An electromagnetic leakage signal classification and identification method based on multi-dimensional feature fusion

Abstract

The application relates to the technical field of electromagnetic compatibility and signal processing, and discloses an electromagnetic leakage signal classification and identification method based on multi-dimensional feature fusion, which comprises the following steps: firstly, preprocessing an original electromagnetic signal to obtain standardized data; secondly, extracting time domain, frequency domain and space domain features to construct a nine-dimensional feature vector; thirdly, constructing an initial feature library and improving the initial feature library through a self-supervision checking mechanism; fourthly, training a classification model by adopting a CNN-LSTM fusion model with an attention mechanism; fifthly, classifying a signal to be identified; sixthly, starting a self-supervision reinforcement learning mechanism to optimize the model according to an F1 score; and seventhly, dynamically optimizing the feature library every month. The application solves the defects of traditional technologies, improves the identification accuracy and the system adaptability, and can be used for cable information security leakage protection.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic compatibility and signal processing technology, specifically a method for classifying and identifying electromagnetic leakage signals based on multi-dimensional feature fusion. Background Technology

[0002] In the information age, the widespread use of electronic devices inevitably leads to electromagnetic leakage during operation. These leakage signals carry sensitive operational information, and once intercepted and analyzed, they directly threaten information security in critical sectors such as finance and energy. Therefore, the accurate classification and identification of electromagnetic leakage signals has become a core aspect of information security protection. It not only requires accurately distinguishing between "normal electromagnetic signals" and "leakage signals," but also further precise identification of the source of the leakage signals, which is of irreplaceable importance for ensuring the safe and stable operation of critical electronic systems.

[0003] However, existing technologies suffer from three major problems, making it difficult to meet practical application needs: First, they rely on single features; most identification methods are based on only one dimension of features in the frequency or time domain, failing to comprehensively represent the complex characteristics of cable leakage signals, resulting in weak anti-interference capabilities and low classification accuracy. Second, the models lack adaptability; traditional models have fixed parameters, and when the electromagnetic environment changes or new types of leakage signals appear, the model's generalization ability is insufficient, making it unable to adapt to changing scenarios. Third, the feature library is static; existing feature libraries only contain the features of initially acquired signals and cannot dynamically add new leakage signal features, leading to a lag in the identification of new types of display screens and cable leakage signals. These problems make existing technologies poorly adaptable to complex real-world scenarios, making it difficult to meet the needs of critical fields for accurate identification of electromagnetic leakage signals.

[0004] This invention aims to provide a method for classifying and identifying electromagnetic leakage signals based on multi-dimensional feature fusion. By fusing time-domain, frequency-domain, and spatial-domain features, a feature vector with strong discriminative power is constructed. A CNN-LSTM model with an attention mechanism is introduced to achieve adaptive feature weighting and recognition. Furthermore, a dynamic feature library and closed-loop feedback mechanism are designed to continuously optimize the model and feature library. This method can effectively improve the accuracy and system adaptability of identifying leakage signals from cables and other sources, providing reliable technical support for high-safety electromagnetic protection. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide an electromagnetic leakage signal classification and identification method based on multi-dimensional feature fusion, which aims to overcome the core problems existing in the current electromagnetic leakage signal identification technology, such as single feature dimension, poor model adaptability, lagging feature library update and weak environmental interference resistance.

