Electric energy meter fault fusion detection method based on VMD-GRU

Abstract

The invention discloses an electric energy meter fault fusion detection method based on VMD-GRU, and the method comprises the following steps: data collection, data preprocessing, VMD feature extraction, IVF-PQ modal importance analysis, and CNN-GRU prediction. VMD (variational mode decomposition) and IVF-PQ (inverted file-product quantization) technologies are introduced, so that the recognition precision and reliability of the model on fault signals are further improved, the method has good performance on key indexes such as accuracy, accuracy, recall rate and F1 score, and efficient classification of normal signals and fault signals is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a VMD-GRU-based electric energy meter fault fusion detection method. Background Art

[0002] Smart energy meters, as an advanced terminal device for collecting electricity information, show the characteristics of large number of services and wide geographical distribution. Such devices play a vital role in the power system. They are not only responsible for collecting users' electricity information in real time and accurately, but also provide a solid foundation for the intelligent management of power supply.

[0003] In recent years, with the rapid development of electronic instrumentation technology and the continuous improvement of manufacturing processes, the functions of smart energy meters have become more diversified. However, this diversification also brings new challenges, especially when facing the frequent failures of emerging electronic components in the field and the traditional reliability prediction manual cannot provide effective information. If these faults can be found and solved in time, it is crucial for the operation of smart energy meters. Because the service life of smart energy meters is closely related to their rotation time, and the number of smart energy meters in my country is huge, the resource consumption involved in the rotation work is extremely considerable. If the potential failures of smart energy meters can be quickly identified and predicted, so as to plan and implement effective rotation measures in advance, it will greatly save the operating costs of the power grid. In addition, the stable operation of smart energy meters is the cornerstone of the efficient operation of power systems. Therefore, the high reliability research on smart energy meters, especially in fault warning and diagnosis, not only has significant economic value, but also carries important social significance. Such research helps to ensure the stability of the power system, reduce service interruptions caused by faults, and thus ensure the normal operation of society. Summary of the invention

[0004] The purpose of the present invention is to provide an electric energy meter fault fusion detection method based on VMD-GRU to solve the problems of current electric energy meter fault detection technology, such as single data source, insufficient environmental adaptability and poor real-time performance of fault detection.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A VMD-GRU-based electric energy meter fault fusion detection method comprises the following steps:

[0007] Collect multi-source data of the running electric energy meter in real time, fuse the collected multi-source data into a unified information flow, extract fault features from the fused data, and construct a data set for electric energy meter fault diagnosis;

[0008] Perform data cleaning, linear normalization, and signal segmentation on the data set used for power meter fault diagnosis to obtain power meter operation data with reliable quality and rich features;

[0009] Perform VMD decomposition on the preprocessed data to extract various modes containing different frequency component characteristics;

[0010] Use inverted file-product quantization to compress each mode containing different frequency component characteristics, and perform importance analysis on the compressed modes;

[0011] The data after IVF-PQ modal importance analysis is input into the CNN-GRU model to predict the fault of the electric energy meter.

[0012] Furthermore, during the data collection process, the multi-source data includes calibration data, power consumption data, operating voltage, operating current, and environmental data of the electric energy meter during operation.

[0013] Furthermore, in the process of data preprocessing, the data cleaning adopts the missing value processing method, the linear normalization is to scale the signal data to a fixed range to eliminate the influence between different dimensions, and the signal segmentation is to segment the long-term continuous signal into short time segments of fixed length, and each fixed segment contains complete feature information.

[0014] Furthermore, in the VMD feature extraction process, VMD is optimized by the alternating direction multiplier method to minimize the width of each mode and separate different frequency components. The formula is expressed as:

[0015]

[0016] Among them, u k (t) is the kth mode, w k is the center frequency of the kth mode, δ(t) is the unit impulse function, j is the imaginary unit, * denotes the convolution operation, and ‖·‖2 denotes the L2 function.

[0017] Furthermore, in the VMD feature extraction process, the following steps are included:

[0018] Set the number of modes to K and initialize the center frequency of each mode and the Lagrange multiplier λ;

[0019] In each iteration, for each mode u k , update u by fixing the other modes and the Lagrange multiplier λ k , the formula is as follows:

[0020]

[0021] in, represents the Fourier transform of the kth mode at the n+1th iteration, is the Fourier transform of the original signal f, is the Fourier transform of the kth mode, α is the penalty parameter;

[0022] Update the center frequency of each mode. The formula is as follows:

[0023]

[0024] in, represents the center frequency of the kth mode at the n+1th iteration, represents the Fourier transform of the kth mode at the n+1th iteration;

[0025] Update the Lagrange multiplier λ to accelerate convergence. The formula is as follows:

[0026]

[0027] in, and is the variable related to the number of iterations;

[0028] The iteration stops and whether the convergence condition is met is determined. If so, the iteration stops.

