Intelligent electric energy meter fault prediction method based on multi-mode sensor fusion

Abstract

The invention provides an intelligent electric energy meter fault prediction method based on multi-modal sensor fusion. The method comprises the following steps: collecting original data of a multi-modal sensor and preprocessing the original data to obtain a standardized data stream; establishing a sensor health state model and monitoring the working state of the sensor; performing feature extraction and pattern recognition on the standardized data to generate an electric energy meter state feature vector; combining a sensor health assessment result and an electric energy meter state feature vector, and applying a double-layer causal reasoning framework to realize separation and identification of a sensor fault and an electric energy meter fault; multi-source uncertainty is quantified, diagnosis reliability is evaluated, and a fault evaluation report is generated. The method can effectively distinguish the sensor fault from the electric energy meter fault, improves the fault prediction accuracy, reduces the false alarm rate, and optimizes the maintenance decision.

Description

Technical Field

[0001] The invention relates to a fault prediction method, in particular to a smart electric energy meter fault prediction method based on multi-modal sensor fusion. Background Art

[0002] With the rapid development of smart grids, smart energy meters, as key terminal devices, undertake multiple functions such as energy metering, data collection, and electricity consumption information interaction. Energy meter failure not only affects the accuracy of energy metering and the fairness of electricity bill settlement, but may also cause local abnormal operation of the distribution network, reduce power supply reliability, and even cause safety accidents. Real-time monitoring and fault prediction of the operating status of smart energy meters are of great significance for improving the safety of power grid operation, reducing economic losses, and optimizing the allocation of operation and maintenance resources. Especially in the context of large-scale deployment of smart grids, active fault prediction technology based on sensor fusion can significantly reduce the failure rate, improve power supply reliability indicators, and reduce the maintenance costs of power grid companies and power outage losses of users.

[0003] At present, the fault prediction of smart electric energy meters mainly adopts three technical routes: threshold monitoring method based on a single sensor, statistical feature analysis method based on time series data, and model-based fault diagnosis method. The threshold monitoring method determines the operating status of the electric energy meter by setting the safety limit of parameters such as voltage, current, and temperature. It is simple to implement and has a small amount of calculation. The statistical feature analysis method uses the statistical characteristics of the electric energy meter operation data, such as mean, variance, kurtosis, etc., combined with fuzzy logic or decision tree methods to identify abnormal patterns, thereby improving the detection accuracy. The model-based fault diagnosis method establishes a mathematical model for the normal operation of the electric energy meter, and predicts potential faults through residual analysis with high accuracy. In recent years, some researchers have tried to use machine learning methods to process single or simple combination of sensor data for fault prediction, such as electric energy meter error prediction based on BP neural network, component aging monitoring based on support vector machine, etc., which also improves the prediction performance.

[0004] The key technical problem faced by existing methods is that it is difficult to effectively distinguish between sensor faults and actual faults of the energy meter. In practical applications, sensor faults such as voltage sensor drift, slow response of current sensor or decreased sensitivity of temperature sensor are often highly similar to actual faults such as aging of the circuit board of the energy meter and abnormality of the metering chip in data performance. This similarity makes it impossible for the system to accurately determine the true source of abnormal data, resulting in a false alarm rate of about 30%-45%, a large number of maintenance personnel are ineffectively dispatched, and resources are seriously wasted. For example, the zero drift of the current sensor may be misjudged as a fault in the sampling circuit of the energy meter, and the decreased sensitivity of the temperature sensor may be misjudged as abnormal heating of the energy meter. Existing technologies mostly use single redundancy or simple thresholds to judge sensor faults, lack a joint evaluation mechanism for the health status of the sensor and the fault status of the energy meter, and cannot achieve accurate fault source tracing. In addition, when multiple sensors work together, the cross-effect of fault effects becomes more complicated, further increasing the difficulty of fault source identification. Summary of the invention

[0005] The purpose of the invention is to provide a smart electric energy meter fault prediction method based on multimodal sensor fusion, in order to solve one of the above-mentioned technical problems existing in the prior art.

[0006] Technical solution: A smart energy meter fault prediction method based on multimodal sensor fusion, comprising the following steps:

[0007] Collect and preprocess multi-modal sensor raw data to obtain standard data;

[0008] Based on standard data, a sensor health status model is established and the sensor working status is monitored to obtain the sensor health assessment results;

[0009] Perform feature extraction and pattern recognition on standard data to generate the state feature vector of the electric energy meter;

[0010] Combining the sensor health assessment results and the energy meter state feature vector, the two-layer causal graph reasoning module is used to perform two-layer fault source separation and identification, and output the fault type and confidence level.

[0011] Based on the fault type and confidence level, multi-source uncertainty is quantified and reliability is evaluated to generate a fault assessment report.

[0012] Preferably, the two-layer causal graph reasoning module includes:

[0013] The upper layer module is used to represent the causal relationship between sensor failure and measurement anomaly.

[0014] The lower module is used to represent the causal relationship between the power meter failure and the abnormal physical parameters.

[0015] Interlayer connections are used to represent the observed relationship between measurement anomalies and physical parameter anomalies.

[0016] When working, it infers measurement anomalies from possible sensor failures from top to bottom, and infers potential faults from observed anomalies from bottom to top, propagates probability information in the two-layer causal graph through the message passing mechanism, and outputs the probability distribution of fault sources;

[0017] Preferably, the process of performing double-layer fault source separation and identification includes:

[0018] Read the observation evidence set as the input of the two-layer causal graph reasoning module to obtain the probability distribution of the fault source;

[0019] Among them, the observation evidence set includes sensor health assessment results, energy meter state feature vectors and current measurement data;

[0020] The fault sources with probability higher than a threshold in the fault source probability distribution are taken as candidate fault sources;

[0021] Three competing hypotheses, namely sensor failure, meter failure and mixed failure, are constructed. The Bayesian model selection method is applied to evaluate the posterior probability of each hypothesis, and the optimal fault hypothesis and candidate fault source set are output.

[0022] Preferably, the process of performing double-layer fault source separation and identification further includes:

[0023] For each candidate fault in the candidate fault source set, the original signal is decomposed into a sensor fault component and an electric energy meter fault component, and a pure electric energy meter state signal is reconstructed;

[0024] Combined with the physical working principle of the electric energy meter, verify whether the state signal of the pure electric energy meter conforms to the physical consistency law, and output the fault verification result;

[0025] Combining the sensor health assessment results, the energy meter state feature vector and the fault verification results, a dynamic decision tree is used to perform hierarchical identification of sensor faults and energy meter faults to obtain the final fault type judgment and fault confidence.

[0026] Preferably, the step of reconstructing the pure electric energy meter state signal comprises:

[0027] Construct a collaborative signal decomposition model for distinguishing different manifestations of sensor fault signatures and electric energy meter fault signatures;

[0028] Construct characteristic basis function libraries for sensor faults and electric energy meter faults respectively, forming a dual-mode adaptive basis function library;

[0029] Based on the collaborative decomposition model and dual-mode adaptive basis function library, the original signal is decomposed into sensor fault component and electric energy meter fault component;

[0030] Calculate the contribution ratio of the sensor fault component and the electric energy meter fault component to the original signal, and generate an energy distribution diagram for evaluating the relative importance of the two types of fault components;

[0031] According to the weight analysis of the component energy distribution diagram, the sensor fault component is eliminated, the separation threshold is adjusted, and the pure electricity meter status signal is reconstructed in combination with the normal signal baseline.

[0032] Preferably, the step of verifying whether the pure electric energy meter status signal conforms to the physical consistency law and outputting the fault verification result includes:

[0033] Based on the working principle of the electric energy meter, a set of physical constraint models of the electric energy meter including electrical model, thermal model and mechanical model is established to form a set of physical constraint equations that can be quantified and verified;

[0034] Substitute the pure electric energy meter state signal into the physical constraint equation group, calculate the physical constraint satisfaction, and generate a physical consistency score;

[0035] Adjust the parameters and thresholds in the physical constraint equation group according to the current working condition type, so that the constraint test adapts to the current working condition, and output the working condition adaptive constraint test results;

[0036] For the fault features identified in the pure electric energy meter status signal, check whether they conform to the physical manifestation rules of the corresponding fault type and generate a physical rationality score for the fault features;

[0037] The above outputs (physical consistency score, working condition adaptive constraint inspection results and fault feature physical rationality score) are integrated to generate fault verification results, including verification pass status and credibility assessment.

[0038] Preferably, the process of establishing a sensor health status model and monitoring the sensor working status to obtain a sensor health assessment result includes:

[0039] Based on the historical normal operation data in the standard data, a normal response characteristic model of the sensor is established;

[0040] Compare the current sensor real-time data with the sensor normal response characteristic model, detect sensor characteristic drift, and generate sensor characteristic drift indicators;

[0041] Perform frequency domain analysis on standard data, extract sensor frequency domain features, and construct sensor frequency domain feature vectors;

[0042] Perform multi-sensor cross-validation to analyze the physical correlation between different sensor data and generate sensor credibility scores;

[0043] By integrating the above outputs (sensor characteristic drift index, sensor frequency domain feature vector and sensor credibility score), a sensor health status model is constructed to output the sensor health assessment results, including the sensor health index and abnormal type identification results.

[0044] Preferably, the process of generating a sensor credibility score comprises:

[0045] Analyze the physical dependencies and associations between sensor data in the standard data to obtain the physical constraint relationship matrix;

[0046] Based on the physical constraint relationship matrix, the consistency check is performed on the standard data, and the residual value of each physical constraint equation is calculated to form the sensor consistency residual vector;

[0047] Decompose the sensor consistency residual vector into individual sensor contributions and output the sensor anomaly contribution metric;

[0048] Identify subsets of sensors that can authenticate each other, build redundant sensor groups, and form a sensor group mapping table;

[0049] For each group in the sensor group mapping table, the complementary information of the sensors in the group is extracted to construct a sensor mutual information matrix;

[0050] The sensor anomaly contribution metric and sensor mutual information matrix are integrated, and the sensor credibility score is calculated and output through the credibility scoring function.

