Electric energy router AC port fault detection method based on multi-modal feature fusion

Abstract

The invention provides an electric energy router AC port fault detection method based on multi-modal feature fusion. The method comprises the following steps: S1, obtaining historical operation data of an electric energy router AC port; s2, carrying out denoising and standardization processing on the acquired multi-mode signal data; s3, time domain feature information and frequency domain feature information in the multi-mode signals are extracted; s4, performing weighted fusion on the multi-modal feature information; s5, constructing a neural network to train the fused feature set to obtain a fault detection model; and S6, inputting real-time monitored operation data of the AC port of the electric energy router into the fault detection model, and outputting a fault detection result. According to the invention, through multi-modal signal fusion and a deep learning model, the spatial-temporal characteristics of the fault can be comprehensively captured, and the accuracy of fault detection is significantly improved. Compared with a traditional fault detection method of a single signal source or a shallow model, the method can effectively recognize a complex fault mode, and improves the robustness of a system in a complex power environment.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault detection, and in particular to a method for detecting AC port faults of an electric energy router based on multimodal feature fusion. Background Art

[0002] With the continuous development of smart grid technology, power routers, as core equipment in power systems, play a vital role, particularly in distributed power access, power dispatch, and power quality control. AC port failures are a common occurrence in residential power routers, and these failures are often difficult to detect in their early stages. If not detected promptly, they can lead to reduced system efficiency and even more serious power equipment failures, jeopardizing the safe operation of the power system.

[0003] Currently, most power router fault detection methods rely on traditional time-frequency analysis or model-based algorithms. For example, CN11854984A uses wavelet transforms and support vector machines (SVMs) to extract features from current signals for fault classification. While these methods can achieve some success in simple fault scenarios, they suffer from low accuracy and robustness in complex power environments due to the nonlinearity and time-varying nature of signal characteristics and interference from multiple signal sources. This is particularly true for low-order harmonic current signals caused by open-circuit faults, where traditional methods have limited feature extraction capabilities and are prone to misdiagnosis and missed diagnoses.

[0004] With the rapid development of deep learning technology, an increasing number of intelligent fault diagnosis methods have been applied. Deep learning models, particularly convolutional neural networks (CNNs) and attention mechanisms, can automatically learn and extract high-dimensional feature information, demonstrating significant advantages in processing multimodal data and complex fault scenarios. Combining EEMD (ensemble empirical mode decomposition) with multimodal data fusion can significantly improve the accuracy and real-time performance of fault detection.

[0005] EEMD, as a signal decomposition technique, effectively processes nonlinear and nonstationary signals and is particularly well-suited for decomposing low-order harmonic currents caused by faults in power routers. By performing EEMD decomposition on multidimensional signals such as current, voltage, and temperature, intrinsic mode functions (IMFs) of varying frequency components can be extracted, effectively revealing fault signatures within the signal. Multimodal data fusion, on the other hand, integrates information from multiple signal sources, providing more comprehensive and rich feature information and enhancing the robustness of fault detection.

[0006] However, existing fault diagnosis methods often rely solely on the features of a single modal signal and mostly employ traditional machine learning models, resulting in issues such as manual feature selection and limited classification effectiveness. Therefore, efficiently detecting AC port faults on power routers through the integration of deep learning techniques, particularly multimodal feature fusion and EEMD decomposition, remains a key technical challenge in the industry. Summary of the Invention

[0007] The present invention proposes a method for detecting AC port faults of power routers based on multimodal feature fusion, which solves the problems in the prior art that fault diagnosis methods only rely on the characteristics of a single modal signal, resulting in low fault detection accuracy and poor robustness.

[0008] The technical solution of the present invention is achieved as follows: A first aspect of the present invention provides a method for detecting AC port faults of an energy router based on multimodal feature fusion, comprising the following steps: S1, multi-modal signal acquisition: Acquire historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals, and temperature signals under normal operation and fault operation; S2, data preprocessing: denoising and standardization of the collected multimodal signal data; S3, feature extraction: using time domain feature extraction and frequency domain feature extraction methods to extract feature information from multimodal signals; S4, multimodal fusion: weighted fusion of multimodal feature information to obtain the fused feature set; S5, model training: build a neural network, input the fused feature set into the neural network for training, and obtain a fault detection model; S6, model prediction: input the current signal, voltage signal and temperature signal of the AC port of the power router monitored in real time into the fault detection model and output the fault detection result.

