Multi-dimensional electrical characteristic fusion screening method of electric energy meter

Abstract

The invention discloses a multi-dimensional electrical characteristic fusion screening method for an electric energy meter, and relates to the technical field of intelligent monitoring of a power system, and the method comprises the steps: carrying out the preprocessing of multi-dimensional electrical characteristic time sequence data of the electric energy meter in the operation process, and obtaining the preprocessed time sequence data; inputting the preprocessed time series data into a bimodal association model, and outputting association rules between features and abnormal states; respectively calculating scores of each feature on the four evaluation indexes according to association rules; calculating a real-time anomaly occurrence rate according to the anomaly labels, and dynamically adjusting the weights of the four evaluation indexes according to the real-time anomaly occurrence rate; calculating a comprehensive score of each feature according to the score of each feature on the four evaluation indexes and the weights of the four evaluation indexes; the dynamic threshold is determined according to the maximum value and the average value of the comprehensive scores of all the features, the features with the comprehensive scores larger than the dynamic threshold are screened out, the key feature set is obtained, and the accuracy of feature screening is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of power systems, and in particular to a multi-dimensional electrical feature fusion screening method for electric energy meters. BACKGROUND

[0002] With the rapid development of smart grids, the stability of the operating state and the accuracy of the metering data of electric energy meters, as core equipment for power metering and monitoring, are directly related to the safe operation of power systems and the fair trade of power markets. During long-term operation, electric energy meters are susceptible to environmental interference, equipment aging, communication fluctuations and other factors, leading to frequent faults such as metering abnormalities and communication abnormalities, which pose challenges to the fine management and reliable power supply of power grids.

[0003] Accurate diagnosis and abnormal detection of the operating state of electric energy meters are the core means to address the above challenges, and the key lies in screening electrical features that are strongly associated with various abnormal states from massive operating data to build a high-precision electric energy meter abnormal state diagnosis model. Traditional electric energy meter feature screening methods often only focus on the single correlation between features and abnormalities, while ignoring the stability of features in long-term operation, the redundancy between features, and the timeliness of features for abnormal early warning, and using fixed thresholds for feature screening, which cannot adapt to the changes in feature score distribution in different data sets or different periods, resulting in poor generalization ability of the finally screened features and difficulty in adapting to complex and diverse actual environments in power grids. SUMMARY

[0004] The technical problem to be solved by the present application is the single feature screening dimension and rigid screening mechanism in the prior art. The present application provides a multi-dimensional electrical feature fusion screening method for electric energy meters, which solves the above problems.

[0005] The present application is implemented by the following technical solutions:

[0006] In a first aspect, the present application provides a multi-dimensional electrical feature fusion screening method for electric energy meters, comprising:

[0007] Pretreating multi-dimensional electrical feature time series data of electric energy meters during operation to obtain pretreated time series data;

[0008] Inputting the pretreated time series data into a dual-modal correlation model to output correlation rules between features and abnormal states; the correlation rules include rule time series, strongly correlated node pairs, primary feature rules, secondary feature rules and abnormal labels;

[0009] According to the correlation rules, the score of each feature on four evaluation indexes is calculated; the four evaluation indexes include correlation, stability, redundancy and timeliness;

[0010] According to the abnormal label, a real-time abnormality occurrence rate is calculated, and weights of the four evaluation indexes are dynamically adjusted according to the real-time abnormality occurrence rate;

[0011] According to the score of each feature on the four evaluation indexes and the weights of the four evaluation indexes, a comprehensive score of each feature is calculated;

[0012] A dynamic threshold is determined according to the maximum value and the average value of the comprehensive scores of all features, features with a comprehensive score greater than the dynamic threshold are screened out, and a key feature set is obtained.

[0013] Optionally, the multi-dimensional electrical feature time series data includes metering abnormality type data and / or communication abnormality type data; the pre-processing of the multi-dimensional electrical feature time series data of the electric energy meter during the running process to obtain the pre-processed time series data includes:

[0014] The dynamic time warping algorithm is used for data correction on the metering abnormality type data, and the missing value filling method based on the attention mechanism is used for data filling on the communication abnormality type data, to obtain the repaired time series data;

[0015] The sliding window mechanism is introduced to update the statistical parameters in real time, and the repaired time series data is normalized to obtain the pre-processed time series data.

[0016] Optionally, the input of the pre-processed time series data into the dual-modal correlation model to output the correlation rule between the features and the abnormal state includes:

[0017] The pre-processed time series data is input into the XGBoost model to output the contribution weight of each feature;

[0018] The pre-processed time series data is divided into transactions according to a preset sliding window; each transaction contains the state of each feature in the window and the abnormal label;

[0019] According to the contribution weight of each feature and the number of transactions, the weighted support of each feature is calculated;

[0020] According to the contribution weight, a support threshold is calculated;

[0021] Based on the weighted support and the support threshold, a feature correlation graph is constructed to perform strong connected component division, and the correlation rule between the features and the abnormal state is output.

[0022] Optionally, the calculation formula of the weighted support is as follows:

[0023]

[0024] Wherein, S iω is the weighted support degree of the ith feature; ω i ω is the contribution weight of the ith feature; N i N is the number of transactions containing the ith feature; N is the total number of transactions.