[0051] Preferably, the process of forming the sensor group mapping table is specifically as follows:

[0052] Calculate the physical correlation distance matrix between all sensor pairs in the standard data;

[0053] Determine the initial cluster center based on the physical constraint relationship matrix, perform principal component analysis on the physical constraint relationship matrix, select the sensor with the largest contribution of each principal component as the candidate center, and obtain the initial cluster center set;

[0054] Using the initial cluster center set as the starting cluster center, each sensor is assigned to the nearest center, and a new center that minimizes the total distance is reselected in each group to obtain the initial division of the sensor group;

[0055] Calculate the internal physical redundancy for each sensor group, adjust the critical group, and output the optimized sensor group;

[0056] Each sensor group is matched with a physical constraint relationship matrix, and it is confirmed that the sensors in the group constitute a meaningful physical association unit, so as to obtain a sensor group mapping table with physical meaning.

[0057] Preferably, the step of generating the electric energy meter state feature vector comprises:

[0058] Perform multi-scale wavelet packet decomposition on standard data, extract time domain, frequency domain and time-frequency joint features, and construct an initial multi-scale feature set;

[0059] Using the electrical characteristics and thermal model of the electric energy meter, the initial multi-scale feature set is enhanced with physical knowledge to generate a physical enhanced feature set.

[0060] Apply feature optimization algorithm and dimensionality reduction processing to the physical enhancement feature set to form an optimized electric energy meter feature vector;

[0061] According to the working condition information of the electric energy meter, the optimized electric energy meter feature vector is dynamically weighted and the working condition adaptive feature vector is output;

[0062] The adaptive characteristic vector of the working condition is identified to generate the characteristic vector of the electric energy meter state and the preliminary fault type determination result.

[0063] Beneficial effect: The present invention can effectively distinguish sensor faults from electric energy meter faults, improve fault prediction accuracy, reduce false alarm rate, and optimize maintenance decisions. The relevant technical effects will be described in detail in conjunction with the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of the present invention.

[0065] Figure 2 This is a flow chart of performing double-layer fault source separation and identification in Embodiment 1 of the present invention.

[0066] Figure 3 This is a flow chart of performing double-layer fault source separation and identification in Embodiment 2 of the present invention.

[0067] Figure 4 It is a flow chart of the present invention for reconstructing the state signal of the pure electric energy meter. DETAILED DESCRIPTION

[0068] According to one aspect of the present application, based on the fault type and confidence level, quantifying multi-source uncertainty and evaluating reliability, the steps of generating a fault assessment report include:

[0069] Based on the sensor health assessment results and historical calibration data, a sensor measurement error model is established to quantify the uncertainty range of each sensor data point;

[0070] Using the Monte Carlo method, uncertainty propagation analysis is performed on the feature extraction process and fault diagnosis algorithm, and the uncertainty index of model reasoning is calculated;

[0071] Analyze the impact of environmental temperature, humidity, electromagnetic interference and other factors on sensor data and diagnostic results, and establish an environmental impact sensitivity model;

[0072] Integrate the sensor measurement error model, model reasoning uncertainty index and environmental impact sensitivity model to build a comprehensive assessment framework for multi-source uncertainty and calculate the confidence interval and risk level of the fault diagnosis results;

[0073] Based on the final fault type determination, fault confidence and comprehensive evaluation results of multi-source uncertainty, a fault assessment report containing fault type, location, severity, confidence interval and reliability score is generated.

[0074] According to one aspect of the present application, the steps of dynamically adjusting the weight of the optimized electric energy meter feature vector according to the electric energy meter working condition information and outputting the working condition adaptive feature vector include:

[0075] For each feature in the optimized electric energy meter feature vector, its ability to distinguish between faults and normal states is calculated under various operating conditions, and a feature-operating condition sensitivity matrix is ​​constructed;

[0076] Based on the operating parameters of the electric energy meter, a working condition pattern identifier is designed to classify the current working condition into predefined categories and output the current working condition type and working condition confidence level;

[0077] For each type of working condition, a feature importance adaptive evaluation function based on the feature-working condition sensitivity matrix is ​​constructed, and the working condition adaptive weight function is output by combining the historical accuracy of fault diagnosis and the environmental adaptability factor;

[0078] The adaptive weight function of the working condition is applied to sort the importance of the optimized electric energy meter feature vectors, and a threshold-based automatic feature selection algorithm is designed to output the working condition optimized feature subset.

[0079] The features in the working condition optimization feature subset are weighted and combined, and the combination coefficient is dynamically adjusted according to the current working condition type and working condition confidence to generate a working condition adaptive feature vector.

[0080] According to one aspect of the present application, a physical consistency checker is designed using the physical working principle of an electric energy meter to verify whether a pure electric energy meter state signal conforms to physical laws. The steps of outputting a fault verification result include:

[0081] Based on the working principle of the electric energy meter, a set of physical constraint models of the electric energy meter including electrical model, thermal model and mechanical model is established to form a set of physical constraint equations that can be quantified and verified;

[0082] Substitute the pure electric energy meter state signal into the physical constraint equation group, calculate the physical constraint satisfaction, and generate a physical consistency score;

[0083] According to the current working condition type, dynamically adjust the parameters and thresholds in the physical constraint equation group to adapt the constraint test to different working conditions, and output the working condition adaptive constraint test results;

[0084] For the fault features identified in the pure electric energy meter status signal, check whether they conform to the physical manifestation rules of the corresponding fault type and generate a physical rationality score for the fault features;

[0085] The physical consistency score, the working condition adaptive constraint inspection results and the physical rationality score of the fault characteristics are integrated to generate the final fault verification result, including the verification pass status and credibility assessment.

[0086] According to one aspect of the present application, a constrained projection algorithm is designed to decompose the sensor consistency residual vector into individual sensor contributions, and the steps of outputting the sensor abnormal contribution metric value include:

[0087] Based on the physical constraint relationship matrix and the current sensor measurement value, the partial derivative of each constraint equation for each sensor is calculated to construct a sensitivity matrix;

[0088] Construct a weighted pseudo-inverse projection matrix to project the sensor consistency residual vector into the sensor space and estimate the sensor measurement error vector;

[0089] Calculate the standardized score of each sensor error, apply the improved CUSUM algorithm to accumulate the scores, and output the cumulative anomaly score of each sensor;

[0090] Set adaptive thresholds, identify abnormal sensors, calculate abnormal contribution, and output a preliminary set of abnormal sensors;

[0091] The residuals are re-evaluated in an iterative manner, the residuals are recalculated based on the currently identified normal sensor subset, residual projection is performed, and new abnormal sensors are identified until the residuals are small enough or no new anomalies are detected, and the final sensor anomaly contribution metric is output.

[0092] According to one aspect of the present application, a bidirectional causal inference algorithm is designed to infer possible fault sources through a message passing mechanism between upper and lower layers, and the steps of outputting a probability distribution of the fault sources include:

[0093] Map the data in the observation evidence set to the corresponding nodes of the sensor-electricity meter double-layer causal graph, and initialize the probability distribution of the evidence nodes;

[0094] Perform top-down reasoning, starting from the potential fault nodes at the sensor layer, calculate their impact on the measurement abnormal nodes through the calibrated causal model parameters, and generate top-down propagation results;

[0095] Perform bottom-up reasoning, starting from the observed abnormal nodes, calculate the probability of possible power meter fault nodes through Bayesian reverse reasoning, and generate bottom-up propagation results;

[0096] Design a cross-layer message passing mechanism to integrate top-down propagation results with bottom-up propagation results, handle inter-layer conflicts, and calculate the integrated failure probability;

[0097] The node probability is updated iteratively until convergence or the maximum number of iterations is reached, and the final fault source probability distribution is output.

[0098] According to one aspect of the present application, for high-probability candidate fault sources, three competitive hypotheses of sensor fault, electric energy meter fault and mixed fault are constructed, the posterior probability is evaluated, and the steps of outputting the optimal fault hypothesis and the candidate fault source set include:

[0099] Based on the probability distribution of fault sources, three competitive fault hypothesis models are constructed: hypothesis 1 is a pure sensor fault, hypothesis 2 is a pure electric energy meter fault, and hypothesis 3 is a mixed fault of sensor and electric energy meter;

[0100] For each hypothesized model, the likelihood value of its explanation of the observed data is calculated based on the sensor-electricity meter double-layer causal graph;

[0101] Introduce model complexity penalty terms, adjust the hypothesis model according to Occam's razor principle, avoid overly complex explanations, and calculate the adjusted hypothesis score;

[0102] The Bayes factor is used to evaluate the relative support between the hypotheses and calculate the posterior probability of each hypothesis;

[0103] The hypothesis with the highest posterior probability is selected as the optimal fault hypothesis, and high-probability fault sources are extracted from the hypothesis to form a set of candidate fault sources.

[0104] According to one aspect of the present application, based on the collaborative decomposition model and the dual-mode adaptive basis function library, the steps of applying a signal decomposition algorithm with sparse constraints to decompose the original signal into a sensor fault component and an electric energy meter fault component include:

[0105] Construct an optimization objective function, including data fitting terms, sparse constraint terms, and physical consistency constraint terms, to form a constrained optimization objective function;

[0106] Design differentiated sparse constraints based on sensor fault and meter fault characteristics, impose stronger sparsity constraints on sensor fault components, and determine sparse regularization parameters;

[0107] Introduce physical model constraints of the electric energy meter to ensure that the decomposition results are consistent with the working principle of the electric energy meter and build a decomposition model with physical constraints;

[0108] The accelerated proximal gradient descent algorithm is used to solve the optimization objective function, and the coefficients of the sensor fault component and the electric energy meter fault component are iteratively optimized to obtain the preliminary decomposition results.

