A method for anomaly diagnosis of power metering terminals based on multidimensional state information

Abstract

A method for anomaly diagnosis of power metering terminals based on multidimensional state information, relating to the field of power metering technology, enables online identification, classification, and maintenance guidance of the terminal's own operating status and abnormal user electricity consumption behavior. First, a real-time screening and triggering mechanism based on change point detection is established. Then, a hierarchical data acquisition and collaborative verification strategy is implemented. Differentiated data acquisition methods are adopted for terminals based on early warning results. Normal terminals maintain periodic acquisition of basic power parameters, while early warning terminals automatically trigger high-frequency sampling commands to obtain complete voltage and current waveform data. A two-layer multidimensional state information is defined, and a dual-branch hybrid deep learning model is designed. Bayesian optimization parameter tuning and decision support are introduced to automatically and efficiently optimize the key hyperparameters of the hybrid model globally. This invention not only achieves high-accuracy anomaly detection but also accurately classifies fault types and outputs a maintenance list sorted by probability.

Description

Technical Field

[0001] This invention relates to the field of electricity metering technology, specifically a method for diagnosing anomalies in electricity metering terminals based on multi-dimensional state information. Background Technology

[0002] With the rapid development of smart grids, the number and widespread distribution of electricity metering terminals have surged, posing a severe challenge to their operational reliability. Currently, maintenance work mainly relies on traditional manual periodic inspections and statistical spot checks. This approach is inefficient and costly when dealing with massive numbers of terminals, and it cannot achieve real-time status monitoring, leading to delayed fault detection. Faults are often only addressed after metering disputes or abnormal line losses occur. Existing remote diagnostic methods, such as simple data comparison or threshold alarms, can detect obvious faults like power outages, but their diagnostic accuracy and sensitivity are insufficient for complex and hidden performance degradation problems within the meter, such as metering chip gain drift, phase errors, and harmonic measurement inaccuracies. More importantly, existing methods often only indicate the existence of anomalies but cannot provide specific fault type identification, abnormal electricity consumption behavior classification, or a fault list ranked by probability. This leaves maintenance personnel without clear goals and directions when working on-site. Therefore, the operation and maintenance field urgently needs a solution to the core operation and maintenance challenge of how to accurately identify not only the faults of the meters themselves from a massive number of operating metering terminals, but also effectively detect abnormal electricity consumption behavior of users and provide a fault list sorted by probability, thus providing clear guidance for on-site operation and maintenance. This is precisely the core problem that this invention aims to solve. Summary of the Invention

[0003] The purpose of this invention is to provide an abnormal diagnosis method for power metering terminals based on multi-dimensional state information. It achieves wide-area screening by using lightweight change point detection at the front end, and performs accurate diagnosis at the back end by using a dual-branch deep learning model that integrates VI trajectory images and key physical features. The model performance is optimized by using Bayesian optimization technology, and finally realizes online identification, classification and operation and maintenance guidance for the terminal's own operating status and abnormal power consumption behavior of users.

[0004] The technical solution adopted by this invention to solve its technical problem is: a method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information, comprising the following steps:

[0005] S1, Real-time monitoring and anomaly warning across the entire area

[0006] S1.1, Constructing the basic data acquisition channel

[0007] The system continuously receives short sequences X of basic electrical parameters uploaded by all online terminals at fixed time intervals. Sequence X is: , (1); where, and Let D be a random variable, and let D be the change in the mean. For index functions, for The average value;

[0008] S1.2 Sequence Change Point Detection and Analysis

[0009] For each terminal sequence, the improved bilateral sliding window algorithm CUSUM is run. CUSUM dynamically calculates the cumulative sum statistic of the sequence, where the positive cumulative sum g0 is... + and statistic g k + for: (2); Negative accumulation g0 - and statistic g k - for: (3); where, The steady-state mean estimated by the sliding window; This refers to noise tolerance parameters;

[0010] S1.3 Early Warning Decision-Making and Triggering

[0011] When the cumulative statistics of any terminal continuously exceed the preset threshold, the terminal is marked as a suspect and a deep diagnostic trigger command is immediately generated.

[0012] S2, Intelligent Hierarchical Data Acquisition and Data Quality Control

[0013] S2.1, High-frequency data acquisition

[0014] A special command is sent to a terminal under suspicion to trigger it to enter a high-frequency sampling mode and collect complete instantaneous voltage and current waveforms.

[0015] S2.2 Multi-level data verification

[0016] A regional collaborative verification mechanism was initiated to conduct a horizontal comparative analysis of the data of the suspected terminal with other terminals in the same power supply area, and at the same time to conduct a vertical verification with the data of the meter at the next higher level to verify the power balance relationship and harmonic transmission characteristics.

[0017] S2.3 Data Standardization Preprocessing

[0018] A pipeline-style preprocessing method is used to preprocess the waveform data that has passed the verification.

[0019] S3. Extraction and Representation of Multidimensional State Information

[0020] S3.1, Image Processing of VI Trajectories

[0021] Generate VI trajectory images and perform coordinate mapping, pixelation, resolution setting, and morphological enhancement on the VI trajectory images;

[0022] S3.2 Feature Extraction of VI Trajectory

[0023] Extract geometric, spectral, and statistical features from the VI trajectory image;

[0024] S4, Hybrid Deep Learning Model Construction

[0025] S4.1 Image Feature Extraction Branch

[0026] The image feature extraction branch uses a convolutional neural network architecture to process the morphological features of the VI trajectory;

[0027] S4.2, Physical Feature Processing Branch

[0028] The physical feature processing branch takes a 15-dimensional key feature vector as input and performs hierarchical processing through three fully connected layers.