[0025] Optionally, the range of the support threshold is as follows:

[0026]

[0027] ω is the maximum value of the contribution weight of all features. max

[0028] Optionally, the calculation formula of the score of the four evaluation indexes is as follows:

[0029] Corr=0.6×MIC+0.4×KCC

[0030] Corr represents the score of the correlation index; MIC represents the maximum information coefficient; KCC represents the kernel correlation coefficient;

[0031] Stab=0.7×(1-CV)+0.3×TSI

[0032] Stab represents the score of the stability index; CV represents the coefficient of variation in the sliding window, CV=σ / μ, σ is the standard deviation of the data in the sliding window, μ is the mean of the data in the sliding window; TSI is the trend stability index, TSI =(1 - |k|), k is the slope obtained by linear regression on the data in the sliding window;

[0033] Red=0.5×I1 +0.5×I2

[0034] Red represents the score of the redundancy; I1 and I2 are indicator functions, I1 =1 when r≥0.8, otherwise I1 =0, r is the Pearson correlation coefficient; I2 =1 when MI≥0.6, otherwise I2 =0, MI is the mutual information entropy;

[0035] Timelinesss=Δt÷T duration

[0036] Timelinesss represents the score of the timeliness index; Δt is the time difference between the time when the feature shows abnormal signs and the time when the electric energy meter actually occurs abnormality; T duration is the duration of the anomaly.

[0037] Optionally, the calculation formula of the real-time anomaly occurrence rate is as follows:

[0038] ρ t =0.3×ρ t-1 ​+0.7 x Ns' / Ns

[0039] wherein, p t is the real-time abnormality occurrence rate of the current period, p t-1 is the real-time abnormality occurrence rate of the last period, N s ' is the sample number of abnormality labels in the current period, N s is the total sample number of the current period.

[0040] Optionally, the dynamically adjusting the weights of the four evaluation indexes according to the real-time abnormality occurrence rate comprises:

[0041] when the real-time abnormality occurrence rate is greater than or equal to a preset rate, increasing the ratio between the weight of the correlation index and the weight of the stability index;

[0042] when the real-time abnormality occurrence rate is less than the preset rate, decreasing the ratio between the weight of the correlation index and the weight of the stability index.

[0043] Optionally, the calculation formula of the dynamic threshold value is as follows:

[0044] T = alpha x S 综合max + (1-alpha) x S 综合mean

[0045] wherein, T is the dynamic threshold value; S 综合max is the maximum value of the comprehensive score of all features; S 综合mean is the average value of the comprehensive score of all features, alpha is a threshold parameter, and the value range is [0.3, 0.7]; the threshold parameter is optimized through a Q-learning algorithm.

[0046] Optionally, after determining the dynamic threshold value according to the maximum value and the average value of the comprehensive score of all features, screening out features with a comprehensive score greater than the dynamic threshold value to obtain a key feature set, the method further comprises:

[0047] generating difficult example samples by using an adversarial generation technology based on different user scenario data sets; the user scenarios include urban resident users, industrial users and remote area users;

[0048] verifying the key feature set by using the difficult example samples and standard samples;

[0049] if the key feature set passes the verification, inputting the key feature set into an abnormality state detection to obtain the state of the electric energy meter.

[0050] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0051] This application provides a multi-dimensional electrical feature fusion and screening method for electricity meters. It constructs a four-dimensional evaluation index system including correlation, stability, redundancy, and timeliness, and dynamically adjusts the weights of each evaluation index based on the real-time anomaly occurrence rate. This solves the problem of single-dimensional feature screening and achieves comprehensive and adaptive feature evaluation. Furthermore, it employs dynamic thresholds for screening, allowing the screening criteria to adaptively align with the comprehensive score distribution of current data features, improving the accuracy and robustness of feature screening. It can automatically filter out the most effective set of key features for anomaly detection from the massive multi-dimensional electrical features generated by electricity meter operation, and subsequently improve the accuracy of electricity meter anomaly detection based on this set of key features. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0053] Figure 1 A flowchart illustrating the multi-dimensional electrical feature fusion screening method for electricity meters provided in this application embodiment;

[0054] Figure 2 A schematic diagram of the structure of the multi-dimensional electrical feature fusion screening device for an electricity meter provided in this application embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0056] To address the problems of limited feature selection dimensions and rigid selection mechanisms in existing technologies, this application provides a multi-dimensional electrical feature fusion selection method for electricity meters. Please refer to... Figure 1 This is a flowchart illustrating the multi-dimensional electrical feature fusion screening method for electricity meters provided in this application embodiment. The following is a further explanation. Figure 1 The multi-dimensional electrical feature fusion screening method for electricity meters is introduced.

[0057] S101. Preprocess the multi-dimensional electrical characteristic time-series data of the electricity meter during operation to obtain preprocessed time-series data.

[0058] In practical implementation, the electricity meter's built-in sensors and edge computing module can be used to collect real-time time-series data of multi-dimensional electrical characteristics during operation. The collected multi-dimensional electrical characteristics include voltage (sampling frequency 50Hz), current (sampling frequency 50Hz), active power, reactive power, power factor, frequency, and harmonic distortion rate. The sampling interval can be set from 1 to 5 minutes according to actual needs.

[0059] In one possible embodiment, the specific steps of S101 include:

[0060] For measurement anomaly data, a dynamic time warping algorithm is used to correct the data, and for communication anomaly data, a missing value imputation method based on an attention mechanism is used to impute the missing value, thus obtaining repaired time series data. By introducing a sliding window mechanism to update statistical parameters in real time, the repaired time series data is normalized to obtain preprocessed time series data.

[0061] In practical implementation, multidimensional electrical characteristic time-series data may include metering anomaly data and / or communication anomaly data. Existing technologies mostly adopt a uniform processing method, failing to consider the characteristic differences of different anomaly types, resulting in poor correction effects for non-stationary data. In the embodiments of this application, different preprocessing methods are adopted for different anomaly types of data, which are described below.

[0062] The first scenario involves measurement anomalies.

