An intelligent diagnosis and tracing method for an electric energy metering anti-electricity-stealing device based on multi-dimensional feature fusion

Abstract

The application provides an intelligent diagnosis and tracing method for an electric energy metering anti-stealing electricity device based on multi-dimensional feature fusion, relates to the technical field of diagnosis and tracing, and synchronously collects and spatiotemporally aligns multi-dimensional information through a multi-source feature fusion layer, extracts features and vectorizes fusion, discriminates rules and statistics, rapidly locates and labels abnormal dimensions through a first layer of an abnormality detection model based on a multi-dimensional metering feature vector, maps nodes and reasons relationships, correlates causes and reasons and reasons root causes through a second layer of a graph neural network topology knowledge graph based on an abnormal dimension set, dynamically selects formulas and performs weighted combination calculation, multi-formula collaborative estimation and confidence interval quantization through a root cause combination and severity model based on an adaptive electric quantity compensation algorithm, and obtains a final compensation electric quantity result, so that the application improves the automation level and accuracy of electric energy metering anti-stealing electricity diagnosis, and reduces the workload and misjudgment risk of manual troubleshooting.

Description

Technical Field

[0001] This invention proposes an intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion, which relates to the field of diagnosis and traceability technology, specifically to the field of intelligent diagnosis and traceability technology for electricity metering anti-theft devices based on multi-dimensional feature fusion. Background Technology

[0002] Electricity metering is a fundamental aspect of electricity marketing, but existing anti-theft diagnostic technologies mostly rely on single data source analysis, such as using only meter event records or analyzing only load curves. They cannot integrate multi-dimensional information such as waveforms, events, environment, and historical verification, making it difficult to detect covert electricity theft. Anomaly diagnosis is mostly based on isolated threshold alarms, lacking the ability to reason about causal relationships between multiple anomalies. It is difficult to automatically trace back to fundamental physical failures such as loose wiring or transformer saturation, and the investigation still relies on manual experience. Electricity compensation generally uses a single correction coefficient, which cannot adapt to complex root cause scenarios, and the calculation results lack reliable quantitative expression. Summary of the Invention

[0003] This invention provides an intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion, in order to solve the above-mentioned problems: This invention proposes an intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion, the method comprising: S1. Multi-dimensional quantitative feature vectors are obtained by synchronously collecting and spatiotemporally aligning multi-dimensional information, extracting features and fusion vectorization through a multi-source feature fusion layer. S2. By using multi-dimensional quantitative feature vectors, rule and statistical discrimination are performed on the first-layer anomaly detection model, and the anomaly dimensions are quickly located and labeled to obtain the set of anomaly dimensions; S3. By using the set of anomalies, perform node mapping and relational reasoning on the topological knowledge graph of the second-layer graph neural network, as well as causal relationships and root cause reasoning among multiple anomalies, to obtain the root cause combination. S4. The adaptive power replenishment algorithm is dynamically selected and weighted by combining root causes and its severity model, and then subjected to multi-formula collaborative estimation and confidence interval quantification to obtain the final power replenishment result.

[0004] Furthermore, the system includes: The acquisition and fusion module is used to synchronously acquire and spatiotemporally align multi-dimensional information, extract features and vectorize them through a multi-source feature fusion layer to obtain multi-dimensional quantitative feature vectors. The anomaly set module is used to perform rule and statistical discrimination on the first-layer anomaly detection model through multi-dimensional quantitative feature vectors, quickly locate and label anomaly dimensions, and obtain an anomaly dimension set. The root cause combination module is used to perform node mapping and relation reasoning, causal association and root cause reasoning among multiple anomalies, and obtain root cause combinations by using the set of anomaly dimensions to perform node mapping and relation reasoning on the topological knowledge graph of the second-layer graph neural network; The selection and quantization module is used to perform dynamic formula selection and weighted combination calculation, multi-formula collaborative estimation and confidence interval quantization on the adaptive power replenishment algorithm through root cause combination and its severity model, so as to obtain the final power replenishment result.

[0005] The beneficial effects of this invention are as follows: This method solves a series of technical problems in traditional methods, such as isolated data analysis, disconnect between anomaly detection and root cause diagnosis, and reliance on single empirical coefficients for supplementary calculations. Multi-source feature fusion effectively addresses the inconsistency in time references and sampling frequencies of heterogeneous data, ensuring that subsequent analysis is based on unified and complete data. The dual-channel discrimination mechanism of statistics and rules resolves the contradiction of high false negative or false positive rates in single discrimination methods, enabling rapid location and credibility assessment of anomaly dimensions. Graph neural network knowledge graph reasoning solves the technical challenge of automatically identifying causal relationships among multiple anomalies, enabling interpretable reasoning from surface anomalies to deep-seated root causes. Adaptive multi-formula weighted supplementation overcomes the shortcomings of traditional single correction coefficient methods in adapting to complex root cause scenarios, making the supplementary results more scientific and credible. Overall, this method improves the automation level and accuracy of electricity metering anti-theft diagnosis, reducing the workload of manual investigation and the risk of misjudgment. Attached Figure Description

[0006] Figure 1 A schematic diagram of an intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion. Detailed Implementation

[0007] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0008] In one embodiment of the present invention, an intelligent diagnosis and tracing method for electricity metering anti-theft devices based on multi-dimensional feature fusion is proposed, the method comprising: S1. Multi-dimensional quantitative feature vectors are obtained by synchronously collecting and spatiotemporally aligning multi-dimensional information, extracting features and fusion vectorization through a multi-source feature fusion layer. S2. By using multi-dimensional quantitative feature vectors, the first-layer anomaly detection model is subjected to rule and statistical discrimination, and the anomaly dimensions are quickly located and labeled to obtain a set of anomaly dimensions that includes anomaly type and anomaly degree. S3. By using the set of anomalies, perform node mapping and relational reasoning on the topological knowledge graph of the second-layer graph neural network, as well as causal relationships and root cause reasoning among multiple anomalies, to obtain the most likely combination of root causes. S4. The adaptive power replenishment algorithm is dynamically formulated and weighted, calculated using a root cause combination and severity model, and then subjected to multi-formula collaborative estimation and confidence interval quantification to obtain the final power replenishment result. Figure 1 As shown.

[0009] The technical explanation of the above technical solution is as follows: By simultaneously collecting and aligning multi-dimensional information in a multi-source feature fusion layer, extracting features and fusion them into vectorized multi-dimensional measurement feature vectors, the first-layer anomaly detection model is subjected to rule and statistical discrimination, rapid localization and labeling of anomaly dimensions to obtain an anomaly dimension set containing anomaly type and degree. The anomaly dimension set is then used to map nodes and infer relationships in the second-layer graph neural network topology knowledge graph, and causal association and root cause reasoning among multiple anomalies are used to obtain the most likely root cause combination. The root cause combination and its severity model are used to dynamically select formulas and perform weighted combination calculations for the adaptive power replenishment algorithm, and multi-formula collaborative estimation and confidence interval quantification to obtain the final power replenishment result, forming a complete intelligent diagnostic and traceability chain from data fusion to anomaly detection to root cause reasoning and finally to power replenishment.

[0010] The system uses a multi-source data acquisition interface to collect real-time waveform data (including voltage, current, and harmonics), internal event records of the electricity meter (including voltage loss, cover opening, and programming), environmental parameters (including temperature, humidity, and load fluctuations), and historical verification data in parallel and synchronously. The timestamps of each source data are normalized and interpolated to obtain a multi-source raw dataset. From this raw dataset, noise filtering, missing value imputation, outlier removal, data quality cleaning, and standardization transformation are performed on various data types to obtain a standardized single-dimensional feature sequence. From this standardized single-dimensional feature sequence, principal component analysis and correlation reduction are performed on each dimension of the features to eliminate redundancy and retain key information, resulting in a core feature subset. Finally, from this core feature subset, vector concatenation and normalization mapping, unified feature vector construction, and dimension alignment are performed on the multi-dimensional features to obtain a multi-dimensional metering feature vector. In the specific implementation of parallel synchronous acquisition and time alignment, real-time waveform data, internal event records of the electricity meter, environmental parameters, and historical verification data are acquired in parallel and synchronously through a multi-source data acquisition interface. The time stamps of each data source are parsed in a unified format and converted in precision to obtain the data and its original timestamp. The time reference of each data source is checked for consistency using the data and its original timestamp. Linear compensation and drift correction of clock deviations of each source are performed with the internal clock of the electricity meter as the main reference to obtain the timestamp sequence of each source. A unified target time grid is determined for all data sources using the timestamp sequence of each source, and the data of each source is resampled and located. High sampling rate data is downsampled, low sampling rate data is encrypted by cubic spline interpolation, and discrete event data is filled with forward nearest neighbor to obtain a multi-source synchronous dataset. Boundary effect checks and alignment quality assessments are performed on the multi-source synchronous dataset for each dimension of data. Boundary points with interpolation errors at both ends exceeding the tolerance threshold are removed to obtain the original multi-source dataset.