[0029] S4.3 Feature Fusion and Classification

[0030] The 256-dimensional feature vector from the image branch is concatenated with the 16-dimensional feature vector from the physical feature branch to form a comprehensive feature representation. The fused features are then subjected to deep semantic learning through a fully connected layer containing 128 neurons, and finally, a Softmax classifier outputs the probability distribution of each fault category.

[0031] S5, Model Training and Optimization

[0032] S5.1 Model Training Configuration

[0033] The cross-entropy loss function is chosen as the optimization objective. The cross-entropy loss function is: (4). In the formula, It is the sample size. It is the number of categories. It is a sample The true label, The model predicts the sample. Category The probability; the optimizer uses the Adam adaptive moment estimation algorithm, with an initial learning rate of 0.001, and the batch size is set to 32 or 64 depending on the computing resources. An early stopping mechanism is used during training, and training is automatically terminated when the validation set loss no longer decreases for 10 consecutive rounds, effectively preventing overfitting.

[0034] S5.2 Bayesian Hyperparameter Optimization

[0035] Bayesian hyperparameter optimization was performed on the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, and Dropout rate.

[0036] S5.3 Model Evaluation Indicators

[0037] Accuracy reflects the overall classification performance, precision measures the prediction accuracy of each category, recall evaluates the model's ability to identify positive examples, F1 score comprehensively balances the performance of precision and recall, and confusion matrix provides a detailed analysis of misclassification among categories, offering specific directions for model improvement.

[0038] Further, in step S1.3, the adaptive threshold (5); where, To determine the minimum detectable mutation magnitude, the minimum abnormal change in electrical parameters that the algorithm can identify is determined; The maximum number of delayed sampling points, The sampling interval is denoted as .

[0039] Furthermore, in step S2.3, a digital filter is used to eliminate high-frequency noise and power frequency interference, an interpolation algorithm is used to ensure strict synchronization of voltage and current data, and normalization processing is performed to eliminate the influence of dimensions.

[0040] Furthermore, in step S3.1, coordinate mapping involves using the synchronously sampled voltage sequence U(t) and current sequence I(t) as X-axis and Y-axis coordinates to draw a VI trajectory diagram on a two-dimensional plane that intuitively reflects the dynamic changes in impedance characteristics.

[0041] Furthermore, in step S3.1, pixelation and resolution setting convert the continuous trajectory map into a grayscale image of 256×256 pixels.

[0042] Furthermore, in step S3.1, morphological enhancement smooths the trajectory through an interpolation algorithm, reducing the jagged effect caused by sampling.

[0043] Further, in step S3.2, geometric features are extracted from the overall shape and spatial structure of the VI trajectory. These geometric features include the trajectory area, principal axis tilt angle, and eccentricity. (6); where, One cycle time; principal axis tilt angle (7); Eccentricity (8).

[0044] Further, in step S3.2, the spectral characteristics are the characteristic parameters obtained after converting the time-domain signal to the frequency domain through Fourier transform. The spectral characteristics include the total harmonic distortion (THD), harmonic power distribution, and spectral centroid. The total harmonic distortion of the current... (9); where, This is the effective value of the fundamental current. It is the first RMS value of subharmonic current It is the highest harmonic order considered.

[0045] Furthermore, in step S3.2, the statistical features are feature parameters extracted from voltage and current time series data through statistical analysis. The statistical features include waveform similarity, peak factor, fluctuation index, mean, and variance.

[0046] Further, in step S4.1, after the input layer receives a 256×256 pixel grayscale image, it sequentially passes through three convolution-pooling modules: the first convolution-pooling module uses 32 3×3 convolutional kernels for feature extraction, introduces non-linearity through the ReLU activation function, and then reduces dimensionality through a 2×2 max pooling layer; the second convolution-pooling module increases to 64 convolutional kernels to deepen feature learning; the third convolution-pooling module further expands to 128 convolutional kernels to capture more complex image patterns; finally, the feature map is converted into a feature vector through a global average pooling layer and connected to a fully connected layer containing 256 neurons, using a Dropout rate of 0.5 to prevent overfitting.

[0047] Furthermore, in step S4.2, the 64 neurons of the first fully connected layer perform nonlinear transformation on the original features, the 32 neurons of the second fully connected layer perform feature compression, and the 16 neurons of the third fully connected layer complete high-level feature encoding.

[0048] Furthermore, in step S5.2, the learning rate is optimized on a logarithmic scale within the interval [1e-5, 1e-2], the batch size is optimally selected from {16, 32, 64, 128}, the number of convolutional kernels is determined within the range of {16, 32, 64, 128}, the number of neurons in the fully connected layer is optimized within {64, 128, 256, 512}, and the Dropout rate is adjusted within the interval [0.1, 0.7]. The optimization process uses a Gaussian process as a surrogate model, with the expected improvement as the acquisition function, and achieves efficient global optimization through 100 iterations.

[0049] The beneficial effects of this invention are: it not only achieves high-accuracy anomaly detection, but also accurately classifies fault types and outputs a maintenance list sorted by probability, successfully transforming the traditional maintenance mode from passive response and blind inspection to proactive early warning and precise intervention, providing solid technical support for improving the efficiency of smart grid metering and maintenance, ensuring data accuracy, and combating electricity theft. Attached Figure Description

[0050] Figure 1 This is a technical roadmap of the present invention;

[0051] Figure 2 Here is a flowchart of the bilateral CUSUM change point detection algorithm;

[0052] Figure 3 For adaptive threshold The calculation principle diagram;

[0053] Figure 4 Branch graph for image feature extraction;

[0054] Figure 5 Extract branch graphs for physical features;

[0055] Figure 6 Sampling plots for different types of data;

[0056] Figure 7 The change point detection map shows three abnormal states;