[0006] To achieve the aforementioned objectives, the technical solution adopted is as follows: A method for classifying and identifying electromagnetic leakage signals based on multi-dimensional feature fusion includes the following steps: S1: Acquire the raw electromagnetic signal using a signal acquisition device, and use a wavelet threshold denoising algorithm with 5-level decomposition based on the db4 wavelet basis to remove Gaussian white noise; filter high-frequency interference using a Butterworth low-pass filter with a cutoff frequency of 50MHz; map the amplitude of the denoised signal to... After the interval is normalized, standardized signal data is obtained, the expression of which is: ; in, The standardized signal value, The original signal value, The minimum value of the original signal. The original signal's maximum value; S2: Based on the characteristics of cable leakage signals, extract time-domain features, frequency-domain features, and spatial-domain features to construct a 9-dimensional feature vector: ; in, For peak factor, For ravine, The root mean square value, For zero-crossing rate, For center frequency, For bandwidth, For spectral entropy; Spatial distribution dispersion, The spatial correlation coefficient is given by the formula for calculating the spatial distribution dispersion. ,in It is a signal sequence. It is the mean. The number of sampling points; the formula for calculating the spatial correlation coefficient is: ; S3: Based on known cable leakage signals and normal noise samples, an initial feature library is constructed after labeling the categories. When a new leakage signal is detected, i.e. the features do not match the feature library and the classification confidence is low, a self-supervised verification mechanism is started. This mechanism includes signal reconstruction verification and cluster consistency analysis to automatically confirm the new signal. If the verification passes, its multi-dimensional features are extracted and added to the feature library to realize the automatic improvement of the feature library. S4: A CNN-LSTM fusion model with attention mechanism is used for training. The CNN module contains 2 convolutional layers and 2 max pooling layers, where the convolutional kernel sizes are respectively... , The activation function is ReLU, and the pooling size is 2×2. The LSTM module contains one hidden layer with 64 neurons and a dropout coefficient of 0.2. The attention layer is located between the CNN output and the LSTM input. It calculates the importance scores of each feature vector extracted by the CNN through a small feedforward neural network, and then normalizes the scores using an exponential function. ; Normalized to adaptive weights, where, For the final attention weight, The total number of feature vectors output by the CNN module. , It is the i-th feature vector extracted by the CNN. , , These are learnable parameters; the final weighted feature vector is input into the LSTM module; this mechanism allows the model to dynamically assign higher weights to key features of cable leakage signals while suppressing interference from irrelevant features; using 70% of the samples in the feature library as the training set and 30% as the validation set, the cross-entropy loss function is employed. ; in, For real labels, To predict probabilities, the Adam optimizer was trained iteratively for 50 epochs to obtain a binary classification model. The parameters of the Adam optimizer were set as follows: learning rate 0.001. , ; S5: Perform the preprocessing and feature extraction steps S1 and S2 on the signal to be identified. After obtaining the feature vector, input it into the trained classification model to output the class probability distribution. Determine the signal type according to the following judgment rules: If the highest probability corresponds to "cable leakage" and the probability At the same time, satisfying kurtosis And peak factor Spectral entropy And bandwidth Spatial correlation coefficient If so, it is determined to be a cable leakage signal; If the model outputs probability If at least two of the above physical criteria are met, it is determined to be a "suspected leakage signal," and a self-monitoring verification mechanism is initiated: Signal reconstruction autoencoder verification: The signal reconstruction error is calculated using a pre-trained autoencoder. ; in, For reconstruction error, The first normalized signal Each sample point value, The first signal after reconstruction by the self-encoder Each sample point value, The total number of signal sampling points, if , Based on the statistics of known leakage signals, it is determined to be a high-confidence leakage signal; Contrastive learning feature enhancement: Time-frequency domain data enhancement is performed on low-confidence signals, and the contrastive loss is calculated using the SimCLR framework to improve feature discriminative power; Automatic cluster consistency labeling: The cosine similarity expression for calculating the signal to be verified and the cluster centers of the feature library is as follows: ; in, For vectors and Cosine similarity between them For vectors and The inner product, , For vectors and The L2 norm, if and Automatically flagged as a leakage signal and added to the feature library; Otherwise, it is considered normal noise; S6: Calculate the confusion matrix of the identification results and calculate the F1 score of the cable leakage signal. The calculation formula is as follows: ; The accuracy rates for the two types of signals are as follows: Recall rate: When the F1 score is lower than P, the self-supervised reinforcement learning mechanism is activated. First, the low-confidence samples are augmented in the time and frequency domain. Then, the SimCLR framework is used to optimize the feature representation. Finally, the feature extraction module is fine-tuned through the BYOL algorithm, and augmented samples are added as appropriate to improve the weak performance of the model in a targeted manner. S7: The feature library is dynamically optimized monthly using self-supervised representation learning. BYOL self-supervised training: Improves the quality of feature representations through interactive learning between an online network and the target network. The loss function is: ; in, For the predictor output of online networks, The feature projection output of the target network. , These represent parameters for the online network and the target network, respectively. Hierarchical clustering purification: hierarchical clustering is performed based on the learned feature space to merge semantically similar features and remove low-variance redundant features; dynamic threshold adaptive self-supervised parameters are automatically adjusted according to the F1 score change trend; S8: Periodically use self-supervised representation learning algorithms to dynamically optimize the feature library, forming an intelligent closed-loop feedback.

[0007] As a further improvement of the present invention, in step S3, when constructing the initial feature library, 100 sets of cable leakage signal samples and 100 sets of normal noise samples are collected, and multi-dimensional features are extracted and labeled to construct the initial feature library.

[0008] As a further improvement of the present invention, in step SS6, the confusion matrix of 1000 consecutive identifications is statistically analyzed, and the F1 score of "cable leakage" is calculated.