[0109] According to the signal decomposition quality and physical constraint satisfaction, the regularization parameters are adaptively adjusted to balance the decomposition accuracy and physical rationality, and the final sensor fault components and electric energy meter fault components are output.

[0110] According to one aspect of the present application, the steps of calculating the contribution ratio of the sensor fault component and the electric energy meter fault component to the original signal, generating a component energy distribution diagram, and evaluating the relative importance of the two types of faults include:

[0111] Calculate the energy values ​​of sensor fault components and electric energy meter fault components in the time domain and frequency domain, and quantify the absolute intensity of each component;

[0112] Analyze the proportion of each component energy to the total energy of the original signal to determine the relative contribution ratio of sensor fault and electric energy meter fault;

[0113] Study the distribution characteristics of each component in the time-frequency domain, identify the time location and frequency characteristics of the fault component, and form the time-frequency domain component characteristics;

[0114] Combined with the fault type knowledge base, the impact of different components on system performance is evaluated and the fault importance index is calculated;

[0115] The energy value, contribution ratio, time-frequency characteristics and importance index are integrated to generate a multi-dimensional component energy distribution diagram to guide the subsequent pure signal reconstruction process.

[0116] In view of the problem that it is difficult to distinguish between the fault of the sensor itself and the actual fault of the electric energy meter, the existing technology mainly uses a single redundant or simple threshold to judge the sensor fault, lacks a joint evaluation mechanism, and leads to a high false alarm rate (30-45%) and a waste of maintenance resources. Therefore, this application uses the sensor feature deviation analysis SCDA method to build a normal working characteristic model of the sensor and detect the deviation. However, after applying SCDA, it is found that the fault effects cross-influence when multiple sensors work together, and it is difficult to identify the source of the fault. Therefore, through the collaborative fault source separation algorithm CFSA, a fault probability propagation model based on Bayesian network and a multi-sensor collaborative state evaluation method are designed. In order to improve the robustness under dynamic working conditions and multi-source uncertainty conditions. Through the two-layer adaptive fault identification framework DAFIF, a hierarchical fault identification structure and a cross-layer constraint mechanism are constructed, and a recursive filter is introduced to improve robustness. Therefore, in this scheme, CFSA and DAFIF are merged to form a complete framework, including a characteristic modeling layer, a deviation analysis layer, a two-layer identification layer, and an uncertainty processing layer, which effectively solves the key technical problems of separating and identifying sensor faults and electric energy meter faults.

[0117] Case 1

[0118] S1: Multimodal sensor data acquisition and preprocessing

[0119] The raw data of the multimodal sensors of the electric energy meter, including voltage sensor data, current sensor data, temperature sensor data, vibration sensor data and acoustic sensor data, are read and synchronized through a timestamp alignment algorithm to obtain time-aligned multimodal sensor data.

[0120] An adaptive wavelet threshold denoising algorithm is applied to the time-aligned multimodal sensor data to identify and mark the sudden outliers, and a local median filter is used for smoothing to output the denoised sensor data.

[0121] For the denoised sensor data, adaptive Z-score normalization is performed according to the range and physical characteristics of each sensor, and the data of different physical quantities are mapped to a unified metric space to obtain a standardized multimodal sensor data stream.

[0122] Analyze the working state characteristics of the electric energy meter in the standardized multimodal sensor data stream, identify different working conditions such as steady-state operation, load change, and startup process, and divide the data stream into data segments marked with working conditions as the basis for subsequent analysis.

[0123] S2: Sensor health status modeling and monitoring

[0124] Based on the historical normal operation data in the standardized multimodal sensor data stream, a sensor normal response characteristic model is established for each type of sensor, including core parameters such as sensitivity, linearity, response time, and noise level.

[0125] Compare the current real-time data of the sensor with the normal response characteristic model of the sensor, use the cumulative sum control chart (CUSUM) method to detect the sensor characteristic drift, and output the sensor characteristic drift index.

[0126] Short-time Fourier transform (STFT) is performed on the standardized multimodal sensor data stream to extract the spectral characteristics of the sensor signal and construct the sensor frequency domain feature vector. By comparing it with the normal spectrum template, the abnormal mode of the sensor is identified.

[0127] A multi-sensor cross-validation algorithm is designed to analyze the physical correlation between different sensor data, verify the consistency of sensor data based on energy conservation and physical constraints, and generate a sensor credibility score.

[0128] By integrating the sensor characteristic drift index, the sensor frequency domain feature vector analysis results and the sensor credibility score, the sensor health status model is constructed using the improved evidence theory to output the health index and abnormal type identification results of each sensor.

[0129] S3: Electricity meter fault feature extraction and pattern recognition

[0130] Multiscale wavelet packet decomposition is performed on the standardized multimodal sensor data stream to extract time domain, frequency domain and time-frequency joint features and construct an initial multiscale feature set.

[0131] The electrical characteristics and thermal model of the electric energy meter are used to enhance the initial multi-scale feature set with physical knowledge. Physical features such as power factor, harmonic distortion, and phase shift are calculated to generate a physically enhanced feature set.

[0132] The recursive feature elimination algorithm combined with principal component analysis is applied to the physical enhanced feature set to remove redundant and low-information features, retain the most discriminative feature subset, and form the optimized electric energy meter feature vector.

[0133] According to the data segment information of the working condition mark, the feature importance evaluation function is designed for different working conditions, the feature weight is dynamically adjusted, and the working condition adaptive feature vector is output.

[0134] The improved support vector machine algorithm is used to classify the adaptive feature vectors of the working conditions, identify the normal state and potential fault modes, and generate the state feature vector of the electric energy meter and the preliminary fault type determination result.

[0135] S4: Two-layer fault source separation and identification

[0136] A Bayesian network model is constructed to express the conditional dependency relationship between the sensor health status model and the energy meter state feature vector, quantify the influence of sensor anomalies on the fault judgment of the energy meter, and output the fault correlation matrix.

[0137] A two-layer causal reasoning framework is designed to transform the fault correlation matrix into a causal graph structure. Through reverse reasoning, it is determined whether the observed anomaly is caused by a sensor fault or an actual fault of the electricity meter, and a set of candidate fault sources is generated.

[0138] For each candidate fault in the candidate fault source set, a collaborative signal decomposition algorithm is applied to decompose the original signal into a sensor fault component and an electric energy meter fault component, and reconstruct the pure electric energy meter state signal after removing the sensor fault.

[0139] By utilizing the physical working principle of the electric energy meter and the inherent constraints between multi-sensor data, a physical consistency checker is designed to verify whether the separated pure electric energy meter status signal conforms to the physical laws and output the fault verification results.

[0140] Combining the sensor health index, the energy meter state feature vector and the fault verification result, a dynamic decision tree structure is adopted to realize the hierarchical identification of sensor faults and energy meter faults, and output the final fault type judgment and fault confidence.

[0141] S5: Multi-source uncertainty quantification and reliability assessment

[0142] Based on the sensor health index and historical calibration data, a sensor measurement error model is established to quantify the uncertainty range of each sensor data point.

[0143] The Monte Carlo method is used to perform uncertainty propagation analysis on the feature extraction process and fault diagnosis algorithm, and the model reasoning uncertainty index is calculated.

[0144] Analyze the impact of environmental temperature, humidity, electromagnetic interference and other factors on sensor data and diagnostic results, establish an environmental impact sensitivity model, and evaluate the degree of interference of environmental conditions on fault judgment.

[0145] By integrating the sensor measurement error model, model reasoning uncertainty index and environmental impact sensitivity model, a multi-source uncertainty comprehensive evaluation framework is constructed to calculate the confidence interval and risk level of the fault diagnosis results.

[0146] Based on the final fault type determination, fault confidence and comprehensive evaluation results of multi-source uncertainty, a fault assessment report containing fault type, location, severity, confidence interval and reliability score is generated.

[0147] In this embodiment, the physical correlation between multiple sensors is used for mutual inspection to realize the health status assessment of the sensor itself. Through the two-layer causal reasoning framework and the collaborative signal decomposition algorithm, the sensor fault and the electric energy meter fault are effectively separated. The feature adaptive selection mechanism based on the working condition solves the problem of feature sensitivity changes under different working conditions. The fault verification method based on physical constraints ensures that the fault separation result conforms to the physical working principle of the electric energy meter.

[0148] In this embodiment, the process of cross-validation and self-diagnosis of sensors to obtain credibility scores is specifically as follows:

[0149] The physical dependency relationship between sensor data in the standardized multimodal sensor data stream is analyzed, and a sensor-related constraint equation group is established, including power balance constraint (voltage × current = power), thermodynamic constraint (relationship between power consumption and temperature rise), electromechanical coupling constraint (relationship between vibration and electrical parameters), etc., and the physical constraint relationship matrix is ​​output.

[0150] Based on the physical constraint relationship matrix, the consistency check is performed on the standardized multimodal sensor data stream, and the residual value of each physical constraint equation is calculated to form a sensor consistency residual vector to quantify the degree of coordination of the data between sensors.

[0151] Improved cluster analysis is used to identify subsets of sensors that can verify each other and build redundant sensor groups. The sensor data in each group can verify or deduce each other to form a sensor group mapping table.

[0152] A constrained projection algorithm is designed to decompose the sensor consistency residual vector into the contribution of each sensor to achieve accurate positioning of abnormal sensors. Specifically, the sensitivity matrix of the residual to each sensor is constructed; a weighted pseudo-inverse operation is performed to calculate the contribution of each sensor to the total residual; and a statistical hypothesis test is applied to identify sensors that significantly contribute to the residual and output sensor abnormal contribution metrics.

[0153] For each group in the sensor group mapping table, the complementary information of the sensors in the group is extracted, and the sensor mutual information matrix is ​​constructed to quantify the information redundancy and complementarity between different sensors.