[0057] Figure 8 The confusion matrix diagram for change point detection;

[0058] Figure 9 This is a diagram of the global system architecture. Detailed Implementation

[0059] This invention proposes a hierarchical anomaly diagnosis method integrating screening, diagnosis, and early warning. The core idea is to construct a hierarchical collaborative diagnostic system: a lightweight sequence change detection mechanism is deployed at the front end, forming a wide-area monitoring layer covering all terminals, responsible for real-time perception and initial anomaly screening of all terminal operational data. A resource-intensive multi-dimensional state information deep learning model is deployed at the back end as a precise diagnostic layer, performing deep feature analysis and state discrimination on the front-end early warning targets. The two are seamlessly connected through an event-driven intelligent triggering and response mechanism, jointly forming an efficient closed-loop diagnostic system. Figure 1As shown, the specific measures include: (1) Establishing a real-time screening and triggering mechanism based on change point detection: The sequence change point detection technology is applied to the monitoring of basic operation data of power metering terminals. An improved bilateral sliding window CUSUM algorithm is used to perform online analysis on the sequence of active power average of terminals collected in a short time period. When the CUSUM statistic of a terminal exceeds the preset threshold, the system automatically marks it as a suspected state and triggers the subsequent in-depth diagnosis process. (2) Implementing a hierarchical data collection and collaborative verification strategy: Differentiated data collection methods are adopted for terminals according to the early warning results. Normal terminals maintain periodic collection of basic power parameters, while early warning terminals automatically trigger high-frequency sampling commands to obtain complete voltage and current waveform data. At the same time, the quality of the collected data is verified through a regional collaborative verification mechanism to provide reliable data support for subsequent analysis. (3) Defining a two-layer multidimensional state information: The first layer is a two-dimensional image representation of the VI trajectory. By converting the time-series voltage and current data into a two-dimensional image, the morphological features of the trajectory are completely preserved. The second layer consists of key physical features extracted from the VI trajectory, covering geometric features, spectral features, and statistical features. These features have clear physical meanings, and together they constitute a multidimensional information set describing the terminal's operating state. (4) Design a dual-branch hybrid deep learning model: In view of the characteristics of multidimensional state information, the model is divided into two functional branches. One branch processes the VI trajectory image through a convolutional neural network to extract deep visual features; the other branch processes key physical features through a fully connected network to complete feature transformation. The features of the two branches are fused at a high level, and finally, the classifier is used to achieve accurate identification and classification of abnormal states. (5) Introduce Bayesian optimization parameter tuning and decision support: The Bayesian optimization framework is adopted to automatically and efficiently optimize the key hyperparameters of the hybrid model globally, replacing the traditional time-consuming manual parameter tuning, and ensuring that the model is always in the best performance state. The system finally outputs a list of fault probabilities arranged in descending order of diagnostic probability, as well as specific operation and maintenance suggestions, providing clear and direct action guidelines for on-site work.

[0060] The following is a detailed description of an anomaly diagnosis method for an energy metering terminal based on multidimensional state information according to the present invention, with reference to the accompanying drawings.

[0061] S1. Real-time monitoring and early warning of anomalies across the entire area.

[0062] This step aims to establish a real-time monitoring system covering all electricity metering terminals across the network, and to achieve early anomaly detection through lightweight algorithms.

[0063] S1.1, Construct basic data acquisition channels.

[0064] The system continuously receives short-sequence data of basic electrical parameters uploaded by all online terminals at fixed time intervals, with a focus on collecting the average active power sequence. The data sequence X is represented as: , (1). Among them, and Let D be a random variable; D represents the change in the mean. For index functions; for The average value; τ is the time when the change occurs.

[0065] S1.2 Sequence change point detection and analysis.

[0066] For the active power sequence of each terminal, the system runs the improved bilateral sliding window algorithm CUSUM in real time. For example... Figure 2 As shown, CUSUM sensitively detects changes in data distribution characteristics by dynamically calculating the cumulative sum and statistics of the sequences. The sliding window mechanism in CUSUM adaptively tracks natural fluctuations in load levels, while the two-sided detection design can simultaneously capture both positive and negative mean shifts. Specifically, the positive cumulative sum g0... + and statistic g k + : (2); Negative accumulation g0 - and statistic g k - : (3). In the formula, The steady-state mean estimated by the sliding window; This is a noise tolerance parameter used to accommodate natural fluctuations in electrical parameters during normal operation and to avoid false alarms caused by random noise.

[0067] S1.3, Early Warning Decision-Making and Triggering.

[0068] The system sets theoretically calculated and practically verified threshold limits for the CUSUM statistic. When the cumulative statistic of any terminal continuously exceeds the preset threshold, the system automatically marks the terminal as suspicious and immediately generates a deep diagnostic trigger command. This mechanism ensures the timeliness of anomaly detection and the accuracy of early warnings. The adaptive threshold... (4). In the formula, To determine the minimum detectable mutation magnitude, the minimum abnormal change in electrical parameters that the algorithm can identify is determined; This represents the maximum number of delayed sampling points. The sampling interval; and The correction factor is , and , .

[0069] like Figure 3As shown, the critical criteria for judging anomalies in the bilateral CUSUM change point detection algorithm are clearly defined, intuitively demonstrating how to find a balance between detection sensitivity and false alarm rate. Here, μ0 represents the steady-state average value of basic electrical parameters such as active power in the power metering terminal, serving as the benchmark value for judging whether the data is normal. The maximum allowable delay time from the occurrence of an anomaly to its detection; h is the dynamically calculated critical threshold curve for anomaly judgment.

[0070] S2, Intelligent hierarchical data acquisition and data quality control.

[0071] This step is a refined data acquisition phase for early warning terminals, balancing diagnostic needs and system load through a tiered data collection strategy.