[0011] A statistical anomaly scoring sequence is obtained by real-time sliding window comparison of the statistical distribution baseline of each feature using multi-dimensional measurement feature vectors, and by calculating the deviation of mean and variance anomalies. The statistical anomaly scoring sequence is then used to trigger multiple preset rule thresholds, including harmonic exceedance limits, undervoltage frequency thresholds, and cap opening count thresholds, through rule matching and limit-crossing marking to obtain a set of rule trigger flags. The rule trigger flag set and the statistical anomaly scoring sequence are then used to perform comprehensive confidence ranking, false alarm elimination, and anomaly priority ranking to obtain a list of high-confidence anomaly dimensions. Finally, the high-confidence anomaly dimension list is used to encode the type and severity of each anomaly dimension (including mild, moderate, and severe), and the anomaly dimension attributes are labeled and structured to obtain a set of anomaly dimensions containing anomaly types and degrees. In the specific implementation of rule trigger determination, the numerical values ​​of multiple preset rule thresholds are compared one by one by statistically analyzing the abnormal scoring sequence. The real-time scores or real-time measurement values ​​of each dimension are compared with the corresponding rule limits, and whether the limits are reached or exceeded is recorded to obtain the preliminary trigger determination result. The preliminary trigger determination result is used to continuously verify each trigger event. Only when the same rule is triggered at multiple consecutive sampling points is it confirmed as a valid trigger and a valid trigger flag is obtained. The valid trigger flag is used to verify the mutual exclusion and association logic of each rule item. Rules with mutual exclusion relationships are selected according to priority rules, and rules with association relationships are merged and marked to obtain a rule trigger set. The rule trigger set is used to encode and structure each trigger item, and the rule number, rule name, trigger time, and corresponding threshold information of each trigger are organized into a unified flag format to obtain a rule trigger flag set.

[0012] By performing anomaly node localization, node activation, and initial feature assignment on a pre-constructed metering device topology knowledge graph containing anomaly types and degrees, an activated set of anomaly nodes and their adjacent subgraphs are obtained. Then, a graph neural network model is subjected to multi-layer graph convolutional propagation, inter-node relationship strength calculation, and attention weight allocation using the anomaly node set and its adjacent subgraphs to obtain the influence probability matrix between nodes. Using the influence probability matrix, path search and causal chain backtracking are performed on multiple anomaly nodes, and root cause candidate nodes are sorted and combined to obtain a root cause candidate combination list and its confidence score. Finally, physical failure mode matching (including loose wiring, current transformer saturation, and electricity theft bypass) is performed on the highest-scoring combination using the root cause candidate combination list and its confidence score, along with root cause name standardization and failure path description generation to obtain root cause combinations and their corresponding physical failure paths. In the specific implementation of graph neural network propagation computation, the first layer of the graph neural network model is summarized by a set of abnormal nodes and their adjacent subgraphs. The features of each node are weighted and averaged with the features of its directly connected nodes to obtain the intermediate features after the first layer summary. The intermediate features are used to activate the nonlinear transformation layer of the graph neural network model, and the negative values ​​in the intermediate features are set to zero to enhance the discriminativeness of the features to obtain the preliminary feature vector. The preliminary feature vector is used to perform a secondary summary of the higher-order neighbor features of the second layer of the graph neural network model. The features of each node and its indirectly connected nodes are weighted and merged according to the strength of the association to obtain the higher-order features after the second layer summary. The higher-order features are used to perform a probability transformation on the output layer of the graph neural network model, and the influence degree between each pair of connected nodes is converted into a probability value between 0 and 1 to obtain the influence probability matrix between each node.

[0013] The candidate supplementary formula set is obtained by performing root cause formula mapping retrieval, applicable formula type screening, and matching degree calculation on the built-in supplementary formula library through root cause combination and its corresponding physical failure path; the weight coefficients of each formula are dynamically assigned through the candidate supplementary formula set and severity model, and the severity correction calculation of the formula effectiveness is obtained to obtain the multi-formula contribution coefficient; the multi-formula contribution coefficient is used to perform parallel calculation and result fusion of multiple sets of supplementary formulas, cross-validation between multiple estimates and outlier removal to obtain the preliminary supplementary power estimate; the preliminary supplementary power estimate is used to perform Monte Carlo sampling and confidence interval construction, upper and lower bound calculation and interval symmetry adjustment on the estimation error distribution of each formula to obtain the final supplementary power result. In the specific implementation of dynamic weighting coefficient assignment, the historical application effect of each candidate formula in the candidate supplementary formula set and severity model is queried. The frequency of use and calculation result deviation of each formula in similar historical problems are statistically analyzed to obtain the historical reliability score of each formula. The basic weight of each formula is calculated based on the historical reliability score and the current severity level. The initial weight value of each formula is obtained by multiplying the historical reliability score by the weight factor corresponding to the severity. The weights of all candidate formulas are normalized based on the initial weight value so that the sum of the weights of all formulas equals 1. Then, the applicability of each formula to the current root cause scenario is reviewed, and the weights of formulas that are not fully applicable are reduced to obtain the weight of each formula. Formulas with too low weights are removed based on the weights of each formula. The weights of the remaining formulas are redistributed and finally balanced to obtain the contribution coefficient of the multi-formula formula.

[0014] The working principle and technical effect of the above technical solution are as follows: A multi-source feature fusion layer is used to uniformly collect and integrate heterogeneous information scattered across different data sources. Real-time waveform data records instantaneous changes in voltage, current, and harmonics using high-frequency sampling. Internal event logs of the electricity meter are stored using discrete triggering to record operation logs such as voltage loss, cover opening, and programming. Environmental parameters are collected using a low-frequency periodic method to collect temperature, humidity, and load fluctuation information. Historical verification data is stored in archive format, preserving past verification records. These data sources differ significantly in time base, sampling frequency, and data format. The fusion layer maps the data from each source to a unified time grid through spatiotemporal alignment, then transforms the original signals into numerical features through feature extraction, and finally concatenates the multi-dimensional features into a fixed-length feature vector through vectorized fusion.

[0015] After receiving the aforementioned feature vectors, the first-layer anomaly detection model employs a dual-channel strategy combining statistical and rule-based discrimination for rapid screening. Statistical discrimination identifies abnormal fluctuations by comparing the deviation of current features from historical normal baselines, while rule-based discrimination makes hard judgments based on preset technical thresholds (such as harmonic exceedance limits, undervoltage frequency thresholds, etc.). After comprehensive confidence evaluation of the outputs from both channels, suspected false alarms with low confidence are eliminated, high-confidence anomaly dimensions are retained, and each anomaly dimension is assigned a type code and severity classification label, forming a structured list of anomaly dimensions.

[0016] The second-layer root cause reasoning model utilizes a graph neural network to perform reasoning calculations on a pre-constructed knowledge graph of metering device topology. The knowledge graph contains device nodes such as electricity meters, transformers, switches, and terminals, along with their electrical connections. Anomalies are mapped to corresponding nodes in the graph and activated. The graph neural network calculates the causal probability between nodes through message propagation across multiple layers, and finally outputs the most probable root cause combination and its physical failure path description through reverse path tracing.

[0017] The adaptive power replenishment algorithm dynamically selects applicable calculation formulas from the built-in replenishment formula library based on the root cause type and severity output by S3. After assigning weight coefficients matching the severity to each formula, it performs parallel calculations, cross-validates and merges multiple sets of calculation results, and finally gives the replenished power result with confidence interval.

[0018] This method addresses a series of technical problems inherent in traditional methods, such as isolated data analysis, disconnect between anomaly detection and root cause diagnosis, and reliance on single empirical coefficients for supplementary calculations. Multi-source feature fusion effectively resolves the inconsistencies in time references and sampling frequencies of heterogeneous data, ensuring subsequent analysis is based on unified and complete data. A dual-channel statistical and rule-based discrimination mechanism resolves the contradiction of high false negative or false positive rates associated with single discrimination methods, enabling rapid location and credibility assessment of anomalies. Graph neural network knowledge graph reasoning overcomes the technical challenge of automatically identifying causal relationships among multiple anomalies, achieving interpretable reasoning from surface anomalies to deep-seated root causes. Adaptive multi-formula weighted supplementation overcomes the limitations of traditional single correction coefficient methods in adapting to complex root cause scenarios, making the supplementary results more scientific and credible. Overall, this method improves the automation and accuracy of electricity metering anti-theft diagnosis, reducing the workload of manual investigation and the risk of misjudgment.