[0063] During the metering process, sudden load switching (such as the start-up of a high-power motor) or the connection of intermittent distributed power sources can cause non-stationary fluctuations or distortions in metering characteristic data such as electricity consumption, voltage, current, and power. This non-stationarity can mask true anomaly patterns and interfere with subsequent correlation analysis and feature evaluation. Therefore, a dynamic time warping algorithm is used to correct anomaly-type metering data. The specific steps are as follows:

[0064] A segment of similar characteristic data under normal operating conditions and stable load periods is selected from historical data as a reference sequence. Non-stationary fluctuating data sequences are extracted from the current real-time data stream. The distance matrix between the reference sequence and the data sequence is calculated. The optimal matching path between the two sequences is found in the distance matrix. Based on the optimal matching path, the non-stationary fluctuating data sequences are time-axis aligned and corrected.

[0065] In this embodiment, for abnormal measurement data, a dynamic time warping algorithm is used to directly correct the root cause of its "non-stationary fluctuations." It is not a simple smoothing or filtering process, but rather it seeks the optimal alignment path on the time axis, thereby restoring the true evolution of the data and effectively eliminating instantaneous distortions caused by load shocks, providing a high-fidelity data foundation for subsequent analysis.

[0066] The second scenario involves data related to communication anomalies.

[0067] Communication anomalies (such as signal interference or connection timeouts) can cause random or continuous data loss in the electricity meter's data stream. When the duration of the missing data is ≤5 minutes, an attention-based missing value imputation method is used to impute the communication anomaly data. The specific steps are as follows:

[0068] For missing segments whose duration does not exceed the first preset duration (e.g., 15 minutes), feature sequences from several (e.g., 3) cycles before and after the missing segment are extracted as reference sequences. The reference sequences are encoded using a Transformer encoder, and the attention weight of each time step in the reference sequence is calculated. The higher the attention weight, the stronger the correlation between the data at that time step and the missing segment. The time step data of the reference sequence are weighted and fused according to the attention weight to generate the imputation value for the missing value.

[0069] In this embodiment, for communication anomaly data, an attention-based imputation method intelligently utilizes the contextual information before and after the missing point. The attention mechanism can automatically learn and assign higher weights to historical periodic data most relevant to the missing segment, thereby achieving "context-aware" and accurate imputation, ensuring the accuracy of the imputed data.

[0070] Furthermore, by introducing a sliding window mechanism to update statistical parameters in real time, the repaired time-series data is normalized. The normalization formula is as follows:

[0071]

[0072] Where x is the original data, μ t Let σ be the mean of the data within the sliding window at time t. t Let x' be the standard deviation of the data within the sliding window at time t, and x' be the normalized standard value.

[0073] In this embodiment, the original time-series data is mapped to a distribution range with a mean of 0 to 1, thereby achieving time-adaptive normalization of the data. This eliminates the influence of differences in data distribution at different time scales. Through time-adaptive normalization, the definition of feature states is ensured to be consistent across different time periods, eliminating time-series distribution deviations. Furthermore, time-adaptive normalization smooths only local noise through window settings, preserving real abnormal fluctuations and providing more reliable input data for the bimodal association model.

[0074] In one possible embodiment, the multidimensional electrical characteristics include stationary characteristics and non-stationary characteristics. The stationary characteristics include voltage and frequency, and the non-stationary characteristics include current, active power and reactive power. The window length of the stationary characteristics is greater than the window length of the non-stationary characteristics.

[0075] In this embodiment, a 2-hour window is used for characteristics with high stability, such as voltage and frequency, to reduce high-frequency noise interference. For characteristics with drastic load variations, such as current and power, a 1-hour window is used to improve the response speed to short-term fluctuations. The sliding step size is consistent with the sampling interval to ensure that the latest local statistics are obtained at each time point. A fixed-length rectangular window is used to avoid parameter bias introduced by weighted windows, ensuring computational simplicity and result interpretability.

[0076] S102. Input the preprocessed time series data into the dual-modal association model and output the association rules between features and abnormal states.

[0077] In one possible embodiment, the specific steps of S102 are as follows:

[0078] S1.1 Input the preprocessed time series data into the XGBoost model and output the contribution weight of each feature.

[0079] In the specific implementation process, the preprocessed time-series data is input into the XGBoost model, which outputs the gain importance of each feature. The gain importance is then standardized to the [0,1] interval to obtain the contribution weight of each feature. The training process of the XGBoost model is as follows:

[0080] First, historical multi-dimensional electrical characteristic time-series data of the electricity meters during operation are collected, and the true status labels of the electricity meters are marked, where 0 represents normal and 1 represents abnormal (such as metering abnormality, communication abnormality, etc.). Second, the historical multi-dimensional electrical characteristic time-series data is preprocessed to obtain preprocessed historical time-series data. Then, using the preprocessed historical time-series data as input features (such as normalized values ​​of voltage, current, and power), and the electricity meter status labels as target variables, a sample set is constructed, divided into training and validation sets in a 7:3 ratio. Next, an XGBoost model is constructed using a tree-structured ensemble model, with key parameters set as follows: tree depth = 5 (to avoid overfitting), learning rate = 0.1, number of iterations = 100, and subsampling rate = 0.8. Finally, the XGBoost model is trained based on the training set, learning the association between features and abnormalities by minimizing the error between the predicted and true status labels of the electricity meters.

[0081] S1.2 Divide the preprocessed time series data into transactions according to a preset sliding window.

[0082] In the specific implementation process, the preprocessed time series data is aligned and divided into transactions according to a preset sliding window to ensure that the feature status contained in each transaction matches the local data distribution of that time period. Each transaction contains the status and abnormal labels of each feature (such as Boolean values ​​such as "current > threshold" and "power fluctuation > 5%) within the window.