[0072] S2.1 High-frequency data acquisition.

[0073] The system issues a special command to the suspected terminal, triggering it to enter high-frequency sampling mode. In this mode, the terminal acquires complete instantaneous voltage and current waveforms at a sampling frequency of no less than 6.4kHz, ensuring that harmonic components of the 13th order and below can be captured, providing data support for subsequent in-depth analysis.

[0074] S2.2 Multi-level data verification.

[0075] The system initiates a regional collaborative verification mechanism, comparing the data from the suspected terminal with other terminals within the same power supply area, including the consistency of current distribution and the similarity of power factor distribution. Simultaneously, it performs a longitudinal verification with data from the next higher-level meter, validating the power balance relationship and harmonic propagation characteristics. This process effectively identifies and eliminates false alarms caused by momentary communication interference or malfunctioning data acquisition equipment.

[0076] S2.3 Data standardization preprocessing.

[0077] The verified waveform data undergoes pipeline-style preprocessing: digital filters are used to eliminate high-frequency noise and power frequency interference, interpolation algorithms ensure strict synchronization of voltage and current data, and normalization is performed to eliminate dimensional effects. These preprocessing operations lay a solid foundation for subsequent feature extraction.

[0078] S3. Extraction and representation of multidimensional state information.

[0079] This step aims to extract discriminative multidimensional state information from preprocessed data. This invention creatively employs a fusion perception strategy, combining macroscopic visual perception with microscopic quantitative perception to comprehensively characterize the terminal state feature system. The core of the terminal state feature system lies in simultaneously utilizing the overall morphological information provided by the VI trajectory image and the precise quantitative parameters provided by key physical features to form a complementary and corroborating chain of feature evidence. Based on the divide-and-conquer approach of fault tree analysis, the extracted features are strongly correlated with specific fault modes to construct a systematic diagnostic path, specifically including the following steps.

[0080] S3.1, Visual processing of VI trajectories.

[0081] The generation of VI trajectory images is a key step in converting one-dimensional time-series signals into two-dimensional spatial patterns. Its purpose is to preserve and enhance the overall shape and spatial vertical information that is extremely sensitive to complex faults and abnormal power consumption behavior but is difficult to describe numerically.

[0082] The VI trajectory image processing workflow includes three key steps: S3.1.1, Coordinate Mapping: Using the synchronously sampled voltage sequence U(t) and current sequence I(t) as X-axis and Y-axis coordinates, a VI trajectory diagram that intuitively reflects the dynamic changes in impedance characteristics is drawn on a two-dimensional plane. S3.1.2, Pixelation and Resolution Setting: The continuous trajectory diagram is converted into a 256×256 pixel grayscale image. The resolution is carefully selected to achieve the best balance between preserving necessary details and controlling computational complexity. S3.1.3, Morphology Enhancement: The trajectory is smoothed through interpolation algorithms to reduce the jagged effect caused by sampling, thereby highlighting its essential geometric shape.

[0083] S3.2 Feature extraction of VI trajectory.

[0084] To construct an interpretable diagnostic system and quickly pinpoint specific faults, this invention simultaneously extracts low-dimensional, highly representative numerical feature vectors with clear physical meaning, including geometric features, spectral features, and statistical features.

[0085] Geometric features are parameters with intuitive physical meaning extracted from the overall shape and spatial structure of the VI trajectory. They are like measuring the shape of the trajectory, directly reflecting the basic operating state of the terminal. Geometric features include trajectory area, principal axis tilt angle, and eccentricity. These geometric parameters transform abstract electrical relationships into quantifiable spatial characteristics, providing an intuitive physical basis for fault diagnosis. Among them, the trajectory area... (5). In the formula, This is the time of one cycle. Main spindle tilt angle. (6). Eccentricity (7). In the formula, Cov(U, I) is the covariance of U and I; Var(U) is the variance of U; Var(I) is the variance of I; a is the semi-major axis length of the ellipse calculated by fitting; b is the semi-minor axis length of the ellipse calculated by fitting. Spectral characteristics are characteristic parameters obtained by converting the time-domain signal to the frequency domain through Fourier transform, and are specifically used to diagnose faults related to waveform distortion and harmonic measurement. Spectral characteristics reveal the intrinsic state of the terminal from the frequency dimension, making up for the insufficiency of time-domain analysis in capturing harmonic characteristics. Spectral characteristics include total harmonic distortion (THD), harmonic power distribution, and spectral centroid. Total harmonic distortion of current (8). In the formula, This is the effective value of the fundamental current. It is the first RMS value of subharmonic current It is the highest harmonic order considered.

[0086] Statistical features are characteristic parameters extracted from voltage and current time-series data through statistical analysis, used to assess the stability and consistency of terminal operation. Statistical features include waveform similarity, peak factor, fluctuation index, mean, and variance. Statistical features provide supplementary evidence for condition assessment from the perspective of data distribution patterns, enhancing the robustness of the diagnostic system.

[0087] S4, Hybrid Deep Learning Model Construction.

[0088] This invention innovatively designs a dual-branch hybrid deep learning model, which achieves complementary advantages by fusing image features and physical features.

[0089] S4.1, Image Feature Extraction Branch.

[0090] The image feature extraction branch employs a convolutional neural network architecture to specifically handle the morphological features of the VI trajectory. For example... Figure 4 As shown, after receiving a 256×256 pixel grayscale image, the input layer sequentially passes through three convolutional-pooling modules: the first convolutional-pooling module uses 32 3×3 convolutional kernels for feature extraction, introduces non-linearity through the ReLU activation function, and then performs dimensionality reduction through a 2×2 max pooling layer; the second convolutional-pooling module increases the number of convolutional kernels to 64, deepening feature learning; the third convolutional-pooling module further expands to 128 convolutional kernels to capture more complex image patterns. Finally, a global average pooling layer converts the feature map into a feature vector, which is then connected to a fully connected layer containing 256 neurons, using a Dropout rate of 0.5 to prevent overfitting.

