensional features. While ensuring the recognition accuracy, the electromagnetic interference signal sample set is effectively reduced in dimension and redundant features are eliminated.

[0054] S4. Construct an electromagnetic interference source identification model based on an optimized neural network based on electromagnetic principles, and train it based on the training set;

[0055] The present invention establishes an electromagnetic interference source identification model based on an electromagnetic principle optimization neural network, and selects a BP neural network as the basic model for electromagnetic interference emission source identification. Due to its unique large-scale parallel structure, distributed storage and other characteristics, the BP neural network not only has powerful function approximation and classification capabilities, but also has good self-organization, adaptability and fault tolerance, and it shows a strong ability in the identification of electromagnetic interference sources. As an implementation method of a feedforward neural network, the BP algorithm is a back propagation algorithm based on a feedforward neural network structure, which is used to train the parameters of the neural network. Its forward transmission is transmitted from the input layer to the output layer by calculating the weighted sum of each layer of neurons and their activation function output. Figure 4 A schematic diagram of the neural network topology is given. The functional relationship of the hidden layer of the BP network is first analyzed. In the hidden layer, each neuron contains an activation function, which converts the input of the neuron into output, so that the neural network can capture more complex patterns and relationships.

[0056] According to the following Figure 4 The back propagation process of the neural network is explained. Assume that the input of the neural network is x1, x2, ..., x n , the output is y, where is the output of the i-th neuron in the r-th hidden layer, m is the number of neurons in the hidden layer, For the i-th set of input variables to the weight of the j-th neuron in the next layer, is the bias of the jth neuron corresponding to the rth hidden layer, b y The neural network used in the present invention is a single hidden layer neural network. For a single hidden layer neural network, r = 1 and m = q. At this time, the forward propagation formula of the neural network model is as follows:

[0057]

[0058] Where g(x) is the activation function. Since the step function is a discontinuous function, it will cause a lot of trouble to the calculation process when performing complex mathematical processing. Neural networks that use continuous functions as excitation functions began to appear. With the further development of neural network technology, the selection of excitation functions has also become more diverse. They introduced nonlinear characteristics to improve the expression ability of the model. The input of each neuron can be represented by the output of the previous neuron, which is a linear combination of the output of the previous layer. Its weight term and bias term are represented by w and b respectively, b u 、b y denote the bias terms of the hidden layer and the output layer respectively, w u 、w y Represent the weight items of the hidden layer and the output layer respectively. The optimization network model in the present invention selects a single hidden layer neural network, wherein the activation function used in the hidden layer is the Sigmoid function, and the function used in the output layer is a linear function, wherein the expression of the Sigmoid function is as follows:

[0059]

[0060] For the back propagation process, the present invention uses the gradient descent method to back propagate the error, and the objective function is the predicted output The deviation from the actual output y, the objective function is defined as L, and its calculation formula is as follows:

[0061]

[0062] Where k represents the serial number of the object sample. For the BP neural network model, the gradient descent method and error back propagation are usually used in combination, which together form the basis of neural network training. In the process of error signal back propagation, the error signal propagates forward layer by layer from the output end, and the network weights are adjusted by error feedback. The actual output of the network is closer to the expected output through continuous correction of the weights. The following formula shows the process of network parameter optimization using the gradient descent method:

[0063]

[0064] in and Respectively represent the bias terms of the hidden layer and the output layer after reverse gradient descent; and Respectively represent the weight items of the hidden layer and the output layer after reverse gradient descent, and the parameter update represents the process of each error back propagation. μ is the learning rate, which controls the step size of each parameter update. At the same time, it determines the degree of adjustment of the parameters in each iteration. If it is too small, it will lead to slow training, and if it is too large, it may cause oscillation or failure to converge. It is necessary to adjust μ according to the specific problem. The present invention uses a single hidden layer neural network. When the activation function used in the hidden layer is the Sigmoid function and the output layer is a linear function, the back propagation formula of the optimized network model is as follows:

[0065]

[0066] The present invention uses an optimization algorithm based on electromagnetic principles to optimize the initial weight and bias of the BP neural network: an optimization algorithm derived from the Maxwell equations. The algorithm optimizes the weight and bias components of the BP neural network basic model to obtain an electromagnetic interference source identification model with higher recognition accuracy and stability. The specific algorithm model establishment flow chart is as follows Figure 5 As shown, the identification process is described below:

[0067] (1) The established combined electromagnetic interference signal sample set is divided into training set, test set and validation set according to the ratio of 8:1:1. The sample data of the training set is normalized as the input of the BP neural network, the sample labels of the training set are used as the network output, and the network topology is determined.