[0083] S1.3 Calculate the weighted support of each feature based on its contribution weight and the number of transactions.

[0084] In practice, the number of transactions containing each feature (i.e., transaction count) is calculated. This count is then multiplied by the contribution weight of the corresponding feature, summed, and divided by the total number of transactions to obtain the weighted support. The core function of weighted support is to quantify the co-occurrence strength of features and anomalies within the same time window. The specific formula for weighted support is as follows:

[0085]

[0086] Among them, S i ω is the weighted support of the i-th feature; i The contribution weight of the i-th feature; N i is the transaction count, i.e., the number of transactions containing the i-th feature; N is the total number of transactions.

[0087] S1.4 Calculate the support threshold based on the contribution weight.

[0088] In the specific implementation process, the support threshold is calculated based on the contribution weight, and the formula for calculating its value range is as follows:

[0089] 0.05×(1+ω max )~0.1×(1+ω max )

[0090] Where, ω max This represents the maximum value of the contribution weights of all features.

[0091] In this embodiment, the support threshold can be adaptively adjusted according to the importance of the feature. The support threshold range corresponding to important features is higher, which avoids misjudgment by weak association rules. Because the weighted support of key features is high, even if the frequency of occurrence is low, the support threshold can still be met, thus avoiding accidental deletion during calculation.

[0092] S1.5 Based on weighted support and support threshold, construct a feature association graph to divide strongly connected components and output the association rules between features and abnormal states.

[0093] In the specific implementation process, the relationship between features and anomalies is first visualized by establishing a feature association graph, and then transformed into computable embedded features to achieve efficient search of frequent itemsets. The graph nodes and edges are defined as follows:

[0094] Each node contains all electrical features (such as voltage and current) and anomaly types (such as metering anomalies, communication anomalies, etc.), totaling m+n nodes, where m is the number of features and n is the number of anomaly types. If the weighted support of a node is lower than the support threshold, it indicates that the feature has a weak overall correlation with the anomaly, and the node and all its connecting edges are removed from the graph.

[0095] The formula for calculating the edge weight between nodes is as follows:

[0096]

[0097] Where P(B|A|) is the probability of feature B appearing when feature A appears, and P(B) is the global probability of feature B appearing; LIFT greater than 1 indicates that feature A and feature B are positively correlated, the larger the value of LIFT, the stronger the correlation between feature A and feature B, and LIFT less than or equal to 1 indicates that feature A and feature B are not correlated or are negatively correlated.

[0098] Furthermore, embedding extraction is performed using a two-layer GCN. Initial feature vectors and adjacency matrices (i.e., edge weights LIFT) are input first. The first-layer GCN aggregates neighbor node information through convolution, and the second-layer GCN further aggregates global information, outputting a 32-dimensional node embedding vector. This embedding vector contains the relationship features between the node and other nodes. The cosine similarity between any two node embedding vectors is calculated. Nodes with a similarity greater than a preset similarity (e.g., 0.8) are grouped into the same strongly connected component. Dynamic confidence is calculated from the strongly connected components, and association rules are output based on a confidence threshold.

[0099] The formula for calculating the confidence threshold is as follows:

[0100] P(B|A|)≥0.7+0.3×ω B

[0101] Where, ω B This represents the contribution weight of feature B; 0.7 and 0.3 are pre-set base thresholds. These base thresholds ensure that the association rules for the features reach a certain level of confidence, preventing meaningless weak rules from being included in the calculation.

[0102] Association rules include rule sequence, strongly correlated node pairs, primary feature rules, secondary feature rules, and anomaly labels. Rule sequence refers to the pattern or regularity of the association between features and abnormal states over time. Strongly correlated node pairs refer to the strong correlation between two features. Primary feature rules are direct association rules between a single basic electrical feature and associated features, reflecting the causality or strong co-occurrence of two electrical features in time sequence. They do not directly correlate with the final anomaly type of the energy meter, but only reflect the linkage relationship at the feature level. Secondary feature rules are mapping rules between combinations of multiple associated features and specific anomaly types of the energy meter. They reflect the causal relationship between feature combinations and the final anomaly type, directly relating to the fault diagnosis results of the energy meter. This is the core output of bimodal association, for example: {current anomaly, power anomaly} → metering anomaly. Anomaly labels are used to indicate whether the energy meter is in an abnormal state.

[0103] In this embodiment, before mining association rules, the powerful supervised learning model XGBoost is first used to learn from the global data, obtaining the contribution weight of each feature to predicting the anomalous target. Continuous time-series data is transformed into a discrete set of transactions, and the contribution weights obtained from XGBoost are introduced into the support calculation of traditional association rules, allowing high-importance features to occupy a larger proportion in rule mining. The support threshold is dynamically adjusted based on the maximum weighted support of all features, allowing important features to stand out even if they appear infrequently, and suppressing frequent noise even if it appears frequently. Finally, by constructing a feature association graph, multi-level association rules including rule time series and strongly associated node pairs are output, providing rich input information for subsequent evaluation.

[0104] S103. Calculate the score of each feature on the four evaluation indicators according to the association rules.

[0105] In the specific implementation process, a four-dimensional evaluation index system including relevance, stability, redundancy, and timeliness is constructed to comprehensively evaluate the electrical characteristics. The four evaluation indexes include relevance, stability, redundancy, and timeliness. The following describes how to calculate the scores of each evaluation index.