[0091] S4.2, Physical Feature Processing Branch.

[0092] The physical feature processing branch is used specifically to analyze feature parameters with explicit physical meaning. For example... Figure 5As shown, this branch takes a 15-dimensional key feature vector as input and processes it hierarchically through three fully connected layers: the 64 neurons in the first fully connected layer perform nonlinear transformation on the original features, the 32 neurons in the second fully connected layer perform feature compression, and the 16 neurons in the third fully connected layer complete high-level feature encoding. Each fully connected layer uses the ReLU activation function to gradually extract deep patterns from the physical features, laying the foundation for subsequent feature fusion.

[0093] S4.3 Feature Fusion and Classification.

[0094] This approach achieves effective integration of multi-source information. The 256-dimensional feature vector from the image branch is concatenated with the 16-dimensional feature vector from the physical feature branch to form a comprehensive feature representation. The fused features undergo deep semantic learning through a fully connected layer containing 128 neurons, and finally, a Softmax classifier outputs the probability distribution of each fault category. This fusion mechanism preserves the spatial features of the image while incorporating the explicit meaning of the physical features, significantly improving the model's expressive power and diagnostic accuracy.

[0095] S5. Model Training and Optimization.

[0096] This step aims to ensure that the model can accurately distinguish between the normal and abnormal states of the electricity metering terminal and precisely locate the specific fault type through carefully designed training strategies and optimization methods.

[0097] S5.1 Model training configuration.

[0098] The model training configuration adopts proven best practices. The cross-entropy loss function is chosen as the optimization objective, effectively measuring the difference between the predicted probability distribution and the true labels. The optimizer uses the Adam adaptive moment estimation algorithm with an initial learning rate of 0.001, improving training efficiency while ensuring convergence stability. The batch size is flexibly set to 32 or 64 based on computational resources, and an early stopping mechanism is employed during training: training automatically terminates when the validation set loss no longer decreases for 10 consecutive epochs, effectively preventing overfitting. The cross-entropy loss function is... (9). In the formula, It is the sample size. It is the number of categories. It is a sample The true label, The model predicts the sample. Category The probability of.

[0099] S5.2 Bayesian hyperparameter optimization.

[0100] Bayesian hyperparameter optimization is the key innovation of this invention. A search space is established for the learning rate, batch size, number of convolutional kernels, number of neurons in the fully connected layer, and Dropout rate: the learning rate is optimized logarithmically within the range [1e-5, 1e-2]; the batch size is optimally selected from {16, 32, 64, 128}; the number of convolutional kernels is determined within the range {16, 32, 64, 128}; the number of neurons in the fully connected layer is optimized within {64, 128, 256, 512}; and the Dropout rate is adjusted within the range [0.1, 0.7]. The optimization process uses a Gaussian process as a surrogate model, with the desired improvement as the acquisition function, achieving efficient global optimization through 100 iterations.

[0101] S5.3 Model evaluation indicators.

[0102] The model evaluation employs a multi-dimensional indicator system to comprehensively measure model performance. Accuracy reflects the overall classification effect, precision measures the prediction accuracy for each class, recall assesses the model's ability to identify positive examples, and the F1 score comprehensively balances the performance of precision and recall. Furthermore, a confusion matrix is ​​used to analyze misclassification among different classes in detail, providing specific directions for model improvement. These indicators together constitute a complete model evaluation framework, ensuring the reliability of the diagnostic system in practical applications.

[0103] The present invention will now be experimentally verified.

[0104] K1. Experimental Design.

[0105] To verify the effectiveness and feasibility of the present invention, existing and publicly available datasets were applied to the experiments of the present invention. Public datasets were used as experimental datasets, and the experiments were designed according to the following steps:

[0106] K1.1, Dataset Selection.

[0107] To fully verify the effectiveness and generalization ability of the present invention, this experiment uses the PLAID public dataset. The PLAID public dataset provides high-quality voltage and current waveform data, and its 30kHz sampling frequency can fully capture the detailed features of power signals. In particular, the dataset construction fully reflects the hierarchical acquisition strategy. For normal operation, short-sequence data based on basic parameters such as electrical quantity and power are mainly collected. For abnormal operation, both short-sequence statistical features and complete voltage and current waveform data are preserved.

[0108] K1.2 Data Preprocessing.

[0109] The data preprocessing stage follows the hierarchical acquisition strategy proposed in this invention, establishing a differentiated data processing pipeline. For normal state data, short-sequence sampling processing is adopted, extracting active power, reactive power, and average power at 15-minute intervals to form a 360-point statistical sequence, significantly reducing data storage and processing overhead. For abnormal state data, a complete data processing workflow is employed, retaining both short-sequence statistical characteristics and long-sequence waveform data to ensure the completeness and accuracy of anomaly analysis. The normal data, data errors, data loss, and data discontinuity of the 360 ​​sampling points are as follows: Figure 6 As shown.

[0110] By analyzing the shape of the data curve, the terminal status can be effectively monitored: when the curve is stable and the data is continuous, it indicates that the terminal is operating normally. If the curve shows obvious jumps, it indicates a fault event, which can be accurately located using the CUSUM algorithm. When data loss lasts for more than 1.5 seconds, it is considered data loss. Discontinuous curves usually indicate an abnormality in hardware connection or communication equipment. The test results for data errors, loss, and intermittency during the acquisition process are as follows: Figure 7 As shown.

