Operation state monitoring method of electric energy metering box, storage medium and equipment

Abstract

The invention discloses a running state monitoring method of an electric energy metering box, a storage medium and equipment, and relates to the technical field of intelligent power grid monitoring. The method comprises the following steps: performing multi-scale frequency domain decomposition on a multi-dimensional electrical signal to generate subsequences with different frequency characteristics; a dynamic window mechanism is adopted to adaptively segment signals and optimize a filtering strategy; extracting spatio-temporal joint features through multi-scale time sequence convolution and a physical constraint graph network; inputting the features into an attention prediction module integrated with a historical knowledge base for knowledge enhancement reasoning; and generating prediction uncertainty estimation based on Monte Carlo Dropout, calculating a composite abnormal index in combination with the distribution deviation degree, and triggering graded early warning. According to the method, accurate detection of transient abnormality and long-term trend is realized, the prediction accuracy is improved through spatio-temporal feature fusion and dynamic knowledge retrieval, the false alarm rate is reduced by using an adaptive threshold strategy, and a multi-dimensional and high-reliability state monitoring solution is provided for the electric energy metering box.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grid monitoring, and particularly relates to a method, storage medium, and device for monitoring the operating state of an electric energy metering box. Background Art

[0002] As a key device in the smart grid, the operating state of the electric energy metering box directly affects the accuracy of power metering and the safety of the power grid. Existing monitoring methods mostly adopt time-domain analysis with a fixed window or single-band feature extraction, making it difficult to effectively capture the composite fault modes of transient anomalies and long-term trends.

[0003] Traditional methods use a fixed window size during signal segmentation, resulting in the loss of high-frequency transient features or insufficient accuracy in low-frequency trend analysis; the feature extraction module mostly focuses on a single dimension of time series or spatial correlation, lacking the modeling of the dynamic coupling relationship between electrical parameters; the prediction model usually relies on static historical data and cannot perform knowledge-enhanced reasoning by combining real-time working conditions; anomaly detection mostly adopts a fixed threshold strategy, with insufficient estimation of uncertainty and prone to false alarms and missed detections.

[0004] In addition, existing technologies have problems such as high computational redundancy and poor real-time performance when dealing with multi-dimensional signals, making it difficult to meet the accurate monitoring requirements in complex power grid environments. Summary of the Invention

[0005] (I) Technical Problems to be Solved To solve the above problems, the present invention proposes a method, storage medium, and device for monitoring the operating state of an electric energy metering box, aiming to solve the problems in the existing state monitoring methods where a fixed window size is used during signal segmentation, resulting in the loss of high-frequency transient features or insufficient accuracy in low-frequency trend analysis; and the feature extraction module mostly focuses on a single dimension of time series or spatial correlation, lacking the modeling of the dynamic coupling relationship between electrical parameters.

[0006] (II) Technical Solutions A method for monitoring the operating state of an electric energy metering box according to the present invention includes: Performing multi-scale frequency-domain decomposition on the multi-dimensional electrical signals of the electric energy metering box to generate subsequences containing different frequency features; Adopting a dynamic window partitioning mechanism to adaptively segment the subsequences, and dynamically adjusting the window size and filtering strategy based on signal features; Performing spatio-temporal joint feature extraction on the segmented signal segments, including capturing local pattern features through multi-scale temporal convolution and constructing a graph network based on physical constraints to model the dynamic association features between parameters, to obtain a spatio-temporal feature set; Inputting the spatio-temporal feature set into an attention prediction module integrating a historical knowledge base, and generating a state prediction result through multi-dimensional interactive weight allocation and knowledge-enhanced reasoning; Generate uncertainty estimates of prediction results based on the Monte Carlo Dropout technique, calculate a composite anomaly index by combining the distribution deviation and historical benchmarks, and trigger a hierarchical early warning mechanism according to the index threshold.