[0106] (1) Correlation indicators:

[0107] Corr = 0.6 × MIC + 0.4 × KCC

[0108] Where Corr represents the score of the correlation index; MIC represents the maximum information coefficient; and KCC represents the kernel correlation coefficient.

[0109] MIC is used to measure linear and nonlinear correlations between features and outliers; KCC is a nonlinear correlation calculated by mapping features and outlier labels to a high-dimensional space using a Gaussian kernel function; the larger the absolute value of KCC, the stronger the nonlinear correlation.

[0110] In this embodiment of the application, the maximum information coefficient and the kernel correlation coefficient are calculated based on the preprocessed time series data (i.e., the time series normalized feature values) and the secondary feature rules output by the bimodal association model. By fusing the maximum information coefficient and the kernel correlation coefficient, the correlation between features and abnormal states can be reflected more comprehensively.

[0111] (2) Stability indicators:

[0112] Stab = 0.7 × (1 - CV) + 0.3 × TSI

[0113] Where Stab represents the stability index score; CV is the coefficient of variation within the sliding window, CV=σ / μ, where σ is the standard deviation of the data within the sliding window and μ is the mean of the data within the sliding window; TSI is the trend stability index, TSI =(1 - |k|), where k is the slope obtained by linear regression on the data within the sliding window.

[0114] In this embodiment, because stable features (such as voltage) exhibit small fluctuations and a gentle trend during normal operation, while unstable features (such as harmonic distortion rate) are easily disturbed and unfavorable for anomaly identification, a dual-dimensional evaluation (CV and TSI) is adopted. A smaller CV indicates more stable short-term fluctuations, and a TSI closer to 1 indicates a more stable trend. When CV ≤ 0.1 and TSI ≥ 0.8, the feature is considered to have high stability because time-series normalization eliminates time-series distribution bias, making CV and TSI comparable across different time periods. The first-level feature rules output by the dual-modal correlation model can help verify the rationality of stability. When the anomaly label shows a strong correlation between two features, but the stability difference is too large, the data quality needs to be checked.

[0115] (3) Redundancy index:

[0116] Red = 0.5 × I1 + 0.5 × I2

[0117] Where Red represents the score of the redundancy index; I1 and I2 are indicator functions, I1=1 when r≥0.8, otherwise I1=0, and r is the Pearson correlation coefficient; I2=1 when MI≥0.6, otherwise I2=0, and MI is the mutual information entropy. The higher the Red value, the higher the redundancy.

[0118] In this embodiment, Pearson correlation coefficient and mutual information entropy are calculated based on the preprocessed time series data (i.e., the time series normalized feature values) and the strongly correlated node pairs output by the bimodal correlation model. The redundancy between features is measured by fusing the Pearson correlation coefficient (with a threshold of 0.8) and the mutual information entropy (with a threshold of 0.6), taking into account both linear and nonlinear redundancy, to more accurately reflect the degree of redundancy between features.

[0119] (4) Timeliness indicators:

[0120] Timelinesss = Δt ÷ T duration

[0121] Wherein, Timelinesss represents the score of the timeliness index; Δt is the time difference between the time when the feature shows abnormal signs and the actual time when the energy meter experiences an abnormality; T duration Δt represents the duration of the anomaly. Δt > 0 indicates an early warning, while Δt < 0 indicates a delayed warning.

[0122] In this embodiment, based on the preprocessed time-series data (i.e., the time difference Δt between the time when the feature shows abnormal signs and the actual time of the abnormality in the electricity meter) and the rule-based time series output by the bimodal correlation model, the ratio of the feature anomaly advance time (i.e., the time difference Δt between the time when the feature shows abnormal signs and the time when the electricity meter actually experiences an abnormality) to the duration of the abnormality is calculated as a timeliness indicator. When a threshold Δt ≥ 30s is set, the timeliness score is ≥ 0.7, ensuring that the selected features can provide early warnings of abnormalities.

[0123] S104. Calculate the real-time anomaly rate based on the anomaly label, and dynamically adjust the weights of the four evaluation indicators based on the real-time anomaly rate.

[0124] In one possible embodiment, the formula for calculating the real-time anomaly occurrence rate is as follows:

[0125] ρ t =0.3×ρ t-1 +0.7× Ns' / Ns

[0126] Where, ρ t ρ represents the real-time anomaly occurrence rate for the current period. t-1 N represents the real-time anomaly occurrence rate of the previous period. s 'N represents the number of samples with abnormal labels in the current period.' s This represents the total number of samples in the current period.

[0127] In this embodiment, the real-time anomaly occurrence rate is calculated using an exponential moving average instead of a simple arithmetic average. This means that the real-time anomaly occurrence rate not only includes anomalies in the current period (70%), but also inherits anomalies from the previous period (30%). This smooths out transient pulse interference (e.g., instantaneous network jitter generating a large number of false anomalies), allowing the real-time anomaly occurrence rate to more robustly reflect the true trend of electricity meter anomalies, rather than instantaneous fluctuations. The update cycle can be 1 hour, which allows the system to respond to state changes in a timely manner while avoiding frequent policy swings caused by excessively short cycles.

[0128] In one possible embodiment, the step of dynamically adjusting the weights of the four evaluation metrics based on the real-time anomaly occurrence rate includes:

[0129] When the real-time anomaly rate is greater than or equal to the preset rate, increase the ratio between the weight of the correlation indicator and the weight of the stability indicator; when the real-time anomaly rate is less than the preset rate, decrease the ratio between the weight of the correlation indicator and the weight of the stability indicator.