[0111] For long-sequence waveform data, preprocessing includes coordinate normalization, contrast enhancement, and size unification. First, voltage and current data are normalized to the [-1, 1] interval to eliminate dimensionality effects. Then, histogram equalization is used to enhance the contrast of trajectory features. Finally, bicubic interpolation is used to unify all images to a resolution of 256×256 pixels. Regarding physical feature extraction, corresponding preprocessing strategies are adopted for different feature types: geometric features are fitted to ellipses using the least squares method; spectral features are processed using the Blackman window function to reduce spectral leakage; and statistical features are normalized using Z-scores to eliminate distribution bias.

[0112] K1.3, Data Partitioning.

[0113] A hierarchical data partitioning method is adopted to ensure that the training set, validation set, and test set have the same category distribution and data type composition. The specific partitioning ratio is 7:2:1, and the partitioning process fully considers the characteristics of hierarchical collection: normal state samples mainly contain short sequence data, while abnormal state samples contain both short and long sequence data.

[0114] The partitioning process employs a time-series-sensitive strategy, using the first 70% of the data as the training set, the middle 20% as the validation set, and the last 10% as the test set, effectively avoiding leakage of time information. Simultaneously, it ensures that data from the same terminal appears in only one set, preventing the model from overfitting based on device features. For regional data, an administrative division-based partitioning strategy is used, with the training set containing data from two regions, and the validation and test sets each containing data from different districts and counties within a different region, fully validating the model's generalization ability.

[0115] K1.4 Model Selection.

[0116] This invention selects a dual-branch hybrid deep learning model as its basic architecture. One branch of the model specifically processes VI trajectory images obtained from hierarchical acquisition, while the other branch processes short-sequence physical features. The features from the two branches are fused at a high-level. For the change-point detection module, an improved CUSUM algorithm based on a sliding window is chosen to adapt to the dynamic characteristics of load level changes in actual power systems.

[0117] K1.5, hyperparameter optimization.

[0118] A Bayesian optimization algorithm for hyperparameter optimization is employed to comprehensively optimize a selected deep learning model to obtain optimal model performance. Through multiple experiments and precise evaluations, the optimal hyperparameter combination is determined, ensuring that the model exhibits superior performance during training and testing, thereby effectively improving the accuracy and efficiency of power system terminal fault diagnosis.

[0119] K2, Experimental Results.

[0120] K2.1 Performance verification of the change point detection module.

[0121] The bilateral sliding window algorithm CUSUM proposed in this invention was independently tested on the PLAID dataset to demonstrate the change point detection module. Experimental results show that this module has excellent detection performance for three typical abnormal states: such as... Figure 8 The confusion matrix results show that the algorithm achieves an accuracy of 91% in data acquisition error detection, with an average response time of 1.1 seconds. For data loss anomalies, the detection accuracy is 88%, but the average response time is slightly longer at 1.8 seconds due to the need to confirm continuous missing states. In detecting intermittent data anomalies, the accuracy reaches a maximum of 95%, with a response time of 0.9 seconds. Of particular note is that by introducing a sliding window mechanism, the algorithm can adaptively track changes in load levels, keeping the overall false alarm rate below 4.3%, meeting the reliability requirements of practical engineering applications.

[0122] K2.2 Performance analysis of the hybrid diagnostic model.

[0123] The proposed dual-branch hybrid model achieves an overall accuracy of 95.8% and a macro F1 score of 0.947, significantly outperforming other methods. This demonstrates that through deep fusion of image and physical features, the model can acquire complementary information from different dimensions, thereby achieving more accurate identification of complex fault modes. Regarding model complexity, the hybrid model has 4.85 million parameters, and the single inference time is controlled within 58 milliseconds, meeting the real-time requirements of practical deployment while ensuring accuracy.

[0124] K2.3, Analysis of Experimental Results.

[0125] Experimental results show that the proposed method for hierarchical diagnosis of anomalies in power metering terminals based on change point detection and multidimensional information fusion has significant advantages in detection accuracy, response speed, resource efficiency, and practicality. Through hierarchical diagnostic architecture and hybrid model design, it achieves early warning of abnormal states, ensures accurate fault diagnosis, and significantly reduces system operation and maintenance costs, providing reliable technical support for the refined operation and maintenance of smart grids.

[0126] The overall system architecture of this invention is as follows: Figure 9 As shown, it includes the following modules.

[0127] (1) Data access layer.

[0128] As the data source of the system, it is responsible for receiving massive amounts of raw data uploaded by smart energy metering terminals from the power grid master station system, including voltage and current instantaneous waveforms, and basic power data of active and reactive power.

[0129] (2) Change point detection layer.

[0130] Deploy the improved bilateral sliding window algorithm CUSUM. Perform rapid screening of basic data to identify suspicious data before initiating full VI data collection. Simultaneously, perform data cleaning and regional verification to ensure data quality and efficiency.

[0131] (3) Feature engineering layer.

[0132] This task involves extracting and constructing features from the complete waveform data of the suspected terminal. Its core objective is to transform the original one-dimensional time-series data into two forms of multi-dimensional state information: first, mapping the synchronously sampled voltage and current sequences to VI trajectories on a two-dimensional plane; and second, calculating a series of numerical features with clear physical meaning from the same data source, including geometric features, spectral features, and statistical features.

[0133] (4) Model service layer.

[0134] A two-branch hybrid model was deployed, with one branch analyzing the image and the other branch analyzing the numerical features, and finally the two branches were fused for anomaly detection and classification. Bayesian optimization was used to automatically adjust the model parameters to maintain optimal performance.

[0135] (5) Application service layer.