[0009] Specifically, step S3 includes the following steps: S301, using the EEMD method, decompose different modal signals separately to obtain different IMF components; S302, using time domain feature extraction and frequency domain feature extraction methods to extract time domain and frequency domain feature information of different IMF components respectively to obtain different feature sets; S303: Use correlation analysis to screen the extracted features, remove redundant features, and retain features that are useful for fault diagnosis.

[0010] Furthermore, step S301 includes: Generate white noise using a pseudorandom number generator: or ( t )= s × randn (1, L ) in, or ( t ) is white noise; s is the set white noise standard deviation, randn (1, L ) means generating 1× L A random matrix, each element in the random matrix obeys the standard normal distribution; Superimpose white noise on the original signal to break the potential spectrum aliasing in the signal. The formula is as follows: x noisy ( t )= x ( t )+ or ( t ) in, x ( t ) is the original signal, x noisy ( t ) is the signal after superimposing white noise; The EEMD method is used to decompose the noisy signal and obtain the IMF component. The formula is as follows:

[0011] in, IMF m ( t ) is the m The IMF components obtained by the decomposition are: r m ( t ) is the residual term; After multiple decompositions, multiple different IMF components are obtained; the decomposed IMF components are averaged to form the final stable IMF output sequence:

[0012] in, IMF j ( t ) is the j The final IMF component signal obtained after decomposing the modal signal is M is the total number of decompositions.

[0013] Furthermore, in step S302, The formulas for extracting time domain features of IMF components include:

[0014] in, m j is the mean, N is the total number of sampling points;

[0015] in, s j is the standard deviation;

[0016] in, RMS j is the root mean square value;

[0017] in, Skewness j is the skewness;

[0018] in, Kurtosis j is the kurtosis; The formulas for extracting frequency domain features of IMF components include:

[0019] in, x j ( f ) is the j-th harmonic component in the spectrum after Fourier transform;

[0020] in, E j ( f ) is the spectrum energy;

[0021] in, X THD is the total harmonic distortion rate, x 1 ( f ) is the fundamental component.

[0022] Furthermore, in step S303, the correlation analysis calculation formula is:

[0023] in, f i For the i eigenvalues, y i For thei The feature values ​​correspond to the value of the fault label; m f is the mean of the eigenvalues, m y is the mean of the labels; r is the correlation coefficient; k is the total number of extracted eigenvalues; Select the correlation coefficient r Features with a value greater than 0.7 are retained.

[0024] Specifically, in step S4, the multimodal fusion method includes: The weighted fusion method is used to fuse multimodal features. The weight of each modality is assigned according to its contribution to fault diagnosis. The weighted fusion formula is as follows:

[0025] in, F p For the p The feature set of a modality, w p For the p The weight coefficient corresponding to the feature set of each mode; F weighted is the fused feature set; The weight coefficients are learned by the gradient descent method. The goal is to minimize the error of the fault detection model through the weighted fusion feature set. The loss function is:

[0026] in, L ( w ) is the loss function, which represents the error of the fault detection model; k 1 is the total number of features in the feature set; l is a regularization parameter used to prevent overfitting; y i For the i The feature values ​​correspond to the value of the fault label; Update weight coefficients by gradient descent w p , the gradient descent formula is:

[0027] in, is the updated weight coefficient; is the weight coefficient before updating; or is the learning rate, which is used to control the step size of each update; is the partial derivative of the loss function with respect to the weight coefficient.

[0028] Furthermore, the principal component analysis method is used to reduce the dimension of the weighted fusion feature set matrix, including the following steps: Normalize the fused feature set matrix:

[0029] in, F weighted is the feature set matrix after weighted fusion, for F weighted The mean of for F weighted variance; F s is the standardized feature set matrix; Calculate the covariance matrix Σ of the standardized feature set matrix:

[0030] in, T Represents the transpose of a matrix; Perform eigenvalue decomposition on the covariance matrix Σ:

[0031] in, oh is the eigenvalue, I for p × p The identity matrix of Solving the equation yields p eigenvalues oh 1, oh 2, oh 3,…, oh p , for each eigenvalue, the corresponding eigenvector is obtained by solving the following linear equations:

[0032] in, oh i For the i eigenvalues, e i For the i The eigenvector corresponding to the eigenvalue; Arrange the eigenvalues ​​and corresponding eigenvectors in descending order, and select the first v The eigenvalues ​​corresponding to the eigenvectors are composed of n × v Matrix w T = [ e1, e 2,…, e v ], project the standardized weighted fusion feature set matrix onto the selected front v On the principal components, we get a new feature matrix: F PCA = w T F s in, F PCA is the feature matrix after dimensionality reduction.