[0154] The sensor anomaly contribution metric and sensor mutual information matrix are integrated to design a credibility scoring function, including: calculating the stability of the residual contribution of each sensor in multiple time windows; analyzing the consistency between the sensor and other sensors in its redundant group; and considering the sensor's historical reliability record and environmental adaptability factors to finally output the credibility score of each sensor.

[0155] In this embodiment, the sensor health status is comprehensively evaluated and the sensor health index is calculated, specifically:

[0156] The sensor characteristic drift index, sensor frequency domain feature vector analysis results and sensor credibility score are integrated to construct a multi-dimensional feature space of sensor health to comprehensively characterize the sensor status.

[0157] An improved Dempster-Shafer evidence theory framework suitable for sensor health assessment is designed. The main improvements include: introducing a basic probability distribution function based on the physical characteristics of the sensor; designing a dynamic adjustment mechanism for feature reliability weights; and developing an evidence fusion model to output conflicting evidence optimization combination rules.

[0158] Based on historical data and expert knowledge, a typical sensor failure mode feature library is established, including: slow response feature mode, zero drift feature mode, sensitivity reduction feature mode, nonlinear distortion feature mode, and abnormal noise level feature mode. Each mode contains time domain and frequency domain feature descriptors.

[0159] The evidence fusion model is applied to fuse the features in the multidimensional feature space of sensor health, calculate the probability mass of each sensor belonging to various states, and derive the probability distribution of sensor health states.

[0160] The sensor health status probability distribution is matched with the sensor typical failure mode feature library to identify the most likely sensor abnormality type. At the same time, the normalized sensor health index is calculated based on the health status probability values, with a quantization range of 0-100, where 100 represents a completely healthy state.

[0161] In this embodiment, the process of generating the working condition adaptive feature vector based on the feature adaptive selection of the working condition is specifically as follows:

[0162] For each feature in the physical enhancement feature set, its ability to distinguish between faults and normal states is calculated under various operating conditions, and the feature-operating condition sensitivity matrix is ​​constructed by applying Fisher discriminant ratio and mutual information analysis.

[0163] Based on the operating parameters of the electric energy meter in the data segment marked with the operating condition, an operating condition pattern identifier is designed to divide the current operating condition into predefined categories (such as no load, light load, heavy load, fluctuating load, etc.), and output the current operating condition type and operating condition confidence.

[0164] For each type of working condition, a feature importance adaptive evaluation function is constructed: the basic weight is set based on the feature-working condition sensitivity matrix; the weight coefficient is adjusted in combination with the historical accuracy of fault diagnosis; the environmental adaptability factor is introduced to compensate for the impact of environmental changes on feature reliability; and the working condition adaptive weight function is output.

[0165] The working condition adaptive weight function is applied to rank the importance of the physical enhancement feature set, and a threshold-based automatic feature selection algorithm is designed to dynamically determine the best feature subset. The redundancy and complementarity between features are considered at the same time, and the working condition optimized feature subset is output.

[0166] The features in the working condition optimization feature subset are weighted and combined to construct composite features. The combination coefficient is dynamically adjusted according to the current working condition type and working condition confidence, and finally a working condition adaptive feature vector is generated.

[0167] In this embodiment, the process of tracing the fault source based on causal reasoning is specifically as follows:

[0168] Based on the fault correlation matrix and the knowledge of the electric energy meter system, a two-layer causal graph structure is constructed: the upper layer represents the causal relationship between sensor fault and measurement anomaly; the lower layer represents the causal relationship between electric energy meter fault and physical parameter anomaly; the inter-layer connection represents the observation relationship between measurement anomaly and physical parameter anomaly; the sensor-electric energy meter two-layer causal graph is output.

[0169] The conditional probability table in the sensor-electricity meter double-layer causal graph is initialized by using historical fault data and expert knowledge. Then, parameter optimization is performed based on the reliably labeled measured data to output the calibrated causal model parameters.

[0170] The sensor health index, the energy meter state feature vector and the current measurement data are integrated to form an observation evidence set as the input for causal reasoning.

[0171] Develop a bidirectional causal inference algorithm: Starting from possible sensor failures, infer the possible measurement anomalies; starting from the observed anomalies, infer the most likely potential failures; propagate probability information in a two-layer causal graph through a message passing mechanism; and output the probability distribution of fault sources.

[0172] For the candidate fault sources with higher probability in the fault source probability distribution, construct competitive hypotheses:

[0173] Assumption 1: All observed anomalies are caused by sensor failure; Assumption 2: All observed anomalies are caused by electric energy meter failure; Assumption 3: Some observed anomalies are caused by sensor failure and some are caused by electric energy meter failure. The Bayesian model selection method is used to evaluate the posterior probability of each hypothesis and output the optimal fault hypothesis and candidate fault source set.

[0174] In this embodiment, the process of fault feature separation and reconstruction is specifically as follows:

[0175] A collaborative signal decomposition model is developed, taking into account the different manifestations of sensor fault characteristics and electricity meter fault characteristics: sensor failures usually manifest as sudden jumps, drifts or increased noise; electricity meter failures usually follow specific physical evolution laws; based on this difference, a matrix decomposition framework with physical constraints is designed to output a collaborative decomposition mathematical model.

[0176] Characteristic basis function libraries are constructed for sensor failure modes and electric energy meter failure modes respectively: sensor failure basis function: including step function, ramp function, oscillation function, etc.; electric energy meter failure basis function: a dual-mode adaptive basis function library is formed by combining characteristic function families designed based on physical models.

[0177] Based on the collaborative decomposition mathematical model and the dual-mode adaptive basis function library, a signal decomposition algorithm with sparse constraints is designed: an optimization objective function is constructed, including data fitting terms, sparse constraint terms and physical consistency constraint terms; an accelerated proximal gradient descent algorithm is used to solve the optimization problem; the decomposition accuracy and physical rationality are balanced through adaptive adjustment of regularization parameters; and the signal decomposition results are output, including sensor fault components and electric energy meter fault components.

[0178] The contribution ratio of the sensor fault component and the electric energy meter fault component to the original signal is calculated, and a component energy distribution diagram is generated to intuitively show which type of fault is dominant.

[0179] The sensor fault component is eliminated and only the meter fault component is retained. Combined with the normal signal baseline, the pure meter status signal is reconstructed to eliminate the interference introduced by the sensor fault.

[0180] In this embodiment, the process of fault verification based on physical constraints includes:

[0181] Based on the working principle of the electric energy meter, a set of physical constraint models of the electric energy meter is established, including: electrical model, specifically the relationship between voltage, current, power, and power factor; thermal model, specifically the relationship between power consumption, temperature rise, and heat dissipation; mechanical model, specifically the vibration transmission and resonance characteristics; output a set of physical constraint equations that can be quantified and verified.

[0182] Substitute the pure electricity meter state signal into the physical constraint equation group, calculate the physical constraint satisfaction, generate a physical consistency score, and quantify the physical rationality of the reconstructed signal.

[0183] According to the current working condition type, the parameters and thresholds in the physical constraint equation group are dynamically adjusted to adapt the constraint test to different working conditions, and the working condition adaptive constraint test results are output.

[0184] For the fault characteristics identified in the fault components of the electric energy meter, check whether they conform to the physical manifestation laws of the corresponding fault type and generate a physical rationality score for the fault characteristics.

[0185] The physical consistency score, the working condition adaptive constraint inspection results and the physical rationality score of the fault characteristics are integrated to generate the final fault verification result, including the verification pass status and credibility assessment.

[0186] In this embodiment, the redundant information sensor group identification process is specifically as follows:

[0187] Introduce the distance metric enhanced by physical knowledge: d_phys(S_i, S_j) = w_1*d_corr(S_i, S_j) +w_2*d_info(S_i, S_j) + w_3*d_constraint(S_i, S_j); where d_corr is the temporal correlation distance, d_info is the mutual information distance, and d_constraint is the physical constraint deviation distance

[0188] Using the physical constraint relationship matrix to guide the initial center selection: identify the sensor subsets that are interrelated in the physical constraint equation; select a representative sensor from each associated subset as the initial cluster center; ensure that the initial clusters have physical meaning rather than purely statistical associations;

[0189] Design a cluster number evaluation function based on physical redundancy: CR(k) = λ_1*WSS(k) + λ_2*BSS(k) +λ_3*PRI(k); WSS is the intra-group sum of squares, BSS is the inter-group sum of squares, and PRI is the physical redundancy index; automatically determine the optimal number of groups by minimizing CR(k);

[0190] When working, first read the standardized multimodal sensor data stream (standard data) and the physical constraint relationship matrix; calculate the physical correlation distance matrix between all sensor pairs: obtain the n×n sensor physical correlation distance matrix D.

[0191] Determine the initial cluster center based on the physical constraint relationship matrix: perform principal component analysis on the physical constraint relationship matrix; select the sensor with the largest contribution of each principal component as the candidate center; apply the maximum and minimum distance criteria to optimize the candidate center set; and obtain the initial cluster center set C_init.

[0192] Use the initial cluster center set C_init as the starting cluster center and iterate: (a) assign each sensor to the center with the closest distance; (b) reselect a new center in each group that minimizes the total distance; (c) stop if the center no longer changes or the maximum number of iterations is reached; output the initial division of the sensor group G_init.

[0193] Calculate the internal physical redundancy PR_i for each group Gi; adjust the critical group and merge the low-redundancy group; split the overly large and highly heterogeneous groups; verify whether each group meets the minimum redundancy requirement. Obtain the optimized sensor group G_opt and sensor group mapping table.