[0019] In one embodiment of the present invention, S1 includes: The multi-source data acquisition interface is used to collect real-time waveform data (voltage, current, harmonics), internal event records of the energy meter (voltage loss, cover opening, programming), environmental parameters (temperature and humidity, load fluctuations) and historical verification data in parallel and synchronously. The timestamps of each source data are normalized and interpolated to obtain a multi-source raw dataset. By performing noise filtering, missing value imputation, outlier removal, data quality cleaning and standardization transformation on various types of data from multiple source raw datasets, a standardized single-dimensional feature sequence is obtained. By performing principal component analysis and correlation reduction, feature redundancy elimination and key information retention on each dimension of the feature sequence through standardized single-dimensional feature sequence, a core feature subset is obtained. By performing vector concatenation and normalization mapping on multi-dimensional features through a subset of core features, as well as constructing unified feature vectors and aligning dimensions, multi-dimensional quantitative feature vectors are obtained.

[0020] The electricity meter records voltage and current data multiple times per second, the temperature sensor only records data every few minutes, and the event records occur randomly, so their timelines are completely misaligned. This step is to unify the timestamps of all the data to the same clock, and use interpolation to fill in the missing time points between densely sampled and sparsely sampled data, so that all the data are aligned on the same time grid.

[0021] The raw data often contains obviously erroneous values, such as voltage suddenly dropping to tens of thousands of volts. This is clearly caused by sensor noise or interference, and these spikes need to be removed. If there are gaps in the data, they should be filled in using the average of the preceding and following times. After cleaning, all dimensions of data should be transformed to the same scale, for example, by converting them to multiples of deviation from normal values, so that data from different units can be compared side-by-side.

[0022] Many features actually describe the same thing; for example, phase A voltage and phase A current are highly correlated, so putting them together would be redundant. This step uses principal component analysis to merge the redundant features, retaining only the most core features that best represent the essence of the problem, which reduces the amount of computation without losing key information.

[0023] Arrange all the remaining features after dimensionality reduction into a long string of numbers in a fixed order, much like filling a bunch of scattered information into a fixed-format table, and finally output a standardized multidimensional feature vector.

[0024] The working principle and technical effects of the above technical solution are as follows: Four types of data are acquired in parallel and synchronously through a multi-source data acquisition interface: real-time waveform data, including instantaneous sampled values ​​of voltage, current, and each harmonic; internal event records of the electricity meter, including timestamps and event type codes for undervoltage events, cover opening events, and programming events; environmental parameters, including periodic readings from temperature and humidity sensors and statistics calculated from load fluctuations; and historical verification data, including error records and correction coefficients from the meter's past verification cycles. These data immediately enter a time alignment process after acquisition. The timestamps of each source data are unified to the same reference clock. Data with higher sampling frequencies are aligned with lower frequency data through downsampling. Data with lower sampling frequencies is supplemented with missing time points through interpolation and encryption. Discrete event-type data is effectively marked within the time period after the event occurs through forward padding, ultimately forming a complete dataset for each dimension on a unified time grid.

[0025] The system performs quality cleaning on the aligned original dataset: median filtering or mean filtering is used to suppress high-frequency measurement noise in the waveform data; missing data points are filled using linear interpolation of adjacent times or the mean of similar days at the same time; and outliers that significantly exceed reasonable ranges (such as voltage sudden changes to several times the rated value) are identified and removed. After cleaning, the data in each dimension is standardized, converting physical quantities of different dimensions into dimensionless deviation values ​​centered on the historical mean and measured in units of historical standard deviation, allowing subsequent processing to compare features of different properties on the same scale.

[0026] The system performs dimensionality reduction on the standardized high-dimensional feature sequences: on the one hand, it calculates the variance contribution rate of each principal component through principal component analysis and selects the principal components based on the cumulative contribution rate; on the other hand, it identifies highly correlated redundant feature pairs through correlation analysis and retains the features that are more strongly associated with the target state. Through dimensionality reduction, the original feature space of tens or hundreds of dimensions is compressed into more than ten core features, significantly reducing the computational complexity of subsequent models while retaining most of the information.

[0027] The system concatenates the dimensionality-reduced features into a one-dimensional vector in a preset fixed order, and then performs a final normalization on the concatenated vector to ensure that the numerical range of each dimension is consistent. The output is a standard multidimensional quantitative feature vector that can be directly input into anomaly detection models.

[0028] This method addresses the technical challenge of directly fusing multi-source heterogeneous data in practical engineering. Parallel synchronous acquisition resolves the asynchrony caused by different data sources due to varying communication protocols and interface standards. Timestamp normalization and interpolation alignment address the inability to directly correlate high-frequency waveform data, low-frequency environmental data, and discrete event data on a time base, enabling subsequent analysis to accurately correlate states across dimensions at the same moment. Noise filtering and outlier removal address data quality issues caused by sensor noise, communication interference, and occasional faults, preventing erroneous data from misleading diagnostic results. Standardization transformation resolves the inability to directly compare and fuse different physical quantities such as voltage, current, and harmonics due to differences in dimensions and numerical ranges. Principal component analysis and correlation dimensionality reduction address the issue of abundant redundant information in high-dimensional feature spaces, effectively reducing model complexity and computational resource consumption while preserving key diagnostic information. Vector concatenation and normalization mapping resolve inconsistencies in format and scale among multi-dimensional features, providing a unified and standardized data input format for subsequent model layers. This method improves the quality and efficiency of multi-source data fusion.

[0029] In one embodiment of the present invention, the method involves parallel and synchronous acquisition of real-time waveform data (voltage, current, harmonics), internal event records of the electricity meter (voltage loss, cover opening, programming), environmental parameters (temperature and humidity, load fluctuations), and historical verification data through a multi-source data acquisition interface. The timestamps of each source data are then normalized and interpolated to obtain a multi-source raw dataset, including: The system uses a multi-source data acquisition interface to collect real-time waveform data, internal event records of the electricity meter, environmental parameters and historical verification data in parallel and synchronously. It performs unified format parsing and precision conversion on the time stamps of each data source to obtain the data from each source and its original timestamp. The consistency of the time base of each data source is checked by using the data from each source and its original timestamp. The internal clock of the electricity meter is used as the main reference for linear compensation and drift correction of the clock deviation of each source, and the timestamp sequence of each source is obtained. By determining a unified target time grid for all data sources through timestamp sequences from each source and resampling and locating the data from each source, high-sampling-rate data is extracted and downsampled, low-sampling-rate data is encrypted using cubic spline interpolation, and discrete event-type data is filled with forward nearest neighbor to obtain a multi-source synchronized dataset. By using a multi-source synchronized dataset, boundary effect checks and alignment quality assessments are performed on the data in each dimension. Boundary points with interpolation errors exceeding the tolerance threshold at both ends are removed, and the original multi-source dataset is obtained.

[0030] The working principle and technical effect of the above solution are as follows: Four types of data are collected in parallel and synchronously through a multi-source data acquisition interface to obtain the raw data output from each data source and its accompanying timestamps. The timestamp formats of different data sources may vary; some use Unix timestamps, some use accumulated milliseconds within the device, and some use standard date and time strings. The system first parses these different timestamp formats into a unified high-precision time format (such as a millisecond-level Unix timestamp), making the time representations of all data sources comparable.

[0031] The system uses the internal clock of the electricity meter as the primary reference clock to verify the consistency of the time base of each data source. Since each data source may use its own independent clock, there will be a fixed time deviation and drift error accumulating over time. The system calculates the deviation between each source clock and the primary reference clock, uses a linear compensation method to eliminate the fixed deviation, and performs dynamic correction by continuously monitoring drift trends, ensuring that the timestamps of each data source accurately correspond under the same time base.

[0032] After unifying the time base of all data sources, the system determines a unified target time grid. The time interval of this grid is typically the reciprocal of the highest sampling frequency among all data sources to ensure that information from the highest frequency data is not lost. The system then repositions the original sampling points from each data source onto this target grid: for data with sampling frequencies higher than the target grid (such as voltage and current waveforms), sampling is performed at equal intervals according to the grid interval; for data with sampling frequencies lower than the target grid (such as temperature and humidity), cubic spline interpolation is used to densify the values ​​at grid points between adjacent sampling points; for discrete event-type data (such as pressure loss and lid opening records), forward nearest neighbor padding is used, that is, the value at the time of the event is moved forward along the time axis until before the next event occurs. Through the above processing, all source data are unified onto the same set of time coordinates.

[0033] The system performs boundary effect checks and alignment quality assessments on the interpolated and aligned data. Interpolation errors are often large at the two ends of the data sequence (the beginning and end segments) due to a lack of sufficient neighboring data. The system calculates the relative interpolation error of each data point in the boundary regions at both ends, removes boundary points with errors exceeding a preset tolerance threshold, and marks dimensions with substandard data quality to ensure that the output multi-source original dataset contains valid data at both ends of the time axis.