[0007] In the present invention, the multi-scale frequency domain decomposition of the multi-dimensional electrical signals of the power metering box includes: Decompose the signal into high-frequency subsequences and low-frequency subsequences. The high-frequency subsequences are used for transient anomaly detection, and the low-frequency subsequences are used for long-term trend analysis; Automatically screen effective frequency bands according to the energy distribution characteristics of sub-bands.

[0008] In the present invention, the dynamic window partitioning mechanism includes: Dynamically adjust the window length according to the signal fluctuation degree. Shorten the window and enhance filtering when the signal changes violently, and expand the window and optimize the smoothing process in the steady state.

[0009] In the present invention, the spatio-temporal joint feature extraction includes: Construct parallel multi-scale time series feature extraction branches and physically constrained graph feature modeling branches to fuse local time series patterns and dynamic associations between parameters.

[0010] In the present invention, the attention prediction module includes: Introduce positional encoding to retain time series information, capture interaction patterns through multi-dimensional attention weight assignment, and perform knowledge-enhanced reasoning by combining with an external knowledge base.

[0011] In the present invention, the calculation of the composite anomaly index includes: Based on the results of multiple Monte Carlo Dropout forward propagations, infer the statistical prediction distribution of the results, and generate a dynamic anomaly index by combining the standardized value of the prediction deviation and the deviation degree from the historical distribution. The dynamic anomaly index is defined as: , where represents the predicted value of the model at time point , represents the actual observed value at time point , is the standard deviation of the predicted value, is the mean of the current prediction distribution, is the mean of the historical normal distribution, is the standard deviation of the historical normal distribution, represents the Kullback-Leibler divergence, which measures the deviation degree between the current prediction distribution and the historical normal distribution , are the weights of the two terms of the expression respectively, and ; Trigger a hierarchical response mechanism according to the comparison result between the exponent and the adaptive threshold.

[0012] In the present invention, the dynamic threshold is defined as: , where is the Shannon entropy of the current predicted distribution, is the maximum entropy value, is a preset multiple coefficient, is an entropy value correction factor.

[0013] In the present invention, it further includes: Construct a two-layer optimization framework, and synchronously optimize the model parameters and hyperparameters through implicit differentiation technology to achieve a balance between training efficiency and model performance; Adopt a mixed data augmentation strategy in the training stage, and improve the generalization ability of the model for marginal working conditions through adversarial sample generation and synthetic data expansion.

[0014] Another computer-readable storage medium of the present invention, on which a computer program is stored, characterized in that when the program is executed by a processor, it implements the operation state monitoring method of the power metering box described in any one of the above technical solutions.

[0015] Another execution device of the present invention, including an execution device body and a controller, characterized in that the controller includes a processor and a computer program stored on a memory and operable on the processor, and when the processor executes the program, it implements the operation state monitoring method of the power metering box described in any one of the above technical solutions.

[0016] (III) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: In the present invention, through wavelet packet decomposition and adaptive frequency band screening, accurate separation of transient anomalies and long-term trends is achieved; the dynamic window mechanism combines signal fluctuation characteristics to adjust the segmentation strategy in real time, taking into account the capture of high-frequency events and the analysis efficiency of low-frequency signals.

[0017] In the present invention, multi-scale temporal convolutional and physically constrained graph networks extract local pattern and parameter correlation features in parallel, significantly improving the completeness of feature representation. The attention mechanism integrating the historical knowledge base enhances the model's reasoning ability for marginal working conditions through multi-dimensional weight distribution and dynamic retrieval. Description of the drawings

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a schematic diagram of the logical flow of the monitoring method; Figure 2 It is a schematic diagram of dynamic window division; Figure 3 It is a spectral schematic diagram of multi-scale frequency domain decomposition; Figure 4 It is a time-domain schematic diagram of dynamic window adjustment; Figure 5 It is a schematic diagram of Monte Carlo Dropout; Figure 6 It is a schematic diagram of the frame structure of the execution device.

[0020] 1. Processor, 2. Memory, 3. Communication interface, 4. Communication bus. Specific implementation manners

[0021] Example 1

[0022] Monte Carlo Dropout is a Bayesian approximation method based on deep learning. It estimates the uncertainty of the prediction results by randomly enabling the Dropout layer multiple times during the model inference stage. It is an extension of the traditional Dropout technique, which extends the randomness from the training stage to the inference stage and is used to quantify the confidence of the model in the input data.