[0033] Specifically, in step S5, the neural network includes an input layer, a multi-layer convolution layer, two LSTM layers, a fully connected layer, and an output layer connected in sequence. The input layer is used to input the weighted fused feature set, and the multi-layer convolution layer is used to extract local features from the multimodal feature set; the LSTM layer is used to process the timing information of the signal, and a Dropout layer is added between the two LSTM layers to prevent model overfitting; the fully connected layer is used to perform feature fusion and classification on the output of the LSTM layer; the output layer uses a Softmax activation function to output the probability distribution of each fault detection category; the loss function of the fault detection model is as follows:

[0034] in, n is the number of samples in the feature set after weighted fusion; y i For the i The feature values ​​correspond to the value of the fault label; p i For the fault detection model to predict i The failure probability of a sample; During model training, the backpropagation algorithm and gradient descent method are used to minimize the loss function. By continuously adjusting the weights in the network, the model learns how to predict faults based on the multimodal features of the input. During model training, the Adam optimizer is selected, and the initial learning rate is set to 0.001. After every 10 training rounds, the learning rate is multiplied by 0.9.

[0035] Preferably, the fault detection method further includes: Step S7, real-time feedback and optimization: By real-time monitoring of the operating data of the AC port of the power router, real-time fault detection is performed based on the current operating data, and the fault detection results are fed back to the central control platform in real time; the central control platform adjusts and optimizes the hyperparameters of the fault detection model based on the feedback results, and retrains the fault detection model using the newly collected operating data.

[0036] A second aspect of the present invention provides an AC port fault detection system for an energy router based on multimodal feature fusion, based on the fault detection method, comprising: Signal acquisition module: used to collect historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals and temperature signals under normal operation and fault operation; Preprocessing module: used to perform denoising and standardization on the collected multimodal signal data; Feature extraction module: used to extract time domain features and frequency domain features from multimodal signals; Feature fusion module: used to perform weighted fusion of multimodal feature information to obtain a fused feature set; Fault detection module: used to build a neural network, input the fused feature set into the neural network for training, and obtain a fault detection model; input the real-time collected operating data of the AC port of the power router into the fault detection model, and output the fault detection results.

[0037] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention can comprehensively capture the temporal and spatial characteristics of faults through multimodal signal fusion (time and frequency domain characteristics of current, voltage, and temperature signals) and deep learning models, significantly improving the accuracy of fault detection. Compared with traditional fault detection methods based on a single signal source or shallow models, the present invention can effectively identify complex fault modes and improve the robustness of the system in complex power environments; (2) This invention adopts a CNN-LSTM combined model, using a convolutional neural network (CNN) to extract local features of multimodal signals and a long short-term memory network (LSTM) to capture the temporal characteristics of the signals. This combined model breaks through the traditional method's neglect of temporal information and can more accurately identify faults, especially when the signal data is complex or time-dependent. (3) This invention introduces a real-time feedback mechanism, enabling the fault detection system to continuously optimize the deep learning model through online learning based on real-time monitoring and feedback data. This innovation enables the system to adapt to dynamic environmental changes and gradually improve the accuracy of fault detection. In particular, when faced with new fault modes and different working environments, the model can be quickly adjusted to maintain efficiency and stability, overcoming the problem of traditional models' poor adaptability in dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 This is a flow chart of a method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to the present invention; Figure 2 This is a schematic diagram of the EEMD decomposition process in an embodiment of the present invention; Figure 3 This is a diagram of the CNN-LSTM network structure in an embodiment of the present invention; Figure 4 Graph showing the change in loss function and accuracy of the training set and test set versus the number of iterations in an embodiment of the present invention; Figure 5 FIG. 4 is a confusion matrix diagram of fault detection in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] Reference Figure 1 The first aspect of the present invention provides a method for detecting AC port faults of an energy router based on multimodal feature fusion, comprising the following steps: S1, multi-modal signal acquisition: Acquire historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals, and temperature signals under normal operation and fault operation; In step S1, in order to accurately identify and determine the AC port fault of the power router, key multimodal signal data needs to be collected. The data mainly comes from the real-time monitoring system and historical database of the power router during operation. The real-time monitoring system collects the operating data of the power router at a sampling interval of 10ms and stores it in CSV format, generating a new data file every hour; the historical database stores the operating data of the past three years in HDF5 format for efficient query and management.

[0042] This embodiment uses sensors to collect the following three types of key signals: Three-phase current signal: As the core indicator for detecting abnormal AC port current, especially when a fault occurs, the current of a specific phase will produce significant amplitude attenuation and low-order harmonic changes; Three-phase voltage signal: Assists in judging the correlation between line voltage anomalies and faults; Temperature signal: monitors the temperature rise of key components in the power router to reflect the system's operating load, heat dissipation capacity, and potential thermal instability issues.

