Smart Electricity Meter Condition Assessment Method Based on Multi-source Data Fusion

Abstract

This paper discloses a smart electricity meter status assessment method based on multi-source data fusion. The method involves multi-source heterogeneous data acquisition and preprocessing, interactive perception dynamic alignment and spatiotemporal feature extraction, evidence flow causal reasoning analysis, and multi-dimensional status assessment and result output. This method effectively improves the accuracy and interpretability of electricity meter status assessment, addressing the technical issues of difficulty in capturing interactive features of multi-source heterogeneous data and the lack of interpretability of status assessment.

Description

Technical Field

[0001] The present invention relates to power electronics technology, and in particular to a smart electric energy meter state assessment method based on multi-source data fusion. Background Art

[0002] With the deepening of smart grid construction, the number of smart electricity meters deployed in power systems has increased dramatically. Their health status is directly related to the accuracy of electricity metering and the reliability of grid operation.

[0003] Current research on smart meter status assessment primarily focuses on analytical methods based on a single data source. Traditional methods rely primarily on periodic manual inspections or simple data threshold monitoring, lacking the comprehensive utilization of multidimensional data. In recent years, some studies have begun to adopt data-driven approaches, such as using traditional machine learning algorithms such as support vector machines and decision trees for status classification, or using statistical models such as autoregressive moving averages to analyze time series characteristics. Other studies have attempted to use artificial neural networks to establish a mapping relationship between meter parameters and status, but most methods remain limited to a single data source or simple data fusion strategies such as feature-level concatenation or weighted averaging.

[0004] However, existing technologies still suffer from two key issues: First, they cannot effectively capture the interactive features between different data sources, especially the interactions between heterogeneous data such as electrical characteristics, environmental factors, and historical operating records, resulting in inaccurate condition assessment results in complex scenarios. Second, condition assessment results lack interpretability and cannot provide the causal links and evidence flow of abnormal conditions. This makes it difficult for maintenance personnel to understand the assessment basis and locate the root cause, making actual maintenance decision-making inefficient. In particular, in scenarios with multiple concurrent faults or nonlinear causal relationships, existing methods have difficulty distinguishing between correlation and causality, resulting in inaccurate anomaly location and an inability to support the formulation of targeted maintenance measures. Summary of the Invention

[0005] The purpose of the invention is to provide a smart electricity meter status assessment method based on multi-source data fusion, in order to solve at least some of the problems existing in the prior art.

[0006] The technical solution provides a method for evaluating the status of smart electricity meters based on multi-source data fusion, which includes the following steps:

[0007] Collect basic data, perform quality assessment, standardization, extract temporal and spatial correlation features, and obtain preprocessed data sets;

[0008] The preprocessed dataset is input into the attention module to extract the data source interaction features, calculate the periodic attention features and residual enhancement features, and generate the fused spatiotemporal features through multi-channel spatiotemporal attention fusion;

[0009] Based on the fusion of spatiotemporal features, an evidence flow network is constructed. Through evidence flow propagation and posterior reasoning, causal strength is evaluated and strong causal links are screened. A strong causal link set and a diagnostic report are generated. Combined with the fusion of spatiotemporal features, multi-dimensional status evaluation indicators are calculated, multi-level status classification and anomaly location are implemented, and comprehensive evaluation results are obtained.

[0010] Preferably, generating fused spatiotemporal features includes:

[0011] Read the preprocessed data set, perform time and space alignment in sequence, generate a time-space aligned data set, and calculate the interaction feature matrix through the interaction operator of the data source pair to form an interaction feature set;

[0012] Applying the period-aware hierarchical attention mechanism to the spatiotemporally aligned dataset, we calculate the intra-day, inter-day, inter-week, and inter-month multi-level attention and generate periodic attention features.

[0013] Based on the periodic attention feature, the normal periodic behavior pattern is predicted, the residual with the actual observation value of the spatially aligned dataset is calculated, the multi-scale residual is decomposed and the attention weight is adjusted to form the residual enhancement feature;

[0014] Combining the above outputs, including periodic attention features, residual enhancement features and interaction feature sets, the weights of each channel are adjusted to generate fused spatiotemporal features.

[0015] Preferably, forming an interactive feature set includes:

[0016] Based on the spatially aligned dataset, a data source pair set and a type mapping matrix are constructed;

[0017] According to the type mapping matrix, the correlation index is calculated for different types of data sources to form a correlation matrix; and the mutual information, conditional entropy and transfer entropy are calculated to form an information theory feature matrix;

[0018] Perform Granger causality test on each pair of data sources to generate a Granger causality matrix;

[0019] Based on the above outputs, including the correlation matrix, information theory feature matrix and Granger causality matrix, interaction operators for electrical quantities, environmental quantities and electrical quantities, and events and measurements are constructed to obtain interaction operators and selection strategies. The interaction operators and selection strategies are then applied to process the set of data source pairs to form an interaction feature set.

[0020] The beneficial effect effectively improves the accuracy and interpretability of electricity meter status assessment, and solves the technical problems of difficulty in capturing interactive features of multi-source heterogeneous data and lack of interpretability of status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the present invention.

[0022] Figure 2 It is a flow chart of the present invention for generating an integrated fusion of spatiotemporal features.

[0023] Figure 3 It is a flow chart of forming an interactive feature set according to the present invention.

[0024] Figure 4 It is a flow chart of forming residual enhancement features of the present invention. DETAILED DESCRIPTION

[0025] like Figures 1 to 4 As shown in FIG, the smart energy meter status assessment method based on multi-source data fusion includes:

[0026] In this embodiment, the basic data includes real-time data of the electric energy meter, grid operation data, environmental data and historical operation records.

[0027] In another embodiment of the present application, specifically:

[0028] S1. Multi-source heterogeneous data collection and preprocessing

[0029] S11. Read the real-time data of the electricity meter (including voltage, current, power factor, active power, reactive power and other measurement values), collect the grid operation data (including grid frequency, phase, harmonic content and other parameters), obtain environmental data (including temperature, humidity, vibration, etc.), and historical operation records (including historical abnormal events, maintenance records, etc.), to form the original multi-source data set.

[0030] S12. Perform a quality assessment on the original multi-source dataset to identify missing values, outliers, and noise. Use a time-windowed conditional mean interpolation method to handle missing values, use an IQR (interquartile range)-based local anomaly detection method to mark outliers, and apply a wavelet transform to remove noise. This yields a quality assessment indicator set Q and a preliminary cleaned dataset.

[0031] S13. For different data sources in the preliminary cleaned data set, corresponding standardization methods are used for processing: continuous measurement data is standardized using Z-score, periodic data is normalized using maximum-minimum normalization, and discrete event data is encoded using one-hot encoding to obtain a standardized data set.

[0032] S14. Segment the standardized dataset into multiple time scales (hours, days, weeks, and months) to extract temporal features. For each time segment, calculate statistical features (mean, variance, skewness, kurtosis), trend features (slope, inflection point), and periodic features (spectral components, autocorrelation coefficient) to form a temporal feature set.

[0033] S15. Based on the geographic location and grid topology, a spatial correlation network of electric energy meters is constructed. The correlation coefficients, phase differences, and load distribution characteristics between adjacent electric energy meters are calculated, and the spatial correlation patterns are extracted to obtain a spatial feature set.

[0034] S16. Combine the temporal feature set and the spatial feature set with the standardized data set for preliminary fusion, and use dimensionality reduction processing based on principal component analysis to remove redundant information and retain key features, and finally generate a preprocessed data set as input for subsequent analysis.

[0035] S2. Interactive perception dynamic alignment spatiotemporal feature extraction

[0036] S21, read the preprocessed data set and calculate the time delay correlation function C between different data sources i,j (τ), quantifies the time lag between data sources i and j. For each pair of data sources (i, j), based on the principle of maximizing mutual information, the optimal time alignment parameter τ* is solved. i,j = argmax τ I(X i (t), X j (t+τ)), where I represents the mutual information function. The optimal time alignment parameters are applied to correct the original data to obtain the time-aligned dataset.

[0037] S22. Construct a unified spatial reference frame for the time-aligned dataset. Design the spatial mapping function S i,j (·) maps spatially distributed data sources into a unified coordinate system. For nodes in the power grid topology, a mapping relationship is established based on electrical distance and physical location. For environmental sensor data, spatial weights are determined based on sensor location and impact range. This ultimately forms a spatially aligned dataset, achieving consistent representation of multi-source data in the spatial dimension.

[0038] S23. Based on the spatially aligned dataset, we design a data source pair interaction operator Θ to capture the mutual influence and interaction features between different data sources. For any data source pair (i, j), we calculate the interaction feature matrix M i,j = X' i Θ X' j , where X' i and X' j is the aligned data source. The specific implementation of the interaction operator Θ includes: correlation coefficient calculation, mutual information measurement, conditional entropy analysis and Granger causality test. The interaction features of all data source pairs are summarized to form the interaction feature set M interact .

[0039] S24. Apply the period-aware hierarchical attention mechanism to the spatially aligned dataset. Divide the time series into segments according to predefined periods (hours, days, weeks, months), calculate the relative position of each time point at each period level, and design a dedicated period encoding. Calculate the self-attention matrix for each period level: intra-day attention A hour (t) = softmax(Q h K h T / √d k )V h ; Temporal Attention A day (t) = softmax(Q d K d T / √d k )V d ; Weekly Attention A week (t) = softmax(Q w K w T / √d k )V w ; Monthly Attention A month (t) = softmax(Q m K m T / √d k )V m ; By learning weight α h , α d , α w , α m The attention results at different levels are integrated to obtain the periodic attention features.

[0040] S25. Predict normal periodic behavior patterns based on periodic attention features and calculate the residual r between the actual observed value and the predicted value t = x t – x* t Apply wavelet transform to the residual signal, decompose it into different frequency components, and calculate the residual energy E of each scale j = Σ t |r j,t | 2 . Design dynamic threshold function τ j (t), when the residual signal is detected to exceed the threshold, the abnormal enhancement mechanism is activated and the attention weight is adjusted to form the residual enhancement feature A residual .

[0041] S26. Design dedicated attention channels for original features and interactive features respectively. Original feature attention A raw Combining periodic attention features and residual enhancement features Aresidual ; Interactive Feature Attention A inter Based on the interactive feature set M interact Calculation. Design a context-aware fusion device, adaptively adjust the weights of each channel according to the current state, and calculate the final fusion feature A final =γ1(c)A raw + γ2(c)A inter , where c represents the current context vector, γ1(·) and γ2(·) are context-dependent weight functions. Finally, the fused spatiotemporal feature F is generated. final , as a key input for status assessment.

[0042] S3. Evidence Flow Causal Reasoning Analysis

[0043] S31. Based on the structure and working principle of the electric energy meter, design the node set V = {v1, v2, ..., v n}, representing the key component states and environmental factors of the energy meter. Design the directed edge set E = {e i,j}, representing the causal relationship between components. Construct the initial directed acyclic graph (DAG) G = (V, E) as the initial causal network G init .

[0044] S32, the initial causal network G init Expanded to the evidence flow network G f = (V, E, W, F), where W is the set of edge weights, F = {f1, f2, ..., f l} is a set of evidence flows flowing through the network. Each evidence flow carries information and uncertainty index: f i = (info i , uncertainty i ). Define the evidence stream merging operator: f i ⊕ f j = g(info i , info j ,uncertainty i , uncertainty j ), which is used to simulate the process of evidence propagation and merging in the network. flow .

[0045] S33, based on historical data and fusion of spatiotemporal features F final , learn the evidence flow network G f Conditional probability table (CPT) in low: P(v i | pa(v i )), where pa(v i) is the node v i The parent node set of . For data sparse areas, the prior probability is set in combination with expert knowledge. Calculate the joint probability distribution of the complete network: P(V) = ∏ i=1 n P(v i | pa(v i )), and get the network parameter set Θ.

[0046] S34. Inject the observation data as evidence E = e into the evidence flow network G flow , triggering the evidence flow propagation process. For each observed variable, an evidence flow is created and propagated along the network edges. When multiple evidence flows converge at a node, a merge operator is applied to fuse them. This process updates the network state and calculates the posterior probability distribution: P(V | E = e). Identify the most likely state path: S* = argmax S P(S | E = e), generates the state reasoning result S i nfer.