[0009] Compared with the prior art, the present invention discloses the following beneficial effects: 1. Construct a full-process technical architecture of "acquisition-extraction-classification-optimization", integrating multi-dimensional features of time domain, frequency domain, and spatial domain with a closed-loop feedback mechanism to provide a systematic solution for the classification and identification of electromagnetic leakage signals in cables, breaking the limitations of traditional single-dimensional and static processing.

[0010] 2. A nine-dimensional discriminative feature vector is constructed, which largely covers the differentiated characteristics of cable leakage signals in terms of time fluctuation, frequency distribution, and spatial propagation. This solves the problems of weak anti-interference and low classification accuracy caused by the reliance on single features in traditional technologies, and significantly improves the comprehensiveness and discriminative power of signal feature expression.

[0011] 3. A CNN-LSTM fusion model with attention mechanism is adopted. The CNN module efficiently extracts key local features of the signal, the LSTM module captures the temporal correlation of features, and the attention layer strengthens key leakage features and suppresses irrelevant interference through adaptive weighting. This achieves the synergistic extraction of "local features + temporal patterns + dynamic weights", which solves the defects of traditional models with fixed parameters and insufficient generalization ability, and enhances the adaptability to complex electromagnetic environments.

[0012] 4. Design a dynamic and scalable feature library and a closed-loop feedback optimization mechanism: Automatically add effective features through a self-supervised verification mechanism, combine F1 score threshold to trigger sample supplementation and model retraining, and complete feature library optimization monthly using self-supervised representation learning to achieve co-evolution of the model and feature library, solving the problem of lagging recognition of new leakage signals caused by the static nature of the existing feature library.

[0013] 5. By integrating multi-source information such as signal physical characteristics, model confidence, and self-supervised reconstruction error, suspected leakage signals are automatically and intelligently verified, forming a multi-layered verification system of "model judgment + physical criteria + self-supervised verification", which greatly reduces the risk of misjudgment and missed judgment and improves the reliability of identification results. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an application scenario diagram for electromagnetic leakage signal classification and identification of cables provided in an embodiment of the present invention; Figure 2 This is an application scenario diagram for electromagnetic leakage signal classification and identification of normal signals (including signals from compliant equipment) provided in an embodiment of the present invention; Figure 3 This is a flowchart of the leakage signal classification and identification method provided in an embodiment of the present invention; Figure 4 This is an example diagram of an electromagnetic leakage signal classification and identification system provided in an embodiment of the present invention. Detailed Implementation

[0015] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] The following is in conjunction with the appendix Figure 1-4 The present invention will be described in further detail below.