[0034] This method solves the technical challenge of accurate time alignment for multi-source data by defining specific implementation steps for timestamp normalization and interpolation alignment. Unified time stamp format parsing resolves the parsing confusion caused by different time representation methods used by different data sources, enabling correlation between data sources within the same time coordinate system. Linear clock skew compensation and drift correction address the technical issues of fixed deviations and cumulative drift between independent clocks of different data sources, ensuring consistency of time references for all data sources under long-term operation. Unified determination of the target time grid and resampling location solve the problem of direct correspondence between data with significantly different sampling frequencies, enabling collaborative analysis of high-frequency, low-frequency, and discrete event data on the same time axis. Cubic spline interpolation encryption ensures the smoothness and accuracy of low-frequency continuous signals during encryption, avoiding errors potentially introduced by simple linear interpolation. Forward nearest neighbor filling addresses the rationality of state representation within event intervals for discrete event data. Boundary effect checking and tolerance threshold removal address the inherent problem of decreased accuracy at both ends of the data in interpolation algorithms, ensuring the overall quality of the output data. This method provides a complete and reliable technical solution for time alignment of multi-source heterogeneous data, ensuring accuracy and consistency of data in the time dimension.

[0035] In one embodiment of the present invention, S2 includes: By performing real-time sliding window comparison of the statistical distribution baseline of each feature using multi-dimensional econometric feature vectors, and calculating the deviation of mean and variance anomalies, a statistical anomaly score sequence is obtained. By statistically analyzing the abnormal scoring sequence, the system performs item-by-item trigger judgment, rule matching, and limit overrun marking on multiple preset rule thresholds (harmonic exceedance limit, undervoltage frequency threshold, lid opening count threshold, etc.) to obtain a set of rule trigger flags. By using a set of rule-triggered flags and a statistical anomaly scoring sequence, the anomaly dimensions are comprehensively sorted by confidence, false alarms are eliminated, and anomalies are prioritized to obtain a list of high-confidence anomaly dimensions. By using a list of high-confidence anomaly dimensions, each anomaly dimension is coded by type and graded by severity (mild / medium / severe), and its attributes are labeled and encapsulated in a structured manner to obtain a set of anomaly dimensions that includes anomaly type and anomaly severity.

[0036] By comparing data from a short period before and after the current moment with data from the same period during the device's normal operation in the past, the deviation of each indicator from the normal state is calculated. For example, if the average voltage is several standard deviations higher than usual, or the fluctuation range is many times larger than usual, these deviations are output in the form of scores.

[0037] We use some hard, clear rules to determine the trigger, such as triggering an alarm if the harmonic distortion rate exceeds 5% of the national standard, triggering an alarm if there are more than three pressure loss events in a day, and triggering an alarm if the cover is opened during non-working hours at night. These rules are as clear and specific as traffic lights, and any trigger will generate an alarm and be recorded.

[0038] The statistical deviation score calculated in the first step and the rule trigger flags in the second step are combined to assign a total confidence score to each anomaly dimension. Anomalies with very low scores are likely false alarms, such as temporary fluctuations caused by thunderstorms, and are directly removed, retaining only high-confidence anomalies.

[0039] Each confirmed anomaly is clearly labeled, such as whether it is a third harmonic problem or a C-phase undervoltage problem, and its severity is assessed as minor, moderate, or severe. All anomalies are compiled into a structured list and directly submitted to the next step.

[0040] The working principle and technical effect of the above technical solution are as follows: A statistical distribution baseline is established for each feature dimension. The baseline data comes from the historical feature values ​​of the meter during the same period in its normal operation (i.e., the time period without known anomalies and event records). For example, if the current time is 10:00 AM on a weekday, the baseline is the historical data of the same period in the past 30 weekdays. The system uses a sliding window method to calculate the mean and variance of each dimension in real time, and compares them with the baseline mean and variance to calculate the mean deviation (reflecting the degree of deviation of the current value from the normal level) and the variance deviation (reflecting the degree of change of the current value fluctuation range relative to the normal fluctuation range). The two deviations are combined to form a statistical anomaly score sequence for each dimension.

[0041] The system sends statistical anomaly scores and some raw measurement values ​​into the rule discrimination channel, comparing them item by item with multiple preset rule thresholds. These rules include, but are not limited to: whether the harmonic content exceeds the limits specified by national or industry standards, whether the frequency of voltage loss events per unit time exceeds a threshold, whether the opening event occurs during unauthorized periods, whether the three-phase current imbalance exceeds a threshold, and whether the power factor undergoes a sudden change. Each rule corresponds to a specific numerical threshold or logical condition. Once the condition is met, the corresponding rule flag is triggered, ultimately forming a set of rule trigger flags containing multiple trigger records.

[0042] The system integrates the statistical anomaly score sequence with the rule trigger flag set. Each anomaly dimension has both a statistical score and a rule trigger status. The system uses a weighted fusion method to calculate the overall confidence level of each dimension, with the statistical score and rule trigger each having a certain weight ratio. The overall confidence level reflects the degree of certainty that anomalies actually exist in that dimension. The system sorts all anomaly dimensions from high to low based on their overall confidence level. Dimensions with confidence levels below a preset lower limit are judged as false alarms (such as temporary fluctuations caused by external disturbances like lightning strikes or high-power equipment switching) and are removed. The system then outputs a list of high-confidence anomaly dimensions after filtering.

[0043] The system annotates each item in the high-confidence anomaly dimension list with attributes: assigning a type code based on the specific manifestation of the anomaly (e.g., H03 for third harmonic anomaly, UV_C for C-phase undervoltage, OC for cap opening event, etc.), and comprehensively assessing the severity level (mild, moderate, severe) based on the magnitude and duration of the deviation from the baseline. All anomaly dimensions after annotation are encapsulated into a structured anomaly dimension set, with each anomaly item containing a dimension identifier, type code, severity, occurrence time, and related feature value summary information, serving as input for S3 root cause inference.

[0044] This method addresses the technical challenge of single anomaly detection methods struggling to balance comprehensiveness and accuracy. Statistical distribution baseline comparison overcomes the limitation of fixed thresholds in adapting to variations in normal fluctuation ranges across different devices and time periods, enabling anomaly detection to be adaptive. Mean and variance anomaly deviation calculations address the issue of focusing solely on average changes while ignoring fluctuation anomalies, resulting in more comprehensive detection. Pre-setting multiple rule thresholds addresses the potential for statistical methods to miss known and definite anomaly patterns, making detection results more reliable. Comprehensive confidence ranking and false alarm elimination address the high false alarm rate caused by complex on-site environments and numerous interference sources, improving the usability of detection results. Type coding and severity grading address the issues of limited anomaly output formats and insufficient information content. This method improves the accuracy, adaptability, and richness of output information in anomaly detection, significantly reducing the false alarm rate while providing high-quality input data for deep root cause inference.

[0045] In one embodiment of the present invention, the step of performing item-by-item trigger determination, rule matching, and limit-crossing marking on preset multiple rule thresholds (harmonic exceedance limit, undervoltage frequency threshold, cover opening count threshold, etc.) by statistically analyzing anomaly scoring sequences to obtain a set of rule trigger flags, including: By statistically analyzing the abnormal scoring sequence, the numerical values ​​of multiple preset rule thresholds are compared one by one. The real-time scores or real-time measurement values ​​of each dimension are compared with the corresponding rule limits and the limits are recorded to obtain the preliminary trigger judgment results. The initial trigger judgment results are used to continuously verify each trigger event. Only when the same rule is triggered at multiple consecutive sampling points is it confirmed as a valid trigger and a valid trigger flag is obtained. By valid trigger flags, mutual exclusion and correlation logic checks are performed on each rule item. For rules with mutual exclusion (such as simultaneously triggering harmonic exceedance and voltage undervoltage in the same dimension), they are selected according to priority rules. For rules with correlation, they are merged and marked to obtain the rule trigger set. By encoding and structurally encapsulating each trigger item through a rule trigger set, the rule number, rule name, trigger time, and corresponding threshold information of each trigger are organized into a unified flag format to obtain a rule trigger flag set.

[0046] The working principle and technical effect of the above technical solution are as follows: The statistical anomaly scoring sequence and real-time measurement values ​​of each dimension are input into the rule discrimination engine. The engine stores multiple preset rules, each containing three basic elements: the applicable feature dimension or calculation expression, the trigger threshold or logical condition, and the rule priority. The system retrieves each rule one by one, comparing the corresponding feature value or statistical score with the rule threshold. If the threshold is reached or exceeded, the trigger status of the rule is marked as "yes"; otherwise, it is marked as "no." After all rules are traversed, a preliminary trigger judgment result is formed.

[0047] To prevent false triggers caused by sporadic noise or transient disturbances, the system performs a continuous verification on rules initially identified as triggered. Specifically, for each triggered rule, the system backtracks and checks its triggering state over several consecutive sampling points (e.g., three consecutive sampling periods). Only if the rule is in a triggered state at the current moment and for several consecutive previous moments is it confirmed as a valid continuous trigger; otherwise, it is considered a sporadic disturbance and not confirmed. This mechanism effectively filters out false triggers caused by non-continuous factors such as measurement noise and transient communication interference.