[0194] Match each group with a physical constraint relationship matrix; confirm that the sensors in each group constitute a meaningful physical association unit; adjust the groups that cannot be physically explained; output sensor redundant groups and mapping tables with physical meaning

[0195] In this embodiment, the abnormal sensor positioning process based on constraint projection is specifically as follows:

[0196] Discretize the m constraint equations in the physical constraint relationship matrix into a time series form;

[0197] For time point t, the jth constraint equation is expressed as: f_j(S_1(t), S_2(t), ..., S_n(t)) = 0; where S_i(t) represents the measurement value of the i-th sensor at time t;

[0198] For each time point t and each constraint equation j, calculate the residual after substituting the actual measurement value: r_j(t) = f_j(S_1(t), S_2(t), ..., S_n(t)); ideally, r_j(t) should be close to zero; deviation from zero indicates constraint violation; for each time window W, collect the residuals of all constraint equations at all time points in the window; construct an m-dimensional residual vector r(W), where each element is the average residual of a constraint equation in the window W; and finally form the sensor consistency residual vector R = [r_1(W), r_2(W), ..., r_m(W)]^T;

[0199] The process of the constrained projection algorithm is:

[0200] Obtain the physical constraint equations and current sensor measurements, calculate the partial derivative of each constraint equation for each sensor: J_ij = df_i / dS_j|current measurement point; obtain the m×n sensitivity matrix J.

[0201] Construct the weighted pseudo-inverse projection matrix: P = (J^T·W·J)^(-1)·J^T·W; where W is the diagonal matrix of the reliability weights of the constraint equation;

[0202] Perform a residual projection δS = P·R; where δS represents the estimated sensor measurement error vector;

[0203] Calculate the standardized score Z_i = |δS_i| / σ_i for each sensor error; where σ_i is the historical error standard deviation of the i-th sensor;

[0204] Apply the improved CUSUM algorithm to obtain the cumulative score C_i(t) = max(0, C_i(t-1) + Z_i(t) - k); where k is the sensitivity parameter; and obtain the cumulative anomaly score C_i of each sensor.

[0205] Set adaptive threshold T_i = α·σ_C_i + β; Identify abnormal sensors. If C_i > T_i, mark sensor i as abnormal; Calculate abnormal contribution A_i = (C_i - T_i) / T_i, if C_i > T_i; A_i = 0; if C_i ≤T_i; Sensor abnormal contribution metric A = [A_1, A_2, ..., A_n]^T.

[0206] Initialize all sensors to normal state; repeat until the residual is small enough or no new anomalies are detected: (a) recalculate the residual based on the current normal sensor subset; (b) perform residual projection to identify new abnormal sensors; (c) update the sensor state set; (d) verify the remaining residual after the abnormal sensor is removed; obtain the final sensor anomaly contribution metric and the abnormal sensor set.

[0207] In this embodiment, the process of decomposing the residual vector is specifically as follows:

[0208] Group the m constraint equations according to the involved sensor subsets; form k relatively independent constraint groups G_1, G_2, ..., G_k; generate independent residual subvectors for each constraint group; construct a local sensitivity matrix J_i for each constraint group G_i; perform local residual projection δS_i = P_i·R_i; merge the projection results of each group, considering the correlation between groups;

[0209] Calculate residual vectors at multiple time scales (short-term, medium-term, and long-term); short-term residuals focus on sudden anomalies, while long-term residuals focus on drift anomalies; fuse multi-scale analysis results to form a comprehensive anomaly assessment;

[0210] Introduce a sparsity regularization term and assume that most sensors work normally; add a smoothness constraint to suppress the oscillation of anomaly detection; optimize the objective function min ||R - J·δS|| 2 _W + λ1||δS||1 + λ2||D·δS|| 2 .

[0211] Case 2 includes the following steps:

[0212] In this embodiment, a three-phase smart energy meter deployed in a smart distribution network is taken as the research object, and five types of sensors are configured: voltage sensor (range 0-300V, accuracy ±0.2%), current sensor (range 0-100A, accuracy ±0.5%), temperature sensor (range -20℃~100℃, accuracy ±0.5℃), vibration sensor (range 0-20g, accuracy ±2%) and acoustic sensor (range 30-130dB, accuracy ±3dB).

[0213] The data acquisition system uses a 24-bit high-precision AD converter with a 10kHz sampling frequency. To ensure data synchronization, a GPS-based time synchronization strategy is used, and the maximum time deviation between sensors is measured to be 43μs. For the original timestamp TS_original(i) of the i-th sensor, the following formula is used for alignment: TS_aligned(i) = TS_original(i) - ΔT(i); where ΔT(i) is the time deviation of the i-th sensor relative to the reference clock. After time alignment, the time consistency error of each sensor data is controlled within ±5μs.

[0214] An adaptive wavelet threshold denoising algorithm is applied to the collected raw data. For the voltage signal, the db4 wavelet basis function is used for 4-layer decomposition, and the threshold λ = σ·√(2·log(N)); where σ is the noise standard deviation, estimated by the median absolute deviation (MAD) of the highest frequency band coefficient: σ = MAD / 0.6745; N is the signal length. In actual measurement, σ is usually in the range of 0.15-0.38V. For the detected sudden abnormal values, a 7-point local median filter is used for processing x_filtered(i) =median{x(i-3), x(i-2), x(i-1), x(i), x(i+1), x(i+2), x(i+3)}; after processing, the signal-to-noise ratio is improved by 8.3dB on average.

[0215] The denoised sensor data is Z-score normalized by x_normalized = (x - μ) / σ, where μ and σ are the mean and standard deviation of the data, respectively. The calculation is based on a 20-minute sliding window and is updated every 10 minutes. For example, for the current sensor data, μ and σ on a typical working day are 43.7A and 12.5A, respectively.

[0216] By analyzing the voltage and current characteristics and identifying the working state of the energy meter, the standardized data is divided into different working condition sections. The working condition classification standards are: light load: load power <20% rated power; medium load: load power between 20%-60% rated power; heavy load: load power >60% rated power; fluctuating load: load power change rate >10% / minute;

[0217] In the measured data, the proportions of various working conditions are: light load 37.2%, medium load 42.1%, heavy load 15.6%, and fluctuating load 5.1%.

[0218] Based on historical data, the normal characteristic model of each sensor is established. Taking the current sensor as an example, its normal model includes: linearity deviation <0.8%, response time <125ms, and noise level <0.3% full scale. The training data set contains 30 days of normal operation data, totaling about 4.32×10^6 sample points.

[0219] The Cumulative Sum Control Chart (CUSUM) method is used to monitor the drift of sensor characteristics. For the j-th sensor characteristic parameter pj, the CUSUM statistic calculation formula is S^+(n) = max{0, S^+(n-1) + (x_n - μ_0 - k)} S^-(n)= max{0, S^-(n-1) + (μ_0 - k - x_n)}; where μ_0 is the normal mean of the parameter and k is the reference value (with a value of 0.5σ). When S^+ or S^- exceeds the threshold h (with a value of 5σ), the drift alarm is triggered. In actual measurements, the zero drift detection sensitivity of the voltage sensor is 0.23% of the full scale.

[0220] Perform short-time Fourier transform (STFT) on the standardized sensor data to extract spectral features. The transformation parameters are set as follows: the window function is a Hanning window, the window length is 1024 points, and the overlap rate is 50%. Construct the sensor frequency domain feature vector FV_i, which contains 5 frequency band energy ratios and 3 harmonic indicators. Taking the temperature sensor as an example, the energy proportion of the low frequency band (0-0.1Hz) is 87.3%±3.2% in normal operation, and this index rises to 96.5%±1.8% when a slow response fault occurs.

[0221] Establish the sensor-related constraint equations, including: power balance constraint P = U·I·cosφ; thermodynamic constraint ΔT = k·P·t / C + T_ambient; electromechanical coupling constraint V_vib = α·I^2 + β;

[0222] Among them, P is active power (W), U is voltage (V), I is current (A), cosφ is power factor, ΔT is temperature rise (℃), k is thermal conductivity (0.42℃·W^-1·min^-1), t is time (min), C is heat capacity, T_ambient is ambient temperature (℃), V_vib is vibration amplitude (g), α and β are equipment characteristic parameters (measured values ​​are 0.0012g·A^-2 and 0.023g, respectively).

[0223] Calculate the residuals of each physical constraint equation to form a consistent residual vector. Taking the power balance constraint as an example, the residual r_power(t) = P_measured(t) - U_measured(t)·I_measured(t)·cosφ_measured(t); during normal operation, r_power fluctuates within the range of ±2.1%.

[0224] The sensor distance matrix d_phys(S_i, S_j) =0.3·d_corr(S_i, S_j) + 0.3·d_info(S_i, S_j) + 0.4·d_constraint(S_i, S_j) is constructed using the improved physical correlation distance metric; where d_corr is the Pearson correlation coefficient distance, d_info is the mutual information distance, and d_constraint is the physical constraint deviation distance. The improved K-medoids clustering algorithm is applied to identify three main sensor groups on the measured data: Group 1: voltage, current, power (physical redundancy 0.86); Group 2: current, temperature, vibration (physical redundancy 0.73); Group 3: vibration, acoustics (physical redundancy 0.68);

[0225] Construct a sensitivity matrix J (dimension m×n, m is the number of constraints, n is the number of sensors), calculate the partial derivative of the i-th constraint on the j-th sensor J_ij = df_i / dS_j|current measurement point; take the partial derivative of the power balance constraint on the current sensor as an example, J_power,I = U·cosφ. Perform residual projection δS = P·R = (J^T·W·J)^-1·J^T·W·R; where W is a diagonal matrix, representing the reliability weight of the constraint equation (the power balance constraint weight is 0.85, and the thermodynamic constraint weight is 0.75).

[0226] Calculate the standardized anomaly score: Z_i = |δS_i| / σ_i; Apply the improved CUSUM algorithm to accumulate the anomaly score: C_i(t) = max{0, C_i(t-1) + Z_i(t) - 0.5}; When C_i exceeds the threshold T_i (set to 4.2), mark sensor i as abnormal and calculate the anomaly contribution: A_i = (C_i - T_i) / T_i; In the measured data, the anomaly contribution caused by the zero drift of the current sensor A_I = 1.75.