[0029] Furthermore, the features extracted by VMD include energy features, mean features, variance features, center frequency and frequency domain features of each mode.

[0030] Furthermore, in the process of IVF-PQ modal importance analysis, modal compression first uses the inverted file method to index the decomposed modal data. The inverted file method divides the data into multiple clusters, each cluster contains similar modalities, and can quickly locate and access modalities with similar features. Then, product quantization is used within each cluster to compress the modal data, decompose the high-dimensional modal vector into low-dimensional sub-vectors, and each sub-vector is independently quantized and encoded.

[0031] Further, the importance analysis of the compressed modes includes the following steps:

[0032] The reconstruction error is calculated, and the clarity of the reconstructed signal is measured by the signal-to-noise ratio. The reconstruction error is used as a quality assessment indicator to evaluate the effectiveness of the compressed mode. The reconstruction error calculation formula is as follows:

[0033]

[0034] Among them, Reconstruction Error is the reconstructed mode With the original mode uk The error between

[0035] The signal-to-noise ratio calculation formula is as follows:

[0036]

[0037] Among them, SNR is the reconstructed mode With the original mode u k The signal-to-noise ratio between

[0038] The modal importance score is calculated by reconstructing the error and signal-to-noise ratio to calculate the contribution of each mode in fault detection. The modal importance score is defined as I k , the calculation formula is as follows,

[0039] I k =α·SNR k -β·Reconstruction Error k

[0040] Among them, α and β are weight parameters used to balance the contribution of SNR and reconstruction error;

[0041] Modal screening, setting threshold screening, scoring according to importance I k , a threshold is set, and the modes with scores higher than the threshold are retained as important modes, and the modes with scores lower than the threshold can be discarded to reduce redundant information and improve computational efficiency.

[0042] Furthermore, CNN-GRU prediction includes the following steps:

[0043] The modality analyzed by IVF-PQ modality importance is input into the input layer of CNN, local features are extracted through convolutional layer and pooling layer, and nonlinearity is introduced through activation function. The batch normalization layer is connected after the convolutional layer to accelerate model training and prevent gradient disappearance.

[0044] The multi-dimensional feature map after extracting local features is input into the flattening layer to flatten it into a one-dimensional vector;

[0045] The flattened one-dimensional vector is input into the GRU layer to capture the dynamic features in the time series, and then processed by the fully connected layer to extract the final high-level features;

[0046] Finally, the final prediction of fault detection is output through the output layer.

[0047] Beneficial effects:

[0048] The present invention further improves the recognition accuracy and reliability of the model for fault signals by introducing variational mode decomposition (VMD) and inverted file-product quantization (IVF-PQ) technology. The present invention performs well in key indicators such as accuracy, precision, recall and F1 score, and achieves efficient classification of normal signals and fault signals. The summary is as follows:

[0049] (1) Improved fault identification accuracy:

[0050] The present invention combines the feature extraction and time series modeling capabilities of CNN-GRU to effectively identify the fault signals of electric energy meters. In the experiment, an accuracy rate of 89% was achieved, and the precision and recall rates also reached a high level, which significantly improved the accuracy of fault detection.

[0051] (2) Data multi-source feature fusion:

[0052] The present invention decomposes the electric energy meter signal into multiple modes through VMD technology, and uses IVF-PQ to screen out the most important mode for analysis, so that the model can still maintain good detection effect when processing complex and noisy multi-source data, thereby improving the robustness of the system.

[0053] (3) Reduce false positives and false negatives:

[0054] The present invention achieves lower false positives and false negatives in fault detection, reduces the risk of false positives and missed negatives by balancing high precision and recall, makes the detection results more reliable, and facilitates accurate maintenance and troubleshooting in practical applications.