[0048] The system performs logical verification on valid trigger flags that pass continuous validation. Different rules may have mutually exclusive or correlated relationships. For example, if the "harmonic exceedance" rule and the "low voltage" rule are triggered simultaneously on the same dimension at the same time, but harmonic exceedance might be a secondary phenomenon caused by low voltage, then priority must be prioritized to retain the more fundamental rule trigger. Similarly, when multiple related rules (such as similar anomalies at different measurement points in the same circuit) are triggered simultaneously, the system merges them into a single comprehensive event to avoid duplicate counting. Through mutual exclusion and correlation logical verification, the rule trigger set is purified and simplified at the semantic level.

[0049] The system encodes and structurally encapsulates each triggering rule after it has undergone persistence and logical verification. Each triggering rule is assigned a unique rule number, and along with key information such as the rule name, trigger time, corresponding feature value or statistical score at the time of triggering, and the threshold used, it is organized into a standardized rule trigger record according to a unified field format. All records are aggregated to form a complete set of rule trigger flags, which serves as one of the inputs for comprehensive confidence assessment.

[0050] This method addresses the technical issues of high false alarm rates and unclear handling of logical conflicts in simple threshold comparisons during engineering applications by refining the specific implementation steps of the rule discrimination channel. Item-by-item numerical comparison and preliminary trigger determination resolve the execution framework issues when multiple rules are judged in parallel, standardizing the rule processing flow. Continuous verification solves the problem of single-point false triggers caused by instantaneous noise or occasional disturbances, significantly improving the reliability of rule triggering. Mutual exclusion and correlation logic verification resolves potential logical conflicts and duplicate counting issues when multiple rules are triggered simultaneously, making the triggering results more semantically accurate and concise. Encoding and structured encapsulation solve the problems of inconsistent triggering result formats and incomplete information, providing standardized data input for subsequent confidence fusion. This method improves the anti-interference capability and logical consistency of the rule discrimination channel, further reduces the false alarm rate, and makes the information in the rule triggering results more complete and usable.

[0051] In one embodiment of the present invention, S3 includes: By using an anomaly dimension set that includes anomaly type and anomaly degree, the pre-constructed metering device topology knowledge graph is used to locate, activate, and assign initial feature values ​​to anomaly nodes, thereby obtaining the activated set of anomaly nodes and its adjacent subgraph. By performing multi-layer graph convolution propagation, calculating the strength of relationships between nodes, and assigning attention weights to the graph neural network model using the set of abnormal nodes and their adjacent subgraphs, the influence probability matrix between each node is obtained. By using the influence probability matrix, path search and causal chain backtracking are performed on multiple anomaly nodes, and root cause candidate nodes are sorted and combined enumerated to obtain a root cause candidate combination list and its confidence score. By using the root cause candidate combination list and its confidence score, the combination with the highest score is matched for physical failure mode (such as loose wiring, CT saturation, power theft bypass, etc.), the root cause name is standardized and the failure path description is generated, and the root cause combination and its corresponding physical failure path are obtained.

[0052] The system has a pre-built topology knowledge graph, which shows all device nodes such as electricity meters, transformers, switches, and terminals, as well as their electrical connections. Now, the anomaly list obtained in the previous step is mapped onto this graph, the corresponding device nodes are found and activated, and the related subgraphs around these nodes are also extracted to narrow down the scope of subsequent reasoning.

[0053] By utilizing graph neural networks to propagate information between activated nodes, each anomalous node transmits its anomalous features to neighboring nodes along the connections. After multiple layers of propagation, the algorithm can calculate the probability of a causal relationship between any two nodes, forming an influence probability matrix.

[0054] Guided by the probability matrix, the propagation path is traced back along the high-probability path, from the apparent anomaly all the way to the source, enumerating all possible combinations of root cause nodes. Then, the most likely root cause candidates are listed according to the confidence level of each combination from high to low.

[0055] The system matches the root cause combination with the highest confidence level against a pre-defined failure mode library, such as loose wiring, current transformer saturation, and power theft bypass. Upon successful matching, the root cause is standardized and named, and a human-readable cause description is generated, such as "Loose C-phase wiring leads to increased contact resistance, causing harmonics and intermittent voltage loss." Finally, a clear root cause combination and its physical failure path are output.

[0056] The working principle and technical effect of the above technical solution are as follows: Anomaly dimension sets are mapped onto a pre-constructed metering device topology knowledge graph. The knowledge graph is pre-constructed before system operation, and the node types included in the graph include, but are not limited to: electricity meter nodes (including each phase measurement unit), current transformer nodes, voltage transformer nodes, terminal block nodes, switch nodes, busbar nodes, etc.; the edges in the graph represent the physical electrical connection relationships between nodes (such as voltage loop connection, current loop series connection, control signal connection, etc.), and spatial location relationships (such as within the same meter box, within the same distribution bay). Each anomaly dimension has a predefined mapping rule with the knowledge graph nodes. For example, the "C-phase voltage loss" anomaly is mapped to the C-phase voltage input terminal node of the electricity meter, and "3rd harmonic exceedance" is mapped to the upstream transformer node connected to the electricity meter. After mapping is completed, the corresponding nodes are activated and assigned anomaly characteristic values ​​(anomaly type and severity). Simultaneously, the system extracts the adjacency subgraph of these activated nodes (including the activated node itself and its adjacent nodes within several hop ranges), narrowing the reasoning scope from the global graph to a local region related to the current anomaly.

[0057] The system performs graph neural network (GNN) computation on the activated set of abnormal nodes and their adjacent subgraphs. GNN is a deep learning method specifically designed for graph-structured data; its core operation is message passing along the edges of the graph. In the first layer, each activated node collects features from its directly connected neighbors, weights its own features with those of its neighbors according to connection weights, and generates a summarized intermediate feature. This intermediate feature is then processed by non-linear activation, setting negative values ​​to zero to enhance its discriminative power. In the second layer, each node continues to collect information from its second-order neighbors updated in the first layer, performing a second-level aggregation to incorporate information from more distant neighbors into each node's features. After two layers of propagation, for each pair of directly connected nodes, the system calculates an influence probability value based on the final feature vectors of the two nodes, representing the likelihood of a causal relationship between them. All probability values ​​are summed to form an influence probability matrix.

[0058] The system utilizes an influence probability matrix for path search and causal chain backtracking. Starting from each terminal anomalous node (i.e., a node that only exhibits an anomalous result and has no subsequent anomalous nodes), it traces backward level by level along edges with influence probabilities higher than a preset threshold to find the source node that may have influenced the anomaly, forming multiple candidate causal paths. Nodes in all paths are statistically analyzed and combined, identifying nodes that frequently appear at the intersections of different paths as potential root cause nodes, and forming multiple root cause combination hypotheses, each accompanied by a comprehensive confidence score.

[0059] The system performs physical failure mode matching on the root cause combination with the highest confidence level. The system's built-in failure mode library stores common physical failure types and their typical characteristics. For example, "loose wiring" typically corresponds to increased contact resistance, intermittent voltage loss, and abnormal harmonics; "current transformer saturation" typically corresponds to metering deviations and specific harmonic characteristics under high current; and "electricity theft bypass" typically corresponds to abnormally low current while voltage remains normal. The system matches the characteristic modes of the root cause combination with the failure mode library. Upon successful matching, it outputs a standardized root cause name and a human-readable failure path description, such as "Loose C-phase wiring leads to increased contact resistance, causing harmonics and intermittent voltage loss."

[0060] This method solves the technical challenge of automatically identifying causal relationships and tracing root causes from multiple concurrent anomalies. Knowledge graph anomaly node localization and activation solves the problem of mapping the association between anomaly information and physical devices, enabling analysis to map from the data space to the physical space. Adjacency subgraph extraction addresses the problem of excessive computational load for reasoning across the entire graph, improving computational efficiency by focusing on local areas. Graph neural network multi-layer propagation solves the problem of difficulty in quantifying the transmission path of anomaly effects in electrical connection relationships, making causal inference computable. Influence probability matrix solves the problem of difficulty in determining the strength and direction of relationships between multiple anomalies, providing quantitative evidence for path tracing. Path search and causal chain backtracking solve the problem of reverse reasoning from manifest anomalies to root causes, making root cause localization interpretable. Physical failure mode matching addresses the problem of overly abstract root cause analysis results, transforming algorithm output into physical descriptions easily understood by engineers. Overall, this method achieves a qualitative leap in anomaly diagnosis, moving from "knowing where the anomaly is" to "knowing why the anomaly is."

[0061] In one embodiment of the present invention, the step of performing multi-layer graph convolutional propagation, calculating the strength of relationships between nodes, and assigning attention weights to a graph neural network model through a set of abnormal nodes and their adjacent subgraphs to obtain an influence probability matrix between nodes includes: The first layer of the graph neural network model is summarized by aggregating the neighbor features of the abnormal node set and its adjacent subgraph. The features of each node are weighted and averaged with the features of its directly connected nodes to obtain the intermediate features after the first layer is summarized. The nonlinear transformation layer of the graph neural network model is activated by intermediate features, and the negative values ​​in the intermediate features are set to zero to enhance the discriminative power of the features, thus obtaining a preliminary feature vector. The second layer of the graph neural network model is summarized by the preliminary feature vectors. The features of each node and its indirectly connected nodes are weighted and merged according to the strength of the association to obtain the high-order features after the second layer summary. By performing probability transformation on the output layer of the graph neural network model using higher-order features, the influence degree between each pair of connected nodes is converted into a probability value between 0 and 1, thus obtaining the influence probability matrix between each node.