[0130] In the specific implementation process, the preset rate is, for example, 5%, when ρ t When ≥5%, adjust the weights of the relevance indicator to 0.4, the stability indicator to 0.2, the redundancy indicator to 0.2, and the timeliness indicator to 0.2; when ρ t When the percentage is less than 5%, the weights of the relevance indicator, stability indicator, redundancy indicator, and timeliness indicator are adjusted to 0.3, 0.3, 0.2, and 0.2 respectively.

[0131] In this embodiment, the weights of redundancy and timeliness remain unchanged. When the real-time anomaly rate is high, the electricity meter may be in an unstable state. In this case, the weight of the correlation index is increased to quickly and accurately pinpoint the key features causing the problem. When the real-time anomaly rate is low, it indicates that the electricity meter is in a stable state. In this case, increasing the weight of the stability index can filter out features with long-term stable performance and small fluctuations.

[0132] S105. Calculate the comprehensive score for each feature based on the scores of each feature on the four evaluation indicators and the weights of the four evaluation indicators.

[0133] In practice, since the scores of the relevance, stability, and timeliness indicators are all between 0 and 1, and all indicate that a higher value indicates a better beneficial effect, standardization is unnecessary. However, although the score of the redundancy indicator is also between 0 and 1, a lower value indicates a better effective effect. Therefore, the score of the redundancy indicator is standardized. The formula for calculating the comprehensive score is as follows:

[0134]

[0135] in, The score represents the overall feature score, while Corr represents the score for the relevance index. Here, represents the weight of the correlation index, and Stab represents the score of the stability index. Here, Red represents the weight of the stability index, and Red represents the score of the redundancy index. Here, represents the weight of the redundancy index, and Timeliness is the score of the timeliness index. The weight of the timeliness indicator.

[0136] S106. Determine the dynamic threshold based on the maximum and average values ​​of the comprehensive scores of all features, and filter out features with comprehensive scores greater than the dynamic threshold to obtain a set of key features.

[0137] In one possible embodiment, the dynamic threshold is calculated using the following formula:

[0138] T=α×S 综合max +(1-α)×S 综合mean

[0139] Where T is the dynamic threshold; S 综合max S is the maximum value of the combined score of all features; 综合mean The average score of all features is α, which is a threshold parameter with a value range of [0.3, 0.7]. The threshold parameter is optimized using the Q-learning algorithm.

[0140] In the specific implementation process, the state space of Q-learning is the current α value and the F1 values ​​of the last 3 times, the action space is the adjustment amount of α ± 0.05, the learning rate ε = 0.1, the discount factor γ = 0.9, and the reward function of Q-learning is the anomaly identification F1 value of the filtered feature set. The goal is to make the F1 value ≥ 0.85 for 5 consecutive cycles to ensure the high recognition performance of the feature set. Combined with the genetic algorithm, the feature set with the highest F1 value is found. With the feature dimension ≤ 8 and the redundancy Red ≤ 0.3 as constraints, the optimal feature set is selected to improve the efficiency and accuracy of feature selection.

[0141] In this embodiment, Smax (maximum value) represents the best-performing feature in the current feature set. Smean (mean value) represents the overall quality distribution of all features. When α approaches 0.7, T is closer to the maximum value, the selection criteria are very strict, and only a few top-performing features with very high scores are selected, ensuring the refinement and high quality of the key feature set. When α approaches 0.3, T is closer to the mean value, the selection criteria are relatively lenient, and more mid-to-high-level features are included, ensuring the comprehensiveness of the key feature set and avoiding the omission of potential features. Therefore, the selection strategy can be flexibly controlled through the α value. The Q-learning algorithm, through learning, can automatically and intelligently select the most suitable α value based on the feature score distribution of the current data, thereby generating an optimal key feature set.

[0142] In one possible embodiment, after obtaining the set of key features, the method further includes:

[0143] Based on datasets from different user scenarios, adversarial generation techniques are used to generate difficult example samples. These user scenarios include urban residential users, industrial users, and users in remote areas. The set of key features is validated using difficult example samples and standard samples. If the set of key features passes validation, it is input into anomaly detection to obtain the status of the electricity meter. If the set of key features fails validation, the weights and dynamic thresholds of the four evaluation indicators are adjusted.

[0144] In the specific implementation process, the selected key feature sets are validated across three heterogeneous scenarios (urban residents, industrial users, and users in remote areas) on datasets. Adversarial validation methods are used to generate difficult sample samples. The Fast Gradient Sign Method (FGSM) is used to add a perturbation of ε=0.01 to the standard samples to simulate abnormal samples under complex working conditions, thereby generating difficult sample samples.

[0145] Furthermore, the accuracy rates for hard examples and standard examples are calculated separately using the FGSM algorithm, i.e., standard example accuracy = number of accurately identified standard examples / total number of standard examples. The F1 score for each scenario is then calculated using the following formula:

[0146] F1 = 2 × (Precision × Recall) / (Precision + Recall)

[0147] Calculate the standard deviation based on the F1 value:

[0148]

[0149] in, σ This represents the standard deviation of the three scenarios. The average F1 score for the three scenarios is given by F1.i Let be the accuracy for the i-th type of scenario.

[0150] The features are mapped to a high-dimensional space using a Gaussian kernel function, and the distribution distance is calculated.

[0151]

[0152] Where X is the feature set of urban residents, Y is the feature set of industrial scenes, MMD(X,Y) represents the distribution distance between feature sets X and Y; n is the sample size of the feature set of urban residents, and m is the sample size of the feature set of industrial scenes. Let x be the sample value of the x-th feature in the feature set X. Let y be the sample value of the y-th feature in the feature set Y.