[0227] Based on the abnormal contribution and group consistency, the trust score Trust_i = (1 - 0.6·A_i)·(0.7·C_i + 0.3·H_i) is calculated; where C_i is the consistency measure with the redundant group and H_i is the historical reliability factor. In actual measurements, the trust score of normal sensors is usually in the range of 0.85-0.98, while the trust score of faulty sensors drops to 0.35-0.60.

[0228] The sensor characteristic drift index, frequency domain characteristics and credibility score are integrated, and the improved Dempster-Shafer evidence theory is used to build a health status model.

[0229] For the state set Θ = {normal, slightly abnormal, severely abnormal} of sensor i, a basic probability distribution function m(·) is constructed based on each feature evidence. For example, for the characteristic drift index m_drift({normal}) = exp(-κ·S_drift) when S_drift < T_drift m_drift({slightly abnormal, severely abnormal}) = 1 - m_drift({normal}); where κ is the sensitivity parameter (value 1.5), S_drift is the drift statistic, and T_drift is the threshold (value 4.2).

[0230] Fusion of multi-source evidence m_1,2(A) = [Σ_{B∩C=A} m_1(B)·m_2(C)] / [1 - Σ_{B∩C=∅} m_1(B)·m_2(C)] through Dempster combination rule;

[0231] The final calculated sensor health index Health_i = 100·m({normal}) + 60·m({slight abnormality}) +20·m({severe abnormality}); in actual measurements, the index of a healthy sensor is usually between 85-100, a slight abnormality is between 60-85, and a severe abnormality is below 60.

[0232] Perform multi-scale wavelet packet decomposition on the standardized sensor data, and use the sym6 wavelet basis function to perform a 4-layer decomposition. Extract the statistical features of each decomposition node, including energy, entropy, kurtosis, skewness, etc., and construct the initial feature set F_init (dimension is 120).

[0233] The enhanced features are calculated using the physical model of the electric energy meter, including: power factor PF = P / S = P / √(P^2+Q^2); total harmonic distortion THD = √(Σ_{i=2}^n I_i^2) / I_1; phase shift φ = arccos(P / S); load imbalance LU = max(|I_a-I_avg|,|I_b-I_avg|,|I_c-I_avg|) / I_avg; the dimension of the enhanced feature set F_enhanced is 152.

[0234] The recursive feature elimination (RFE) algorithm was combined with principal component analysis (PCA) to optimize the feature set, including: using random forest to calculate feature importance scores; iteratively removing features with importance lower than a threshold (0.01); performing PCA on the remaining features and retaining principal components with an explained variance rate of 95%; and reducing the dimension of the optimized feature vector F_opt to 28.

[0235] Calculate the Fisher discriminant ratio FDR_i,j = |μ_i,1 - μ_i,0|^2 / (σ_i,1^2 + σ_i,0^2) of each feature under different working conditions; where μ_i,1 and μ_i,0 are the means of feature i in faulty and normal samples, respectively, and σ_i,1 and σ_i,0 are the corresponding standard deviations. Construct the feature-working condition sensitivity matrix S. For example, the FDR of the total harmonic distortion (THD) feature under light load conditions is 1.82, which increases to 4.37 under heavy load conditions.

[0236] Based on the voltage, current and power factor parameters, the working condition identifier is designed. The support vector machine (SVM) classifier is used, the kernel function is RBF, the penalty parameter C=10, γ=0.1. The measured working condition identification accuracy rate reaches 96.3%.

[0237] For working condition j, the adaptive weight calculation formula of feature i is w_i,j = S_i,j·(1 + α·Acc_i,j)·(1 - β·E_i,j); where S_i,j is the feature-working condition sensitivity, Acc_i,j is the historical diagnosis accuracy gain, E_i,j is the environmental interference sensitivity coefficient, and α and β are balance parameters (taken as 0.3 and 0.2 respectively).

[0238] Based on the adaptive weights, the features are ranked and the feature subset with cumulative importance reaching 90% is selected. The number of features selected under different working conditions is: 16 for light load, 19 for medium load, 14 for heavy load, and 23 for fluctuating load.

[0239] Perform weighted fusion of selected features F_adaptive = Σ_i w_i,j·F_i;

[0240] Smooth transition when working condition changes: w_i,t = (1-α)·w_i,prev + α·w_i,current; where α is the transition rate (value is 0.3).

[0241] The improved support vector machine (SVM) algorithm is used to identify the state of the electric energy meter. The kernel function selects RBF, and the grid search is used for parameter optimization. The optimal parameters are C=8.5 and γ=0.07. The adaptive feature vector F_adaptive of the working condition is classified, and the fault probability and preliminary type judgment are output. The training set contains 2500 normal samples and 1800 fault samples (covering 8 typical faults), and the cross-validation accuracy rate reaches 89.7%.

[0242] A Bayesian network model is constructed to express the conditional dependency between the health status of the sensor and the status of the electric energy meter. The network consists of three layers: sensor layer (5 nodes), observation layer (12 nodes) and electric energy meter layer (8 nodes). The conditional probability table (CPT) represents the fault propagation relationship, for example: the conditional probability that the current sensor fault leads to the abnormal harmonic detection P (harmonic abnormality | current sensor fault) = 0.83. The initial CPT is constructed based on expert knowledge and historical data, and the parameters are optimized by maximum likelihood estimation using 1200 sets of labeled data.

[0243] Generate a fault correlation matrix R, which represents the correlation between sensor fault and electric energy meter fault, with a value range of [0,1]. For example, the correlation coefficient between voltage sensor and power circuit fault is 0.87.

[0244] Construct a two-layer causal graph structure, including:

[0245] Upper layer: represents the causal relationship between 5 sensor nodes (S1-S5) and 12 measurement anomaly nodes (M1-M12);

[0246] Lower layer: represents the causal relationship between 8 power meter fault nodes (F1-F8) and 12 physical parameter abnormal nodes (P1-P12);

[0247] Inter-layer connection: indicates the correspondence between measurement abnormal nodes and physical parameter abnormal nodes.

[0248] Initialize the conditional probability table, the probability of sensor failure affecting measurement anomaly P(M_j=abnormal|S_i=fault) = exp(-d_ij / 1.5); where d_ij is the impact distance parameter. Optimize the parameters using the expectation maximization (EM) algorithm based on 800 sets of labeled data, and the average parameter convergence error is 0.053.

[0249] The sensor health index, the energy meter state vector, and the current measurement data are integrated to form the observation evidence set E. For example, in a certain observation, the current sensor health index is 57.3, and the power factor measurement value deviates from the normal range by 3.2 standard deviations.

[0250] Design a top-down and bottom-up bidirectional message transmission mechanism: Initialize the probability distribution of each node; Top-down propagation: Infer the measurement anomaly from the sensor node m_down(M_j) = Σ_i P(M_j|S_i)·P(S_i); Bottom-up propagation: Infer the fault source from the observed anomaly m_up(S_i) = P(S_i)·Π_j [P(M_j|S_i) / P(M_j)]^I(M_j∈E); Message integration: P_new(S_i) = α·P(S_i)·Π_j m_j→i P_new(F_k) = α·P(F_k)·Π_j m_j→k; where α is the normalization factor and I(·) is the indicator function. Iterate until convergence (usually 3-5 iterations).

[0251] For high-probability fault sources, three competing hypotheses are constructed: H1: all abnormalities are caused by sensor failure; H2: all abnormalities are caused by electric energy meter failure; H3: some abnormalities are caused by sensor failure and some are caused by electric energy meter failure;

[0252] Calculate the posterior probability of the hypothesis P(H_i|E) = P(E|H_i)·P(H_i) / Σ_j P(E|H_j)·P(H_j); where P(E|H_i) is the likelihood and P(H_i) is the prior probability (set to 0.35, 0.35, 0.30 respectively). Select the hypothesis with the maximum posterior probability and generate the candidate fault source set C. In actual measurement, fault sources with a probability greater than 0.25 can be included in the candidate set.

[0253] Design a decomposition model to decompose the original signal x(t): x(t) = x_baseline(t) + x_sensor(t) +x_meter(t) + ε(t); where x_baseline is the baseline signal, x_sensor is the sensor fault component, x_meter is the electric energy meter fault component, and ε is the residual.

[0254] A dual-mode adaptive basis function library is constructed, in which the sensor fault basis functions include step functions, ramp functions, exponential functions, etc., which are used to represent sensor drift, slow response and other characteristics. The electric energy meter fault basis function is a characteristic function designed based on the physical model of the electric energy meter, which is used to represent circuit faults, component aging and other characteristics. Library size: 25 sensor fault basis functions, 32 electric energy meter fault basis functions.

[0255] Construct the optimization objective function min ||x - x_baseline - D_s·α_s - D_m·α_m|| 2 + λ1||α_s||1 + λ2||α_m||1 + λ3||Φ(α_m)|| 2 ; Where D_s and D_m are the feature dictionary matrices of sensor fault and electric energy meter fault respectively, α_s and α_m are the corresponding sparse coefficient vectors, and Φ(·) is the physical consistency constraint function. Regularization parameter settings: λ1=0.35, λ2=0.25, λ3=0.4. The accelerated proximal gradient descent algorithm is used for solution, the maximum number of iterations is set to 200, and the convergence threshold is 10^-4.

[0256] Calculate the energy proportion of each component E_sensor = ||D_s·α_s|| 2 / ||x - x_baseline|| 2 ;E_meter = ||D_m·α_m|| 2 / ||x - x_baseline|| 2 In actual measurements, for current sensor zero drift failure, the sensor component energy accounts for 87.3%; for energy meter metering chip failure, the energy meter component energy accounts for 82.1%.

[0257] Eliminate the sensor fault component and reconstruct the pure electricity meter status signal x_pure(t) = x_baseline(t) +x_meter(t); use an adaptive threshold (β=0.2) to control the degree of separation: when E_sensor / (E_sensor+E_meter) >0.8, it is considered that the anomaly is mainly caused by sensor failure; when the ratio is <0.2, it is considered that the anomaly is mainly caused by electricity meter failure.