[0047] S35, the state reasoning result S i For each causal link in nfer, we apply intervention theory to evaluate causal strength. We calculate the causal strength metric: CS(A→B) = P(B | do(A)) - P(B), where do(A) represents the intervention operation on variable A. We further verify the correlation and causality using counterfactual analysis: CF(A→B) = P(B | A) - P(B | ¬A). We set a causal strength threshold τ. cs , filter strong causal relationships and form a strong causal link set C Strong .

[0048] S36, based on strong causal link set C Strong and the evidence flow network G flow , generating a three-layer explanation architecture:

[0049] Overview layer: extracts 2-3 key abnormal factors and their impact paths; Detailed layer: displays the complete fault propagation chain and the strength of evidence in each link; Expert layer: provides complete probability analysis and hypothesis verification tools to design a color coding system,

[0050] The intensity and uncertainty of evidence are expressed through color. Combined with natural language templates, professional interpretation text is generated, and finally an interpretable diagnostic report is formed. e xplain.

[0051] S4. Multi-dimensional status assessment and result output

[0052] S41. Integration and fusion of spatiotemporal features F final and the set of strong causal links CStrong , build a multi-dimensional state assessment framework. Design a weighted fusion algorithm, dynamically adjust the weight ratio of spatiotemporal features and causal reasoning according to the needs of different scenarios, and generate a comprehensive assessment feature F combined .

[0053] S42, based on comprehensive evaluation feature F combined , calculates a variety of status evaluation indicators: health index (HI): quantifies the overall health status of the electricity meter; risk factor (RF): evaluates the risk of potential failure and its severity; remaining useful life (RUL): predicts the remaining reliable working time of the electricity meter; measurement accuracy index (MAI): evaluates the metering accuracy level of the electricity meter; these indicators are combined to form the status evaluation indicator set I assessment .

[0054] S43, according to the state evaluation index set I assessment And preset thresholds, the electricity meter status is divided into multiple levels (such as normal, slight abnormality, moderate abnormality, serious fault, etc.). For abnormal status, combined with the strong causal link set C Strong Perform anomaly location, determine the specific component or subsystem of the anomaly, and generate the status classification result S class and the abnormal location result L anomaly .

[0055] S44, based on the state classification result S class , abnormal location result L anomaly and interpretable diagnostic reports R e xplain, designed with intuitive visual interfaces. Includes: status dashboard: displays key status indicators and their historical trends; anomaly map: highlights the location of abnormal components in the electricity meter structure; causal chain diagram: interactively displays the fault propagation path; time evolution view: displays the change process of status over time; generates the final visual status report V report , providing maintenance personnel with comprehensive status assessment information.

[0056] S45, according to the state evaluation index set I assessment and the set of strong causal links C Strong , combined with the maintenance experience knowledge base, generate targeted maintenance decision suggestions. For different types of abnormalities, provide corresponding processing solutions, maintenance steps or replacement suggestions to form decision suggestions D advice , assisting maintenance personnel in making efficient decisions.

[0057] S46, visualize the status report V report and decision-making recommendations D advice Integrate into the final comprehensive evaluation result R final, displayed to maintenance personnel through the user interface, or transmitted to the upper management system through the API interface to complete the entire status assessment process.

[0058] According to one aspect of the present application, S23, extracting interactive features from data sources, includes:

[0059] S231. Read the spatially aligned dataset and construct a data source index set I = {1, 2, ..., N}, where N is the total number of data sources. Based on the data type (continuous / discrete) and physical meaning (electrical quantity / environmental quantity / event record), establish a type mapping matrix T for each pair of data sources (i, j), where T[i, j] represents the interaction type between data sources i and j. Output the data source pair set P = {(i, j) | i, j ∈ I, i ≠ j} and the type mapping matrix T.

[0060] S232. For each pair of data sources (i, j) in the data source pair set P, extract the corresponding time series X from the spatial alignment dataset. i and X j Select the appropriate correlation measurement method based on the type in the type mapping matrix T: Pearson correlation coefficient is used for continuous-continuous pairs; point biserial correlation is used for continuous-discrete pairs; Cramer's V coefficient is used for discrete-discrete pairs. Calculate the correlation index of all data source pairs to form the correlation matrix C corr , where C corr [i,j] represents the correlation strength between data sources i and j.

[0061] S233. For each pair of data sources (i, j) in the data source pair set P, calculate the information theory characteristics according to the type mapping matrix T: for continuous data, apply the adaptive binning technique to discretize the data, and the number of bins is determined according to the Freedman-Diaconis rule; calculate the mutual information I(X i ; X j ) = ∑∑ p(x i ,x j ) log(p(x i ,x j ) / (p(x i )p(x j ))); Calculate the conditional entropy H(X i |X j ) = -∑∑ p(x i ,x j ) log(p(x i |x j )); calculate the transfer entropy TE(X i →X j ) = ∑∑ p(x j^t,x j ^{t-1},x i ^{t-1}) log(p(x j ^t|x j ^{t-1},x i ^{t-1}) / p(x j ^t|x j ^{t-1})); Output information theory feature matrix I i nfo, which contains three sub-matrices: mutual information, conditional entropy and transfer entropy.

[0062] S234. Read the spatially aligned dataset and data source pair set P, and perform a Granger causality test on each pair of data sources (i, j): Construct an autoregressive model AR j :X j ^t = ∑_{k=1}^pa k X j ^{tk} + ε t ; Build model AR containing data source i ij :X j ^t = ∑_{k=1}^pa k X j ^{tk} + ∑_{k=1}^pb k X i ^{tk} + ε t ; Compare the performance of the two models by F test or likelihood ratio test and calculate the p value; if the p value is less than the significance level α (default 0.05), then X i To X j There is Granger causality; repeat the above process for all data source pairs to generate the Granger causality matrix G cause , where G cause [i,j] represents the Granger causality strength from i to j.

[0063] S235, based on the correlation matrix C corr , information theory characteristic matrix I i nfo and Granger causal matrix G cause , design the data source interaction operator Θ: define the mathematical form of the interaction operator: X i Θ X j = f(X i , X j , C corr [i,j], I i nfo[i,j], G cause[i,j]); for the interaction between electrical quantities, we focus on phase difference, harmonic relationship and power factor, and design the electrical interaction operator Θ e For the interaction between environmental quantities and electrical quantities, we focus on the effect of temperature on resistance and the effect of humidity on insulation performance, and design the environmental interaction operator Θ env For the interaction between events and measurements, we design an event impact attenuation function based on a time window and construct an event interaction operator Θ evt ; Automatically select appropriate sub-operators according to data types and comprehensively calculate interaction features; Output interaction operator definition Θ = {Θ e , Θ env , Θ evt} and select strategy S Select .

[0064] S236. For each pair of data sources (i, j) in the data source pair set P, apply the interaction operator Θ to calculate the interaction feature matrix: extract the time series X of data sources i and j from the spatially aligned dataset i and X j ; According to the type mapping matrix T and selection strategy S Select Select appropriate sub-operators; calculate the interaction feature matrix M i,j = X i Θ X j Extract the main features of the interaction feature matrix, including eigenvalue distribution, principal components and singular values; integrate the interaction features of all data source pairs to form a complete interaction feature set M i nteract, serving as an important input for subsequent analysis.

[0065] According to one aspect of the present application, S25, residual signal analysis and anomaly enhancement, includes:

[0066] S251. Read the periodic attention features and build a multi-period prediction model: for each time point t, extract the position information at each period scale: hour position h t , intraday position d t , position w within the week t , position m within the month t ; Based on the cycle encoding and attention features, the prediction function of each cycle scale is constructed: f h (h t ), f d (d t ), f w (w t ), f m (m t ); Use weighted fusion to generate the final prediction value: x* t = α h ·fh (h t ) + α d ·f d (d t ) + α w ·f w (w t ) + α m ·f m (m t ); the weight coefficient α is automatically optimized by minimizing the prediction error on historical data; the output prediction value sequence X pred and the prediction model parameters Θ pred .

[0067] S252, based on the predicted value sequence X pred And the actual observations, calculate the residual sequence: read the actual observations X in the spatial alignment dataset obs ; Calculate the residual sequence: r t = X obs [t] - X pred [t]; Apply multi-resolution wavelet transform to decompose the residual sequence:

[0068] Select an appropriate mother wavelet function ψ(t) (Daubechies-4 wavelet is used by default); perform J-layer wavelet decomposition (J is 4 by default); obtain the wavelet coefficients of each scale: {W1, W2, ..., W j , V j}; where W j Represents the detail coefficient of the jth layer, V j Represents the approximate coefficients of the Jth layer; calculates statistical characteristics of the wavelet coefficients of each layer, including energy, peak, variance and kurtosis; outputs the residual sequence R and the multi-scale wavelet coefficient set W = {W1, W2, ..., W j , V j}.

[0069] S253: Multi-scale energy analysis performs energy analysis on the multi-scale wavelet coefficient set W:

[0070] Calculate the residual energy at each scale: E j = ∑ t |W j [t]| 2 , j = 1,2,...,J; calculate the approximate coefficient energy: E a = ∑ t |V j [t]| 2 ; Construct the energy distribution vector: E = [E1, E2, ..., E j , E a]; Calculate the normalized energy distribution: E norm = E / ∑E; compared with the reference energy distribution E_ref under normal operating conditions, calculate the energy deviation ΔE = E norm - E_ref; Output energy analysis results {E, E norm , ΔE}, as the basic features for anomaly detection.

[0071] S254. Design a dynamic threshold function based on historical data and current operating status:

[0072] Read the energy analysis result set in the historical operation data and extract the statistical characteristics of energy deviation; for each scale j, fit the energy deviation ΔE j Probability distribution model (generalized extreme value distribution GEV is used by default); based on the expected false alarm rate FPR t arget (default 0.001), calculate the initial threshold: τ j ^0 = GEV j ^{-1}(1-FPR t arget);

[0073] Introduce a dynamic adjustment factor λ(t) to consider time, environment and load factors: λ(t) = f(time f actors,env f actors, load factors );time factors Including time period, date type (weekday / weekend / holiday); env f Actors include environmental factors such as temperature and humidity; load factors Including load level, volatility, etc.; calculate the final dynamic threshold: τ j (t) = τ j ^0 · λ(t); Output dynamic threshold function set τ = {τ1(t), τ2(t), ..., τ j (t)}.

[0074] S255, based on the residual sequence R, the multi-scale wavelet coefficient set W and the dynamic threshold function set τ, perform anomaly detection: for each time point t and scale j, calculate the absolute value of the wavelet coefficient |W j [t]|; Calculate the normalized deviation: δ j (t) = |W j [t]| / τ j (t); take the maximum normalized deviation: δ max (t) = max j (δ j(t)); set the detection threshold γ (default is 1.0), if δ max (t)>γ, it is marked as an abnormal point; calculate the severity of the abnormality: s(t) = log(1 + δ max (t) - γ), if δ max (t) ≤γ, then s(t) = 0; output abnormality detection result {δ max (t), anomaly flag (t), s(t)}.

[0075] S256. Based on the anomaly detection results, design an anomaly enhancement factor:

[0076] Define the basic enhancement function: β0(t) = sigmoid(α·s(t) - θ); α is the enhancement strength coefficient (default is 5.0); θ is the enhancement threshold parameter (default is 0.5); sigmoid(x) = 1 / (1+e^{-x}) is the S-type activation function; considering the impact of anomaly duration, introduce a time accumulation factor: calculate the anomaly persistence in the past window: d(t) = ∑_{i=tw}^{t} anomaly f lag(i) / w; w is the sliding window size (default is 24); introduce anomaly mutation quantitative indicators: calculate the rate of change of anomaly severity in the past window: v(t) = (s(t) - s(tw)) / w; mutation coefficient: jump(t) = max(0, tanh(v(t))); comprehensively calculate the final enhancement factor: β(t) = β0(t) · (1 + c1·d(t) + c2·jump(t)); c1 is the persistence weight (default is 0.3); c2 is the mutation weight (default is 0.5); output the anomaly enhancement factor sequence β for dynamic adjustment of the subsequent attention mechanism.