[0023] As Figures 1 - 5 shown, a method for monitoring the operating state of an electric energy metering box includes the following steps: S100. Perform multi-scale frequency domain decomposition on the multi-dimensional electrical signals of the electric energy metering box to generate subsequences containing different frequency characteristics.

[0024] Preprocess the original voltage, current and other multivariate time series, including removing baseline drift, power frequency interference and sensor noise. Adopt the wavelet packet decomposition method, select the Daubechies-4 wavelet basis function to perform multi-scale frequency domain decomposition on the signal, and divide the original signal into subsequences of multiple frequency bands. Due to its compact support, orthogonality and symmetry, this basis function can effectively balance the time-frequency resolution and is especially suitable for extracting the transient characteristics of non-stationary electrical signals. Among them, the high-frequency subsequences are used to detect millisecond-level transient anomalies, and the low-frequency subsequences are used to analyze the hour-level gradual change trend. For example, define the 7th and 8th layers as high-frequency subsequences for detecting transient anomalies such as voltage spikes and harmonic distortions, and the 1st and 2nd layers as low-frequency subsequences for analyzing the gradual change trends such as equipment temperature rise and insulation aging.

[0025] Specifically, the present disclosure performs a three-layer wavelet packet decomposition on the multivariate time series, recursively dividing the original signal into subsequences in eight frequency bands. Among them, the first-layer decomposition divides the original signal into a low-frequency approximation component A1 and a high-frequency detail component D1; the second-layer decomposition performs a secondary decomposition on A1 and D1 respectively to generate A2, D2, AD2, and DD2; the third-layer decomposition further decomposes each component of the second layer, and finally obtains eight sub-bands S1 - S8. Among them, the high-frequency subsequences S7 - S8 correspond to the frequency range of , where is the sampling rate; the low-frequency subsequences S1 - S2 correspond to .

[0026] After obtaining the sub-bands, calculate the relative energy entropy of each sub-band: , where is the relative energy entropy of the th sub-band, which is used to measure the complexity or information content of the energy distribution within this band. The smaller the entropy value, the more concentrated the energy distribution in this band; the larger the entropy value, the more dispersed the energy distribution. represents the signal component of the th sub-band at the discrete point , usually obtained through wavelet packet decomposition, can represent a frequency point or a time point, depending on the signal analysis method. represents the energy value of the th sub-band at the point, represents the total energy of all eight sub-bands at the point , which is used to normalize the energy proportion of each sub-band, is used to quantify the uncertainty of information. When , it is determined that this band is a noise-dominated band and is automatically masked to reduce computational redundancy, where is the total energy.

[0027] Each subsequence is divided into segments of length 128 through a sliding window and normalized to zero mean and unit variance. The inverse wavelet packet transform is performed on the remaining valid sub-bands and reconstructed into a time series subsequence as the input value for subsequent feature extraction.

[0028] S200. Adopt a dynamic window division mechanism to adaptively segment the subsequence, and dynamically adjust the window size and filtering strategy based on signal characteristics.

[0029] As Figure 2 shown, where the black curve represents the signal fluctuation, the dashed box represents shortening the window during severe fluctuations, and the solid box represents expanding the window in a stable state.

[0030] Monitor the signal fluctuation characteristics in real time, construct multi-dimensional quantization indicators based on the variance, root mean square value, and local range of the signal within the window, and combine the dynamic threshold library of historical data to determine the degree of drastic change in the signal state. When the fluctuation index exceeds a preset multiple of the historical mean, trigger the window shortening mechanism, adjust the window length to the minimum size to capture high-frequency transient anomalies. For example, when detecting voltage spikes, the window can be shortened to millisecond-level resolution, and at the same time, enable a Butterworth bandpass filter to suppress low-frequency interference and retain nanosecond-level pulse characteristics.