[0043] In this embodiment, the data acquisition device includes: High-precision current sensors with an accuracy of 0.1% of full scale and a bandwidth of 0-20kHz are used. They can meet the requirements of capturing the amplitude attenuation and low-order harmonic changes (frequency range is usually 2-20kHz) generated by three-phase current during faults. Voltage transformer: Using a voltage transformer with a transformation ratio of 1000:1 and an accuracy of 0.2, it can convert high voltage into a low voltage signal suitable for measurement. At the same time, it has good electrical isolation performance, which can effectively prevent the measuring circuit from affecting the measured circuit and accurately assist in judging the correlation between line voltage anomalies and faults. Embedded or attached temperature sensors use embedded thermistor temperature sensors with a measurement accuracy of ±0.5°C and a response time of less than 1s. They can quickly and continuously measure temperature fluctuations of the heat sink or housing, and reflect the system's operating load, heat dissipation capacity, and potential thermal instability issues in real time. The multi-channel acquisition module ensures that all of the above signals can be collected synchronously, is resistant to electromagnetic interference, and supports long-term online monitoring.

[0044] S2, data preprocessing: denoising and standardization of the collected multimodal signal data; To ensure data quality, the collected multimodal signals need to be preprocessed. In this embodiment, the signal preprocessing includes: Remove missing values, obvious outliers or non-physical range data to remove noise from the collected signals; Standardize or normalize the signals from different signal sources to unify the signal scale and ensure the consistency of feature extraction; Align data with 1ms accuracy based on timestamps and sort data in chronological order; The data samples are reconstructed in chronological order with a sample length of 1 second and an overlap rate of 0.5 seconds between adjacent samples to ensure that subsequent analysis has complete time series characteristics and can fully utilize the data information.

[0045] S3, feature extraction: using time domain feature extraction and frequency domain feature extraction methods to extract feature information from multimodal signals; The specific steps include: S301, using the ensemble empirical mode decomposition (EEMD) method to decompose different modal signals separately to obtain different intrinsic mode function (IMF) components; Generate white noise using a pseudorandom number generator: or ( t )= s × randn (1, L ) in, or ( t ) is white noise; s is the set white noise standard deviation, randn (1, L ) means generating 1× L A random matrix, each element in the random matrix obeys the standard normal distribution; Superimpose white noise on the original signal to break the potential spectrum aliasing in the signal. The formula is as follows: x noisy ( t )= x ( t )+ or ( t ) in, x ( t ) is the original signal, x noisy ( t ) is the signal after superimposing white noise; The EEMD method is used to decompose the noisy signal and obtain the (intrinsic mode function) IMF component. The formula is as follows:

[0046] in, IMF m ( t ) is the m The IMF components obtained by the decomposition are: r m ( t ) is the residual term; After multiple decompositions, multiple different IMF components are obtained, such as Figure 2 As shown, this embodiment undergoes 5 decompositions. Figure 2In the graph, the horizontal axis represents the time series, and the vertical axis represents the amplitude of the signal / IMF component. Original signal represents the complete waveform of the original signal (voltage, current, or temperature). IMF 1-IMF5 represent the five intrinsic mode functions (IMFs) decomposed from the original signal using EEMD. The IMFs are arranged in order from high to low frequency (or from fine to coarse time scale): IMF 1: represents the highest frequency (finest scale) component; IMF 2 - IMF 4: represent oscillatory components of intermediate frequencies / scales; IMF 5: represents the lowest frequency (coarsest scale) component.

[0047] The decomposed IMF components are averaged to form the final stable IMF output sequence:

[0048] in, IMF j ( t ) is the j The final IMF component signal obtained after decomposing the modal signal is M is the total number of decompositions.

[0049] S302, using time domain feature extraction and frequency domain feature extraction methods, by performing multi-dimensional analysis on the signal, extracting time domain and frequency domain feature information that can effectively reflect the fault characteristics from different IMF components, and obtaining different feature sets; The formulas for extracting time domain features of IMF components include: 1) Mean m j :

[0050] in, m j is the mean, N is the total number of sampling points; 2) Standard deviation s j :

[0051] in, s j is the standard deviation; 3) Root mean square value RMS j :

[0052] in, RMS j is the root mean square value; 4) Skewness Skewness j :

[0053] in, Skewness j is the skewness; 5) Kurtosis Kurtosis j :

[0054] in, Kurtosis j is the kurtosis; The formulas for extracting frequency domain features of IMF components include: 6) Fourier transform x j ( f ):

[0055] in, x j ( f ) is the j-th harmonic component in the spectrum after Fourier transform; 7) Spectral energy E j ( f ):

[0056] in, E j ( f ) is the spectrum energy; 8) Total harmonic distortion X THD :

[0057] in, X THD is the total harmonic distortion rate, x 1 ( f ) is the fundamental component.