[0055] (4) Real-time and applicability of the model:

[0056] The combination of CNN and GRU in the present invention not only enables the model to efficiently process time series data, but also ensures computational efficiency, making the model have the potential for real-time detection. The method can adapt to a variety of working environments and different types of electric energy meters, providing technical support for the widespread deployment of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the overall flow chart of the present invention;

[0058] Figure 2 is a line graph of a normal signal and a fault signal in an embodiment of the present invention;

[0059] Figure 3 is a line graph after VMD modal decomposition in an embodiment of the present invention;

[0060] Figure 4 It is a bar chart of IVF-PQ importance analysis in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0062] Example

[0063] A VMD-GRU-based electric energy meter fault fusion detection method, such as Figure 1 As shown, the following steps are included:

[0064] S100, real-time acquisition of multi-source data of the electric energy meter in operation, data fusion processing of the acquired multi-source data, integration into a unified information flow, and then extraction of fault features in the fused data to construct a data set for electric energy meter fault diagnosis. The multi-source data includes calibration data, power consumption data, operating voltage, operating current, and environmental data of the electric energy meter during operation. These multi-source information can fully reflect the operating status of the electric energy meter and provide a reliable data basis for fault diagnosis.

[0065] During data collection, multi-source data of each electric energy meter is collected in real time through multi-sensor technology. The collected data includes multiple electric energy parameters of each electric energy meter in different time periods, which can effectively capture potential fault characteristics during equipment operation. For the acquired multi-source data, the consistency judgment method is used for data fusion processing, and the data from different sources are integrated into a unified information flow to ensure the reliability and consistency of the data. After data fusion, the fault features in the fused data are extracted to construct a data set for electric energy meter fault diagnosis. This data set covers the key parameters of the electric energy meter under various working conditions and can provide sufficient training and test data for machine learning algorithms.

[0066] S200 performs data cleaning, linear normalization, and signal segmentation on a data set used for electric energy meter fault diagnosis to obtain electric energy meter operation data with reliable quality and rich features.

[0067] The data cleaning includes missing value processing. During the data collection process, some data may be missing due to sensor failure or signal transmission problems, so missing value processing is required. Common missing value processing methods include:

[0068] Delete missing values. For a small number of missing values, you can directly delete the records containing missing values.

[0069] Missing value filling: For missing values ​​with a large number, interpolation method is used to fill them and restore the integrity of the data.

[0070] Noise processing,Signal data may be mixed with various noises, such as environmental noise, electromagnetic interference, etc.,The signal is decomposed into components of different scales using wavelet transform, retaining the main features of the signal and removing the noise components.

[0071] The linear normalization is to scale the signal data to a fixed range (such as 0 to 1 or -1 to 1) to eliminate the influence between different dimensions. The formula of linear normalization is as follows:

[0072]

[0073] Where X is the original data, Xmin and Xmax are the minimum and maximum values ​​of the data respectively.

[0074] The signal segmentation is to segment a long continuous signal into short time segments of fixed length, which is convenient for subsequent feature extraction and analysis, and each fixed segment contains complete feature information.

[0075] Specifically, the appropriate segment length and overlap rate are set to ensure the continuity of the signal and the integrity of the features. Generally speaking, the segment length should be determined according to the frequency component and change rate of the signal, and the overlap rate can be set to 50% to increase the amount of data and improve the stability of feature extraction.

[0076] S300, performing VMD decomposition on the preprocessed data to extract various modes containing different frequency component features.

[0077] Variational mode decomposition (VMD) is an adaptive signal decomposition technique that can decompose complex signals into several intrinsic mode functions (IMFs), each of which contains the characteristics of different frequency components. This method can effectively extract the main features of the signal, especially for non-stationary and nonlinear time series data. Through VMD, multi-modal decomposition of the signal is performed to extract key information from complex power meter data. This method can distinguish and isolate different frequency components in the power meter data, providing a more detailed data perspective for fault detection.

[0078] S400, uses inverted file-product quantization to compress each mode containing different frequency component characteristics, and performs importance analysis on the compressed modes.

[0079] In conventional fault detection, few systems consider the impact of data quality during compression and reconstruction. The present invention analyzes the reconstruction quality of the compressed mode and selects the mode that is most critical to fault detection to ensure the accuracy and effectiveness of fault detection.

[0080] S500 inputs the data after IVF-PQ modal importance analysis into the CNN-GRU model to predict the fault of the electric energy meter. The introduction of GRU can effectively capture the time dependency in the modal data, accurately predict the dynamically changing state of the electric energy meter, and improve the real-time and accuracy of fault diagnosis.