[0031] Conversely, in the signal steady state stage, automatically expand the window to the maximum size to cover the long-term operation cycle, and use the Savitzky-Golay algorithm to smooth the low-frequency subsequence to extract gradual trends such as equipment aging. Introduce a progressive scaling strategy during the window adjustment process, gradually transition the window length through linear interpolation, avoid feature breakpoint problems caused by sudden changes, and ensure temporal continuity. The filtering strategy dynamically matches the window size. For example, configure the order and cut-off frequency of the high-pass filter in the short window mode, and optimize the low-pass filtering parameters in the long window mode to achieve a balance between noise suppression and feature retention.

[0032] In addition, by pre-computing historical statistics and using a lightweight recursive filtering algorithm, compress the single-window processing time to below the sampling interval to meet the real-time requirements. The window overlap mechanism dynamically adjusts the overlap ratio according to the signal change rate. For example, adopt 50% overlap in the fast fluctuation stage to capture the complete waveform, and reduce it to 20% in the steady state stage to reduce redundant calculations. Through the closed-loop linkage of signal features and processing strategies, this mechanism takes into account both the accuracy of high-frequency anomaly capture and the efficiency of long-term trend analysis, providing a dynamically adaptable technical basis for the multi-scale state monitoring of the power metering box.

[0033] S300. Perform spatio-temporal joint feature extraction on the segmented signal fragments, including capturing local pattern features through multi-scale temporal convolution and constructing a graph network based on physical constraints to model the dynamic correlation features between parameters, to obtain a spatio-temporal feature set.

[0034] In the spatio-temporal joint feature extraction stage, first construct a multi-scale temporal convolution network for the segmented signal fragments, and parallelly extract local pattern features through convolution kernels of different time scales. Specifically, use dilated convolution layers with different dilation rates to capture short-term high-frequency fluctuations (such as instantaneous current mutations), medium-term periodic changes (such as power fluctuation periods), and long-term trend features (such as equipment temperature rise curves) respectively. The outputs of each convolution layer are fused through channel concatenation and normalization operations to form a temporal feature vector with multi-scale perception ability, effectively covering the feature expressions from millisecond-level transient events to hour-level progressive faults.

[0035] Meanwhile, a dynamic graph convolutional network is constructed based on physical constraints to model the correlation relationships between electrical parameters. Parameters such as voltage, current, and power factor are used as nodes, and the initial edge weights are defined according to physical rules such as Kirchhoff's law and Ohm's law to represent the theoretical correlation strength between parameters. Further, the calculation of mutual information within a sliding window is introduced to dynamically correct the edge weights to reflect the statistical correlation under actual working conditions. For example, when there is a strong correlation between the current harmonic distortion rate and the temperature rise rate during a certain period, the edge weight between the corresponding nodes will be adaptively enhanced. The node features are iteratively updated through spectral graph convolution operations to generate a graph feature matrix reflecting the dynamic coupling relationships between parameters.

[0036] To fuse the temporal and graph features, a bidirectional feature interaction mechanism is designed: the multi-scale feature vectors output by temporal convolution are mapped to the graph node space as supplementary information for node attributes; at the same time, the correlation features extracted by graph convolution are back-projected to the temporal dimension as the attention guiding signal for temporal convolution. Through cross-domain feature cross-enhancement, a spatio-temporal feature set is formed, which contains both the detailed information of local temporal patterns and the semantic constraints of global parameter correlations. Finally, the spatio-temporal feature set undergoes residual connection and layer normalization processing to eliminate the feature scale differences and provide a highly discriminative input representation for the subsequent prediction module.

[0037] S400. Input the spatio-temporal feature set into the attention prediction module integrating the historical knowledge base, and generate the state prediction result through multi-dimensional interactive weight allocation and knowledge-enhanced reasoning.

[0038] When inputting the spatio-temporal feature set into the attention prediction module integrating the historical knowledge base, first perform multi-modal fusion processing on the input features. A learnable 64-dimensional position encoding vector is used to embed the temporal information. The encoding matrix is initialized by a sine function, and the encoding weights are dynamically optimized during the training process, enabling the model to adaptively capture the sequence dependencies at different time steps.