[0058] S303, using correlation analysis to screen the extracted features, remove redundant features, and retain features useful for fault diagnosis; After extracting features in the time, frequency, and time-frequency domains, the features of each signal source (mode) are combined into a single feature set. To improve the accuracy and computational efficiency of the fault detection model, the extracted features are selected and redundant features are removed. Correlation analysis is performed to retain the most useful features for fault diagnosis. The correlation analysis calculation formula is:

[0059] in, f i For the i eigenvalues, y i For the i The feature value corresponds to the value of the fault label, label y i A value of 0 indicates that the AC port of the power router is operating normally. y i A value of 1 indicates a fault on the AC port of the power router; m f is the mean of the eigenvalues, m y is the mean of the labels; r is the correlation coefficient; k is the total number of extracted eigenvalues; Select the correlation coefficient r The features with a value greater than 0.7 are retained. After feature extraction and selection, the effective feature sets of multiple signal sources such as current, voltage and temperature are finally obtained.

[0060] S4, multimodal fusion: weighted fusion of multimodal feature information to obtain the fused feature set; In this example, the current, temperature, and voltage signal features extracted through EEMD decomposition are fused to improve the accuracy and robustness of fault detection. The fusion method used is weighted fusion and principal component analysis (PCA). By rationally combining the features of each signal source, the final fault detection signature is obtained.

[0061] In one embodiment, a weighted fusion method is used to fuse multimodal features based on the different characteristics of current, voltage, and temperature signals. The characteristics of each signal source are assigned different weights based on their contribution to fault diagnosis, thereby obtaining the final fused feature set. The weighted fusion formula is as follows:

[0062] in, F p For the p The feature set of a modality,w p For the p The weight coefficients corresponding to the feature sets of each mode can be optimized based on the system feedback optimization. F weighted is the fused feature set; In order to ensure that each signal source plays the greatest role in fault detection, the weight coefficient is learned by gradient descent method. The goal is to minimize the error of the fault detection model through the weighted fusion feature set. The loss function is:

[0063] in, L ( w ) is the loss function, which represents the error of the fault detection model; k 1 is the total number of features in the feature set after removing redundancy; l is a regularization parameter used to prevent overfitting; Update weight coefficients by gradient descent w p , optimize the objective function, and the gradient descent formula is:

[0064] in, is the updated weight coefficient; is the weight coefficient before updating; or is the learning rate, which is used to control the step size of each update; is the partial derivative of the loss function with respect to the weight coefficient.

[0065] The optimized weight coefficient w p , perform weighted fusion on the features of each signal source to obtain the final fusion feature set F weighted , F weighted is a n × k 1 matrix, n is the sample size, k 1 is the number of features.

[0066] After obtaining the weighted fusion feature set matrix F weighted Finally, principal component analysis (PCA) is used for dimensionality reduction, which includes the following steps: 1) Standardize the fused feature set matrix:

[0067] in, Fweighted is the feature set matrix after weighted fusion, for F weighted The mean of for F weighted variance; F s is the standardized feature set matrix; 2) Calculate the covariance matrix Σ of the standardized feature set matrix:

[0068] in, T Represents the transpose of a matrix; 3) Perform eigenvalue decomposition on the covariance matrix Σ:

[0069] in, oh is the eigenvalue, I for p × p The identity matrix of Solving the equation yields p eigenvalues oh 1, oh 2, oh 3,…, oh p , for each eigenvalue, the corresponding eigenvector is obtained by solving the following linear equations:

[0070] in, oh i For the i eigenvalues, e i For the i The eigenvector corresponding to the eigenvalue; Arrange the eigenvalues ​​and corresponding eigenvectors in descending order, and select the first v The eigenvalues ​​corresponding to the eigenvectors are composed of n × v Matrix w T = [ e 1, e 2,…, e v ], project the standardized weighted fusion feature set matrix onto the selected front v On the principal components, we get a new feature matrix: F PCA = wT F s in, F PCA is the feature matrix after dimensionality reduction.

[0071] S5, model training: build a neural network, input the fused feature set into the neural network for training, and obtain a fault detection model; In this embodiment, a deep learning model is used to extract the feature matrix through multimodal feature fusion. F PCA Through deep learning, the model can automatically learn the spatiotemporal relationships extracted from the fused features of current, voltage, and temperature signals, further improving the accuracy and robustness of fault detection.