[0077] According to one aspect of the present application, S32, evidence flow network modeling, includes:

[0078] S321, based on the initial causal network G init , refine the node definition:

[0079] Identify the core components of the energy meter, including: metering chip (MS), voltage sensor (VS), current sensor (CS), clock circuit (TC), communication module (CM), display unit (DU), power management unit (PM); define environmental factor nodes: temperature (TE), humidity (HU), electromagnetic interference (EMI); define grid-related nodes: grid voltage stability (GV), grid frequency stability (GF), harmonic content (HC); define load-related nodes: load level (LL), load change rate (LR), load type (LT); define status output nodes: measurement accuracy (MA), communication reliability (CR), display correctness (DC), overall health status (HS); output detailed node set V = {v1, v2, ..., v n} and node type mapping T node .

[0080] S322. Based on domain knowledge and the working principle of the electricity meter, design the causal relationship between nodes:

[0081] Define the set of direct causal relationship edges E direct :Sensor to meter: (VS→MS), (CS→MS); Environmental factors to components: (TE→MS), (TE→VS), (TE→CS), (HU→MS), (EMI→CM); Grid factors to sensors: (GV→VS), (GF→TC), (HC→CS); Component to state: (MS→MA), (CM→CR), (DU→DC); Combine the node set V and the edge set E direct , build the basic DAG graph structure; introduce edge weight W base , represents the a priori strength of causal relationship, and the initial value is set based on expert knowledge; the output edge set E = {e i,j} and basic edge weight W base .

[0082] S323. Design evidence flow and its attributes:

[0083] Define the evidence flow data structure: f = (ID, type, source, strength, uncertainty, path h istory); ID: unique identifier; type: evidence type (direct measurement, indirect inference, expert knowledge); source: evidence source node; strength: evidence strength (0-1); uncertainty: uncertainty measure (0-1); path history: transmission path history; Design evidence types and characteristics: Direct evidence: direct measurements from sensors or monitoring equipment, with low uncertainty; Indirect evidence: indirect evidence obtained through reasoning or calculation, with medium uncertainty; Prior evidence: prior information based on expert knowledge or historical patterns, with high uncertainty; Initialize the evidence flow set F init , including the initial evidence flow of all observation nodes; output the evidence flow structure definition and the initial evidence flow set F init .

[0084] S324. Design an evidence stream merging operator ⊕ to handle the fusion of multiple evidence streams at a node: Based on the Dempster-Shafer evidence theory framework, design a basic trust assignment (BPA) conversion function: convert the strength and uncertainty of the evidence stream f into BPA: m(A), m(¬A), m(Θ); m(A) = strength·(1-uncertainty): evidence supporting hypothesis A; m(¬A) = (1-strength)·(1-uncertainty): evidence supporting hypothesis ¬A; m(Θ) = uncertainty: uncertain evidence; design Dempster combination rules to handle high-conflict evidence: for evidence stream f i and f j , calculate the conflict degree K = ∑_{X∩Y= } m i (X)·m j (Y); if K <K threshold (default 0.7), use the standard Dempster combination rule; if K ≥ K threshold , use the PCR5 rule to redistribute conflicting evidence; define the uncertainty calculation method of the merge result: uncertainty combined = g(uncertainty i , uncertainty j , K); g is the uncertainty propagation function, taking into account the original uncertainty and conflict degree; define the complete merging operator: f i ⊕ f j = combine d S(f i , f j ,K threshold ); output the evidence merging operator ⊕ and its parameter settings.

[0085] S325. Adding dynamic features to the evidence flow network:

[0086] Design the propagation attenuation function of the evidence flow on the edge: define the attenuation factor α(ei,j ), represents the attenuation degree of the evidence propagating from node i to node j; α(e i,j ) = f d ecay(W base [i,j], dist(i,j), type compatibility (i,j)); dist(i,j) represents the topological distance, type compatibility Indicates node type compatibility; Design evidence timeliness function: Define time decay function λ(t), which decays exponentially with time; λ(t) = e -t / τ , where τ is a time constant that depends on the type of evidence; Design feedback loop processing mechanism: Identify the feedback loops {L1, L2, ..., L k}; For each loop L i Design stability constraints to ensure that evidence will not be amplified in an infinite loop; output dynamic characteristic parameter set {f d ecay, λ(t), loop_constraints}.

[0087] S326. Integrate the above design elements to build a complete evidence flow network:

[0088] Combined node set V, edge set E, basic edge weight W base , build a static network structure; combine the evidence flow structure definition, evidence merging operator ⊕ and dynamic characteristic parameter set to build a dynamic evidence flow propagation model; initialize the edge propagation matrix T prop , define the propagation characteristics of evidence from source nodes to target nodes; design the network state update equation to define a complete evidence flow propagation cycle; build the evidence flow injection interface to inject new observation evidence into the network; output the complete evidence flow network G f low = (V, E, W, F, ⊕, T prop ), serving as the basic framework for subsequent evidential reasoning.

[0089] According to one aspect of this application, S34, evidence flow propagation and posterior reasoning, includes:

[0090] S341, read the observation data and convert it into evidence stream format: from the fusion spatiotemporal feature F final Extract the state data of the electric energy meter to be evaluated; for each observation variable v obs , determine the corresponding network node v i ; Calculate the initial strength of evidence based on the deviation between the observed value and the normal value init = sigmoid(α·|obs value - normal value| / σ - β); α and β are adjustment parameters, and σ is the normal fluctuation range; based on the data quality and reliability, estimate the initial uncertainty: uncertainty init =f uncertainty (data_quality, measure_reliability, noise_level); data_quality is evaluated based on the quality evaluation indicator set Q; for each observation node, create an initial evidence flow instance: f init (v i ) = (ID new , "direct",v i , strength init , uncertainty init , [v i ]); Output the initial evidence flow set F obs .

[0091] S342, injecting initial evidence into the network and performing the first round of propagation: reading the evidence flow network G f low and the initial evidence flow set F obs ;

[0092] For each initial evidence stream f init (v i )∈F obs :f init (v i ) is added to node v i In the evidence pool; identify v i All child nodes child(v i ) = {v j | e i,j ∈E}; for each child node v j ∈child(v i ), calculate the evidence flow after propagation: strength j = strength i ·α(e i,j ); uncertainty j = uncertainty i + (1-uncertainty i )·(1-α(e i,j ));path h istory j = path history-i + [v j ]; Create a new evidence flow instance f prop (v j), add to v j ; update the network status, record the evidence distribution after the first round of propagation; output the network status G after the first round of propagation f low^1 and propagation evidence set F prop 1 .

[0093] S343. Perform multiple rounds of evidence propagation until the network reaches a stable state:

[0094] Initialize the iteration counter iter = 1 and the maximum number of iterations max i ter (default is 10); define the network stability metric δ(G flow iter, G flow iter-1 ), calculate the state change between two iterations; set the stability threshold ε (default is 0.01);

[0095] Iterative propagation process: For each node v in the network with new evidence i : Merge all evidence streams of the current node: f merged (v i ) = f1 ⊕ f2 ⊕ ... ⊕ f k ; Update the evidence state of the node; Propagate the updated evidence to all child nodes; Increase the iteration count: iter = iter + 1; Calculate the network stability: δ current = δ(G f low^{iter},G f low^{iter-1}); if δ current <ε or iter ≥ max i ter, terminate the iteration; record the final number of iterations iter final and the convergence process {δ1, δ2, ..., δ_{iter final -1}}; Output the final network state G flow final and iterative convergence records.

[0096] S344, aggregate evidence for the final network state and calculate the state of each node: flow Each node v in i :Collect all evidence flows reaching this node: F(v i ) = {f1, f2, ..., f m}; Apply the evidence merging operator to calculate the aggregate evidence: f agg (v i ) = f1 ⊕ f2 ⊕ ... ⊕ fm ; Extract the strength and uncertainty of aggregate evidence: strength(v i ), uncertainty(v i ); calculate the state probability distribution of the node: P(v i = true) =strength(v i )·(1-uncertainty(v i )); P(v i = false) = (1-strength(v i ))·(1-uncertainty(v i )); P(v i = unknown) = uncertainty(v i ); For key output nodes (such as MA, CR, DC, HS), calculate their status evaluation index: index(v i ) = f i ndex(P(v i = true), P(v i = false), P(v i =unknown));f index is the state evaluation function, which maps the probability distribution to the evaluation index; output node state set S node , which contains the probability distribution and state evaluation index of each node.

[0097] S345, based on the node state set S node , identify the most likely state path: construct the global state space S = {S1,S2, ..., S k}, where S k Represent a possible global state configuration; calculate the joint probability of each global state: P(S k ) = ∏ i P(v i = S k [i]); using the maximum a posteriori probability (MAP) criterion to identify the most likely state: S* = argmax S P(S| E = e); use the greedy search algorithm and expand from the high probability node; apply the pruning strategy to only retain the nodes with a probability greater than the threshold θ prune Status branch; keep at most top k (Default is 5) most likely global state configurations; for the identified top k A global state, calculates its confidence score; outputs the most likely state path S_path and state confidence scores.

[0098] S346. Perform counterfactual verification on the identified state path to improve the reliability of reasoning:

[0099] For the key node v in the most likely state path S_path i , construct a counterfactual query: suppose v i The state of the node changes, and the evidence is re-propagated; the changes in the state of other nodes in the network are calculated; the consistency of the counterfactual query and the actual observation is evaluated; according to the counterfactual verification result, the original reasoning result is corrected: if the counterfactual query is highly inconsistent with the actual observation, the confidence of the original reasoning is enhanced; if the counterfactual query is consistent with the actual observation, the confidence of the original reasoning is reduced; the most likely state path S is updated. path And state confidence scores; output the final state reasoning result S i nfer, contains the corrected state path and confidence.

[0100] According to one aspect of this application, S35, causal strength assessment and screening, includes:

[0101] S351, based on the state reasoning result S infer , identify preliminary causal relationships: infer the result S from the state infer Extract the node set V with abnormal status abnormal = {v i | v i The state is abnormal}; for each abnormal node v i ∈V a bnormal, based on the evidence flow network G f low identifies its potential cause node set: direct cause: parent(v i ) = {v j | e_{j,i}∈E and v j The status is abnormal}; indirect cause: ancestor(v i ) = {v j | Existence from v j to v i A directed path with v j The status is abnormal}; construct a preliminary causal graph C init = (V abnormal , E causal ), where E causal Indicates the cause and effect between abnormal nodes.

[0102] S352, preliminary causal diagram C init Each causal link (v j →v i ) to assess causal strength:

[0103] Based on Pearl's intervention theory framework, a computational approximation method for the do-operator is designed:

[0104] Constructing a simplified structural equation model (SEM): v i = f i (pa(v i ), ε i ); do-operator do(v j =x) is realized by disconnecting v j All incoming edges of , and fix v j The value of x; collect enough historical data samples D = {d1, d2, ..., d n}; For each sample d k , calculate the result P(v i =true|d k ); calculate the result P(v i =true|do(v j =true), d k {v j}); average intervention effect ATE(v j →v i ) = (1 / N)∑_{k=1}^N [P(v i =true|do(v j =true), d k {v j}) - P(v i =true|d k )]; calculate the conditional probability P(v i |do(v j )), indicating that v j After intervention i The state distribution of CS(v j →v i ) = P(v i =true|do(v j =true)) - P(v i =true); Output intervention causal strength matrix CS, where CS[j,i] represents the intervention causal strength matrix from v j to v i causal strength.

[0105] S353, Based on Lewis's counterfactual theory, design a counterfactual query calculation method: for causal relationships (v j →v i ), construct a counterfactual query: "If v jNot an abnormal state, v i Will it still be an abnormal state? "; Use the principle of minimum intervention and only modify v j The state of , keeping other variables unchanged. Using the evidence flow network G flow The structure and parameters of the counterfactual conditional probability P(v i =true|v j =false); Based on the observed evidence E, calculate the posterior distribution P(U|E) of the exogenous variable U; Action: Modify v j The state is false, and the modified model M' is obtained; Prediction: Under M', use P(U|E) to calculate v i Probability distribution of counterfactual causal strength CF(v j →v i ) = P(v i =true|v j =true) - P(v i =true|v j =false) ;output counterfactual causal strength matrix CF, where CF[j,i] represents the counterfactual causal strength matrix from v j to v i The counterfactual causal strength.