[0258] Establish the physical constraint model set of the electric energy meter, including the electrical model: power balance constraint S 2 =P 2 +Q 2 ; Thermal model: temperature rise ΔT=k·P·t / C; Mechanical model: vibration and current relationship V_vib=f(I); A total of 9 core constraint equations are constructed.

[0259] Substitute the pure electricity meter state signal into the physical constraint equation group and calculate the constraint satisfaction C_i = exp(-|r_i| / τ_i); where r_i is the residual of the i-th constraint equation and τ_i is the tolerance parameter (ranging from 0.05 to 0.15).

[0260] The comprehensive score of physical consistency C_physics = Σ_i w_i·C_i / Σ_i w_i; w_i is the constraint importance weight. In actual measurements, the C_physics of normal signals is usually greater than 0.85, while the C_physics of abnormal signals drops to 0.35-0.75.

[0261] According to the current working condition, adjust the constraint parameters and thresholds τ_i,j = τ_i,base·(1 + δ_i,j); where τ_i,base is the base tolerance and δ_i,j is the adjustment factor of working condition j for constraint i. For example, the adjustment factor of the power balance constraint under fluctuating load conditions is 0.3.

[0262] Check the physical rationality of the fault feature and calculate the matching degree with the typical fault mode M_k = Σ_i w_i·sim(f_i, p_k,i) / Σ_i w_i; where sim(·) is the similarity function, f_i is the observed feature, and p_k,i is the typical feature of fault mode k.

[0263] The final verification result V_result = β1·C_physics + β2·C_adaptive + β3·M_best is generated by integrating the physical consistency score, the working condition adaptive constraint test results and the physical rationality score of the fault feature; where β1=0.4, β2=0.35, β3=0.25, and M_best is the matching degree of the best matching fault mode. When V_result is lower than the threshold T_V (set to 0.65), the current fault hypothesis is rejected; when V_result is higher than 0.85, the fault judgment has high credibility.

[0264] Based on the sensor health assessment results, the energy meter state feature vector and fault verification results, a dynamic decision tree structure is used to realize fault separation and identification: the sensor state is preliminarily judged based on the comparison between the sensor health index and the threshold T_H (set to 70); the possibility of the energy meter fault is evaluated based on the consistency between the energy meter state feature vector and the model prediction; the final fault type is determined by considering the component energy distribution and physical verification results; the fault confidence is calculated: Conf = α·P(F) + (1-α)·V_result, where α=0.6; and the fault type judgment and confidence are finally output.

[0265] Based on the sensor health index and calibration data, the measurement error model σ_i = σ_i,base·(1 +γ·(100-Health_i) / 100) is established; where σ_i,base is the reference uncertainty and γ is the sensitivity coefficient (valued at 2.5). In the actual measurement, the σ_base of the voltage sensor is 0.2%. When the health index is 60, the measurement uncertainty rises to 0.5%.

[0266] The Monte Carlo method was used to evaluate uncertainty propagation. Random perturbations that matched the measurement uncertainty distribution were added to the input data, and 1000 simulations were performed to calculate the statistical distribution characteristics of the model output.

[0267] The inference uncertainty index U_model = σ_output / μ_output; in actual measurements, U_model is about 0.12 under normal working conditions and rises to 0.28 when the sensor is abnormal.

[0268] Analyze the impact of ambient temperature, humidity and electromagnetic interference on the diagnosis results. Establish the environmental impact sensitivity model S_env = [dF / dT, dF / dH, dF / dE]; where dF / dT is the sensitivity of the diagnosis results to temperature. Measured data show that for every 10°C change in temperature, the diagnostic accuracy changes by an average of 3.5 percentage points. d is the partial derivative character.

[0269] Integrate various uncertainty sources to build a comprehensive evaluation framework U_total = √(w1·U_sensor 2 + w2·U_model 2 +w3·U_env 2 ); where weight coefficients w1=0.5, w2=0.3, w3=0.2.

[0270] Calculate the confidence interval of the fault diagnosis result CI = [F - k·U_total, F + k·U_total]; where k is the coverage factor (the value is 1.96, corresponding to a 95% confidence level), F is the fault diagnosis result, and U_total is the comprehensive uncertainty.

[0271] The risk level is divided into four levels: low risk: U_total < 0.15, confidence level > 0.90; medium-low risk: 0.15≤ U_total < 0.25, 0.80 < confidence level ≤ 0.90; medium-high risk: 0.25 ≤ U_total < 0.35, 0.70< confidence level ≤ 0.80; high risk: U_total ≥ 0.35, confidence level ≤ 0.70;

[0272] In the measured cases, the comprehensive uncertainty of the current sensor zero drift failure is 0.17, corresponding to medium-low risk; and the comprehensive uncertainty of the electricity meter metering chip failure is 0.23, also corresponding to medium-low risk.

[0273] Based on the fault type determination, fault confidence and comprehensive uncertainty assessment results, a fault assessment report is generated, which contains the following key information: Fault type: clearly distinguish between sensor faults and electric energy meter faults; Fault location: specific to the component level; Fault severity: divided into four levels (minor, moderate, severe, critical); Confidence interval: 95% confidence interval of the fault quantitative parameter; Reliability score: 0-100, reflecting the reliability of the diagnosis result; Maintenance recommendation: disposal recommendation based on the fault type and reliability score.

[0274] For example, the report content of a typical case is: Fault type: current sensor zero drift (sensor failure); Fault location: Phase A current sensor; Fault severity: Medium (offset value 2.7%, exceeding the allowable range of 1.5%); Confidence interval: Offset value 2.7% ± 0.46%; Reliability score: 87 (medium-high reliability); Maintenance recommendation: Schedule calibration or replacement of Phase A current sensor during the next routine maintenance.

[0275] 100 smart energy meters were selected for experiments in the smart distribution network of a provincial power company, including three-phase smart energy meters of different brands, models and ages. The data collection period is 6 months, the sampling frequency is 10kHz, the collection window is 30s / time, and the interval is 10 minutes. A comprehensive data set containing 20,736,000 sample points was constructed, including 2,500 groups of marked fault samples, covering 12 typical fault modes.

[0276] The performance of separating and identifying sensor faults and electric energy meter faults was evaluated. 1,200 groups of samples (including 600 groups of sensor faults, 400 groups of electric energy meter faults and 200 groups of mixed faults) were selected from the data set for testing. The main performance indicators are shown in Table 1:

[0277] Table 1 Fault separation and recognition performance

[0278] Fault type Number of samples Correct identification number Recognition rate (%) Average confidence Sensor failure 600 558 93.0 0.86 Energy meter failure 400 356 89.0 0.82 Mixed failure 200 165 82.5 0.78 total 1,200 1,079 89.9 0.83

[0279] Compared with traditional methods, this method has significant advantages in fault source separation and identification, as shown in Table 2:

[0280] Table 2 Comparison of fault source identification performance of different methods

[0281] method Overall recognition rate (%) False alarm rate (%) Missing reporting rate (%) Threshold judgment method 61.7 42.3 15.2 SVM Classification 74.3 23.6 11.8 Bayesian Network Method 82.1 18.2 9.5 This method 89.9 8.6 7.2

[0282] In order to verify the stability of this method under different working conditions, the fault identification performance under four typical working conditions is tested. The results are shown in Table 3:

[0283] Table 3 Fault identification performance under different working conditions

[0284] Working condition type Recognition rate (%) Sensor fault recognition rate (%) Electricity meter fault recognition rate (%) Light load 90.6 94.2 87.3 Medium Load 91.8 93.6 89.5 Overload 88.5 91.2 84.7 Fluctuating load 86.2 89.5 82.3

[0285] The results show that this method performs best under light load and medium load conditions. The performance decreases slightly under fluctuating load conditions, but the overall recognition accuracy is maintained at a high level, verifying the effectiveness of the working condition adaptation mechanism.

[0286] The uncertainty assessment of the fault diagnosis results was verified by randomly selecting 300 groups of fault samples and comparing the relationship between the predicted confidence interval and the actual error. The results show that 92.7% of the actual errors fall within the predicted 95% confidence interval, indicating the effectiveness of the uncertainty quantification method. The average uncertainty indicators of various types of faults are shown in Table 4:

[0287] Table 4 Uncertainty indexes for different fault types

[0288] Fault type Sensor uncertainty Model uncertainty Environmental uncertainty Comprehensive uncertainty Sensor zero drift 0.23 0.14 0.11 0.17 Sensor sensitivity decreases 0.25 0.18 0.13 0.19 Energy meter chip failure 0.18 0.28 0.15 0.23 Energy meter power circuit failure 0.16 0.25 0.17 0.21

[0289] The method was tested in 10 distribution stations of a provincial power company for three months, covering 1,200 smart meters. During the test, the system issued 124 fault warnings, of which 113 were accurate warnings verified on site, with an accuracy rate of 91.1%. Compared with the original system, the false alarm rate was reduced from 42.3% to 8.9%, significantly reducing the ineffective dispatch of maintenance personnel.

[0290] The application results are mainly reflected in the following aspects: improving fault prediction accuracy: from the original 62.5% to 91.1%; reducing maintenance costs: reducing by about 37.2%, saving about 215,000 yuan in maintenance costs each year; extending equipment service life: by discovering and handling potential faults in advance, extending the average service life of the energy meter by about 1.2 years; improving power supply reliability: the power outage time caused by smart energy meter-related faults was reduced by about 42.6%.

[0291] Experimental results show that this method has achieved an accuracy of 89.9% in fault source separation and identification, reducing the false alarm rate from 42.3% to 8.6%. It has strong working condition adaptation ability and uncertainty quantification ability, and can provide reliable decision support for power grid operation and maintenance management. The successful application of this method has important practical value for improving the safe and reliable operation of smart grids and reducing maintenance costs.