[0039] The encoded feature vector and the original spatio-temporal features are concatenated along the channel dimension to form an enhanced input representation. Subsequently, a gated multi-head spatio-temporal attention mechanism is constructed, with 8 independent attention heads. Each head divides the feature subspace for different physical dimensions such as voltage phase, harmonic distortion rate, and environmental temperature and humidity. The similarity between the query vector and the key vector is calculated through scaled dot product to generate an interactive weight matrix, focusing on the non-linear coupling effect between power mutation events and temperature rise rate. The gated network consists of two fully connected layers with a hidden layer dimension of 128. The activation function uses GeLU, and a Dropout layer is introduced to prevent overfitting. The gated output is the correlation score between the current feature and the historical state of the external memory bank. The Top-K selection strategy is used to retain the top 10% of the high-weight features, and low-confidence noise signals are filtered through a soft threshold.

[0040] The external knowledge base adopts a distributed key-value storage architecture. The key is a 128-dimensional feature embedding vector, which is reduced to 32 dimensions through principal component analysis (PCA) and then a hierarchical navigable small world index is constructed. The value range includes fault type labels, disposal strategies, associated operating condition parameters, and historical occurrence frequencies. Using the IVF-PQ index strategy of the FAISS library, the high-dimensional vector space is divided into 1024 inverted lists, and each subspace is compressed to 8 bytes through product quantization to achieve fast approximate nearest neighbor search. The cosine similarity matching threshold is set to 0.85. When the similarity between the feature vector of real-time data and the knowledge base entry exceeds the threshold, the knowledge-enhanced reasoning mode is triggered. The matching results are dynamically fused through attention weights, and the specific formula is: where is the model prediction confidence,[[]]END is the normalized weight of the matching entry,[[]]END is the similarity score,[[]]END is the temperature coefficient, defaulting to 0.1,[[]]END is the historical fault disposal suggestion. The fused prediction result is mapped to the output space through a fully connected layer, and the prediction mean and standard deviation are generated synchronously.[[]]END

[0041] The knowledge base adopts a dynamic update mechanism. For newly added fault modes, two-stage verification is required: First, calculate the maximum cosine similarity with existing entries. If it is lower than 0.7 and confirmed by domain experts, the incremental learning process is triggered. During the update process, the parameters of the feature embedding layer are constrained by the elastic weight consolidation (EWC) algorithm to prevent catastrophic forgetting. The knowledge base capacity is maintained through a sliding window, retaining the most recent 5000 high-frequency fault records, and old entries are eliminated according to the LRU strategy. To ensure real-time performance, knowledge retrieval and model inference share the GPU video memory pool, while knowledge update tasks are asynchronously executed in a dedicated CPU thread pool, and a double-buffer queue is used to avoid read-write conflicts.[[]]END

[0042] In addition, the model synchronously generates a diagnostic report during the reasoning process, recording the JAI index of abnormal events, matching historical cases, and the weight distribution of disposal suggestions. The report displays the temperature rise curve, harmonic spectrum, and heat map of associated parameters through a visualization interface to assist maintenance personnel in quickly locating latent faults such as insulation aging and poor contact.[[]]END

[0043] S500. Generate uncertainty estimates of the prediction results based on the Monte Carlo dropout technique, calculate a composite anomaly index by combining the distribution deviation and historical benchmarks, and trigger a hierarchical warning mechanism according to the index threshold.[[]]END

[0044] ​​​​​​​​In the specific implementation of the Monte Carlo Dropout technique, the Dropout layer is mainly applied to the fully connected layer and the last convolutional layer of the neural network, with a dropout rate of 20%. To improve the computational efficiency, the GPU parallel computing technology is adopted to batch process 50 forward propagation processes. The single-inference time is reduced from 120 milliseconds in serial computing to 18 milliseconds. The Gaussian kernel density estimate selects the Epanechnikov kernel function, and the bandwidth parameter is adaptively determined by the Silverman rule. The formula is , where is the sample standard deviation, is the number of forward propagation times.