[0018] S1: Acquire the raw electromagnetic signal using a signal acquisition device, and use a wavelet threshold denoising algorithm with 5-level decomposition based on the db4 wavelet basis to remove Gaussian white noise; filter high-frequency interference using a Butterworth low-pass filter with a cutoff frequency of 50MHz; map the amplitude of the denoised signal to... After the interval is normalized, standardized signal data is obtained, the expression of which is: ; in, The standardized signal value, The original signal value, The minimum value of the original signal. The original signal's maximum value; S2: Based on the characteristics of cable leakage signals, extract time-domain features, frequency-domain features, and spatial-domain features to construct a 9-dimensional feature vector: ; in, For peak factor, For ravine, The root mean square value, For zero-crossing rate, For center frequency, For bandwidth, For spectral entropy; Spatial distribution dispersion, The spatial correlation coefficient is given by the formula for calculating the spatial distribution dispersion. ,in It is a signal sequence. It is the mean. The number of sampling points; the formula for calculating the spatial correlation coefficient is: ; S3: Based on known cable leakage signals and normal noise samples, an initial feature library is constructed after labeling the categories. When a new leakage signal is detected, i.e. the features do not match the feature library and the classification confidence is low, a self-supervised verification mechanism is started. This mechanism includes signal reconstruction verification and cluster consistency analysis to automatically confirm the new signal. If the verification passes, its multi-dimensional features are extracted and added to the feature library to realize the automatic improvement of the feature library. S4: A CNN-LSTM fusion model with attention mechanism is used for training. The CNN module contains 2 convolutional layers and 2 max pooling layers, where the convolutional kernel sizes are respectively... , The activation function is ReLU, and the pooling size is 2×2. The LSTM module contains one hidden layer with 64 neurons and a dropout coefficient of 0.2. The attention layer is located between the CNN output and the LSTM input. It calculates the importance scores of each feature vector extracted by the CNN through a small feedforward neural network, and then normalizes the scores using an exponential function. ; Normalized to adaptive weights, where, For the final attention weight, The total number of feature vectors output by the CNN module. , It is the i-th feature vector extracted by the CNN. , , These are learnable parameters; the final weighted feature vector is input into the LSTM module; this mechanism allows the model to dynamically assign higher weights to key features of cable leakage signals while suppressing interference from irrelevant features; using 70% of the samples in the feature library as the training set and 30% as the validation set, the cross-entropy loss function is employed. ; in, For real labels, To predict probabilities, the Adam optimizer was trained iteratively for 50 epochs to obtain a binary classification model. The parameters of the Adam optimizer were set as follows: learning rate 0.001. , ; S5: Perform the preprocessing and feature extraction steps S1 and S2 on the signal to be identified. After obtaining the feature vector, input it into the trained classification model to output the class probability distribution. Determine the signal type according to the following judgment rules: If the highest probability corresponds to "cable leakage" and the probability At the same time, satisfying kurtosis And peak factor Spectral entropy And bandwidth Spatial correlation coefficient If so, it is determined to be a cable leakage signal; If the model outputs probability If at least two of the above physical criteria are met, it is determined to be a "suspected leakage signal," and a self-monitoring verification mechanism is initiated: Signal reconstruction autoencoder verification: The signal reconstruction error is calculated using a pre-trained autoencoder. ; in, For reconstruction error, The first normalized signal Each sample point value, The first signal after reconstruction by the self-encoder Each sample point value, The total number of signal sampling points, if , Based on the statistics of known leakage signals, it is determined to be a high-confidence leakage signal; Contrastive learning feature enhancement: Time-frequency domain data enhancement is performed on low-confidence signals, and the contrastive loss is calculated using the SimCLR framework to improve feature discriminative power; Automatic cluster consistency labeling: The cosine similarity expression for calculating the signal to be verified and the cluster centers of the feature library is as follows: ; in, For vectors and Cosine similarity between them For vectors and The inner product, , For vectors and The L2 norm, if and Automatically flagged as a leakage signal and added to the feature library; Otherwise, it is considered normal noise; S6: Calculate the confusion matrix of the identification results and calculate the F1 score of the cable leakage signal. The calculation formula is as follows: ; The accuracy rates for the two types of signals are as follows: Recall rate: When the F1 score is lower than P, the self-supervised reinforcement learning mechanism is activated. First, the low-confidence samples are augmented in the time and frequency domain. Then, the SimCLR framework is used to optimize the feature representation. Finally, the feature extraction module is fine-tuned through the BYOL algorithm, and augmented samples are added as appropriate to improve the weak performance of the model in a targeted manner. S7: The feature library is dynamically optimized monthly using self-supervised representation learning. BYOL self-supervised training: Improves the quality of feature representations through interactive learning between an online network and the target network. The loss function is: ; in, For the predictor output of online networks, The feature projection output of the target network. , These represent parameters for the online network and the target network, respectively. Hierarchical clustering purification: hierarchical clustering is performed based on the learned feature space to merge semantically similar features and remove low-variance redundant features; dynamic threshold adaptive self-supervised parameters are automatically adjusted according to the F1 score change trend; S8: Periodically use self-supervised representation learning algorithms to dynamically optimize the feature library, forming an intelligent closed-loop feedback.

[0019] In step S3, when constructing the initial feature library, 100 sets of cable leakage signal samples and 100 sets of normal noise samples are collected, and multi-dimensional features are extracted and labeled to construct the initial feature library.

[0020] In step S6, the confusion matrix of 1000 consecutive identifications is statistically analyzed, and the F1 score of "cable leakage" is calculated.

[0021] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, component splitting or combination, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for classifying and identifying electromagnetic leakage signals based on multi-dimensional feature fusion, characterized in that, Includes the following steps: S1: Acquire the raw electromagnetic signal using a signal acquisition device, and use a wavelet threshold denoising algorithm with 5-level decomposition based on the db4 wavelet basis to remove Gaussian white noise; filter high-frequency interference using a Butterworth low-pass filter with a cutoff frequency of 50MHz; map the amplitude of the denoised signal to... After the interval is normalized, standardized signal data is obtained, the expression of which is: ； in, The standardized signal value, The original signal value, The minimum value of the original signal. The original signal's maximum value; S2: Based on the characteristics of cable leakage signals, extract time-domain features, frequency-domain features, and spatial-domain features to construct a 9-dimensional feature vector: ； in, For peak factor, For ravine, The root mean square value, For zero-crossing rate, For center frequency, For bandwidth, For spectral entropy; Spatial distribution dispersion, The spatial correlation coefficient is given by the formula for calculating the spatial distribution dispersion. ,in It is a signal sequence. It is the mean. The number of sampling points; the formula for calculating the spatial correlation coefficient is: ; S3: Based on known cable leakage signals and normal noise samples, an initial feature library is constructed after labeling the categories. When a new leakage signal is detected, i.e. the features do not match the feature library and the classification confidence is low, a self-supervised verification mechanism is started. This mechanism includes signal reconstruction verification and cluster consistency analysis to automatically confirm the new signal. If the verification passes, its multi-dimensional features are extracted and added to the feature library to realize the automatic improvement of the feature library. S4: A CNN-LSTM fusion model with attention mechanism is used for training. The CNN module contains 2 convolutional layers and 2 max pooling layers, where the convolutional kernel sizes are respectively... , The activation function is ReLU, and the pooling size is 2×2. The LSTM module contains one hidden layer with 64 neurons and a dropout coefficient of 0.