[0106] S354. Calculate the comprehensive causal strength based on the intervention causal strength matrix CS and the counterfactual causal strength matrix CF:

[0107] Design comprehensive scoring function Score(v j →v i ) = w1·CS(v j →v i ) + w2·CF(v j →v i ) + w3·Conf(v j →v i ); w1, w2, w3 are weight coefficients, the default values are 0.4, 0.4, 0.2 respectively; Conf(v j →v i ) is an additional confidence term that takes into account data quality and model reliability. For electrical relationships (e.g., VS → MS), the intervention term is weighted; for environmental impacts (e.g., TE → VS), the counterfactual term is weighted; for load relationships (e.g., LL → CS), the two measurement methods are balanced. Uncertainty is quantified, and confidence intervals are calculated for each causal strength score. The output is a comprehensive causal strength matrix (Score) and a confidence interval matrix (CI).

[0108] S355. Identify key causal chains based on the comprehensive causal strength matrix Score:

[0109] Set the causal strength screening threshold τ cs (default is 0.15), remove weak causal relationships: E Strong = {(v j →v i )| Score(v j →v i )>τ cs}; Form a strong causal graph G Strong = (V abnormal , E Strong ); define a causal chain as the path from the root cause to the final effect: p = [v1→v2→...→v k ]; Design chain strength scoring function: ChainScore(p) = f({Score(v i →v_{i+1}) | i=1,2,...,k-1}); By default, the geometric mean is used: ChainScore(p) = (∏ i=1 ^{k-1} Score(v i →v i+1 )) 1 / (k-1) ; Apply a modified version of Yen's K-shortest path algorithm to find the K strongest causal chains; Consider the chain length penalty factor to avoid excessively long paths: ChainScore'(p) = ChainScore(p)·e^{-λ·(len(p)-1)}; λ is the length penalty coefficient (default is 0.1); Remove redundant links to ensure the simplicity of causal interpretation; Output the set of strong causal links C Strong , containing the sorted K strongest causal chains.

[0110] S356, identify the strong causal link set C Strong Matching and verification with historical failure mode libraries:

[0111] Read the pre-built electricity meter fault mode library M f ault, including common failure modes and their typical causal chains; for each identified causal chain p∈C Strong , calculate the similarity with each mode in the fault mode library: sim(p, m i ) =pattern S simlarity(p, m i .pattern); pattern S Similarity is calculated based on causal structure similarity and node state consistency; if there is a high similarity match (sim(p, m i )>τ Sim), increase the credibility of the causal chain; for the causal chain that fails to match the known pattern, mark it as a "newly discovered pattern" and perform additional verification; based on the verification results, update the strong causal link set C Strong Sorting and credibility scoring of ; output the final strong causal link set C Strong and pattern matching result M match , providing a basis for subsequent interpretation generation.

[0112] According to one aspect of the present application, S42, calculating a status assessment indicator, includes:

[0113] S421, from the comprehensive evaluation feature F combined Extract the basic features for index calculation: Extract the electrical feature set F elec , including measurement parameters such as voltage, current, and power factor; extracting the measurement feature set F metering , including measurement accuracy and stability indicators; extract hardware feature set F hardware , including temperature, clock deviation, communication status, etc.; extract historical feature set F history , including historical faults, maintenance records, etc.; calculate the statistical description of each feature subset: central trend: mean, median, mode; dispersion: standard deviation, interquartile range, range; distribution characteristics: skewness, kurtosis, entropy; identify outliers and trend change points; output classification feature set {F elec , F metering , F hardware , F history} and statistical description set Stats.

[0114] S422, based on the classification feature set and strong causal link set C Strong , calculate the health index (HI) of the electricity meter:

[0115] Design a hierarchical health assessment model: underlying component health: HI comp (c i ) = g comp (F comp (c i ),anomaly(c i ), severity(c i )); Middle-level subsystem health: HI S ub(s j ) = g S ub({HI comp (c i ) | c i ∈s j}, interaction(s j)); Top-level overall health: HI = g_global({HI S ub(s j )}, C Strong ) ; Component health calculation: Normalize component characteristics: F norm (c i ) = normalize(F comp (c i )); Apply the baseline model to calculate the deviation: dev(c i ) = distance(F norm (c i ), F_ref(c i )); Combined with the anomaly detection results, calculate the component health: HI comp (c i ) = f comp (dev(c i ), anomaly(c i )).

[0116] Overall health: Consider component importance weight w i :HI = ∑ i w i ·HI comp (c i ); Apply nonlinear mapping to normalize the health level to the interval [0,1]; Set the health level threshold: {τ excellent , τ good , τ fair , τ poor ,τ critical}; Output health index HI, component health set {HI comp} and health level Level h I.

[0117] S423, based on the health index HI, strong causal link set C Strong and historical data, calculate the risk factor (RF):

[0118] Identify key risk sources: Internal risks: metering chip degradation, sensor drift, clock deviation, etc.; Environmental risks: temperature exceeding the limit, excessive humidity, electromagnetic interference, etc.; Grid risks: voltage fluctuation, harmonic pollution, frequency instability, etc.; Load risks: overload, impact load, unbalanced load, etc.

[0119] Design risk assessment model: Failure probability P(f i ): Based on component health and historical failure rates; failure impact I(f i ): Based on the degree of influence of fault on measurement performance and reliability; fault detectability D(fi ): Based on the visibility and detection difficulty of the fault; Single risk: RPN(f i ) = P(f i ) × I(f i ) × D(f i );

[0120] Calculation of comprehensive risk factor: RF = h risk ({RPN(f i )}, C Strong , context);h risk It is a comprehensive risk function that considers key risk items, causal relationships, and operating environment; context includes contextual information such as season, climate, and power grid status; and outputs risk factor RF and sub-item risk score {RPN}.

[0121] S424. Predict the remaining useful life (RUL) based on the health index HI, risk factor RF, and historical operating data:

[0122] Construct a degradation model: Collect historical electricity meter degradation data, including normal degradation and accelerated degradation samples; extract key degradation features and establish a feature-lifetime mapping relationship; design a multi-stage Wiener process degradation model: X(t) = X(0) + μ(t)·t + σ(t)·B(t); where μ(t) is the drift coefficient, σ(t) is the diffusion coefficient, and B(t) is the standard Brownian motion.

[0123] Parameter estimation and calibration: Estimate model parameters based on current health and historical degradation curves; establish specific degradation path models for different failure modes; combine strong causal link set C Strong , optimize parameter estimation;

[0124] RUL calculation and uncertainty quantification: defining the failure threshold X f , indicating that the health value is lower than this value and is considered as failure; the time T when the failure threshold is first reached is predicted f ; RUL = T f - T current Generate RUL probability distribution and confidence interval through Monte Carlo simulation; output remaining service life RUL, RUL confidence interval CI_RUL and degradation curve prediction curve pred .

[0125] S425. Calculate the measurement accuracy index (MAI) based on the metrological features in the classification feature set:

[0126] Design accuracy evaluation model: relative error analysis: compare the deviation between the measured value and the reference value; repeatability analysis: evaluate the consistency of measurement under the same conditions; linearity analysis: evaluate the stability of measurement accuracy under different loads; temperature dependence: evaluate the impact of temperature changes on measurement accuracy;

[0127] Multi-dimensional accuracy rating: Active power measurement accuracy: MAI active = f active (error active ,repeatability active , linearity active ); Reactive power measurement accuracy: MAI reactive = f reactive (error reactive , repeatability reactive , linearity reactive ); Voltage / current measurement accuracy: MAI_VI =f_VI(error_V, error i , linearity_V, linearity i );

[0128] Comprehensive accuracy index calculation: MAI = w active ·MAI active + w reactive ·MAI reactive + w_VI·MAI_VI; weight is determined according to application requirements, default: w active =0.6, w reactive =0.3, w_VI=0.1; output measurement accuracy index MAI and item accuracy score {MAI active , MAI reactive , MAI_VI}.

[0129] S426. Integrate the above-calculated indicators to form a complete set of status assessment indicators:

[0130] Standardize each indicator to ensure comparability: HI norm = normalize(HI, [0,1]); RF norm =normalize(RF, [0,1]); RUL norm = normalize(RUL, [0, expected_lifetime]);MAI norm =normalize(MAI, [0,1]);

[0131] Calculate the comprehensive evaluation score: Score overall = c1·HI norm - c2·RF norm + c3·RUL norm + c4·MAI norm ; The coefficients {c1, c2, c3, c4} are adjusted according to the application scenario.

[0132] Construct a multi-dimensional evaluation radar chart to intuitively display the status of each dimension; generate time trend analysis and compare the changes in historical evaluation results; output the final status evaluation indicator set I assessment = {HI, RF, RUL, MAI, Score overall}, as the basis for state classification and decision making.

[0133] According to one aspect of the present application, S43, multi-level status classification and abnormality location, includes:

[0134] S431, based on the state evaluation index set I assessment And historical data, set the multi-level status classification threshold:

[0135] Analyze historical electricity meter status data to determine the statistical distribution of indicators: Statistically analyze the indicator values of electricity meters in normal operating conditions to obtain the baseline distribution; analyze the indicator values of typical abnormal conditions to determine the indicator characteristics of abnormal patterns; set the initial threshold based on the percentile method: the excellent state threshold τ excellent = percentile(HI, 95%); good state threshold τ good = percentile(HI, 80%); general state threshold τ fair = percentile(HI, 60%); poor state threshold τ poor = percentile(HI, 40%); dangerous state threshold τ critical = percentile(HI,20%); adjust the threshold by considering factors such as the type and age of the meter; fine-tune the threshold setting by combining expert knowledge. Implement a dynamic threshold adjustment mechanism to automatically update the threshold based on new data; output the state classification threshold set T = {τ e xcellent, τ good ,τ fair , τ poor , τ critical}.

[0136] S432, based on the state evaluation index set I assessment And the state classification threshold set T, to achieve multi-level state classification:

[0137] Design multi-indicator fusion classification rules: Main indicator rule: preliminary classification based on health index HI; auxiliary indicator adjustment: consider risk factor RF, remaining service life RUL and measurement accuracy index MAI; example rule: if HI>τ good And RF <RF t hreshold and RUL>RUL t hreshold, the status is "good";

[0138] Implement the fuzzy classification algorithm to handle the situation where the indicator is near the threshold boundary: Define the fuzzy membership function μ i (x), indicating the degree to which the index x belongs to state i; calculating the comprehensive fuzzy membership: μ comp rehensive = f fuzzy ({μ i (HI), μ i (RF), μ i (RUL), μ i (MAI)}); Apply decision rules to determine the final state category; Calculate classification confidence to evaluate the reliability of the classification result; Output state classification result S class , including state category, classification confidence and fuzzy membership.

[0139] S433, based on strong causal link set C Strong and component health set {HI comp}, realize exception positioning:

[0140] Identifying abnormal components: From the strong causal link set C Strong Extract the root node set root cause s = {v i | v i is the starting node of a causal chain}; for each root cause node v i , assess its abnormality: anomaly(v i ) = 1 - HI comp (v i ); Filter components whose severity exceeds the threshold: severe comp onents = {v i | anomaly(v i )>τ anomaly};

[0141] Locate key abnormal components: Calculate the impact of component abnormalities on the overall status: impact (v i ) =influence(v i ) × anomaly(v i);influence(v i ) represents component v i Importance in the system, determined based on network centrality and expert knowledge; sorted by influence, identifying top k (Default is 3) key exception components;

[0142] Exception type identification: For each exception component v i , identify the abnormal type type(v according to its characteristic pattern i ); abnormal types include: hardware failure, parameter drift, environmental impact, power grid interference, etc.; type judgment is performed based on historical case library and feature matching; output abnormal location result L anomaly , containing a list of abnormal components, abnormality degree, impact score, and abnormality type.