[0081] Preferably, in the VMD feature extraction process, VMD is optimized by the alternating direction multiplier method to minimize the width of each mode and separate different frequency components. The basic idea of ​​VMD is to decompose the original signal f(t) into K modes u k (t), where each mode has a center frequency w k , and as small as possible in bandwidth, the minimized objective function is as follows:

[0082]

[0083] Among them, u k (t) is the kth mode, w k is the center frequency of the kth mode, δ(t) is the unit impulse function, j is the imaginary unit, * denotes the convolution operation, and ‖·‖2 denotes the L2 function.

[0084] In the VMD feature extraction process, VMD mainly includes the following steps:

[0085] Initialize, set the number of modes to K, and initialize the center frequency of each mode and the Lagrange multiplier λ;

[0086] Modality update, in each iteration, for each modality u k , update u by fixing the other modes and the Lagrange multiplier λ k , the formula is as follows:

[0087]

[0088] in, represents the Fourier transform of the kth mode at the n+1th iteration, is the Fourier transform of the original signal f, is the Fourier transform of the kth mode, α is the penalty parameter;

[0089] Center frequency update: Update the center frequency of each mode. The formula is as follows:

[0090]

[0091] in, represents the center frequency of the kth mode at the n+1th iteration, represents the Fourier transform of the kth mode at the n+1th iteration;

[0092] The Lagrange multiplier is updated to accelerate convergence. The formula is as follows:

[0093]

[0094] in, and is the variable related to the number of iterations;

[0095] The iteration stops, and it is determined whether the convergence condition (for example, whether the update amplitude of each mode is less than the set threshold) is met. If so, the iteration is stopped.

[0096] Preferably, each mode u obtained by VMD decomposition k (t) Extract the following features:

[0097] Energy characteristics, the energy of each mode can be calculated by the following formula:

[0098] E k =∫|u k (t)| 2 dt

[0099] Mean feature, calculate the mean feature of each mode, reflecting the overall level of the mode, the formula is as follows:

[0100]

[0101] Variance feature, calculates the variance of each mode, reflects the change of modal signal, the formula is as follows:

[0102]

[0103] Center frequency, the center frequency of each mode w k It is part of the VMD algorithm and is directly used as one of the features.

[0104] Frequency domain features, perform Fourier transform on each mode and calculate spectrum features, such as peak frequency, bandwidth, etc., to further describe the frequency components of each mode.

[0105] Preferably, in the IVF-PQ modality importance analysis process, IVF index construction and PQ quantification are required.

[0106] IVF index construction: First, the inverted file (IVF) method is used to index the decomposed modal data. IVF divides the data into multiple clusters, each of which contains similar modalities, so that modalities with similar features can be quickly located and accessed.

[0107] PQ quantization: Product Quantization (PQ) is used within each cluster to compress modal data, decompose high-dimensional modal vectors into low-dimensional sub-vectors, and independently quantize and encode each sub-vector. This can significantly reduce storage and computing costs while ensuring data characteristics.

[0108] Preferably, performing importance analysis on the compressed modes comprises the following steps:

[0109] The reconstruction error is calculated, and the clarity of the reconstructed signal is measured by the signal-to-noise ratio. The reconstruction error is used as a quality assessment indicator to evaluate the effectiveness of the compressed mode. For each mode, the error between the reconstructed mode and the original mode is calculated by the following formula:

[0110]

[0111] Among them, Reconstruction Error is the reconstructed mode With the original mode u k The smaller the reconstruction error, the more original information the compressed mode retains and the greater its contribution to fault diagnosis.

[0112] Signal-to-noise ratio is another commonly used quality assessment indicator, which is used to measure the clarity of the reconstructed signal. The signal-to-noise ratio can be calculated by the following formula:

[0113]

[0114] Among them, SNR is the reconstructed mode With the original mode u k The higher the SNR, the better the quality of the reconstructed signal and the greater its contribution to fault detection.

[0115] The modal importance score is calculated by reconstructing the error and signal-to-noise ratio to calculate the contribution of each mode in fault detection. The modal importance score is defined as I k , the calculation formula is as follows,

[0116] I k =α·SNR k -β·Reconstruction Error k

[0117] Among them, α and β are weight parameters used to balance the contribution of SNR and reconstruction error. The higher the contribution, the more important the mode is in fault detection.

[0118] Modal screening, setting threshold screening, scoring according to importance I k , a threshold is set, and the modes with scores higher than the threshold are retained as important modes, and the modes with scores lower than the threshold can be discarded to reduce redundant information and improve computational efficiency.