[0068] (2) Using the weights and biases of the neural network as individual corresponding parameters, setting the population size, initializing the position and velocity of the MEDO population, and using the recognition accuracy of the neural network as the fitness function value.

[0069] (3) Calculate the fitness function, update the optimal parameters of the individuals in the neural network, and adjust the position and speed of the individuals.

[0070] (4) Determine whether the termination condition of the training error or the maximum number of iterations has been reached. If so, stop the iteration; if not, return to (3) to continue training.

[0071] (5) The weights and biases optimized by the MEDO algorithm are used as the initial values ​​for neural network training, and the network model is trained using the sample data of the training set to obtain an optimized neural network model, i.e., the one in the present invention.

[0072] (6) The characteristic values ​​of the training set are used as the input of the trained optimized neural network model to obtain the label prediction results of the network model for the training set, and the predicted category results of the training set are compared with the actual sample categories to calculate the recognition accuracy, thereby verifying the electromagnetic interference source recognition model in the present invention.

[0073] S5. Use the trained electromagnetic interference source identification model to identify the electromagnetic interference source.

[0074] The application part of the present invention is to identify unknown electromagnetic interference sources. The present invention respectively selects a verification set and a test set in a combined electromagnetic interference signal set as unknown interference sources, and uses the characteristics of the unknown interference sources as the input of an electromagnetic interference source identification model. The label prediction results are obtained through the identification model, and the predicted category results of the training set are compared with the actual sample categories. The recognition accuracy is calculated to complete the identification of unknown electromagnetic interference sources.

[0075] The following is a detailed description of the optimization principle of the parameters of the neural network and the process of identifying unknown electromagnetic interference sources using the optimization algorithm based on electromagnetic principles in the present invention:

[0076] MEDO is an optimization algorithm derived from the physical laws of electromagnetic principles and has been applied in many practical cases of electromagnetic compatibility. This algorithm draws conclusions by analyzing the current distribution of the coaxial model in a time-varying electromagnetic field. The derivation of the algorithm principle first simplifies the coaxial circuit into a parallel circuit, and uses the law of conservation of charge and Faraday's law to calculate the current on a specific conductor. Then, the left-hand rule is used to convert the current in the constant magnetic field into the Ampere force on the conductor. During the calculation process, in order to move the conductor to the minimum point of the objective function, the voltage source needs to be set to the negative gradient of the objective function at the current point, and the current difference between adjacent points is used for approximate calculation. The iterative formula for the speed and position of MEDO is as follows:

[0077]

[0078] x=x+v·Δt

[0079] new cur new

[0080] where vnew and v cur Respectively represent the conductor movement speed value before and after each generation update, x new and x cur They represent the position values ​​of the conductor before and after each generation of update. In the process of optimizing the parameters of the network model, the position value x represents the parameters corresponding to the weight and bias of the network model. When the value of the objective function f(x) reaches the minimum within the number of iterations, the position parameter x corresponding to the conductor is the initial weight and bias parameter of the optimized target network model. t is the update time, which represents the number of iterations in the optimization process.

[0081] In order to obtain the updated iterative position parameter x new , we need to calculate the speed parameter v of the update iteration new At this time, it is necessary to set the values ​​of other parameters in the formula according to the principles of physics, including gravitational acceleration g, magnetic flux density B0, conductor density ρ, conductor cross-sectional area S bar and the current i flowing through the conductor GH . The current i GH The expression is as follows:

[0082]

[0083] For i GH The parameters in the formula, such as Z3 and di2, are calculated based on the Maxwell derived equation model. In the process of optimizing the neural network model based on electromagnetic principles, the recognition accuracy f of the electromagnetic interference signal sample set is p (x) is used as the objective function f(x), and the optimal weights and parameter values ​​of the network model are calculated iteratively, and finally a trained neural network model based on electromagnetic principles is obtained. The identification process of electromagnetic interference sources can be summarized as a classification problem. The parameters of the neural network are trained through the features and labels of the training set, and finally the optimized model parameters are obtained, so as to establish an electromagnetic interference source identification model. Since the model is optimized by calculating the classification accuracy during the model establishment process, the accuracy of the electromagnetic interference signal sample classification of the neural network model based on electromagnetic principles is to calculate the proportion of the total samples that are correctly classified in the samples identified as positive examples by the classification model. The accuracy of this classification is the objective function f p (x), which is calculated as follows:

[0084]

[0085] Where P TP Represents the number of correctly classified samples classified as positive examples by the recognition model, P FP Indicates the number of misclassified samples classified as positive examples by the model, so f p(x) represents the ratio of the number of correctly classified samples to the total number of samples, that is, the accuracy of electromagnetic interference source classification based on the neural network model optimized based on electromagnetic principles, which is used as the objective function of the optimization algorithm.

[0086] The performance evaluation of the electromagnetic interference source identification model based on the electromagnetic principle is based on the recognition effect of the electromagnetic interference source. The algorithm model can obtain the recognition accuracy of the overall electromagnetic interference signal sample, and at the same time, the recognition accuracy of the model for each category of electromagnetic interference signal samples can be obtained, and the confusion matrix can be drawn to obtain the recognition of the model for all categories of electromagnetic interference signal samples. For this multi-classification problem, the confusion matrix of the algorithm model can show which categories are more confused, helping to improve the classification performance of the model. At the same time, the algorithm model can draw the ROC curve of various types of electromagnetic interference signal samples and calculate the AUC value corresponding to each curve. The closer the ROC curve is to the upper left corner, the better the performance of the electromagnetic interference source identification model is, and the closer it is to the diagonal line, the poor performance of the electromagnetic interference source identification model. The performance of the interference source identification model can be evaluated by the curve trend. At the same time, the AUC value is the size of the area under the ROC curve, and the maximum value is 1. It can be used to more intuitively and accurately reflect the evaluation of the ROC curve on the performance of the electromagnetic interference source identification model. The closer its value is to 1, the better the performance of the classification model is, and the closer it is to 0.5, the poor performance of the classification model is indicated by the prediction effect of the classifier. Through the model output of the recognition accuracy, confusion matrix, ROC curve and AUC value corresponding to the ROC curve of various electromagnetic interference signal samples in the training set, validation set and test set, the recognition effect of the electromagnetic interference source identification model can be comprehensively and multi-dimensionally evaluated, and finally it is concluded that the network model has excellent recognition performance.

[0087] The electromagnetic interference source identification technology based on electromagnetic principles proposed in the present invention improves the identification accuracy of electromagnetic interference sources, and at the same time has a certain universality, and can automatically identify different electromagnetic environments. This technology especially solves the existing defects of low accuracy in electromagnetic interference source identification and complex identification process, and has application prospects in the engineering industry.

[0088] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for identifying electromagnetic interference sources based on an optimized neural network based on electromagnetic principles, characterized in that: The following steps are involved: S1. Construct a combined electromagnetic interference signal set under electromagnetic environments with different noise intensities and perform time-frequency transformation; S2. Perform multi-dimensional feature extraction on the constructed combined electromagnetic interference signal set, extract typical characteristic indicators in the time domain, frequency domain and energy domain respectively, and pre-process the characteristic indicator data; S3. Randomly select the feature index data extracted in S2 to form a small sample set, perform feature selection on the processed data in the small sample set, reduce the feature dimension, and prevent overfitting in the process of establishing the recognition model; S4. Construct an electromagnetic interference source identification model based on an optimized neural network based on electromagnetic principles, and train it based on the training set: S401. Divide the established electromagnetic interference signal sample set into a training set, a test set and a validation set, normalize the sample data of the training set as the input of the neural network model, use the sample label value of the training set obtained in S3 as the network output, and determine the network topology; S402. Using the weight and bias of the neural network as the individual corresponding parameters, setting the population size, initializing the position and speed of the MEDO population, and using the recognition accuracy of the neural network as the fitness function value; S403. Calculate the fitness function, update the optimal parameters of the individuals in the neural network, and adjust the position and speed of the individuals; S404. Determine whether the termination condition of the training error or the maximum number of iterations is reached, if so, stop the iteration, if not, return to S303 to continue training; S405. According to the optimal initial values ​​of weights and biases, the network model is trained using sample data of the training set to obtain an optimized neural network model; S406. Using the test set and the validation set as inputs of the trained optimized neural network model, obtaining the recognition accuracy of the network model for the test set and the validation set, and verifying the electromagnetic interference source recognition model; S5. Use the trained electromagnetic interference source identification model to identify the electromagnetic interference source: The characteristics of the unknown interference source are used as the input of the electromagnetic interference source identification model, and the label prediction result is obtained through the electromagnetic interference source identification model.

2. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 1 is characterized by: The step S1 comprises: S101. The electromagnetic emission contains four basic electromagnetic emission signals, including pulse signals, mismatch signals, analog signals and digital signals. The four types of signals are randomly convexly combined to form a basic signal set of combined electromagnetic interference; S102. Adding background noise to the sample signal of the constructed basic signal set of combined electromagnetic interference to obtain a combined electromagnetic interference signal, which is used as an identification signal of the electromagnetic interference source identification technology based on electromagnetic principles; S103. Transform the time-frequency domain of the identified samples to obtain a time-domain waveform diagram and a two-dimensional time-frequency diagram of the sample set, showing the time-domain and frequency-domain characteristics of the sample signal.

3. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 2 is characterized in that: The step S101 includes: Four basic electromagnetic emission signals are used as individuals of the combined signal, and the four basic signals are numbered 1-4 in the order of pulse signal, mismatch signal, analog signal and digital signal; Extract any one, two, three or all four types of signals from the four basic signals numbered 1-4, and perform random convex combination on the extracted single and multiple signals to form a combined signal. The specific method of constructing a combined signal through random convex combination is as follows: First, the basic signals of categories 1 to 4 are represented as {h i (t)}, i = 1, 2, 3, 4, there is a real number u i Satisfy the condition 0.2≤u i ≤0.8 and u i Randomly select values ​​under this condition, where n is the number of signal categories of the sample vector, and the sample vector {w m (t)}, m = 1, 2, ..., 15 is a combination signal. There are 15 types of combination signals. The expression of the basic combination signal is as follows: Where p, q = 1, 2, 3, 4, and the number of categories of basic signals contained in the combined signal is n = q-p+1. The 15 combined signals constructed are numbered 1-15 respectively, and the numbers correspond to the sample label values ​​of each category of sample signals.

4. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 2 is characterized by: The step S102 includes: According to S101, the basic interference signal set is used to generate the signal, and Gaussian white noise is used as the environmental background noise ε during the signal generation. m , construct a combined electromagnetic interference signal x containing background noise m (t), m=1,2,…,15, its expression is as follows: x m (t)=w m (t)+ε m Let the basic combination signal w m The effective power of (t) is P n , environmental background noise ε m The effective power is P s , the power ratio of the interference signal to the background noise is S, and the interference signals when S is 3dB, 6dB and 10dB are selected to form the electromagnetic interference signal sample set of the model. The calculation formula is as follows: S=10lg(P s / P n )。 5. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 3 is characterized by: In step S2, the time domain feature extraction process uses a time domain analysis method to extract peak features, trend features, margin features and pulse features in the time domain; wherein the peak features in the time domain include main peak features and peak factor features; The process of frequency domain feature extraction uses the methods of fast Fourier transform and wavelet transform to extract trend features, peak features, centroid frequency features, mean square frequency features and frequency variance features in the frequency domain; among which, the peak features in the frequency domain include main peak features and sub-peak features, and the trend features in the frequency domain include main trend features and trend position features; The entropy analysis method is used in the process of energy domain feature extraction to extract the power spectrum entropy features and energy entropy features in the energy domain; The extracted l features are recorded as X j ,j=1,2,…,l.

6. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 5 is characterized by: In step S2, the extracted characteristic index data is preprocessed using a standard score detection method, outliers are determined based on the deviation between the data points and the mean, and the outliers are eliminated.

7. The electromagnetic interference source identification method based on electromagnetic principle optimization neural network according to claim 1 is characterized by: The step S3 comprises: The sample data of electromagnetic interference source identification technology based on electromagnetic principles is processed, wherein the characteristic value of each sample in the sample data is the preprocessed feature extracted in S2, and the sample label value contained is the sample number specified in S101; it is necessary to perform feature dimensionality reduction on the sample data, select random forest as the model for feature dimensionality reduction, and perform dimensionality reduction on the sample features: First, it is necessary to extract some samples from the sample features, and then use the extracted samples as a small sample set to pre-train the network model to eliminate redundant features and prevent overfitting of the optimized network model during the training process.

Images