[0143] S434, based on strong causal link set C Strong and the abnormal location result L anomaly , analyze the abnormal propagation path:

[0144] Construct anomaly propagation graph G prop = (V abnormal , E prop ): node set V abnormal Represents the detected abnormal components; edge set E prop Represents the abnormal propagation relationship, from the strong causal link set C Strong Extracted from

[0145] Propagation path feature extraction: path length: the number of steps from the root cause to the final impact; propagation speed: the time delay for the anomaly to spread from the source component to other components; impact range: the number and importance of affected components;

[0146] Identification of key propagation paths: Calculation of path importance index: importance(p) = ∑_{v i ∈p} impact(v i ) × CS(v i →v_{next(i)}); filter the top K propagation paths ranked by importance; output the abnormal propagation analysis result P prop agation, including key transmission pathways, transmission characteristics, and impact assessment.

[0147] S435, based on the abnormal positioning result L anomaly And the abnormal propagation analysis results P propagation,Evaluate the severity of anomalies: Design a multi-dimensional severity evaluation model: Functional impact dimension: evaluate the impact of anomalies on the function of the electricity meter; Measurement impact dimension: evaluate the impact of anomalies on measurement accuracy; Reliability impact dimension: evaluate the impact of anomalies on system reliability; Safety dimension: evaluate the safety risks that may be caused by anomalies;

[0148] Calculate the comprehensive severity index: Severity = w f unc·S func + w meter· S meter + w reliable ·S reliable + w Safety ·S Safety ;The weight is adjusted according to the application scenario and the type of electricity meter;

[0149] Severity classification: Set the severity level threshold set {τ S 1, τ S 2, τ S 3, τ S 4}; Map the severity index to 5 severity levels: minor, moderate, severe, critical, catastrophic; output the abnormal severity assessment result S S everity, including the scores of each dimension, the comprehensive severity index and the severity level.

[0150] S436. Integrate the above analysis results to form a complete anomaly location output:

[0151] Combined anomaly location result L anomaly , Abnormal propagation analysis results P propagation and abnormal severity assessment results S Severity ; Construct a hierarchical anomaly representation structure: top layer: overall status assessment and severity level; middle layer: list of abnormal components and their anomaly types; bottom layer: detailed anomaly characteristics and propagation paths; generate anomaly summary reports, providing an overview of key information; output the final anomaly location results L anomaly , providing complete data for subsequent visualization and report generation.

[0152] According to one aspect of the present application, in another embodiment, S3, adaptive collision perception evidence flow reasoning analysis, may also be:

[0153] S31. Based on the structure and working principle of the electric energy meter, design the node set V = {v1, v2, ..., v n}, representing the key component states and environmental factors of the energy meter. Design the directed edge set E = {e i,j}, representing the causal relationship between components. Construct the initial directed acyclic graph (DAG) G = (V, E) as the initial causal network G init .

[0154] S32, the initial causal network G init Extended to collision-aware evidence flow network G cf = (V, E, W, F, C), where W is the set of edge weights, F = {f1, f2, ..., f l} is the set of "evidence flows" flowing through the network, C = {c1, c2, ...,c m} is the set of evidence collision points. Each evidence stream carries triple information: f i = (info i , uncertainty i ,source_credibility i ), a new source_credibility function is added to quantify the credibility of the data source. Based on the traditional evidence stream merging operator, a collision handling mechanism is designed: when two evidence streams collide at node v (evidence conflict), the collision resolution function is applied: C resolve (f i , f j , v) = {f' i , f' j , f new}, where f' i and f' j is the adjusted original flow, f new Is a new evidence flow that may be generated. Forming a collision-aware evidence flow network G cf .

[0155] S33, based on historical data and fusion of spatiotemporal features F f inal, learning collision-aware evidence flow network G cf The innovation lies in the introduction of the data source trust weight matrix W t Rust, dynamically adjust the contribution weight of each data source: P w eighted(v i | pa(v i )) = ∑ k w k · P k (v i | pa(v i )), where w k is the trust weight of data source k, P kis the conditional probability calculated based on the data source k. The trust weight is updated in real time based on the historical accuracy and current signal quality. The weighted joint probability distribution is calculated to obtain the trust weighted network parameter set Θ w .

[0156] S34. Inject the observation data as evidence E = e into the collision-aware evidence flow network G. cf . Innovative design of dependency-aware propagation algorithm, using conditional path tracking: when evidence along path p i Spread and affect node v j After that, explicitly mark this path as "activated" and consider this activation state in subsequent propagation: P(v k | e, p iactivated ) ≠ P(v k | e, p inotactivated ). This solves the limitation of traditional Bayesian networks assuming path independence. It realizes the context sensitivity of evidence propagation and generates dependent path sensitive reasoning results S dep .

[0157] S35. Design a sparse evidence enhancement mechanism based on the sparse abnormal signal characteristics of the electricity meter. Apply the signal amplification function f' to the low-frequency but high-importance evidence stream. S parse = Amplify(f Sparse , context), where context includes historical anomaly patterns and current system state. Based on the enhanced evidence flow, an improved causal strength measure is calculated: CS e nhanced(A→B) = α·P(B | do(A)) + (1-α)·MaxImpact(A→B), where MaxImpact evaluates the potential impact in the worst case and α is a dynamically adjusted balance factor. This forms the enhanced causal link set C. enhanced .

[0158] S36, based on enhanced causal link set C enhanced and collision-aware evidence flow network G cf , generating a three-layer adaptive explanation architecture: Overview layer: extracts key abnormal factors and influencing paths, focusing on identifying evidence collision points; Detailed layer: displays the complete fault propagation chain, including evidence dependencies and collision analysis results; Expert layer: provides in-depth analysis of collision points and hypothesis verification tools.

[0159] 100 smart electricity meters in a distribution station area are taken as research objects. The state assessment method of the present invention is applied by collecting multi-source heterogeneous data.

[0160] S1: Multi-source heterogeneous data collection and preprocessing

[0161] S11: Multi-source data acquisition

[0162] Four key data types are collected from the target smart energy meter: real-time energy meter data, including voltage (U), current (I), power factor (PF), active power (P), and reactive power (Q), with a sampling frequency of 15 minutes; grid operation data, including grid frequency (f), phase angle (φ), and harmonic content (HD), with a sampling frequency of 1 hour; environmental data, including ambient temperature (T), humidity (H), and vibration intensity (V), with a sampling frequency of 30 minutes; historical operation records, including historical alarm events (E_alarm), maintenance records (E_repair), and calibration records (E_calibration);

[0163] The following is an example of raw data from an electricity meter (ID: MT-10086) between 10:00 and 10:15 on May 10, 2023: Real-time meter data: {U: 220.5V, I: 5.32A, PF: 0.92, P: 1082.3W, Q: 448.9var}; Grid operating data: {f: 49.98Hz, φ: 1.57rad, HD: 2.3%}; Environmental data: {T: 32.5°C, H: 65.3%, V: 0.05g}; Recent historical records: {E_alarm: "2023-05-08 08:15 - Voltage out of limit", E_repair: "2023-04-15 - Replace display unit", E_calibration: "2023-01-10 - Calibrate the current loop"}; form the original multi-source data set D raw . ;

[0164] S12: D raw Perform a quality assessment and use the conditional mean interpolation method for time windows to handle missing values. For example, in the data for May 10th, it was found that the ambient temperature data for the period of 10:30-11:00 was missing. The missing value was calculated as follows: T_missing = (1 / 6) * (∑(i=1 to 3) T_(ti) + ∑(j=1 to 3) T_(t+j)) = (1 / 6) *((32.5+32.3+32.0) + (32.8+33.0+33.2)) = 32.63°C;

[0165] The interquartile range (IQR) method is used to detect outliers. The calculation process is as follows: Calculate the first quartile Q1 = 219.5V and the third quartile Q3 = 221.3V of the voltage data (U); Calculate the IQR = Q3 - Q1 = 221.3 - 219.5 = 1.8V; Define the outlier threshold: lower bound = Q1 - 1.5IQR = 219.5 - 1.51.8 = 216.8V, upper bound = Q3 + 1.5IQR = 221.3 + 1.51.8 = 224.0V; Detect that the voltage value at 14:15 on May 10 is 229.8V, which exceeds the upper bound threshold and is marked as an outlier;

[0166] For noise processing, a five-layer decomposition and reconstruction based on the DB4 wavelet is used to remove high-frequency noise. For example, the process for processing current data is as follows: 24 hours of current data (96 sampling points) is extracted; a five-layer DB4 wavelet decomposition is applied to obtain approximate and detail coefficients; the coefficients of the noise frequency band (primarily in the detail coefficients d1 and d2) are set to 0; the signal is reconstructed to obtain denoised current data. Finally, the quality assessment indicator set Q and the preliminary cleaned data set D_clean are obtained.

[0167] S13: Apply corresponding normalization methods to different data sources in D_clean: Apply Z-score normalization to continuous measurement data (such as voltage and current): Z(U) = (U - μ_U) / σ_U = (220.5 - 220.0) / 2.1 = 0.238; Apply maximum-minimum normalization to periodic data: Norm(T) = (T - T_min) / (T_max - T_min) = (32.5 - 20.0) / (45.0 - 20.0) = 0.5; Apply one-hot encoding to discrete event data, such as alarm event type encoding: E_alarm_type = {overvoltage: [1,0,0,0], undervoltage: [0,1,0,0], overcurrent: [0,0,1,0], communication abnormality: [0,0,0,1]}; Get the standardized data set D_norm.

[0168] S14: Divide D_norm into multiple time scales and extract time features: Divide the voltage data into hourly segments and calculate statistical features: mean μ_U_hour = (220.5 + 220.8 + 220.4 + 220.6) / 4 = 220.58V; variance σ 2 _U_hour = 0.026V 2 ; skew_U_hour = 0.21; kurtosis_U_hour = 2.85;

[0169] Extract trend features: Calculate the linear regression slope slope_U_day = 0.015V / h; Identify turning points turning_points_U_day = [5:30, 13:45, 21:15].

[0170] Extract periodic features: Apply FFT to analyze power data and obtain the main frequency components: freq_P = [24h -1 ,12h -1 , 1h -1 ]; calculate the autocorrelation coefficients AC_P(lag=24h) = 0.85, AC_P(lag=48h) = 0.76; form the time feature set T_features.

[0171] S15: Construct a spatial association network of electric energy meters based on geographic location and grid topology: Identify the set of neighboring electric energy meters of the target electric energy meter MT-10086: Neighbors(MT-10086) = {MT-10085, MT-10087, MT-10092, MT-10093};

[0172] Calculate the correlation coefficients between adjacent meters: ρ(MT-10086, MT-10085) = 0.86 ρ(MT-10086, MT-10087) = 0.79 ρ(MT-10086, MT-10092) = 0.68 ρ(MT-10086, MT-10093) = 0.72;

[0173] Calculate the phase difference characteristics: φ_diff(MT-10086, MT-10085) = 0.02 rad; φ_diff(MT-10086, MT-10087) = 0.03 rad; φ_diff(MT-10086, MT-10092) = 0.12 rad φ_diff(MT-10086, MT-10093) = 0.09 rad.

[0174] Extract load distribution features: regional load density LD_area = 0.73; regional load imbalance LU_area = 0.15; target meter load ratio LP_MT-10086 = 0.25; obtain the spatial feature set S_features.

[0175] S16: Combine T_features and S_features with D_norm for preliminary fusion:

[0176] Merge features to get the initial high-dimensional feature vector x_i nit = [x_1, x_2, ..., x_42], dimension is 42;

[0177] Apply principal component analysis (PCA) to reduce dimensionality: Calculate the covariance matrix Σ = Cov(x_i nit ); solve the eigenvalues and eigenvectors: λ = [5.62, 3.45, 2.87, ...], v = [v_1, v_2, v_3, ...]; retain the first 12 principal components with a cumulative explained variance of 90%; transform to principal component space: x_pca = [PC_1, PC_2, ..., PC_12] = [2.41, 1.53, -0.87, ...]; finally generate the preprocessed dataset D_prep as input for subsequent analysis.

[0178] S2: Interaction-aware dynamic alignment spatiotemporal feature extraction

[0179] S21: Read D_prep and calculate the time delay correlation between different data sources:

[0180] Taking voltage (U) and temperature (T) as examples, the time delay correlation function C_U,T(τ) = ∑(t) [(U(t) -μ_U) * (T(t+τ) - μ_T)] / (σ_U * σ_T) is calculated. The mutual information I(U(t), T(t+τ)) = ∑∑ p(U(t),T(t+τ)) * log(p(U(t),T(t+τ)) / (p(U(t))*p(T(t+τ)))) is calculated for different τ values (ranging from -5h to +5h). The optimal time alignment parameter τ*_U,T is solved: τ*_U,T = argmax_τ I(U(t), T(t+τ)) = 1.5h, indicating that temperature changes lead voltage changes by approximately 1.5 hours. The optimal delay τ_U,f = 0.25h, τ_I,P = 0h, and τ*_T,PF = 2h is calculated for other data source pairs.