[0119] Preferably, CNN-GRU prediction includes the following steps,

[0120] The modality analyzed by IVF-PQ modality importance is input into the input layer of CNN, local features are extracted through convolutional layers and pooling layers, and nonlinearity is introduced through activation functions. The batch normalization layer is followed by the convolutional layer to accelerate model training and prevent gradient disappearance.

[0121] The multi-dimensional feature map after extracting local features is input into the flattening layer to flatten it into a one-dimensional vector.

[0122] The flattened one-dimensional vector is input into the GRU layer to capture the dynamic features in the time series, and then processed by the fully connected layer to extract the final high-level features.

[0123] Finally, the final prediction of fault detection is output through the output layer.

[0124] The specific process of CNN-GRU prediction is shown in Table 1.

[0125] Table 1

[0126]

[0127]

[0128] Input layer: The shape of the input data is usually (batch_size, seq_len, features), where seq_len represents the length of the time series and features represents the number of features at each time step.

[0129] Convolutional layer: The first two convolutional layers extract spatial features through one-dimensional convolution operations, contain 64 and 128 filters respectively, and use ReLU as the activation function to ensure the nonlinear expression ability of the model.

[0130] Batch Normalization Layer: Each convolutional layer is followed by a batch normalization layer to speed up model training and prevent gradient disappearance.

[0131] Pooling layer: A pooling layer (MaxPooling1D) is added after each convolutional layer to reduce the size of the feature map, thereby reducing the amount of computation required by the model and improving the model’s robustness to small changes.

[0132] Flattening layer: The data output by the pooling layer is flattened into a one-dimensional vector to provide input for the GRU layer.

[0133] GRU layer: The GRU layer is used to capture dynamic features in time series. This layer is able to remember long-term temporal dependencies and effectively process long sequence data.

[0134] Fully connected layer: After passing through the GRU layer, the data is passed into a fully connected layer to extract the final high-level features.

[0135] Output layer: The number of neurons in the output layer is the number of categories (num_classes). Depending on the task, softmax (for multi-classification) or sigmoid (for binary classification) can be selected as the activation function to generate the final prediction for fault detection.

[0136] In this embodiment, in order to verify the effectiveness of the energy meter fault diagnosis method, a PZ72-E4 / C AC energy meter of a power company is selected as the research object. This type of energy meter can measure the current, voltage, power and other energy parameters of different users in the power system through AC sampling technology, and realize line connection through the RS485 communication interface. It has multiple functions such as input switching quantity, output switching quantity and analog data quantity conversion, which is very suitable for the research of energy meter fault diagnosis.

[0137] In this experiment, 100 PZ72-E4 / C AC energy meters were set up for data collection. These energy meters were installed in a specific power supply area and were responsible for collecting energy data in that area. Finally, 200 samples were collected, and the multi-source fusion information was analyzed using the method proposed above. This was used as the basis for energy meter fault diagnosis and the fault diagnosis performance of the PZ72-E4 / C energy meter was evaluated. Figure 1 As shown in the figure, the difference between the normal signal and the fault signal characteristics during the experiment is shown. Figure 2 As shown in the figure, the results of VMD mode decomposition during the experiment are shown, such as Figure 3 As shown, the results of IVF-PQ importance analysis during the experiment are shown, and from the results it can be seen that mode 1 is the most important, and finally mode 1 is used for classification.

[0138] Based on the experimental results of these 200 data, the confusion matrix shown in Table 2 can be constructed:

[0139] Table 2

[0140] Prediction Normal(0) Predicting anomalies (1) Actual normal (0) 90 10 Actual exception (1) 12 88

[0141] The experimental results are shown in Table 3:

[0142] Table 3

[0143]

[0144]

[0145] This experiment shows that the CNN-GRU based model can effectively detect the normal and fault signals of the electric energy meter, and performs well in key indicators such as accuracy, precision and F1 score, which can meet the needs of practical applications. Through further optimization, the detection performance of the model can be further improved, making it more practical and reliable in the fault detection task of the power system.

[0146] Although the embodiments of the present invention are described in the specification, these embodiments are only for reference and should not limit the protection scope of the present invention. Various omissions, substitutions and changes within the scope of the present invention should be included in the protection scope of the present invention.