[0153] When the accuracy of the standard sample is ≥0.9 (the accuracy of the three scenarios of urban residential users, industrial users, and remote area users is greater than or equal to 0.9), the accuracy of the difficult sample is ≥0.8, σ≤0.05 and MMD≤0.1, it means that the key feature set has been validated. Inputting the key feature set into the abnormal state detection model will output the state of the electricity meter, where 0 represents normal and 1 represents abnormal.

[0154] In this embodiment of the application, by performing cross-scenario verification and adversarial verification on the key feature set in heterogeneous scenarios, the stability and accuracy of the key feature set under different scenarios and complex working conditions are ensured, and the generalization ability of the key feature set is significantly improved.

[0155] In one possible implementation, if the accuracy of the standard sample is <0.9, the accuracy of the difficult sample is <0.8, σ>0.05, or MMD<0.1, it indicates that the key feature set has not passed the validation. The weights of the four evaluation indicators are adjusted or the dynamic threshold is lowered to incorporate more potential features for calculation, thereby completing the optimization.

[0156] In summary, this application provides a multi-dimensional electrical feature fusion and screening method for electricity meters, employing different preprocessing methods for different anomaly types. By constructing a bimodal association model, the preprocessed data is passed through an XGBoost model to output the contribution weights of each feature. Then, transactions are divided into transactions using a sliding window based on time-series data, and weighted support is calculated based on the contribution weights. A dynamic support threshold is calculated based on the weighted support. A feature association graph is constructed to visualize the relationship between features and anomalies, and strongly connected components are partitioned. Dynamically adjusted support and confidence thresholds make the association rules more targeted and reliable. The hierarchical rule set provides more association information for feature evaluation. A four-dimensional evaluation index system considering relevance, stability, redundancy, and timeliness comprehensively considers the performance of features across different dimensions. The index weights are dynamically adjusted based on the real-time anomaly occurrence rate, making the evaluation results more consistent with actual working conditions and accurately reflecting the actual value of the features. A dynamic threshold screening mechanism combined with a genetic algorithm is used to optimize feature subsets, selecting the globally optimal set of key features, improving the effectiveness and relevance of the features. By conducting cross-scenario verification and adversarial verification in heterogeneous scenarios, the stability and accuracy of the key feature set under different scenarios and complex working conditions are ensured, significantly improving the generalization ability of the key feature set and making it suitable for identifying abnormal states of electricity meters for different types of users.

[0157] Based on the same inventive concept, please refer to Figure 2 This application also provides a multi-dimensional electrical characteristic fusion screening device for electricity meters, the device comprising:

[0158] The preprocessing module is used to preprocess the multi-dimensional electrical characteristic time-series data of the electricity meter during operation to obtain preprocessed time-series data;

[0159] The association analysis module is used to input preprocessed time series data into a bimodal association model and output association rules between features and abnormal states. The association rules include rule time series, strongly associated node pairs, first-level feature rules, second-level feature rules, and anomaly labels.

[0160] The feature evaluation module is used to calculate the score of each feature on four evaluation indicators according to the association rules. The four evaluation indicators include relevance, stability, redundancy and timeliness. The module calculates the real-time anomaly occurrence rate based on the anomaly label and dynamically adjusts the weights of the four evaluation indicators based on the real-time anomaly occurrence rate. The module calculates the comprehensive score of each feature based on the score of each feature on the four evaluation indicators and the weights of the four evaluation indicators.

[0161] The feature filtering module is used to determine a dynamic threshold based on the maximum and average values ​​of the comprehensive scores of all features, and to filter out features with comprehensive scores greater than the dynamic threshold to obtain a set of key features.

[0162] Optionally, the multi-dimensional electrical characteristic time-series data includes metering anomaly data and / or communication anomaly data; the preprocessing module is specifically used for:

[0163] Dynamic time warping algorithm is used to correct measurement anomaly data, and attention-based missing value imputation method is used to impute communication anomaly data to obtain repaired time series data.

[0164] By introducing a sliding window mechanism to update statistical parameters in real time, the repaired time series data is normalized to obtain preprocessed time series data.

[0165] Optionally, the association analysis module is specifically used for:

[0166] The preprocessed time series data is input into the XGBoost model, and the contribution weight of each feature is output.

[0167] The preprocessed time series data is divided into transactions according to a preset sliding window; each transaction contains the status of each feature within the window and anomaly labels;

[0168] Calculate the weighted support of each feature based on its contribution weight and the number of transactions.

[0169] Calculate the support threshold based on the contribution weight;

[0170] Based on weighted support and support threshold, a feature association graph is constructed to divide strongly connected components and output the association rules between features and abnormal states.

[0171] Optionally, the formula for calculating weighted support is as follows:

[0172]

[0173] Among them, S i ω is the weighted support of the i-th feature; i The contribution weight of the i-th feature; N i is the transaction count, i.e., the number of transactions containing the i-th feature; N is the total number of transactions.

[0174] Optionally, the support threshold range is as follows:

[0175]

[0176] Where, ω max This represents the maximum value of the contribution weights of all features.

[0177] Optionally, the formulas for calculating the scores of the four evaluation indicators are as follows:

[0178] Corr = 0.6 × MIC + 0.4 × KCC

[0179] Where Corr represents the score of the correlation index; MIC represents the maximum information coefficient; and KCC represents the kernel correlation coefficient.

[0180] Stab = 0.7 × (1 - CV) + 0.3 × TSI

[0181] Where Stab represents the stability index score; CV represents the coefficient of variation within the sliding window, CV=σ / μ, where σ is the standard deviation of the data within the sliding window and μ is the mean of the data within the sliding window; TSI is the trend stability index, TSI =(1 - |k|), where k is the slope obtained by linear regression on the data within the sliding window;

[0182] Red = 0.5 × I1 + 0.5 × I2

[0183] Where Red represents the redundancy score; I1 and I2 are indicator functions, I1=1 when r≥0.8, otherwise I1=0, and r is the Pearson correlation coefficient; I2=1 when MI≥0.6, otherwise I2=0, and MI is the mutual information entropy.