[0181] Apply the optimal time alignment parameters to correct the original data and obtain the time-aligned dataset D_time_aligned.

[0182] S22: Spatial Mapping and Reference Frame Construction

[0183] Construct a unified spatial reference frame: Design a spatial mapping function S_i,j(•) with the meter locations as the reference: the location of meter MT-10086 is taken as the reference point (0, 0); the relative location of the adjacent meter MT-10085 is (-1, 0), with a physical distance of 10m; the relative location of the adjacent meter MT-10087 is (1, 0), with a physical distance of 12m; the relative location of the adjacent meter MT-10092 is (0, 1), with a physical distance of 15m; the relative location of the adjacent meter MT-10093 is (0, -1), with a physical distance of 11m;

[0184] Calculate the spatial weight matrix W_space(i,j) = exp(-d_ij based on the electrical distance 2 / σ 2 ), where d_ij is the electrical distance between meters i and j, and σ is the scale parameter. For example: W_space(MT-10086, MT-10085) = exp(-(10) 2 / 100) = 0.368;

[0185] For environmental sensor data, spatial weights are determined based on sensor location and influence range: temperature sensor T1 is located at (0.5, 0.3) with an influence range of 20m; humidity sensor H1 is located at (0.2, -0.4) with an influence range of 15m; vibration sensor V1 is located at (-0.1, 0.2) with an influence range of 5m;

[0186] Calculate the weight of the environmental sensor on the meter W_env(T1, MT-10086) = exp(-d_T1,MT-10086 2 / (2*20 2 )) = 0.982; finally, the spatial alignment dataset D_space_aligned is formed.

[0187] S23: Based on D_space_aligned, we design a data source pair interaction operator Θ to capture the interaction characteristics between different data sources:

[0188] Construct the data source pair set P and type mapping matrix T: electrical quantity and electrical quantity pairs: (U,I), (U,P), (I,P), etc., type T[U,I] = 1; electrical quantity and environmental quantity pairs: (U,T), (I,H), etc., type T[U,T] = 2; environmental quantity and environmental quantity pairs: (T,H), (T,V), etc., type T[T,H] = 3; event and measurement value pairs: (E_alarm,U), (E_repair,I), etc., type T[E_alarm,U] = 4;

[0189] Calculate the correlation index for the data source pair (U,I): Pearson correlation coefficient: ρ_U,I = Cov(U,I) / (σ_U*σ_I) = 0.78; calculate the point biserial correlation: for (U,E_alarm), rpb = 0.35.

[0190] Compute information theory features: Adaptive binning is applied to the voltage (U) and temperature (T) data, with the number of bins k = 2.0 * IQR * n^(-1 / 3) = 10; calculate the mutual information I(U;T) = 0.42; calculate the conditional entropy H(U|T) = 0.65; and calculate the transfer entropy TE(T→U) = 0.28;

[0191] Granger causality test: Construct autoregressive model AR_U: U^t = 0.75U^(t-1) + 0.15U^(t-2) + ε_t, RSS_U = 4.32; construct model AR_TU including temperature: U^t = 0.70U^(t-1) + 0.12U^(t-2)+ 0.25T^(t-1) + 0.10T^(t-2) + ε_t, RSS_TU = 3.15; F statistic: F = ((RSS_U - RSS_TU) / 2) / (RSS_TU / (n-4)) = 5.86, p-value = 0.004<0.05. Conclusion: There is a Granger causal relationship between temperature and voltage, G_cause[T,U] = 0.75

[0192] Design interaction operators: electrical interaction sub-operator Θ_e: M_U,I = U Θ_e I = [ρ_U,I, cos(φ_U,I), P / (U*I), I(U;I)]; environmental interaction sub-operator Θ_env: M_T,U = T Θ_env U = [ΔU / ΔT, TE(T→U), H(U|T)]; event interaction sub-operator Θ_evt: M_E,U = E Θ_evt U = [U_after - U_before, recovery_time, oscillation_amplitude];

[0193] Example interaction feature matrix calculation: M_U,I = U Θ_e I = [0.78, 0.92, 0.93, 0.55] = [correlation coefficient, power factor, rated power ratio, mutual information] M_T,U = T Θ_env U = [0.12V / °C, 0.28, 0.65] = [temperature sensitivity coefficient, transfer entropy, conditional entropy] M_E_alarm,U = E_alarm Θ_evt U = [6.5V, 35min, 1.2V] = [voltage change after the event, recovery time, oscillation amplitude]. The interaction features of all data source pairs are integrated to form the complete interaction feature set M_interact.

[0194] S24: Apply cycle-aware hierarchical attention mechanism to D_space_aligned:

[0195] The voltage time series is segmented into predefined periods: intraday period (24 points, one point per hour): x_hour = [x_1, x_2, ..., x_24]; daily period (7 points, one point per day): x_day = [x_1, x_2, ..., x_7]; weekly period (4 points, one point per week): x_week = [x_1, x_2, x_3, x_4]; monthly period (12 points, one point per month): x_month = [x_1, x_2, ..., x_12];

[0196] Calculate intraday attention:

[0197] Construct the query matrix Q_h, key matrix K_h, and value matrix V_h, all of shape (24, d_k), where d_k is the feature dimension; compute the attention score S_h = Q_h•K_h^T / √d_k; apply the softmax function A_h = softmax(S_h); and weight the value matrix A_hour(t) = A_h•V_h.

[0198] For example, for the voltage time point t=10h: attention score S_h

[10] = [0.01, 0.02, ...,0.25, ..., 0.03]; after softmax, A_h

[10] = [0.01, 0.02, ..., 0.15, ..., 0.02]; weighted result A_hour(10) = ∑(j=1 to 24) A_h

[10] [j]•V_h[j] = [0.52, 0.63, ...]; similarly, attention at other period levels A_day(t), A_week(t), A_month(t) are calculated;

[0199] Learn the fusion weights α_h = 0.4, α_d = 0.3, α_w = 0.2, α_m = 0.1; fuse multi-level attention: A_periodic(t) = α_h•A_hour(t) + α_d•A_day(t) + α_w•A_week(t) + α_m•A_month(t) = 0.4•[0.52, 0.63, ...] + 0.3•[0.45, 0.58, ...]+ 0.2•[0.41, 0.55,...] + 0.1•[0.39, 0.52, ...]= [0.47, 0.59, ...]; obtain the periodic attention feature A_periodic.

[0200] S25: Predicting normal periodic behavior based on A_periodic: Construct a multi-period prediction model to predict voltage values: hourly prediction function f_h(h_t) = 220.0 + 2.0*sin(2π•h_t / 24 + 0.5); daily prediction function f_d(d_t) = 220.0 + 1.0*sin(2π•d_t / 7 + 0.2); weekly prediction function f_w(w_t) = 220.0 + 0.5*sin(2π•w_t / 4 + 0.1); monthly prediction function f_m(m_t) = 220.0 + 0.3*sin(2π•m_t / 12 + 0.3); predicted value x̂_t = α_h•f_h(h_t) + α_d•f_d(d_t) + α_w•f_w(w_t) + α_m·f_m(m_t) = 0.4·221.5 + 0.3·220.8 + 0.2·220.3 + 0.1·220.2 = 221.0V;

[0201] Calculate the residual sequence r_t = x_t - x ∧ _t = 220.5V - 221.0V = -0.5V; Apply 5-layer db4 wavelet transform to decompose the residual sequence: Obtain detail coefficients: [W_1, W_2, W_3, W_4, W_5] = [[0.21,0.15,...],[-0.12,0.08,...], [0.05,-0.03,...], [-0.02,0.01,...], [0.01,-0.01,...]]; Obtain approximate coefficient: V_5 = [-0.03,0.02,...].

[0202] Calculate the residual energy at each scale: E_1 = ∑_t |W_1[t]| 2= 0.15; E_2 = ∑_t |W_2[t]| 2 =0.08; E_3 = ∑_t |W_3[t]| 2 = 0.04; E_4 = ∑_t |W_4[t]| 2 = 0.02; E_5 = ∑_t |W_5[t]| 2 = 0.01; E_a = ∑_t |V_5[t]| 2 = 0.03;

[0203] Calculate the normalized energy distribution E_norm = E / ∑E = [0.15, 0.08, 0.04, 0.02, 0.01, 0.03] / 0.33 = [0.45, 0.24, 0.12, 0.06, 0.03, 0.09]; compare with the reference energy distribution of normal operation: E_ref = [0.30, 0.25, 0.15, 0.12, 0.08, 0.10] ΔE = E_norm - E_ref = [0.15, -0.01, -0.03, -0.06, -0.05, -0.01]

[0204] Design dynamic threshold function: initial threshold τ_j^0 = GEV_j^(-1)(0.999) = [0.12, 0.10, 0.08, 0.06, 0.04]; dynamic adjustment factor λ(t) = 1 + 0.2•sin(2π•t / 24) + 0.1•(T(t)-25) / 10; final dynamic threshold τ_j(t) = τ_j^0•λ(t) = [0.13, 0.11, 0.09, 0.07, 0.04];

[0205] Calculate the anomaly detection results: δ_j(t) = |W_j[t]| / τ_j(t) = [1.62, 0.73, 0.56, 0.29, 0.25]; δ_max(t) = max_j(δ_j(t)) = 1.62>1.0; mark as an outlier, s(t) = log(1 + 1.62 -1.0) = 0.69;

[0206] Calculate the abnormal enhancement factor: basic enhancement function β_0(t) = sigmoid(5•0.69 - 0.5) = 0.968; time accumulation factor d(t) = 3 / 24 = 0.125; mutation coefficient jump(t) = max(0, tanh(0.69 / 24)) = 0.029; final enhancement factor β(t) = 0.968•(1 + 0.3•0.125 + 0.5•0.029) = 1.01; form the residual enhancement feature A_residual.

[0207] S26: Calculate the raw feature attention A_raw = γ_A_periodic + (1-γ)•A_residual =0.7•[0.47,0.59,...] + 0.3•[0.52,0.67,...]= [0.49,0.61,...]; calculate the interactive feature attention A_inter = f_inter(M_interact) = [0.56,0.64,...]; design the context-aware fusion: context vector c= [peak_load=0, high_temp=1, event_recent=0]; context-dependent weight function γ_1(c) = 0.6,γ_2(c) = 0.4;

[0208] Calculate the final fusion feature A_final = γ_1(c)•A_raw + γ_2(c)•A_inter = 0.6•[0.49,0.61,...] + 0.4•[0.56,0.64,...] = [0.52,0.62,...]. Generate the fused spatiotemporal feature F_final.

[0209] S3: Evidence Flow Causal Reasoning Analysis

[0210] S31: Based on the structure and working principle of the electric energy meter, design the node set:

[0211] Core component nodes: metering chip (MS): responsible for core metering functions; voltage sensor (VS): measures voltage values; current sensor (CS): measures current values; clock circuit (TC): provides time reference; communication module (CM): responsible for data transmission; display unit (DU): displays metering results; power management unit (PM): provides stable power supply.

[0212] Environmental factor nodes: Temperature (TE): ambient temperature; Humidity (HU): ambient humidity; Electromagnetic interference (EMI): ambient electromagnetic interference level;

[0213] Grid-related nodes: Grid voltage stability (GV): input voltage stability; Grid frequency stability (GF): grid frequency stability; Harmonic content (HC): grid harmonic pollution level;

[0214] Load-related nodes: Load level (LL): connected load size; Load change rate (LR): load change speed; Load type (LT): load nature (resistive / inductive / capacitive);

[0215] State output node: Measurement accuracy (MA): measurement accuracy status; Communication reliability (CR): communication quality status; Display correctness (DC): display function status; Overall health status (HS): comprehensive health status; Design a directed edge set E to represent the causal relationship between nodes:

[0216] Examples of direct causal relationships: VS→MS: voltage sensor affects metering chip; CS→MS: current sensor affects metering chip; TE→VS: temperature affects voltage sensor; HU→MS: humidity affects metering chip; EMI→CM: electromagnetic interference affects communication module; GV→VS: grid voltage stability affects voltage sensor; GF→TC: grid frequency stability affects clock circuit; MS→MA: metering chip affects measurement accuracy; CM→CR: communication module affects communication reliability. Construct a basic directed acyclic graph G = (V, E) as the initial causal network G_i nit .