[0072] To meet the requirements of fault detection tasks, this paper adopts a deep learning architecture that combines convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). CNNs are used to extract local features from multimodal features, while LSTMs are used to process the temporal information of signals.

[0073] like Figure 3 As shown, this embodiment optimizes the design of the convolutional neural network (CNN) layer. Based on the frequency characteristics of the multimodal signal and the local characteristics of the fault characteristics, a convolution kernel of size 3×3 is selected. While effectively capturing local features, a convolution kernel of this size can reasonably balance the amount of computation and the range of the receptive field. In order to extract advanced local features more deeply, the number of convolution layers is expanded from a single layer to three layers. The number of convolution kernels in the first two layers is set to 16 and 32 respectively, and the number of convolution kernels in the last layer is 64. After each convolution operation, a batch normalization layer is added. Its main function is to normalize the input data to make the data distribution more stable, thereby accelerating the convergence speed of the model and significantly improving the stability of the training process.

[0074] The features extracted from the CNN layer are input to the LSTM layer. Based on the length and complexity of the signal's time series, the number of LSTM units is set to 128 to ensure the model can fully learn the long-term and short-term dependencies in the signal. To further enhance the ability to capture complex time series features, the number of LSTM layers is increased to two. Furthermore, a Dropout layer is added between the two LSTM layers, with a dropout probability of 0.2. The dropout layer randomly drops neurons during training, effectively preventing overfitting and improving the model's generalization ability.

[0075] After receiving the outputs of the CNN and LSTM, the fully connected layer performs feature fusion and classification. To enhance feature fusion, the number of hidden layer nodes is increased. The first hidden layer has 128 nodes, and the second hidden layer has 64 nodes. Both hidden layers use ReLU as the activation function. The output layer uses the Softmax activation function, outputting the probability distribution of samples belonging to each category.

[0076] The loss function of the fault detection model is as follows:

[0077] in, n for F PCA The number of samples in the feature matrix; y i For the i The feature values ​​correspond to the value of the fault label; p i For the fault detection model to predict i The failure probability of a sample; Furthermore, during the model training process, the backpropagation algorithm and gradient descent method are used to minimize the loss function. By continuously adjusting the weights in the network, the model learns how to predict faults based on the multimodal features of the input. During the model training process, the Adam optimizer is selected, and the initial learning rate is set to 0.001. After every 10 rounds of training, the learning rate is multiplied by 0.9.

[0078] The trained CNN-LSTM model can be used to predict new test data through forward propagation. The model will classify the input signal based on the fused features to determine whether there is a fault.

[0079] S6, model prediction: input the current signal, voltage signal and temperature signal of the AC port of the power router monitored in real time into the fault detection model and output the fault detection result; The fault detection model outputs the probability of a fault occurring. If the probability exceeds a set threshold, a fault is determined to have occurred. In actual applications, the fault detection model will monitor the working status of the power router in real time. Once a fault is detected, the system will automatically provide feedback and activate the corresponding alarm mechanism, prompting operation and maintenance personnel to inspect and repair the power router.

[0080] Step S7, real-time feedback and optimization: Through real-time monitoring and feedback mechanisms, the performance of the fault detection model is continuously optimized to improve the system's accuracy, robustness, and real-time responsiveness. This optimization process involves adjusting model parameters through real-time feedback, optimizing feature extraction, and continuously updating the model based on new data to accommodate changes in the power router's operating environment.

[0081] By monitoring the AC port operating data of the power router in real time, the system performs real-time fault detection based on this data and provides real-time feedback to the central control platform. Based on this feedback, the fault detection system can self-optimize the deep learning model to ensure high performance in a changing operating environment. The optimization process includes the following aspects: Model parameter update: The system continuously adjusts the model's hyperparameters based on feedback data to improve the model's convergence speed and accuracy. The system performs online learning based on newly acquired signal data, allowing updates and adjustments even for data not seen during the training phase. Using incremental learning methods, the system can gradually adapt to new data patterns, improving its ability to predict future faults. As the system operates over time, fault patterns may change, requiring the fault detection model to be highly adaptable. Through continuous feedback and data accumulation, the system continuously improves the accuracy and robustness of fault detection.