[0136] The system's value output end. It translates diagnostic results into concrete operational and maintenance actions, such as generating precise meter replacement plans, lists of suspected electricity theft perpetrators, and repair recommendations, directly guiding on-site work and achieving precise operational and maintenance.

[0137] Traditional methods for detecting anomalies in electricity metering terminals have the following drawbacks: (1) Manual periodic inspections and spot checks; maintenance personnel go to the site to inspect and verify the meters according to the plan. This method is extremely inefficient, costly, and has limited coverage. Fault detection is severely delayed and is usually only dealt with after metering disputes or line loss anomalies occur. (2) Simple remote threshold alarms: The main station system alarms based on a preset single threshold, such as voltage phase loss or current overload faults. This method has a single diagnostic dimension and insufficient sensitivity. It cannot identify the implicit performance degradation problems of metering chip gain drift, phase error, and harmonic measurement inaccuracy. It can only indicate anomalies but cannot locate the cause of the fault. The common drawbacks of traditional methods are that they are passive, extensive, and inefficient, and cannot meet the needs of efficient and accurate operation and maintenance of a large number of terminals.

[0138] The advantages of this invention are: (1) Diagnostic dimension: It upgrades from a single threshold judgment to a comprehensive analysis that integrates VI trajectory image morphology and multi-dimensional physical features, which can capture hidden and complex faults that traditional methods cannot detect. (2) Diagnostic capability: With the help of a hybrid deep learning model that integrates Bayesian optimization, the system can not only detect anomalies, but also accurately identify specific fault types and output a list sorted by probability, providing clear guidance for on-site operation and maintenance. (3) Operation and maintenance mode: Through a hierarchical architecture and event-driven mechanism of front-end lightweight screening + back-end precise diagnosis, it realizes the transformation from passive response to proactive early warning and precise intervention, which greatly reduces the system communication and computing load, optimizes the configuration of operation and maintenance resources, and improves overall efficiency.

[0139] This invention not only solves the core pain points in current electricity metering operation and maintenance, but will also give rise to a series of innovative applications, with broad and far-reaching application prospects. Specifically: (1) Precision operation and maintenance: This invention can change the traditional operation and maintenance mode and promote the transformation of power grid operation and maintenance from regular inspection and passive emergency repair to proactive early warning and precise intervention. Through the fault probability list output by the system, a precise on-site meter replacement and maintenance plan can be formulated, significantly saving operation and maintenance costs and improving service efficiency. (2) Electricity management: In terms of anti-electricity theft analysis, this invention can keenly identify abnormal VI trajectory characteristics caused by various electricity theft methods. At the same time, by identifying and eliminating erroneous data generated by faulty terminals, it can greatly improve the data quality of the electricity information collection system and provide reliable guarantees for core businesses such as electricity bill settlement and line loss analysis. (3) Equipment management: This invention can be extended to the full life cycle management of electricity metering equipment. Equipment manufacturers can use this technology to conduct quality assessment and performance optimization of products leaving the factory; operation and maintenance units can establish a predictive maintenance system based on the health status of equipment to realize the transformation from post-fault repair to pre-fault early warning. This invention successfully constructs an intelligent diagnostic system for anomalies in electricity metering terminals based on multi-dimensional state information and a hierarchical diagnostic architecture. The core innovation of this system lies in its collaborative mechanism of lightweight front-end screening and precise back-end diagnosis. The front-end utilizes an improved bilateral CUSUM variable point detection to achieve real-time, efficient initial screening of all terminals, while the back-end initiates in-depth analysis of early warning terminals. It creatively integrates the macroscopic morphological features and key physical characteristics of VI trajectory images, employing a dual-branch hybrid deep learning model for complementary advantages. Furthermore, this invention introduces Bayesian optimization to automatically tune the model's hyperparameters, ensuring continuously optimal diagnostic performance. Experimental verification shows that this method not only achieves high-accuracy anomaly detection but also accurately classifies fault types and outputs a probability-ranked maintenance list. It successfully transforms the traditional maintenance model from passive response and blind inspection to proactive early warning and precise intervention, providing solid technical support for improving the efficiency of smart grid metering maintenance, ensuring data accuracy, and combating electricity theft.