[0217] S32: G_i nit Extended to collision-aware evidence flow network:

[0218] Define the evidence flow structure: Evidence flow triple: f = (info, uncertainty, source_credibility); example: f_TE = (32.5, 0.05, 0.95), indicating a temperature value of 32.5°C, an uncertainty of 0.05, and a source credibility of 0.95.

[0219] Design an evidence collision handling mechanism: collision metric K(f_i, f_j) = |info_i - info_j| / (uncertainty_i + uncertainty_j); for example, the measurement values of temperature sensors T1 and T2 are f_T1 = (32.5, 0.05, 0.95) and f_T2 = (31.8, 0.1, 0.9), respectively; collision metric K(f_T1, f_T2) = |32.5 -31.8| / (0.05 + 0.1) = 0.7 / 0.15 = 4.67>3.0 (high conflict).

[0220] Example of collision resolution function: When K < 3.0 (low collision): use weighted average f_new = ((info_i • source_credibility_i) / (uncertainty_i) + (info_j • source_credibility_j) / (uncertainty_j)) / ((source_credibility_i) / (uncertainty_i) + (source_credibility_j) / (uncertainty_j));

[0221] When K ≥ 3.0 (high conflict): keep high-credibility evidence f_new = f_i if (source_credibility_i / uncertainty_i)>(source_credibility_j / uncertainty_j) else f_j; for high-conflict examples: f_new = f_T1, because (0.95 / 0.05) = 19>(0.9 / 0.1) = 9.

[0222] Design evidence flow propagation attenuation function α(e_{i,j}) = exp(-L_{i,j} / d_0)•compatibility(i,j; where L_{i,j} is the topological distance, d_0 is the feature attenuation distance (value is 10); compatibility(i,j) is the node type compatibility, with a value range of [0,1]. Example: The attenuation factor from the temperature (TE) node to the voltage sensor (VS) node α(e_{TE,VS}) = exp(-1 / 10)•0.9 = 0.9•0.905 = 0.814.

[0223] Design the evidence timeliness function: λ(t) = exp(-t / τ), where t is the evidence age and τ is the time constant; for temperature data: τ_TE = 2h, for temperature data 4 hours ago; λ(4h) = exp(-4 / 2) = exp(-2) = 0.135; form a complete collision-aware evidence flow network G_cf = (V, E, W, F, C).

[0224] S33: Based on historical data and F_final, learn the conditional probability table in G_cf: Create the data source trust weight matrix W_trust: the energy meter's own data source: W_trust[self] = 0.9; the neighboring meter data source: W_trust[neighbor] = 0.7; the environmental sensor data source: W_trust[env] = 0.8; the historical event data source: W_trust[event] = 0.6;

[0225] Example of calculating weighted conditional probabilities: For example, to calculate the probability of "voltage sensor abnormality" under the condition of "high temperature", calculate P_self(VS=abnormal|TE=high) = 0.25 from the meter's own data; calculate P_neighbor(VS=abnormal|TE=high) = 0.30 from neighboring meter data; and calculate P_event(VS=abnormal|TE=high) = 0.35 from historical event data. The weighted conditional probability P_weighted(VS=abnormal|TE=high) = (0.9•0.25 + 0.7•0.30 + 0.6•0.35) / (0.9 + 0.7 + 0.6) = 0.566 / 2.2 = 0.257.

[0226] Dynamically update the trust weight: Update W_trust_new[source] = 0.8•W_trust[source] + 0.2•accuracy[source] based on the most recent prediction accuracy; for example, the environmental sensor’s most recent prediction accuracy is 0.75: W_trust_new[env] = 0.8•0.8 + 0.2•0.75 = 0.64 + 0.15 = 0.79; obtain the trust-weighted network parameter set Θ_w.

[0227] S34: Inject observation data into G_cf and perform dependency-aware propagation:

[0228] Initial evidence flow conversion: Observed voltage value U = 230.5V (abnormally high), normal range is 220±5V; strength calculation strength_i nit = sigmoid(5•|230.5-220| / 5 - 1) = sigmoid(5•2.1 - 1) = sigmoid(9.5) = 0.9999; uncertainty_i nit = 0.05 (based on the accuracy of the voltage measurement equipment); source credibility source_credibility_i nit = 0.95 (based on device calibration status); create initial evidence flow f_i nit(VS)= ("high", 0.9999, 0.05, 0.95, [VS]).

[0229] Path condition propagation example: f_i nit (VS) is added to the VS node and marked as active. The child nodes of VS are identified: child(VS) = {MS}. The propagated evidence flow is calculated: strength_MS = 0.9999•0.9 = 0.8999 (assuming the edge weight is 0.9). uncertainty_MS = 0.05 + (1-0.05)•(1-0.9) = 0.05 + 0.095 = 0.145. path_history_MS = [VS, MS]; a new evidence flow is created: f_prop(MS) = ("affected", 0.8999, 0.145, 0.95, [VS, MS]). The path VS→MS is marked as active.

[0230] Multi-path dependency processing: Assume that MS is affected by both CS (current sensor) and VS (voltage sensor); Evidence flow from VS: f_VS→MS = ("affected_by_voltage", 0.8999, 0.145, 0.95, [VS,MS]); Evidence flow from CS: f_CS→MS = ("affected_by_current", 0.75, 0.12, 0.90, [CS,MS]); Path-dependent evidence merging: f_merged(MS) = f_VS→MS ⊕_dep f_CS→MS = ("affected_by_both", 0.95, 0.08, 0.93, [VS, CS, MS]); Difference (comparison) from traditional independent path merging: f_merged_trad(MS) = f_VS→MS ⊕ f_CS→MS = ("affected", 0.83, 0.10, 0.92, [VS,CS, MS]).

[0231] Complete iterative propagation: Execute 5 rounds of iterations, with the network stability metric δ = 0.008 < ε(0.01). The final number of iterations is iter_final = 5. Generate the dependency path sensitive inference result S_dep.

[0232] S35: Perform sparse evidence enhancement and causal strength evaluation on S_dep:

[0233] Sparse evidence enhancement processing: Only two current fluctuation events were detected, but they were of high importance: f_sparse(CS) = ("fluctuation", 0.65, 0.15, 0.85, [CS]); applying the signal amplification function f'_sparse(CS) = ("significant_fluctuation", 0.86, 0.10, 0.88, [CS]) increases the amplification strength from 0.65 to 0.86, and reduces the uncertainty from 0.15 to 0.10;

[0234] Calculation of intervention effect: Calculate the conditional probability under intervention P(MS=abnormal|do(VS=abnormal)) = 0.75 P(MS=abnormal) = 0.15; intervention causal strength CS(VS→MS) = 0.75 - 0.15 = 0.60;

[0235] Counterfactual analysis: Calculate the counterfactual conditional probability P(MS=abnormal|VS=abnormal) = 0.75 P(MS=abnormal|VS=normal) = 0.10; the counterfactual causal strength CF(VS→MS) = 0.75 - 0.10 = 0.65.

[0236] Comprehensive causal strength score Score(VS→MS) = 0.4•0.60 + 0.4•0.65 + 0.2•0.90 = 0.24+ 0.26 + 0.18 = 0.68;

[0237] Identification of strong causal links: Set the threshold τ_cs = 0.15 to screen strong causal relationships; form a strong causal graph G_strong = (V_abnormal, E_strong); identify the strongest causal chain: p_1 = [TE→VS→MS→MA], ChainScore(p_1) = 0.61 p_2 = [GV→VS→MS→MA], ChainScore(p_2) = 0.58 p_3 = [HU→PM→MS→MA], ChainScore(p_3) = 0.45; finally, form the strong causal link set C_strong and the diagnostic report R_explain.

[0238] S4: Multi-dimensional status assessment and result output

[0239] S41. Integrate F_final and C_strong to build a multi-dimensional status assessment framework:

[0240] Design a weighted fusion algorithm: spatiotemporal feature weight w_time_space = 0.6; causal inference weight w_causal = 0.4;

[0241] Generate comprehensive evaluation features F_combined = w_time_space•F_final + w_causal•feature(C_strong) = 0.6•[0.52,0.62,...] + 0.4•[0.61,0.58,...]= [0.56,0.60,...];

[0242] S42: Based on F_combined, calculate multiple state evaluation indicators:

[0243] Health Index (HI) calculation: Component health calculation: VS component health HI_comp(VS) = 1 - dev(VS) / max_dev = 1 - 0.32 / 1.0 = 0.68; CS component health HI_comp(CS) = 0.85; MS component health HI_comp(MS) = 0.75; other component health [TC: 0.92, CM: 0.95, DU: 0.89, PM: 0.86]; overall health HI = ∑_i w_i•HI_comp(c_i) = 0.25•0.68 + 0.20•0.85 + 0.30•0.75+ 0.10•0.92 + 0.05•0.95 + 0.05•0.89 + 0.05•0.86 = Health level determination: threshold settings τ_excellent = 0.9, τ_good = 0.8, τ_fair = 0.7, τ_poor = 0.6, τ_critical = 0.5. Result: HI = 0.789 < τ_good, the health level is "fair".

[0244] Risk factor (RF) calculation: Individual risk scores: VS drift risk P(f_VS) = 0.35, I(f_VS) = 0.8, D(f_VS) = 0.7, RPN(f_VS) = 0.35•0.8•0.7 = 0.196; MS precision risk: P(f_MS) = 0.25, I(f_MS) = 0.9, D(f_MS) = 0.6, RPN(f_MS) = 0.25•0.9•0.6 = 0.135; Temperature effect risk: P(f_TE) = 0.45, I(f_TE) = 0.6, D(f_TE) = 0.8, RPN(f_TE) = 0.45•0.6•0.8 = 0.216; Comprehensive risk factor RF = h_risk({RPN(f_i)}, C_strong, context) = 0.4•0.196 +0.3•0.135 + 0.2•0.216 + 0.1•0.05 = 0.0784 + 0.0405 + 0.0432 + 0.005 = 0.167; Risk level determination: RF = 0.167, the risk level is "medium risk".

[0245] Remaining useful life (RUL) prediction: Degradation model parameter estimation μ(t) = 0.005•t + 0.001•t 2 , represents the drift coefficient σ(t) = 0.002•√t, and represents the diffusion coefficient. Define the failure threshold X_f = 0.5, indicating that a health below 0.5 is considered a failure. Current health X_current = 1 - HI = 1 - 0.789 = 0.211. Predict the time when the failure threshold is first reached: Solve for 0.211 + 0.005•T + 0.001•T 2 = 0.5, which yields T ≈ 24.6 months. RUL = 24.6 months, CI_RUL = [20.1 months, 29.3 months].

[0246] Measurement Accuracy Index (MAI) calculation: Active power measurement accuracy: Relative error error_active = 0.015; Repeatability repeatability_active = 0.008; Linearity linearity_active = 0.012; MAI_active = 1 - (0.5•0.015 + 0.3•0.008 + 0.2•0.012) = 1 - 0.0127 = 0.987. Reactive power measurement accuracy MAI_reactive = 0.975. Voltage / current measurement accuracy MAI_VI = 0.960. Overall accuracy index: MAI = 0.6•0.987 + 0.3•0.975 + 0.1•0.960 = 0.5922 + 0.2925 + 0.096 = 0.981. Generate the status assessment indicator set I_assessment = {HI: 0.789, RF: 0.167, RUL: 24.6 months, MAI:0.981, Score_overall: 0.835}.

[0247] S43: Based on I_assessment and C_strong, perform multi-level status classification and anomaly location:

[0248] Status classification results: Health index (HI = 0.789) corresponds to "Fair" status. Risk factor (RF = 0.167) corresponds to "Medium Risk." Remaining Life (RUL = 24.6 months) corresponds to "Normal Life." Measurement accuracy (MAI = 0.981) corresponds to "High Accuracy." Overall score (Score_overall = 0.835) corresponds to "Good" status. The final status classification for the MT-10086 electricity meter is "Good but Needs Attention," with a classification confidence of 0.87.