[0045] When generating the uncertainty estimate of the prediction result based on the Monte Carlo Dropout technique, first, the activation state of the Dropout layer is retained during the model inference stage, and a set of prediction results is generated through 50 random forward propagations. In each forward propagation, 20% of the neurons in the neural network are randomly discarded to obtain the probability distribution of the predicted values. The Gaussian kernel density estimate is performed on multiple prediction results to calculate the predicted mean and the standard deviation , where directly reflects the prediction uncertainty of the model at the current time point . At the same time, based on the sliding window, the prediction data under normal conditions in the past 720 hours are statistically analyzed to establish the historical baseline distribution , which is used as a reference for judging anomalies.

[0046] The calculation of the composite anomaly index integrates both the instantaneous deviation and the long-term distribution shift factors. Define the joint anomaly index: , where the first term quantifies the instantaneous anomaly intensity through the standardized prediction deviation, and the second term uses the KL divergence to measure the deviation between the current prediction distribution and the historical baseline. represents the predicted value of the model at the time point , reflecting the prediction result of the power metering box status. represents the actual observed value at the time point , that is, the real measurement data such as voltage and current. is the standard deviation of the predicted value, characterizing the prediction uncertainty of the model at the time point . The larger the value, the less reliable the prediction. is the mean of the current prediction distribution, representing the prediction central tendency of the model. is the mean of the historical normal distribution, calculated based on the past normal condition data. is the standard deviation of the historical normal distribution, reflecting the data fluctuation range under normal conditions. represents the Kullback-Leibler divergence, measuring the current prediction distribution​ The degree of deviation from the historical normal distribution The larger the value, the more significant the deviation of the current state from the normal mode. are the weights of the two terms of the expression, and .

[0047] The establishment of the historical benchmark distribution adopts a dynamic sliding window mechanism with a window length of 720 hours, which slides and updates once per hour. The data within the window is filtered for anomalies, excluding marked fault periods such as lightning strike records and manual maintenance periods. And it is dynamically updated through the exponentially weighted moving average (EWMA) and , with the weight decay coefficient set to 0.99 to balance the contributions of historical data and new data. The KL divergence calculation adopts a piecewise approximation strategy, discretizing the continuous distribution into 100 intervals, and statistically calculating the probability density through histograms to reduce the real-time calculation complexity while ensuring a low error rate.

[0048] In this disclosure, the weight coefficients and are determined through grid search to balance the detection sensitivity of short-term mutation and long-term gradual change faults. The dynamic threshold is set to , where is the Shannon entropy of the current prediction distribution, and the calculation formula is , which is used to quantify the uncertainty of the prediction result. The larger the entropy value, the higher the uncertainty. Usually take , where is the number of prediction categories, which is used to normalize the entropy value range. For example take (corresponding to 5 typical fault modes), is the entropy value correction factor, and in this disclosure . The entropy value correction term makes the threshold adaptively relax as the uncertainty of the model increases, reducing false alarms caused by noise interference.

[0049] The design of the entropy value correction term is based on the information entropy theory, calculated through the probability distribution of the prediction categories. Assuming that the model output is the probabilities of 5 types of faults, then .

[0050] When Trigger hierarchical warnings as follows: If the overlimit amplitude is within the range of 1-2 times the threshold, record an exception log and mark it for review; if the overlimit is 2-3 times, generate a work order and push it to the operation and maintenance system to start the equipment status review process; if the overlimit is more than 3 times, immediately cut off the high-risk circuit and activate the standby power supply, and simultaneously send an emergency alarm to the monitoring center. The mapping relationship between the warning level and the response strategy is dynamically configured through an expert experience database, supporting flexible adjustment according to the equipment type and operating environment. This mechanism ensures a high detection rate while significantly reducing the false alarm rate, which is significantly better than the fixed threshold scheme.