2. The attention layer is located between the CNN output and the LSTM input. It calculates the importance scores of each feature vector extracted by the CNN through a small feedforward neural network, and then normalizes the scores using an exponential function. ； Normalized to adaptive weights, where, For the final attention weight, The total number of feature vectors output by the CNN module. , It is the i-th feature vector extracted by the CNN. , , These are learnable parameters; the final weighted feature vector is input into the LSTM module; this mechanism allows the model to dynamically assign higher weights to key features of cable leakage signals while suppressing interference from irrelevant features; using 70% of the samples in the feature library as the training set and 30% as the validation set, the cross-entropy loss function is employed. ； in, For real labels, To predict probabilities, the Adam optimizer was trained iteratively for 50 epochs to obtain a binary classification model. The parameters of the Adam optimizer were set as follows: learning rate 0.

001. , ; S5: Perform the preprocessing and feature extraction steps S1 and S2 on the signal to be identified. After obtaining the feature vector, input it into the trained classification model to output the class probability distribution. Determine the signal type according to the following judgment rules: If the highest probability corresponds to "cable leakage" and the probability At the same time, satisfying kurtosis And peak factor Spectral entropy And bandwidth Spatial correlation coefficient If so, it is determined to be a cable leakage signal; If the model outputs probability If at least two of the above physical criteria are met, it is determined to be a "suspected leakage signal," and a self-monitoring verification mechanism is initiated: Signal reconstruction autoencoder verification: The signal reconstruction error is calculated using a pre-trained autoencoder. ； in, For reconstruction error, The first normalized signal Each sample point value, The first signal after reconstruction by the self-encoder Each sample point value, The total number of signal sampling points, if , Based on the statistics of known leakage signals, it is determined to be a high-confidence leakage signal; Contrastive learning feature enhancement: Time-frequency domain data enhancement is performed on low-confidence signals, and the contrastive loss is calculated using the SimCLR framework to improve feature discriminative power; Automatic cluster consistency labeling: The cosine similarity expression for calculating the signal to be verified and the cluster centers of the feature library is as follows: ； in, For vectors and Cosine similarity between them For vectors and The inner product, , For vectors and The L2 norm, if and Automatically flagged as a leakage signal and added to the feature library; Otherwise, it is considered normal noise; S6: Calculate the confusion matrix of the identification results and calculate the F1 score of the cable leakage signal. The calculation formula is as follows: ； The accuracy rates for the two types of signals are as follows: Recall rate: When the F1 score is lower than P, the self-supervised reinforcement learning mechanism is activated. First, the low-confidence samples are augmented in the time and frequency domain. Then, the SimCLR framework is used to optimize the feature representation. Finally, the feature extraction module is fine-tuned through the BYOL algorithm, and augmented samples are added as appropriate to improve the weak performance of the model in a targeted manner. S7: The feature library is dynamically optimized monthly using self-supervised representation learning. BYOL self-supervised training: Improves the quality of feature representations through interactive learning between an online network and the target network. The loss function is: ； in, For the predictor output of online networks, The feature projection output of the target network. , These represent parameters for the online network and the target network, respectively. Hierarchical clustering purification: hierarchical clustering is performed based on the learned feature space to merge semantically similar features and remove low-variance redundant features; dynamic threshold adaptive self-supervised parameters are automatically adjusted according to the F1 score change trend; S8: Periodically use self-supervised representation learning algorithms to dynamically optimize the feature library, forming an intelligent closed-loop feedback.

2. The electromagnetic leakage signal classification and identification method based on multi-dimensional feature fusion according to claim 1, characterized in that: In step S3, when constructing the initial feature library, 100 sets of cable leakage signal samples and 100 sets of normal noise samples are collected, and multi-dimensional features are extracted and labeled to construct the initial feature library.

3. The electromagnetic leakage signal classification and identification method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step SS6, the confusion matrix of 1000 consecutive identifications is statistically analyzed, and the F1 score of "cable leakage" is calculated.

Images