[0249] Anomaly Location Results: Key anomaly components: Voltage sensor (VS): Anomaly severity 0.32, impact score 0.28, anomaly type "temperature-sensitive drift." Metering chip (MS): Anomaly severity 0.25, impact score 0.22, anomaly type "parameter offset." Temperature environment (TE): Anomaly severity 0.30, impact score 0.15, anomaly type "sustained high temperature."

[0250] The abnormal propagation path p_priority = [TE→VS→MS→MA], with an importance index of 0.61, indicates that high temperature causes voltage sensor drift, which in turn causes parameter shifts in the metering chip, ultimately affecting measurement accuracy.

[0251] Anomaly severity assessment: Functional impact S_func = 0.15. Metering impact S_meter = 0.25. Reliability impact S_reliable = 0.20. Safety: S_safety = 0.05. Combined severity = 0.4•0.15 + 0.4•0.25 + 0.15•0.20 + 0.05•0.05 = 0.06 + 0.10 + 0.03 + 0.0025 = 0.1925. Severity level: "Medium."

[0252] S44: Based on the above analysis results, generate a visual status report:

[0253] Design Status Dashboard: Health Index: Displays 0.789, yellow (Fair). Risk Factor: Displays 0.167, yellow (Medium Risk). Remaining Life: Displays 24.6 Months, green (Normal). Measurement Accuracy Index: Displays 0.981, green (High Accuracy).

[0254] Abnormal map display: On the electricity meter structure diagram, the voltage sensor (VS) area is highlighted in red. The metering chip (MS) area is in orange. Other component areas are in green.

[0255] Causal Link Diagram: Interactively displays the primary fault path [TE→VS→MS→MA]. Node size indicates the degree of anomaly, and edge thickness indicates causal strength. The temperature (TE) node is labeled "32.5°C (exceeds the normal range)." The voltage sensor (VS) node is labeled "Offset +2.5%."

[0256] Time Evolution View: Displays health trends over the past three months. Key events are highlighted, such as the calibration on May 1st, which briefly improved health. Abnormal patterns are highlighted, such as the downward trend in health over the past seven days, which coincides with rising ambient temperatures. Generates a visual status report (V_report).

[0257] S45: Decision suggestion generation

[0258] Based on I_assessment and C_strong, combined with the maintenance experience knowledge base, targeted maintenance decision recommendations are generated:

[0259] Regarding voltage sensor (VS) drift, recommended action: Recalibrate the voltage sensor. Expected improvement: Improve health by approximately 0.08, extending service life by 3-5 months. Priority: High, recommended within 2 weeks.

[0260] For high ambient temperature (TE): Recommended Action: Improve the ventilation of the meter box or install an auxiliary heat sink. Expected Improvement: Reduce the drift rate of temperature-sensitive components and minimize measurement errors. Priority: Medium. Recommended Action: Implement within one month.

[0261] For metrology chip (MS) parameter drift: Recommended Action: Perform system parameter recalibration. Expected Improvement: Improves measurement accuracy and corrects drift trends. Priority: Medium. Recommended after resolving temperature issues.

[0262] Comprehensive maintenance strategy: Short-term (within one month): Improve environmental conditions and calibrate the voltage sensor. Medium-term (within three months): Recalibrate system parameters and monitor whether the abnormality transmission chain is under control. Long-term (over six months): If the problem persists, consider replacing the voltage sensor assembly. Generate decision advice (D_advice).

[0263] S46: Integrate V_report and D_advice to output the final comprehensive evaluation result R_final, including:

[0264] Energy meter basic information: ID: MT-10086; Model: DT-FS350; Installation location: Cabinet 3, Area A, distribution station; Service life: 3.5 years;

[0265] Multi-dimensional status assessment results: Health: 0.789 (Fair); Risk factor: 0.167 (Medium risk); Remaining service life: 24.6 months; Measurement accuracy: 0.981 (High precision); Overall score: 0.835 (Good); Status classification: "Good but needs attention";

[0266] Abnormal diagnosis results: Main abnormality: Voltage sensor temperature-sensitive drift; Root cause: Sustained high ambient temperature (32.5°C); Transmission path: High temperature → Voltage sensor drift → Metering parameter offset → Impact on measurement accuracy; Evidence strength: 0.61 (high confidence);

[0267] Maintenance Decision Recommendation: Priority Action: Improve ventilation and calibrate the voltage sensor; Expected Effect: Increase the health level to above 0.87, extending service life; Subsequent Monitoring: Focus on the relationship between temperature and voltage sensor output;

[0268] Visual Report: This includes a status dashboard, anomaly map, causal chain diagram, and time evolution view. It provides an interactive query interface for in-depth exploration of detailed data. The comprehensive evaluation result R_final is presented to maintenance personnel through the user interface, completing the entire status assessment process.

Claims

1. A method for evaluating the state of a smart electric energy meter based on multi-source data fusion, characterized in that: The following steps are involved: Collect basic data, perform quality assessment, standardization, extract temporal and spatial correlation features, and obtain preprocessed data sets; The preprocessed dataset is input into the attention module to extract the data source interaction features, calculate the periodic attention features and residual enhancement features, and generate the fused spatiotemporal features through multi-channel spatiotemporal attention fusion; Based on the fusion of spatiotemporal features, an evidence flow network is constructed. Through evidence flow propagation and posterior reasoning, causal strength is evaluated and strong causal links are screened. A strong causal link set and diagnostic report are generated. Combined with the fusion of spatiotemporal features, multi-dimensional status evaluation indicators are calculated, multi-level status classification and anomaly location are implemented, and comprehensive evaluation results are obtained. The preprocessed dataset is input into the attention module to extract data source interaction features, including: According to the data type and physical meaning, a type mapping matrix T is established for each pair of data sources; Calculate the correlation index of all data source pairs to form the correlation matrix C corr ; According to the type mapping matrix T, the information theory characteristics are calculated to obtain the information theory characteristic matrix I info , which contains three sub-matrices: mutual information, conditional entropy and transfer entropy; Read the spatial alignment dataset and data source pair set P, perform Granger causality test on each pair of data sources, and generate the Granger causality matrix G cause ; Based on the correlation matrix C corr , information theory characteristic matrix I info and the Granger causal matrix G cause , design the data source pair interaction operator Θ; Apply the interaction operator Θ to calculate the interaction feature matrix and extract the features of the interaction feature matrix, including eigenvalue distribution, principal components and singular values; form a complete interaction feature set M interact .

2. The method according to claim 1, characterized in that Generate fused spatiotemporal features, including: Read the preprocessed data set, perform time and space alignment in sequence, generate a time-space aligned data set, and calculate the interaction feature matrix through the interaction operator of the data source pair to form an interaction feature set; Applying the period-aware hierarchical attention mechanism to the spatiotemporally aligned dataset, we calculate the intra-day, inter-day, inter-week, and inter-month multi-level attention and generate periodic attention features. Based on the periodic attention feature, the normal periodic behavior pattern is predicted, the residual with the actual observation value of the spatially aligned dataset is calculated, the multi-scale residual is decomposed and the attention weight is adjusted to form the residual enhancement feature; Combining the periodic attention features, residual enhancement features and interaction feature sets, the weights of each channel are adjusted to generate fused spatiotemporal features.

3. The method according to claim 2, characterized in that Form an interactive feature set, including: Based on the spatially aligned dataset, a data source pair set and a type mapping matrix are constructed; According to the type mapping matrix, the correlation index is calculated for different types of data sources to form a correlation matrix; and the mutual information, conditional entropy and transfer entropy are calculated to form an information theory feature matrix; Perform Granger causality test on each pair of data sources to generate a Granger causality matrix; Based on the correlation matrix, information theory characteristic matrix and Granger causality matrix, interaction operators of electrical quantities, environmental quantities and electrical quantities, events and measurements are constructed to obtain interaction operators and selection strategies. The interaction operators and selection strategies are then applied to process the set of data source pairs to form an interaction feature set.

4. The method according to claim 2, characterized in that Forming residual enhancement features, including: A multi-period prediction model is constructed based on the periodic attention feature to generate a sequence of predicted values and calculate the residual between the predicted value and the actual observed value. The residual is decomposed by wavelet transform to obtain a set of multi-scale wavelet coefficients. Calculate the residual energy distribution of each scale, compare it with the reference distribution of normal operation, output the energy analysis results and design the dynamic threshold accordingly; Calculate the normalized deviation of the wavelet coefficients and combine it with the dynamic threshold to give the anomaly detection result; The enhancement factor is designed according to the anomaly detection results, and the anomaly enhancement factor sequence is obtained as the residual enhancement feature.

5. The method according to claim 2, characterized in that Adjust the weights of each channel to generate fused spatiotemporal features, including: Combine the periodic attention features and the residual enhancement features to calculate the original feature attention; Calculate interactive feature attention based on the interactive feature set; According to the context vector and the pre-stored weight function, the original feature attention and the interactive feature attention are fused to generate the fused spatiotemporal features.

6. The method according to claim 1, characterized in that Generates a strong set of causal links and diagnostic reports, including: Based on the structure and working principle of the electric energy meter, an initial causal network representing the status of the key components and environmental factors of the electric energy meter is constructed and expanded into a collision-aware evidence flow network. Read the observation data and inject it into the collision-aware evidence flow network as evidence, and apply the dependency-aware propagation algorithm to generate dependency path-sensitive reasoning results; Perform sparse enhancement on the dependency path sensitive reasoning results, calculate the enhanced causal strength, and form an enhanced causal link set; Based on the enhanced causal link set and collision-aware evidence flow network, a hierarchical adaptive evidence interpretation is generated to obtain a diagnosis report and a strong causal link set.

7. The method according to claim 6, characterized in that Collision-aware evidence flow network, including: Point set and node type mapping, based on the core components of the energy meter, environmental factors, grid-related factors, and load-related factors; The causal relationship edge set and basic edge weights between nodes are designed based on domain knowledge and the working principle of electricity meters; Evidence flow triplet attributes, including information content, uncertainty, and source credibility; The collision resolution function handles conflicting evidence and designs a collision handling mechanism based on the evidence conflict measure; The propagation attenuation function, timeliness function and feedback loop processing mechanism of evidence flow.

8. The method according to claim 6, characterized in that Apply dependency-aware propagation algorithms, including: Read the observation data and convert it into evidence stream format, estimate uncertainty and source credibility based on data quality and reliability, create an initial evidence stream set and inject it into the collision-aware evidence stream network for the first round of evidence propagation; Mark and track activation paths during propagation, consider path dependencies during evidence propagation, and form conditional propagation states; Aggregate the evidence flow of each node, handle evidence collisions, calculate the probability distribution of the node state, and identify the most likely state path and state confidence based on this; Combined with the activation path information, the state path is dependency verified and corrected, and the dependency path sensitive reasoning result is output.

9. The method according to claim 6, characterized in that Sparse evidence enhancement and causal strength assessment, including: Identify abnormal state nodes from the dependent path sensitive reasoning results, construct a preliminary causal graph and perform sparse evidence enhancement processing to obtain an enhanced evidence set, comprehensively consider the intervention effect and the maximum potential impact, and calculate the enhanced causal strength. Based on the enhanced causal strength, the enhanced causal link set is identified and optimized.

10. The method according to claim 1, characterized in that Obtain comprehensive assessment results, including: Integrate and fuse spatiotemporal features and strong causal link sets to build a multi-dimensional state assessment framework and generate comprehensive assessment features; Based on the comprehensive assessment characteristics, the health index, risk factor, remaining service life and measurement accuracy index are calculated to form a set of status assessment indicators; Based on the state evaluation indicator set and the strong causal link set, the electric energy meter state is divided into multiple levels, and the specific components or subsystems with abnormalities are determined to generate state classification results and abnormality location results; Generate a visual status report based on status classification results, abnormality location results and diagnosis reports; Generate targeted maintenance decision suggestions based on the condition assessment indicator set and strong causal link set, combined with the maintenance experience knowledge base; Integrate visual status reports and decision recommendations to output comprehensive assessment results.

Images