[0051] The hierarchical warning mechanism is deeply integrated with the operation and maintenance system. The work order generation module calls the work order interface of the operation and maintenance platform through the REST API, and the pushed information includes the exception timestamp, value, predicted fault type, and recommended handling measures such as insulation detection, harmonic analysis, etc. The emergency circuit cut-off communicates with the intelligent circuit breaker through the Modbus / TCP protocol, and the response delay is controlled within 50 milliseconds to ensure that the power-off operation is completed within 1 second after the third-level warning is triggered.

[0052] Embodiment 2

[0053] The embodiment of the present invention provides a computer-readable storage medium.

[0054] The computer program stored on the computer-readable storage medium provided by the embodiment of the present invention can implement the steps of any one of the above-mentioned operation status monitoring methods of the power metering box when executed by a processor.

[0055] The computer-readable storage medium may include: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0056] For the introduction of the computer-readable storage medium provided by the embodiment of the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

[0057] The embodiments in the specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0058] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of the examples have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0059] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of the two. The software modules can be placed in a random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0060] Embodiment 3

[0061] An embodiment of the present invention provides an execution device.

[0062] Please refer to Figure 6 , Figure 6 , which is a schematic structural diagram of an execution device provided by the present invention. The execution device may include: A memory for storing a computer program; A processor, when executing the computer program, can implement the steps of any one of the above-mentioned methods for monitoring the operating state of an electric energy metering box.

[0063] As Figure 6 shown, it is a schematic structural diagram of the composition of an execution device. The execution device may include: a processor 5, a memory 6, a communication interface 7, and a communication bus 8. The processor 5, the memory 6, and the communication interface 7 all complete communication with each other through the communication bus 8.

[0064] In the embodiment of the present invention, the processor 5 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field programmable gate array, or other programmable logic devices, etc.

[0065] The processor 5 can call the program stored in the memory 6. Specifically, the processor 5 can execute the operations in the embodiment of the button switch fault detection method.

[0066] The memory 6 is used to store one or more programs. The program may include program codes, and the program codes include computer operation instructions. In the embodiment of the present invention, the memory 6 stores at least a program for implementing the following functions: Perform multi-scale frequency domain decomposition on the multi-dimensional electrical signals of the electric energy metering box to generate subsequences containing different frequency characteristics; Adopt a dynamic window division mechanism to adaptively segment the subsequences, and dynamically adjust the window size and filtering strategy based on signal characteristics; Perform spatio-temporal joint feature extraction on the segmented signal segments, including capturing local pattern features through multi-scale temporal convolution, and constructing a graph network based on physical constraints to model the dynamic association features between parameters, to obtain a spatio-temporal feature set; Input the spatio-temporal feature set into the attention prediction module integrating the historical knowledge base, and generate a state prediction result through multi-dimensional interactive weight allocation and knowledge-enhanced reasoning; Generate uncertainty estimation of the prediction result based on the Monte Carlo Dropout technology, calculate a composite anomaly index by combining the distribution deviation degree and the historical benchmark, and trigger a hierarchical early warning mechanism according to the index threshold.

[0067] In a possible implementation manner, the memory 6 may include a program storage area and a data storage area. Among them, the program storage area may store an operating system, as well as application programs required for at least one function, etc.; the data storage area may store the data created during use.

[0068] In addition, the memory 6 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device or other volatile solid-state storage devices.

[0069] The communication interface 7 may be an interface of a communication module, and is used to connect to other devices or systems.

[0070] Of course, it should be noted that Figure 6 The structure shown does not constitute a limitation on the execution device in the embodiment of the present invention. In practical applications, the execution device may include more or fewer components than Figure 6 those shown, or combine some components.