Claims

1. A VMD-GRU-based electric energy meter fault fusion detection method, characterized by: The following steps are included: Collect multi-source data of the running electric energy meter in real time, fuse the collected multi-source data into a unified information flow, extract fault features from the fused data, and construct a data set for electric energy meter fault diagnosis; Perform data cleaning, linear normalization, and signal segmentation on the data set used for power meter fault diagnosis to obtain power meter operation data with reliable quality and rich features; Perform VMD decomposition on the preprocessed data to extract various modes containing different frequency component characteristics; Use inverted file-product quantization to compress each mode containing different frequency component characteristics, and perform importance analysis on the compressed modes; The data after IVF-PQ modal importance analysis is input into the CNN-GRU model to predict the fault of the electric energy meter.

2. The electric energy meter fault fusion detection method based on VMD-GRU according to claim 1 is characterized in that: During the data collection process, the multi-source data includes the calibration data, power consumption data, operating voltage, operating current, and environmental data of the electric energy meter during operation.

3. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: During the data preprocessing process, the data cleaning adopts the missing value processing method, the linear normalization is to scale the signal data to a fixed range to eliminate the influence between different dimensions, and the signal segmentation is to segment the long-term continuous signal into short time segments of fixed length, and each fixed segment contains complete feature information.

4. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: In the VMD feature extraction process, VMD is optimized by the alternating direction multiplier method to minimize the width of each mode and separate different frequency components. The formula is expressed as: Among them, u k (t) is the kth mode, w k is the center frequency of the kth mode, δ(t) is the unit impulse function, j is the imaginary unit, * denotes the convolution operation, and ‖·‖2 denotes the L2 function.

5. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: The VMD feature extraction process includes the following steps: Set the number of modes to K and initialize the center frequency of each mode and the Lagrange multiplier λ; In each iteration, for each mode u k , update u by fixing the other modes and the Lagrange multiplier λ k , the formula is as follows: in, represents the Fourier transform of the kth mode at the n+1th iteration, is the Fourier transform of the original signal f, is the Fourier transform of the kth mode, α is the penalty parameter; Update the center frequency of each mode. The formula is as follows: in, represents the center frequency of the kth mode at the n+1th iteration, represents the Fourier transform of the kth mode at the n+1th iteration; Update the Lagrange multiplier λ to accelerate convergence. The formula is as follows: in, and is the variable related to the number of iterations; The iteration stops and whether the convergence condition is met is determined. If so, the iteration stops.

6. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: The features extracted by VMD include energy features, mean features, variance features, center frequency and frequency domain features of each mode.

7. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: In the process of IVF-PQ modal importance analysis, modal compression first uses the inverted file method to index the decomposed modal data. The inverted file method divides the data into multiple clusters, each cluster contains similar modalities, and can quickly locate and access modalities with similar features. Then, product quantization is used within each cluster to compress the modal data, decompose the high-dimensional modal vector into low-dimensional sub-vectors, and each sub-vector is independently quantized and encoded.

8. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: The importance analysis of the compressed modes includes the following steps: The reconstruction error is calculated, and the clarity of the reconstructed signal is measured by the signal-to-noise ratio. The reconstruction error is used as a quality assessment indicator to evaluate the effectiveness of the compressed mode. The reconstruction error calculation formula is as follows: Among them, Reconstruction Error is the reconstructed mode With the original mode u k The error between The signal-to-noise ratio calculation formula is as follows: Among them, SNR is the reconstructed mode With the original mode u k The signal-to-noise ratio between The modal importance score is calculated by reconstructing the error and signal-to-noise ratio to calculate the contribution of each mode in fault detection. The modal importance score is defined as I k , the calculation formula is as follows, I k =α·SNR k -b·Reconstruction Error k Among them, α and β are weight parameters used to balance the contribution of SNR and reconstruction error; Modal screening, setting threshold screening, scoring according to importance I k , a threshold is set, and the modes with scores higher than the threshold are retained as important modes, and the modes with scores lower than the threshold can be discarded to reduce redundant information and improve computational efficiency.

9. The VMD-GRU-based electric energy meter fault fusion detection method according to claim 1 is characterized in that: CNN-GRU prediction includes the following steps, The modality analyzed by IVF-PQ modality importance is input into the input layer of CNN, local features are extracted through convolutional layer and pooling layer, and nonlinearity is introduced through activation function. The batch normalization layer is connected after the convolutional layer to accelerate model training and prevent gradient disappearance. The multi-dimensional feature map after extracting local features is input into the flattening layer to flatten it into a one-dimensional vector; The flattened one-dimensional vector is input into the GRU layer to capture the dynamic features in the time series, and then processed by the fully connected layer to extract the final high-level features; Finally, the final prediction of fault detection is output through the output layer.

Images