[0082] In one embodiment, the dataset contains data collected under five different working conditions, with 400 sets collected under each working condition, 200 sets under normal conditions and 200 sets under fault conditions, totaling 10 labels. The data is divided into a training set and a test set in a ratio of 9:1. The training set is used for model training and parameter adjustment, and the test set is used for testing. Figure 4 This is the transformation diagram of the loss function and accuracy of the model's training set and test set as the number of iterations increases. Figure 5 The following is a confusion matrix diagram for the 10 data sets shown. The ordinate represents the actual labels, the abscissa represents the predicted labels, the diagonal values ​​represent the number of correct predictions, and the off-diagonal values ​​represent the number of incorrect predictions. The closer the diagonal values ​​are to 20, the more correct predictions there are, indicating a higher model accuracy. As shown in the figure, there were a small number of errors in the 3, 4, 7, and 9 categories of labels. 193 data sets were correctly identified, while 7 were incorrectly identified, resulting in an accuracy rate of 96.5%. This demonstrates the feasibility and effectiveness of the proposed method for diagnosing AC port faults in power routers.

[0083] A second aspect of the present invention provides an AC port fault detection system for an energy router based on multimodal feature fusion, based on the fault detection method, comprising: Signal acquisition module: used to collect historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals and temperature signals under normal operation and fault operation; Preprocessing module: used to perform denoising and standardization on the collected multimodal signal data; Feature extraction module: used to extract time domain features and frequency domain features from multimodal signals (current, voltage, temperature); Feature fusion module: used to perform weighted fusion and dimensionality reduction on multimodal feature information to obtain the fused feature set; Fault detection module: used to build a CNN-LSTM neural network, input the fused feature set into the neural network for training to obtain a fault detection model; input the real-time operating data of the power router AC port into the fault detection model, and output the fault detection results.

[0084] The fault detection method of this invention is applicable to power routers and other power equipment requiring fault detection. Its intelligent, real-time feedback, and multimodal fusion capabilities enable widespread application in smart grids, power automation, distributed energy systems, and other fields, ensuring the stability and security of power systems.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting AC port faults in power routers based on multimodal feature fusion, characterized in that: The following steps are involved: S1, multi-modal signal acquisition: Acquire historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals, and temperature signals under normal operation and fault operation; S2, data preprocessing: denoising and standardization of the collected multimodal signal data; S3, feature extraction: using time domain feature extraction and frequency domain feature extraction methods to extract feature information from multimodal signals; S4, multimodal fusion: weighted fusion of multimodal feature information to obtain the fused feature set; S5, model training: build a neural network, input the fused feature set into the neural network for training, and obtain a fault detection model; S6, model prediction: input the current signal, voltage signal and temperature signal of the AC port of the power router monitored in real time into the fault detection model and output the fault detection result.

2. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 1, characterized in that: Step S3 includes the following steps: S301, using the EEMD method, decompose different modal signals separately to obtain different IMF components; S302, using time domain feature extraction and frequency domain feature extraction methods to extract time domain and frequency domain feature information of different IMF components respectively to obtain different feature sets; S303: Use correlation analysis to screen the extracted features, remove redundant features, and retain features that are useful for fault diagnosis.

3. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 2, wherein: Step S301 includes: Generate white noise using a pseudorandom number generator: η ( t )= σ × randn (1, L ) in, η ( t ) is white noise; σ is the set white noise standard deviation, randn (1, L ) means generating 1× L A random matrix, each element in the random matrix obeys the standard normal distribution; Superimpose white noise on the original signal to break the potential spectrum aliasing in the signal. The formula is as follows: x noisy ( t )= x ( t )+ η ( t ) in, x ( t ) is the original signal, x noisy ( t ) is the signal after superimposing white noise; The EEMD method is used to decompose the noisy signal and obtain the IMF component. The formula is as follows: in, IMF m ( t ) is the m The IMF components obtained by the decomposition are: r m ( t ) is the residual term; After multiple decompositions, multiple different IMF components are obtained; the decomposed IMF components are averaged to form the final stable IMF output sequence: in, IMF j ( t ) is the j The final IMF component signal obtained after decomposing the modal signal is M is the total number of decompositions.

4. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 3, wherein: In step S302, The formulas for extracting time domain features of IMF components include: in, μ j is the mean, N is the total number of sampling points; in, σ j is the standard deviation; in, RMS j is the root mean square value; in, Skewness j is the skewness; in, Kurtosis j is the kurtosis; The formulas for extracting frequency domain features of IMF components include: in, x j ( f ) is the j-th harmonic component in the spectrum after Fourier transform; in, E j ( f ) is the spectrum energy; in, X THD is the total harmonic distortion rate, x 1( f ) is the fundamental component.

5. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 2, wherein: In step S303, the correlation analysis calculation formula is: in, f i For the i eigenvalues, y i For the i The feature values ​​correspond to the value of the fault label; μ f is the mean of the eigenvalues, μ y is the mean of the labels; r is the correlation coefficient; k is the total number of extracted eigenvalues; Select the correlation coefficient r Features with a value greater than 0.7 are retained.

6. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 1, wherein: In step S4, the multimodal fusion method includes: The weighted fusion method is used to fuse multimodal features. The weight of each modality is assigned according to its contribution to fault diagnosis. The weighted fusion formula is as follows: in, F p For the p The feature set of a modality, w p For the p The weight coefficient corresponding to the feature set of each mode; F weighted is the fused feature set; The weight coefficients are learned by the gradient descent method. The goal is to minimize the error of the fault detection model through the weighted fusion feature set. The loss function is: in, L ( w ) is the loss function, which represents the error of the fault detection model; k 1 is the total number of features in the feature set; λ is a regularization parameter used to prevent overfitting; y i For the i The feature values ​​correspond to the value of the fault label; Update weight coefficients by gradient descent w p , the gradient descent formula is: in, is the updated weight coefficient; is the weight coefficient before updating; η is the learning rate, which is used to control the step size of each update; is the partial derivative of the loss function with respect to the weight coefficient.

7. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 6, characterized in that: The principal component analysis method is used to reduce the dimension of the weighted fusion feature set matrix, which includes the following steps: Normalize the fused feature set matrix: in, F weighted is the feature set matrix after weighted fusion, for F weighted The mean of for F weighted variance; F s is the standardized feature set matrix; Calculate the covariance matrix Σ of the standardized feature set matrix: in, T Represents the transpose of a matrix; Perform eigenvalue decomposition on the covariance matrix Σ: in, ω is the eigenvalue, I for p × p The identity matrix of Solving the equation yields p eigenvalues ω 1, ω 2, ω 3,…, ω p , for each eigenvalue, the corresponding eigenvector is obtained by solving the following linear equations: in, ω i For the i eigenvalues, e i For the i The eigenvector corresponding to the eigenvalue; Arrange the eigenvalues ​​and corresponding eigenvectors in descending order, and select the first v The eigenvalues ​​corresponding to the eigenvectors are composed of n × v Matrix w T = [ e 1, e 2,…, e v ], project the standardized weighted fusion feature set matrix onto the selected front v On the principal components, we get a new feature matrix: F PCA = w T F s in, F PCA is the feature matrix after dimensionality reduction.

8. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 7, wherein: In step S5, the neural network includes an input layer, a multi-layer convolution layer, two LSTM layers, a fully connected layer, and an output layer connected in sequence. The input layer is used to input the weighted fused feature set, and the multi-layer convolution layer is used to extract local features from the multimodal feature set. The LSTM layer is used to process the timing information of the signal. A Dropout layer is added between the two LSTM layers to prevent model overfitting. The fully connected layer is used to perform feature fusion and classification on the output of the LSTM layer. The output layer uses a Softmax activation function to output the probability distribution of each fault detection category. The loss function of the fault detection model is as follows: in, n is the number of samples in the feature set after weighted fusion; y i For the i The feature values ​​correspond to the value of the fault label; p i For the fault detection model to predict i The failure probability of a sample; During model training, the backpropagation algorithm and gradient descent method are used to minimize the loss function. By continuously adjusting the weights in the network, the model learns how to predict faults based on the multimodal features of the input. During model training, the Adam optimizer is selected, and the initial learning rate is set to 0.

001. After every 10 training rounds, the learning rate is multiplied by 0.

9.

9. The method for detecting AC port faults of an electric energy router based on multimodal feature fusion according to claim 1, wherein: The fault detection method further includes: Step S7, real-time feedback and optimization: By real-time monitoring of the operating data of the AC port of the power router, real-time fault detection is performed based on the current operating data, and the fault detection results are fed back to the central control platform in real time; the central control platform adjusts and optimizes the hyperparameters of the fault detection model based on the feedback results, and retrains the fault detection model using the newly collected operating data.

10. A power router AC port fault detection system based on multimodal feature fusion, based on the fault detection method according to any one of claims 1 to 9, characterized in that: include: Signal acquisition module: used to collect historical operating data of the AC port of the power router, including three-phase current signals, three-phase voltage signals and temperature signals under normal operation and fault operation; Preprocessing module: used to perform denoising and standardization on the collected multimodal signal data; Feature extraction module: used to extract time domain features and frequency domain features from multimodal signals; Feature fusion module: used to perform weighted fusion of multimodal feature information to obtain a fused feature set; Fault detection module: used to build a neural network, input the fused feature set into the neural network for training, and obtain a fault detection model; input the real-time collected operating data of the AC port of the power router into the fault detection model, and output the fault detection results.

Images