[0292] By constructing a double-layer causal structure of upper-layer sensor fault and lower-layer electric energy meter fault, and connecting the two layers through observation relationship, the effective separation and identification of sensor fault and electric energy meter fault is realized. This structure enables the system to consider the health status of the sensor itself and the state of the measured object at the same time, so that the real fault source can be accurately traced when facing a fault signal. Through the sensor group identification based on physical constraints and the constraint projection anomaly location algorithm, the system can make full use of the physical correlation between sensors for mutual inspection and complementation. This method not only relies on statistical correlation, but also introduces physical knowledge such as energy conservation and thermodynamic laws, making sensor anomaly detection more reliable. By decomposing the mixed fault signal into sensor fault component and electric energy meter fault component, and combining the dual-mode adaptive basis function library and physical constraint sparse decomposition, the system can distinguish which anomalies in the signal are caused by sensor faults and which are caused by real faults of the electric energy meter. This separation capability greatly improves the accuracy of fault diagnosis. By constructing the feature-operating condition sensitivity matrix and adaptive weight function, the feature importance is dynamically adjusted with the operating condition. Since the fault characteristics of the electric energy meter under different operating conditions vary greatly, static feature selection often cannot adapt to such changes. Through adaptive adjustment of multi-condition constraints and verification of fault physical characteristics, the fault separation results are combined with the physical working principle of the electric energy meter for secondary verification, effectively avoiding erroneous judgments that violate physical laws. The scientific fusion of multi-source uncertain information is achieved through the basic probability distribution function based on the physical characteristics of the sensor and the conflict evidence optimization combination rule. The traditional Dempster-Shafer evidence theory often produces counterintuitive results when faced with highly conflicting evidence, while the improved framework significantly improves the rationality of evidence fusion by introducing sensor physical characteristics weights and conflict evidence processing mechanisms. By quantifying the information redundancy and complementarity between different sensors, the system can use multi-sensor information more intelligently. Unlike simple correlation analysis, the mutual information matrix considers the nonlinear information dependency between sensors and can capture more complex data association patterns. Through the top-down and bottom-up bidirectional message passing mechanism, full information interaction between the sensor layer and the electric energy meter layer is achieved. Traditional Bayesian networks usually have only one-way reasoning, which makes it difficult to simultaneously handle the two processes of "sensor fault affecting observation results" and "inferring the fault source based on observations". The bidirectional propagation mechanism enables the system to reason about the source of the fault more comprehensively and coordinately, considering various possible causal paths. By weightedly combining the time series correlation distance, mutual information distance, and physical constraint deviation distance, a distance metric that can better reflect the true association between sensors is formed. Traditional distance metrics are mainly based on statistical similarity, ignoring the physical association between sensors, resulting in a lack of physical meaning in the clustering results. The distance metric enhanced by physical knowledge makes sensor clustering more interpretable and practical, and the constructed sensor group conforms to the physical working principle of the electric energy meter.

[0293] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for predicting faults of smart electric energy meters based on multimodal sensor fusion, characterized in that: include: Collect and preprocess multi-modal sensor raw data to obtain standard data; Based on standard data, a sensor health status model is established and the sensor working status is monitored to obtain the sensor health assessment results; Perform feature extraction and pattern recognition on standard data to generate the state feature vector of the electric energy meter; Combining the sensor health assessment results and the energy meter state feature vector, the two-layer causal graph reasoning module is used to perform two-layer fault source separation and identification, and output the fault type and confidence level. Based on the fault type and confidence level, multi-source uncertainty is quantified and reliability is evaluated to generate a fault assessment report.

2. The method according to claim 1, characterized in that The two-layer causal graph reasoning module includes: The upper layer module is used to represent the causal relationship between sensor failure and measurement anomaly. The lower module is used to represent the causal relationship between the power meter failure and the abnormal physical parameters. Interlayer connections are used to represent the observed relationship between measurement anomalies and physical parameter anomalies. When working, it infers measurement anomalies from possible sensor failures from top to bottom, and infers potential faults from observed anomalies from bottom to top. It propagates probability information in the two-layer causal graph through the message passing mechanism and outputs the probability distribution of fault sources.

3. The method according to claim 2, characterized in that The process of performing two-layer fault source isolation and identification includes: Read the observation evidence set as the input of the two-layer causal graph reasoning module to obtain the probability distribution of the fault source; Among them, the observation evidence set includes sensor health assessment results, energy meter state feature vectors and current measurement data; The fault sources with probability higher than a threshold in the fault source probability distribution are taken as candidate fault sources; Three competing hypotheses, namely sensor failure, meter failure and mixed failure, are constructed. The Bayesian model selection method is applied to evaluate the posterior probability of each hypothesis, and the optimal fault hypothesis and candidate fault source set are output.

4. The method according to claim 3, characterized in that The process of performing two-layer fault source separation and identification also includes: For each candidate fault in the candidate fault source set, the original signal is decomposed into a sensor fault component and an electric energy meter fault component, and a pure electric energy meter state signal is reconstructed; Combined with the physical working principle of the electric energy meter, verify whether the state signal of the pure electric energy meter conforms to the physical consistency law, and output the fault verification result; Combining the sensor health assessment results, the energy meter state feature vector and the fault verification results, a dynamic decision tree is used to perform hierarchical identification of sensor faults and energy meter faults to obtain the final fault type judgment and fault confidence.

5. The method according to claim 4, characterized in that The steps of reconstructing the pure electric energy meter status signal include: Construct a collaborative signal decomposition model for distinguishing different manifestations of sensor fault signatures and electric energy meter fault signatures; Construct characteristic basis function libraries for sensor faults and electric energy meter faults respectively, forming a dual-mode adaptive basis function library; Based on the collaborative decomposition model and dual-mode adaptive basis function library, the original signal is decomposed into sensor fault component and electric energy meter fault component; Calculate the contribution ratio of the sensor fault component and the electric energy meter fault component to the original signal, and generate an energy distribution diagram for evaluating the relative importance of the two types of fault components; According to the weight analysis of the component energy distribution diagram, the sensor fault component is eliminated, the separation threshold is adjusted, and the pure electricity meter status signal is reconstructed in combination with the normal signal baseline.

6. The method according to claim 4, characterized in that The steps of verifying whether the pure electric energy meter status signal conforms to the physical consistency law and outputting the fault verification result include: Based on the working principle of the electric energy meter, a set of physical constraint models of the electric energy meter including electrical model, thermal model and mechanical model is established to form a set of physical constraint equations that can be quantified and verified; Substitute the pure electric energy meter state signal into the physical constraint equation group, calculate the physical constraint satisfaction, and generate a physical consistency score; Adjust the parameters and thresholds in the physical constraint equation group according to the current working condition type, so that the constraint test adapts to the current working condition, and output the working condition adaptive constraint test results; For the fault features identified in the pure electric energy meter status signal, check whether they conform to the physical manifestation rules of the corresponding fault type and generate a physical rationality score for the fault features; The above outputs are integrated to generate fault verification results, including verification pass status and credibility assessment.

7. The method according to claim 1, characterized in that The process of establishing a sensor health status model and monitoring the sensor working status to obtain the sensor health assessment results includes: Based on the historical normal operation data in the standard data, a normal response characteristic model of the sensor is established; Compare the current sensor real-time data with the sensor normal response characteristic model, detect sensor characteristic drift, and generate sensor characteristic drift indicators; Perform frequency domain analysis on standard data, extract sensor frequency domain features, and construct sensor frequency domain feature vectors; Perform multi-sensor cross-validation to analyze the physical correlation between different sensor data and generate sensor credibility scores; The above outputs are integrated to build a sensor health status model and output sensor health assessment results, including sensor health index and abnormal type identification results.

8. The method according to claim 7, characterized in that The process of generating a sensor confidence score includes: Analyze the physical dependencies and associations between sensor data in the standard data to obtain a physical constraint relationship matrix; Based on the physical constraint relationship matrix, the consistency check is performed on the standard data, and the residual value of each physical constraint equation is calculated to form the sensor consistency residual vector; Decompose the sensor consistency residual vector into individual sensor contributions and output the sensor anomaly contribution metric; Identify subsets of sensors that can authenticate each other, build redundant sensor groups, and form a sensor group mapping table; For each group in the sensor group mapping table, the complementary information of the sensors in the group is extracted to construct a sensor mutual information matrix; The sensor anomaly contribution metric and sensor mutual information matrix are integrated, and the sensor credibility score is calculated and output through the credibility scoring function.

9. The method according to claim 7, characterized in that The process of forming the sensor group mapping table is as follows: Calculate the physical correlation distance matrix between all sensor pairs in the standard data; Determine the initial cluster center based on the physical constraint relationship matrix, perform principal component analysis on the physical constraint relationship matrix, select the sensor with the largest contribution of each principal component as the candidate center, and obtain the initial cluster center set; Using the initial cluster center set as the starting cluster center, each sensor is assigned to the nearest center, and a new center that minimizes the total distance is reselected in each group to obtain the initial division of the sensor group; Calculate the internal physical redundancy for each sensor group, adjust the critical group, and output the optimized sensor group; Each sensor group is matched with a physical constraint relationship matrix, and it is confirmed that the sensors in the group constitute a meaningful physical association unit, so as to obtain a sensor group mapping table with physical meaning.

10. The method according to claim 1, characterized in that The steps of generating the state feature vector of the electric energy meter include: Perform multi-scale wavelet packet decomposition on standard data, extract time domain, frequency domain and time-frequency joint features, and construct an initial multi-scale feature set; Using the electrical characteristics and thermal model of the electric energy meter, the initial multi-scale feature set is enhanced with physical knowledge to generate a physical enhanced feature set. Apply feature optimization algorithm and dimensionality reduction processing to the physical enhancement feature set to form an optimized electric energy meter feature vector; According to the working condition information of the electric energy meter, the optimized electric energy meter feature vector is dynamically weighted and the working condition adaptive feature vector is output; The adaptive characteristic vector of the working condition is identified to generate the characteristic vector of the electric energy meter state and the preliminary fault type determination result.

Images