[0062] The working principle and technical effect of the above technical solution are as follows: The activated set of abnormal nodes and their adjacent subgraphs are input into the first layer of the graph convolutional network of the graph neural network model. In this layer, the network performs a neighborhood feature aggregation operation on each node: for each node in the graph, its own feature vector and the feature vectors of all its directly connected neighboring nodes are weighted and summed or weighted averaged according to the weights of the connecting edges. The edge weights are pre-set according to the physical connection attributes between nodes (such as wire impedance, mutual inductance, electrical distance, etc.), reflecting the differences in the strength of different connection relationships when propagating the impact of anomalies. After weighted aggregation, each node obtains an intermediate feature vector that integrates its own state and first-order neighbor information. The dimension of this feature vector is the same as the dimension of the input feature vector.

[0063] The system feeds the intermediate feature vector output from the first layer into a nonlinear transformation layer for activation. The nonlinear transformation layer uses the ReLU activation function, which mathematically preserves positive values ​​and sets all negative values ​​to zero. This operation allows positive values ​​in the feature vector aggregated by the first layer to be retained and propagated, while negative values ​​are suppressed, thereby enhancing the discriminative information in the feature representation. Simultaneously, it introduces nonlinear mapping capabilities into the network, enabling subsequent propagation calculations to express more complex relational patterns. After activation, the system obtains the preliminary feature vector output from the first layer of the graph convolutional network.

[0064] The system inputs the preliminary feature vectors output from the first layer into the second layer of the graph neural network model, the graph convolutional network. In this layer, the network performs a neighborhood feature aggregation operation again. However, since each node's features now include information from its first-order neighbors, the second aggregation effectively incorporates information from second-order neighbors into the current node's feature representation. Unlike the first layer, the second layer introduces an attention mechanism during aggregation: the system calculates an attention coefficient for the connections between a node and its neighbors. This coefficient reflects the importance of the connection for information propagation in the current anomaly scenario. After softmax normalization, the attention coefficient is used as the weight for weighted aggregation, ensuring that neighbors more closely related to the current anomaly have a greater weight in the aggregation. After processing by the second layer, the feature vector of each node incorporates higher-order topological information from its second-order neighborhood.

[0065] The system feeds the high-order feature vectors output from the second layer into the output layer of the graph neural network model. For each pair of nodes with a direct connection, the output layer concatenates the final feature vectors of the two nodes and feeds them into a multilayer perceptron for mapping, outputting a probability value between 0 and 1, representing the probability that the influence relationship between nodes holds true in the final feature state of that node. The system iterates through all pairs of nodes with connections, calculates the influence probability for each pair, and fills all the calculation results into the corresponding probability matrix positions according to the node pair indices. For node pairs without a direct connection, the probability is forcibly set to zero, ultimately forming a complete probability matrix of influence between nodes.

[0066] This method refines the specific computational steps of graph neural network (GNN) inference, addressing the technical challenge of effectively propagating and calculating the impact of anomalies on the topological graph of metering devices. The first layer, neighbor feature aggregation, quantifies and transmits the impact of anomalies between directly electrically connected nodes, making local propagation relationships computable. The nonlinear activation layer addresses the insufficient expressive power of linear transformations, enhancing the model's ability to discriminate complex anomaly patterns. The second layer, higher-order neighbor feature aggregation and an attention mechanism, solves the modeling problem of long-range dependencies between indirectly connected nodes, allowing inference to extend beyond directly adjacent nodes. The attention mechanism also enables the model to dynamically focus on more critical propagation paths within the current anomaly scenario. The output layer probability transformation addresses the difficulty in uniformly quantifying the strength of relationships between nodes, transforming abstract causal relationships into probability values ​​between 0 and 1, facilitating subsequent path search and threshold selection. Overall, this method provides a complete technical solution for the implementation of graph neural networks in root cause inference in metering devices, making the inference results interpretable and computable.

[0067] In one embodiment of the present invention, S4 includes: By using root cause combinations and their corresponding physical failure paths, the built-in supplementary formula library is used to perform root cause formula mapping retrieval, applicable formula type screening, and matching degree calculation to obtain a set of candidate supplementary formulas. By dynamically assigning weight coefficients to each formula using a candidate supplementary formula set and a severity model, and by calculating the correction of the formula's effectiveness based on severity, the contribution coefficients of multiple formulas are obtained. By using the contribution coefficient of multiple formulas to perform parallel calculations and result fusion of multiple supplementary formulas, cross-validation between multiple estimates and outlier removal, a preliminary estimate of the supplementary power capacity is obtained. Monte Carlo sampling and confidence interval construction, upper and lower bound calculation and interval symmetry adjustment are performed on the estimation error distribution of each formula by preliminary supplementary power estimates to obtain the final supplementary power results.

[0068] Root cause to formula mapping retrieval. The system has a built-in supplementary formula library containing various power supplementary calculation formulas for different fault types. Now, based on the diagnosed root cause type, the system searches the formula library for the formulas with the highest historical matching degree, and selects a set of candidate formulas applicable to the current situation.

[0069] Different root causes have varying degrees of severity, and their impact on the follow-up results also differs. This step assigns different weight coefficients to each candidate formula based on a severity model; the higher the severity, the greater the weight of the corresponding formula. Finally, the contribution coefficient of each formula is output.

[0070] Parallel calculation and fusion of multiple formulas. The weighted formulas from the previous step are simultaneously substituted into the actual operating data for calculation, and each formula will calculate a supplementary power consumption value. These results are then combined, outliers that significantly deviate from the target population are removed, and the average or median is taken to obtain a preliminary estimated supplementary power consumption value after fusion.

[0071] Confidence interval construction. Each formula has its own historical error distribution. Based on these distributions, Monte Carlo random simulations are performed to generate a large number of possible replenishment power values, thereby constructing the probability distribution of the final estimate. Finally, the upper and lower bounds at a 95% confidence level are calculated, and after adjusting the interval symmetry, the final replenishment power result with confidence intervals is output.

[0072] The working principle and technical effect of the above technical solution are as follows: Based on the root cause combination and its corresponding physical failure path, a search is performed in the built-in supplementary formula library. The supplementary formula library is pre-built and stores various electrical compensation calculation formulas for different root cause types. For example, for harmonic distortion problems, there may be multiple correction formulas based on different harmonic orders and distortion rates; for voltage loss problems, there may be different estimation formulas based on the voltage loss time ratio and load curve; for loose wiring problems, there may be different compensation formulas based on contact resistance measurements. Each formula has its applicable root cause type label and applicable condition description. The system calculates the matching degree between the current root cause combination and the applicable conditions of each formula in the formula library, filtering out all formulas with a matching degree higher than a preset threshold to form a candidate supplementary formula set.

[0073] The system assigns weights to each formula in the candidate set. The weights are determined by two factors: first, the formula's historical reliability score, which reflects its computational bias in historical applications targeting similar root causes; smaller bias indicates higher reliability. Second, the severity of the currently diagnosed root cause; higher severity corresponds to a larger weight factor, as follow-up formulas for severe anomalies should have higher decision weights. The system multiplies the historical reliability score by the severity weight factor to obtain the initial weight for each formula. Then, it normalizes the weights of all candidate formulas to a sum of 1. Finally, it verifies whether each formula fully meets the applicability conditions in the current root cause scenario, reducing the weight of formulas that do not fully meet the conditions. The system then outputs the contribution coefficient of each formula.

[0074] The system simultaneously substitutes various candidate formulas into actual operating data for parallel computation. Each formula independently calculates an estimated supplementary power consumption based on its own mathematical model and input electrical parameters. Since each formula is based on different physical assumptions and correction logic, the calculation results will exhibit some dispersion. The system aggregates all calculation results and performs cross-validation, identifying and eliminating outliers that deviate significantly from other results. The remaining results are then weighted and averaged or fused using the median to obtain a preliminary, fused estimate of the supplementary power consumption. This fusion process effectively suppresses the extreme biases that may exist with a single formula.

[0075] The system constructs confidence intervals for the initial estimated replenishment power. It pre-calculates the estimation error distribution characteristics (such as the mean and standard deviation of the error) of each candidate formula in similar application scenarios using historical data. Based on these statistical characteristics, the system employs a Monte Carlo simulation method: extensive random sampling is performed on the error distribution, with each sampling yielding a simulated error value, which is then superimposed on the initial estimate to generate a large number of possible simulated replenishment power values. These simulated values ​​constitute an empirical probability distribution for the replenishment power. The system extracts the upper and lower quantiles corresponding to the 95% confidence level from this distribution as the upper and lower bounds of the confidence interval, and outputs the final replenishment power result and its confidence interval after symmetry adjustment.