Claims

1. A method for diagnosing abnormality of an electric energy metering terminal based on multi-dimensional state information, characterized in that, Includes the following steps: S1, Real-time monitoring and anomaly warning across the entire area S1.1, Constructing the basic data acquisition channel The basic electrical parameter short sequence X data uploaded by all online terminals is received continuously in a fixed time period, and the sequence X is: , (1); wherein, and are random variables, D is the change of the average value, is an index function, is the average value; S1.2 Sequence Change Point Detection and Analysis The improved two-sided sliding window algorithm CUSUM is run on each terminal's sequence, CUSUM computes a cumulative sum statistic dynamically for the sequence, where a positive cumulative g0 + and a negative cumulative g0 k + are given by: (2); and a positive cumulative g0 - and a negative cumulative g0 k - are given by: (3); where, is the steady-state mean of the sliding window estimate; is the noise tolerance parameter; S1.3 Early Warning Decision-Making and Triggering When the cumulative statistics of any terminal continuously exceed the preset threshold, the terminal is marked as a suspect and a deep diagnostic trigger command is immediately generated. S2, Intelligent Hierarchical Data Acquisition and Data Quality Control S2.1, High-frequency data acquisition A special command is sent to a terminal under suspicion to trigger it to enter a high-frequency sampling mode and collect complete instantaneous voltage and current waveforms. S2.2 Multi-level data verification A regional collaborative verification mechanism was initiated to conduct a horizontal comparative analysis of the data of the suspected terminal with other terminals in the same power supply area, and at the same time to conduct a vertical verification with the data of the meter at the next higher level to verify the power balance relationship and harmonic transmission characteristics. S2.3 Data Standardization Preprocessing A pipeline-style preprocessing method is used to preprocess the waveform data that has passed the verification. S3. Extraction and Representation of Multidimensional State Information S3.1, Image Processing of VI Trajectories Generate VI trajectory images and perform coordinate mapping, pixelation, resolution setting, and morphological enhancement on the VI trajectory images; S3.2 Feature Extraction of VI Trajectory Extract geometric, spectral, and statistical features from the VI trajectory image; S4, Hybrid Deep Learning Model Construction S4.1 Image Feature Extraction Branch The image feature extraction branch uses a convolutional neural network architecture to process the morphological features of the VI trajectory; S4.2, Physical Feature Processing Branch The physical feature processing branch takes a 15-dimensional key feature vector as input and performs hierarchical processing through three fully connected layers. S4.3 Feature Fusion and Classification The 256-dimensional feature vector of the image branch and the 16-dimensional feature vector of the physical feature branch are concatenated to form a comprehensive feature representation. The fused features are then subjected to deep semantic learning through a fully connected layer containing 128 neurons, and finally the probability distribution of each fault category is output through a Softmax classifier. S5, Model Training and Optimization S5.1 Model Training Configuration The cross-entropy loss function is chosen as the optimization objective. The cross-entropy loss function is: (4); where, It is the sample size. It is the number of categories. It is a sample The true label, The model predicts the sample. Category The probability; the optimizer uses the Adam adaptive moment estimation algorithm, with an initial learning rate of 0.001, and the batch size is set to 32 or 64 depending on the computing resources. An early stopping mechanism is used during training, and training is automatically terminated when the validation set loss no longer decreases for 10 consecutive rounds, effectively preventing overfitting. S5.2 Bayesian Hyperparameter Optimization Bayesian hyperparameter optimization was performed on the learning rate, batch size, number of convolutional kernels, number of neurons in fully connected layers, and Dropout rate. S5.3 Model Evaluation Indicators Accuracy reflects the overall classification performance, precision measures the prediction accuracy of each category, recall evaluates the model's ability to identify positive examples, F1 score comprehensively balances the performance of precision and recall, and confusion matrix provides a detailed analysis of misclassification among categories, offering specific directions for model improvement.

2. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 1, characterized in that, In step S1.3, the adaptive threshold (5); where, To determine the minimum detectable mutation magnitude, the minimum abnormal change in electrical parameters that the algorithm can identify is determined; The maximum number of delayed sampling points, The sampling interval is denoted as .

3. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 2, characterized in that, In step S2.3, a digital filter is used to eliminate high-frequency noise and power frequency interference, an interpolation algorithm is used to ensure strict synchronization of voltage and current data, and normalization processing is performed to eliminate the influence of dimensions.

4. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 3, characterized in that, In step S3.1, coordinate mapping uses the synchronously sampled voltage sequence U(t) and current sequence I(t) as X-axis and Y-axis coordinates to draw a VI trajectory diagram on a two-dimensional plane that intuitively reflects the dynamic changes of impedance characteristics; pixelation and resolution setting convert the continuous trajectory diagram into a grayscale image of 256×256 pixels. Morphology enhancement uses interpolation algorithms to smooth the trajectory and reduce the jagged effect caused by sampling.

5. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 4, characterized in that, In step S3.2, geometric features are extracted from the overall shape and spatial structure of the VI trajectory. These geometric features include the trajectory area, principal axis tilt angle, and eccentricity. (6); where, One cycle time; principal axis tilt angle (7); Eccentricity (8).

6. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 5, characterized in that, In step S3.2, the spectral characteristics are the characteristic parameters obtained after converting the time-domain signal to the frequency domain through Fourier transform. The spectral characteristics include the total harmonic distortion (THD), harmonic power distribution, and spectral centroid. The total harmonic distortion of the current... (9); where, This is the effective value of the fundamental current. It is the first RMS value of subharmonic current It is the highest harmonic order considered.

7. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 6, characterized in that, In step S3.2, Statistical features are characteristic parameters extracted from voltage and current time series data through statistical analysis. Statistical features include waveform similarity, peak factor, fluctuation index, mean, and variance.

8. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 7, characterized in that, In step S4.1, after the input layer receives a grayscale image of 256×256 pixels, it passes through three convolution-pooling modules in sequence: the first convolution-pooling module uses 32 3×3 convolution kernels for feature extraction, introduces non-linearity through the ReLU activation function, and then reduces the dimensionality through a 2×2 max pooling layer; the second convolution-pooling module increases the number of convolution kernels to 64 to deepen feature learning; The third convolutional-pooling module is further expanded to 128 convolutional kernels to capture more complex image patterns; finally, the feature map is converted into a feature vector through a global average pooling layer and connected to a fully connected layer containing 256 neurons, using a dropout rate of 0.5 to prevent overfitting.

9. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 8, characterized in that, In step S4.2, the 64 neurons of the first fully connected layer perform nonlinear transformation on the original features, the 32 neurons of the second fully connected layer perform feature compression, and the 16 neurons of the third fully connected layer complete high-level feature encoding.

10. The method for diagnosing anomalies in an energy metering terminal based on multi-dimensional state information according to claim 9, characterized in that, In step S5.2, the learning rate is optimized on a logarithmic scale within the interval [1e-5, 1e-2], the batch size is optimally selected from {16, 32, 64, 128}, the number of convolutional kernels is determined within the range of {16, 32, 64, 128}, the number of neurons in the fully connected layer is optimized within {64, 128, 256, 512}, and the Dropout rate is adjusted within the interval [0.1, 0.7]. The optimization process uses a Gaussian process as a surrogate model, with the expected improvement as the acquisition function, and achieves efficient global optimization through 100 iterations.

Images