[0071] The above-described embodiments are only used to describe the preferred implementation manners of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various variations and improvements made by those of ordinary skill in the art to the technical solution of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. A method for monitoring the operating state of an electric energy metering box, characterized in that, Including: Performing multi-scale frequency-domain decomposition on the multi-dimensional electrical signals of the electric energy metering box to generate subsequences containing different frequency characteristics; Adopting a dynamic window partitioning mechanism to adaptively segment the subsequences, and dynamically adjusting the window size and filtering strategy based on signal characteristics; Performing spatio-temporal joint feature extraction on the segmented signal segments, including capturing local pattern features through multi-scale temporal convolution and constructing a graph network based on physical constraints to model the dynamic correlation features between parameters, to obtain a spatio-temporal feature set; Inputting the spatio-temporal feature set into an attention prediction module integrating a historical knowledge base, and generating a state prediction result through multi-dimensional interactive weight assignment and knowledge-enhanced reasoning; Generating an uncertainty estimate of the prediction result based on the Monte Carlo Dropout technique, calculating a composite anomaly index by combining the distribution deviation degree and the historical benchmark, and triggering a hierarchical early warning mechanism according to the index threshold.

2. The operating state monitoring method of the electric energy metering box according to claim 1, characterized in that The multi-scale frequency-domain decomposition of the multi-dimensional electrical signals of the electric energy metering box includes: Decomposing the signal into a high-frequency subsequence and a low-frequency subsequence, where the high-frequency subsequence is used for transient anomaly detection and the low-frequency subsequence is used for long-term trend analysis; Automatically screening effective frequency bands according to the energy distribution characteristics of sub-bands.

3. The operating state monitoring method of the electric energy metering box according to claim 2, characterized in that, The dynamic window partitioning mechanism includes: Dynamically adjusting the window length according to the signal fluctuation degree, shortening the window and enhancing filtering when the signal changes violently, and expanding the window and optimizing the smoothing process in a stable state.

4. The method for monitoring the operating state of the electric energy metering box according to claim 3, characterized in that, The spatio-temporal joint feature extraction includes: Constructing a parallel multi-scale temporal feature extraction branch and a physically constrained graph feature modeling branch to fuse local temporal patterns and dynamic correlations between parameters.

5. The method for monitoring the operating state of the electric energy metering box according to claim 4, characterized in that, The attention prediction module includes: Introducing positional encoding to retain temporal information, capturing interaction patterns through multi-dimensional attention weight assignment, and performing knowledge-enhanced reasoning by combining with an external knowledge base.

6. The method for monitoring the operating state of the electric energy metering box according to claim 5, characterized in that, The calculation of the composite anomaly index includes: Based on the forward propagation results of multiple Monte Carlo Dropouts, infer the statistical prediction distribution of the results, and generate a dynamic anomaly index by combining the standardized value of the prediction deviation and the deviation degree from the historical distribution. The dynamic anomaly index is defined as: , where represents the predicted value of the model at time point , represents the actual observed value at time point , is the standard deviation of the predicted value, is the mean of the current prediction distribution, is the mean of the historical normal distribution, is the standard deviation of the historical normal distribution, represents the Kullback-Leibler divergence, which measures the deviation degree between the current prediction distribution and the historical normal distribution , are the weights of the two terms of the expression respectively, and ; Triggering a hierarchical response mechanism according to the comparison result between the exponent and the adaptive threshold.

7. The method for monitoring the operating state of the electric energy metering box according to claim 6, characterized in that, The dynamic threshold is defined as: , where is the Shannon entropy of the current prediction distribution, is the maximum entropy value, is the preset multiple coefficient, is the entropy value correction factor.

8. The method for monitoring the operating state of the electric energy metering box according to claim 7, wherein Also including: Constructing a two-layer optimization framework, synchronously optimizing model parameters and hyperparameters through implicit differentiation technology to achieve a balance between training efficiency and model performance; Adopting a hybrid data augmentation strategy in the training stage to improve the generalization ability of the model for marginal working conditions through adversarial sample generation and synthetic data expansion.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the method for monitoring the operating state of the electric energy metering box according to any one of claims 1-8.

10. An execution device, comprising an execution device body and a controller, characterized in that, The controller includes a processor and a computer program stored in a memory and executable on the processor. When the processor executes the program, it implements the method for monitoring the operating state of the electric energy metering box according to any one of claims 1-8.

Images