[0076] This method addresses a series of technical problems in traditional power tracking methods, such as the reliance on a single formula, the inability to reflect root cause differences, and the lack of reliable expression in the results, by limiting the specific implementation steps of S4. Root cause formula mapping retrieval solves the problem of difficulty in automatically matching different root cause types with different tracking formulas, making formula selection more targeted; the dual weighting of historical reliability scoring combined with severity grading solves the problem of varying applicability of different formulas in different scenarios, making weight allocation more scientific and reasonable; parallel computation of multiple formulas and cross-validation fusion solve the problem of unreliable results from a single formula, improving the robustness of the estimates through mutual verification of multiple formulas; outlier removal solves the problem of extreme bias caused by mismatched assumptions in some formulas, avoiding misleading the final results; and Monte Carlo simulation confidence interval construction solves the problem of the lack of error range expression in tracking results, giving the output results a probabilistically reliable expression.

[0077] In one embodiment of the present invention, the step of dynamically assigning weight coefficients to each formula through a set of candidate supplementary formulas and a severity model, and calculating the correction of the formula effectiveness based on severity, to obtain the contribution coefficients of multiple formulas, includes: By querying the historical application effect of each candidate formula in the candidate supplementary formula set and severity model, the frequency of use and calculation result deviation of each formula in similar historical problems are statistically analyzed to obtain the historical reliability score of each formula. The basic weights of each formula are calculated using historical reliability scores and current severity levels. The historical reliability scores are multiplied by the weight factors corresponding to the severity levels to obtain the initial weight values ​​for each formula. The weights of all candidate formulas are normalized by using the initial weight values ​​so that the sum of all formula weights equals 1. Then, the applicability of each formula to the current root cause scenario is reviewed and the weights of formulas that are not fully applicable are reduced to obtain the weights of each formula. By eliminating formulas with excessively low weights based on their respective weights, the remaining formula weights are then normalized and redistributed, and a final balancing process is performed to obtain the contribution coefficients of multiple formulas.

[0078] The working principle and technical effects of the above solution are as follows: The system queries the historical application effects of each formula in the candidate follow-up formula set. A formula application history database is maintained, recording the usage history of each formula in each actual application, including the root cause type, severity, actual follow-up results, and deviations in post-application verification. The system statistically analyzes the usage frequency and average calculation deviation rate of each formula in similar historical problems, and calculates the standard deviation of the deviation rate to assess its stability. Based on these statistics, the system calculates a historical reliability score for each formula, which comprehensively reflects the accuracy and stability of the formula in similar scenarios.

[0079] The system acquires the severity level (mild, moderate, severe) of the currently diagnosed root cause and converts the level into corresponding weight factors according to a preset level weight mapping table. Different root cause types may correspond to different level weight mapping rules; for example, the impact of some root causes on follow-up results increases linearly with increasing severity, while others may increase exponentially. The system multiplies the historical reliability score of each formula with the severity weight factor corresponding to that formula to obtain the initial weight value of each formula. This value reflects both the inherent reliability of the formula and the importance that the formula should obtain at the current severity level.

[0080] The system normalizes the initial weights of all candidate formulas by dividing the initial weight of each formula by the sum of the initial weights of all formulas, ensuring that the total weight of all formulas is exactly 1. After normalization, the system performs a secondary verification of the physical applicability conditions for each formula. This verification includes, but is not limited to: whether the formula's input parameters are fully available in the current operating data; whether the formula's assumptions are consistent with the physical characteristics of the current root cause; and whether the formula's applicable scope covers the current load level. For formulas found not to fully meet the applicability conditions during the verification, the system applies a reduction factor between 0 and 0.5 to their weights, with the reduction amount depending on the severity of the non-compliance.

[0081] The system checks the weights of each formula after review and correction, removing formulas with weights below a preset lower threshold (these formulas are considered to have too low a contribution value in the current scenario). The weights of the remaining formulas are then renormalized, and the original weights of the removed formulas are proportionally redistributed to the remaining valid formulas, ensuring that the sum of the weights of the valid formulas is again exactly equal to 1. After this round of simplification and rebalancing, the system outputs the final multi-formula contribution coefficients.

[0082] This method addresses the technical problem of lacking objective basis and dynamic adaptability in weight determination during multi-formula weighted fusion by refining the specific implementation steps of dynamic weight assignment. Historical application effect query and reliability scoring resolve the difficulty in quantifying the performance differences of various formulas in different scenarios, ensuring that weight determination is based on historical data rather than subjective experience. Severity grading and weight factor mapping calculation address the need for the same formula to have different decision weights at different severity levels, allowing weight allocation to be dynamically adjusted according to actual conditions. Normalization processing resolves the issue of inconsistent weight scales among formulas, making the weights comparable and interpretable. Secondary verification and reduction correction of physical applicability conditions address the problem that some formulas, while highly reliable in historical data, may not fully match the applicability conditions in the current scenario, preventing the incorrect application of formulas to inapplicable scenarios. Removal of low-weight formulas and re-normalization allocation resolve the issue that retaining too many low-contribution formulas may introduce noise interference, making the final set of formulas more concise and effective. Overall, this method improves the scientific rigor and dynamic adaptability of multi-formula weighted fusion, making the weight allocation of each formula more reasonable and accurate.

[0083] In one embodiment of the present invention, the system includes: The acquisition and fusion module is used to synchronously acquire and spatiotemporally align multi-dimensional information, extract features and vectorize them through a multi-source feature fusion layer to obtain multi-dimensional quantitative feature vectors. The anomaly set module is used to perform rule and statistical discrimination on the first-layer anomaly detection model through multi-dimensional quantitative feature vectors, and to quickly locate and label anomaly dimensions to obtain an anomaly dimension set containing anomaly type and anomaly degree. The root cause combination module is used to perform node mapping and relation reasoning, causal association and root cause reasoning among multiple anomalies, and obtain the most likely root cause combination by using the set of anomaly dimensions to perform node mapping and relation reasoning on the topological knowledge graph of the second-layer graph neural network; The selection and quantization module is used to perform dynamic formula selection and weighted combination calculation, multi-formula collaborative estimation and confidence interval quantization on the adaptive power replenishment algorithm through root cause combination and its severity model, so as to obtain the final power replenishment result.

[0084] The working principle and technical effects of the above technical solution are as follows: This system uses a multi-source feature fusion layer to uniformly collect and integrate heterogeneous information scattered across different data sources. Real-time waveform data records instantaneous changes in voltage, current, and harmonics using high-frequency sampling. Internal event logs of the electricity meter store operation logs such as voltage loss, cover opening, and programming using discrete triggering. Environmental parameters collect temperature, humidity, and load fluctuation information using a low-frequency periodic method. Historical verification data is stored in archive format, preserving past verification records. These data sources differ significantly in time base, sampling frequency, and data format. The fusion layer maps the data from each source to a unified time grid through spatiotemporal alignment, then transforms the original signals into numerical features through feature extraction, and finally concatenates the multi-dimensional features into a fixed-length feature vector through vectorized fusion.

[0085] This system addresses a series of technical problems inherent in traditional methods, such as isolated data analysis, disconnect between anomaly detection and root cause diagnosis, and reliance on single empirical coefficients for retrospective calculations. Multi-source feature fusion effectively resolves the inconsistencies in time references and sampling frequencies of heterogeneous data, ensuring subsequent analysis is based on unified and complete data. A dual-channel statistical and rule-based discrimination mechanism resolves the contradiction of high false negative or false positive rates associated with single discrimination methods, enabling rapid location and credibility assessment of anomalies. Graph neural network knowledge graph reasoning overcomes the technical challenge of automatically identifying causal relationships among multiple anomalies, achieving interpretable reasoning from surface anomalies to deep-seated root causes. Adaptive multi-formula weighted retrospective calculations overcome the limitations of traditional single correction coefficient methods in adapting to complex root cause scenarios, making the retrospective results more scientific and credible.

[0086] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent diagnosis and traceability of electricity metering anti-theft devices based on multi-dimensional feature fusion, characterized in that, The method includes: S1. Multi-dimensional quantitative feature vectors are obtained by synchronously collecting and spatiotemporally aligning multi-dimensional information, extracting features and fusion vectorization through a multi-source feature fusion layer. S2. By using multi-dimensional quantitative feature vectors, rule and statistical discrimination are performed on the first-layer anomaly detection model, and the anomaly dimensions are quickly located and labeled to obtain the set of anomaly dimensions; S3. By using the set of anomalies, perform node mapping and relational reasoning on the topological knowledge graph of the second-layer graph neural network, as well as causal relationships and root cause reasoning among multiple anomalies, to obtain the root cause combination. S4. The adaptive power replenishment algorithm is dynamically selected and weighted by combining root causes and its severity model, and then subjected to multi-formula collaborative estimation and confidence interval quantification to obtain the final power replenishment result.

2. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion as described in claim 1, characterized in that, S1 includes: The multi-source raw dataset is obtained by parallel and synchronously acquiring real-time waveform data, internal event records of the electricity meter, environmental parameters and historical verification data through a multi-source data acquisition interface, and by normalizing and interpolating the timestamps of each source data. By performing noise filtering, missing value imputation, outlier removal, data quality cleaning and standardization transformation on various types of data from multiple source raw datasets, a standardized single-dimensional feature sequence is obtained. By performing principal component analysis and correlation reduction, feature redundancy elimination and key information retention on each dimension of the feature sequence through standardized single-dimensional feature sequence, a core feature subset is obtained. By performing vector concatenation and normalization mapping on multi-dimensional features through a subset of core features, as well as constructing unified feature vectors and aligning dimensions, multi-dimensional quantitative feature vectors are obtained.

3. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion as described in claim 2, characterized in that, The process involves parallel and synchronous acquisition of real-time waveform data, internal event records of the electricity meter, environmental parameters, and historical verification data through a multi-source data acquisition interface. Timestamps of each source data are normalized and interpolated to obtain a multi-source raw dataset, including: The system uses a multi-source data acquisition interface to collect real-time waveform data, internal event records of the electricity meter, environmental parameters and historical verification data in parallel and synchronously. It performs unified format parsing and precision conversion on the time stamps of each data source to obtain the data from each source and its original timestamp. The consistency of the time base of each data source is checked by using the data from each source and its original timestamp. The internal clock of the electricity meter is used as the main reference for linear compensation and drift correction of the clock deviation of each source, and the timestamp sequence of each source is obtained. By determining a unified target time grid for all data sources through the timestamp sequences of each source, and resampling and locating the data from each source, high sampling rate data is extracted and downsampled, low sampling rate data is encrypted using cubic spline interpolation, and discrete event-type data is filled with forward nearest neighbor to obtain a multi-source synchronized dataset. By using a multi-source synchronized dataset, boundary effect checks and alignment quality assessments are performed on data of each dimension. Boundary points with interpolation errors exceeding the tolerance threshold at both ends are removed, and the original multi-source dataset is obtained.

4. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion as described in claim 1, characterized in that, S2 includes: By performing real-time sliding window comparison of the statistical distribution baseline of each feature using multi-dimensional econometric feature vectors, and calculating the deviation of mean and variance anomalies, a statistical anomaly score sequence is obtained. By statistically analyzing the abnormal scoring sequence, multiple preset rule thresholds are triggered item by item, and rule matching and over-limit marking are performed to obtain a set of rule triggering flags. By using a set of rule-triggered flags and a statistical anomaly scoring sequence, the anomaly dimensions are comprehensively sorted by confidence, false alarms are eliminated, and anomalies are prioritized to obtain a list of high-confidence anomaly dimensions. By using a list of high-confidence anomaly dimensions, each anomaly dimension is coded for type and graded for severity, and its attributes are labeled and encapsulated in a structured manner to obtain a set of anomaly dimensions.

5. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion according to claim 4, characterized in that, The step involves statistically analyzing anomaly scoring sequences to determine each preset rule threshold, performing rule matching and exceeding-limit marking, to obtain a set of rule trigger flags, including: By statistically analyzing the abnormal scoring sequence, the numerical values ​​of multiple preset rule thresholds are compared one by one. The real-time scores or real-time measurement values ​​of each dimension are compared with the corresponding rule limits and the limits are recorded to obtain the preliminary trigger judgment results. The initial trigger judgment results are used to continuously verify each trigger event. Only when the same rule is triggered at multiple consecutive sampling points is it confirmed as a valid trigger and a valid trigger flag is obtained. By validally triggering flags, mutual exclusion and association logic checks are performed on each rule item. Rules with mutual exclusion relationships are selected according to priority rules, and rules with association relationships are merged and marked to obtain a rule trigger set. By encoding and structurally encapsulating each trigger item through a rule trigger set, the rule number, rule name, trigger time, and corresponding threshold information of each trigger are organized into a unified flag format to obtain a rule trigger flag set.

6. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion as described in claim 1, characterized in that, S3 includes: By using the set of abnormal dimensions, abnormal nodes are located, activated, and initial features are assigned to the pre-constructed topological knowledge graph of the metering device, resulting in the set of abnormal nodes and its adjacent subgraph. By performing multi-layer graph convolution propagation, calculating the strength of relationships between nodes, and assigning attention weights to the graph neural network model using the set of abnormal nodes and their adjacent subgraphs, the influence probability matrix between each node is obtained. By using the influence probability matrix, path search and causal chain backtracking are performed on multiple anomaly nodes, and root cause candidate nodes are sorted and combined enumerated to obtain a root cause candidate combination list and its confidence score. By using the root cause candidate combination list and its confidence score, the combination with the highest score is matched for physical failure mode, the root cause name is standardized and the failure path description is generated, thus obtaining the root cause combination and its corresponding physical failure path.

7. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion according to claim 6, characterized in that, The process involves performing multi-layer graph convolutional propagation, calculating the strength of relationships between nodes, and assigning attention weights to the graph neural network model using a set of abnormal nodes and their adjacent subgraphs to obtain the influence probability matrix between each node. This includes: The first layer of the graph neural network model is summarized by aggregating the neighbor features of the abnormal node set and its adjacent subgraph. The features of each node are weighted and averaged with the features of its directly connected nodes to obtain the intermediate features after the first layer is summarized. The nonlinear transformation layer of the graph neural network model is activated by intermediate features, and the negative values ​​in the intermediate features are set to zero to enhance the discriminative power of the features, thus obtaining a preliminary feature vector. The second layer of the graph neural network model is summarized by the preliminary feature vectors. The features of each node and its indirectly connected nodes are weighted and merged according to the strength of the association to obtain the high-order features after the second layer summary. By performing probability transformation on the output layer of the graph neural network model using higher-order features, the influence degree between each pair of connected nodes is converted into a probability value between 0 and 1, thus obtaining the influence probability matrix between each node.

8. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion according to claim 1, characterized in that, S4 includes: By using root cause combinations and their corresponding physical failure paths, the built-in supplementary formula library is used to perform root cause formula mapping retrieval, applicable formula type screening, and matching degree calculation to obtain a set of candidate supplementary formulas. By dynamically assigning weight coefficients to each formula using a candidate supplementary formula set and a severity model, and by calculating the correction of the formula's effectiveness based on severity, the contribution coefficients of multiple formulas are obtained. By using the contribution coefficient of multiple formulas to perform parallel calculations and result fusion of multiple supplementary formulas, cross-validation between multiple estimates and outlier removal, a preliminary estimate of the supplementary power capacity is obtained. Monte Carlo sampling and confidence interval construction, upper and lower bound calculation and interval symmetry adjustment are performed on the estimation error distribution of each formula by preliminary supplementary power estimates to obtain the final supplementary power results.

9. The intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion according to claim 8, characterized in that, The process involves dynamically assigning weight coefficients to each formula using a candidate supplementary formula set and a severity model, and calculating the severity's impact on formula effectiveness to obtain multi-formula contribution coefficients, including: By querying the historical application effect of each candidate formula in the candidate supplementary formula set and severity model, the frequency of use and calculation result deviation of each formula in similar historical problems are statistically analyzed to obtain the historical reliability score of each formula. The basic weights of each formula are calculated using historical reliability scores and current severity levels. The initial weight values ​​of each formula are obtained by multiplying the historical reliability scores by the weight factors corresponding to the severity levels. The weights of all candidate formulas are normalized by using initial weight values ​​so that the sum of all formula weights equals 1. Then, the applicability of each formula to the current root cause scenario is reviewed and the weights of formulas that are not fully applicable are reduced to obtain the weights of each formula. By eliminating formulas with excessively low weights based on their respective weights, the remaining formula weights are then normalized and redistributed, and a final balancing process is performed to obtain the contribution coefficients of multiple formulas.

10. A system for implementing the intelligent diagnosis and traceability method for electricity metering anti-theft devices based on multi-dimensional feature fusion as described in claim 1, characterized in that, The system includes: The acquisition and fusion module is used to synchronously acquire and spatiotemporally align multi-dimensional information, extract features and vectorize them through a multi-source feature fusion layer to obtain multi-dimensional quantitative feature vectors. The anomaly set module is used to perform rule and statistical discrimination on the first-layer anomaly detection model through multi-dimensional quantitative feature vectors, quickly locate and label anomaly dimensions, and obtain an anomaly dimension set. The root cause combination module is used to perform node mapping and relation reasoning, causal association and root cause reasoning among multiple anomalies, and obtain root cause combinations by using the set of anomaly dimensions to perform node mapping and relation reasoning on the topological knowledge graph of the second-layer graph neural network; The selection and quantization module is used to perform dynamic formula selection and weighted combination calculation, multi-formula collaborative estimation and confidence interval quantization on the adaptive power replenishment algorithm through root cause combination and its severity model, so as to obtain the final power replenishment result.

Images