[0184] Timelinesss = Δt ÷ T duration

[0185] Wherein, Timelinesss represents the score of the timeliness index; Δt is the time difference between the time when the feature shows abnormal signs and the actual time when the energy meter experiences an abnormality; T duration The duration of the abnormality.

[0186] Optionally, the formula for calculating the real-time anomaly incidence rate is as follows:

[0187] ρ t =0.3×ρ t-1 +0.7× Ns' / Ns

[0188] Where, ρ t ρ represents the real-time anomaly occurrence rate for the current period. t-1 N represents the real-time anomaly occurrence rate of the previous period. s 'N represents the number of samples with abnormal labels in the current period.' s This represents the total number of samples in the current period.

[0189] Optionally, the feature evaluation module is specifically used for:

[0190] When the real-time anomaly rate is greater than or equal to the preset rate, increase the ratio between the weight of the correlation index and the weight of the stability index.

[0191] When the real-time anomaly rate is less than the preset rate, reduce the ratio between the weight of the correlation index and the weight of the stability index.

[0192] Optionally, the formula for calculating the dynamic threshold is as follows:

[0193] T=α×S 综合max +(1-α)×S 综合mean

[0194] Where T is the dynamic threshold; S 综合max S is the maximum value of the combined score of all features; 综合mean The average score of all features is α, which is a threshold parameter with a value range of [0.3, 0.7]. The threshold parameter is optimized using the Q-learning algorithm.

[0195] Optionally, the device also includes a cross-scene verification module, which is specifically used for:

[0196] A dynamic threshold is determined based on the maximum and average comprehensive scores of the features. Features with comprehensive scores greater than the dynamic threshold are selected. After obtaining the key feature set, adversarial generation techniques are used to generate difficult example samples based on datasets from different user scenarios. User scenarios include urban residents, industrial users, and users in remote areas.

[0197] The set of key features was validated using difficult and standard samples.

[0198] If the key feature set passes the verification, the key feature set is input into the abnormal state detection to obtain the status of the electricity meter.

[0199] It should be noted that each module in the multi-dimensional electrical feature fusion screening device of the electricity meter in this embodiment corresponds one-to-one with each step in the multi-dimensional electrical feature fusion screening method of the electricity meter in the aforementioned embodiment. Therefore, the specific implementation of this embodiment can refer to the implementation of the multi-dimensional electrical feature fusion screening method of the electricity meter mentioned above, and will not be repeated here.

[0200] Based on the same inventive concept, this application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory. The computer program is executed by the processor to implement the aforementioned multi-dimensional electrical feature fusion screening method for electricity meters.

[0201] Based on the same inventive concept, this application also provides a computer storage medium storing a computer program, which is executed by a processor to implement the aforementioned multi-dimensional electrical feature fusion screening method for electricity meters.

[0202] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a device including one or any combination of the above-mentioned memories. The computer may be a variety of computing devices, including smart terminals and servers.

[0203] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0204] As an example, executable instructions may, but do not necessarily, correspond to files in the file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborative files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0205] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0206] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0207] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0208] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-dimensional electrical feature fusion screening method for electricity meters, characterized in that, include: The multi-dimensional electrical characteristic time-series data of the electricity meter during operation are preprocessed to obtain preprocessed time-series data; The preprocessed time-series data is input into a bimodal association model, which outputs association rules between features and abnormal states. The association rules include rule time series, strongly associated node pairs, first-level feature rules, second-level feature rules, and abnormal labels. Each feature is scored on four evaluation metrics based on the association rules; the four evaluation metrics include relevance, stability, redundancy, and timeliness. The real-time anomaly rate is calculated based on the anomaly labels, and the weights of the four evaluation indicators are dynamically adjusted based on the real-time anomaly rate. Calculate the overall score for each feature based on the scores of each feature on the four evaluation indicators and the weights of the four evaluation indicators; A dynamic threshold is determined based on the maximum and average scores of all features. Features with a comprehensive score greater than the dynamic threshold are then selected to obtain a set of key features.

2. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The multidimensional electrical characteristic time-series data includes metering anomaly data and / or communication anomaly data; The preprocessing of multi-dimensional electrical characteristic time-series data of the electricity meter during operation to obtain preprocessed time-series data includes: The measurement anomaly data is corrected using a dynamic time warping algorithm, and the communication anomaly data is filled with missing values ​​using an attention-based missing value imputation method to obtain repaired time series data. By introducing a sliding window mechanism to update statistical parameters in real time, the repaired time series data is normalized to obtain preprocessed time series data.

3. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The step of inputting the preprocessed time-series data into a bimodal association model and outputting association rules between features and abnormal states includes: The preprocessed time series data is input into the XGBoost model, and the contribution weight of each feature is output. The preprocessed time-series data is divided into transactions according to a preset sliding window; each transaction includes the status and anomaly labels of each feature within the window; Calculate the weighted support of each feature based on its contribution weight and the number of transactions. Calculate the support threshold based on the contribution weight; Based on the weighted support and the support threshold, a feature association graph is constructed to perform strongly connected component partitioning, and the association rules between features and abnormal states are output.

4. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 3, characterized in that, The formula for calculating the weighted support is as follows: ； Among them, S i ω is the weighted support of the i-th feature; i The contribution weight of the i-th feature; N i is the transaction count, i.e., the number of transactions containing the i-th feature; N is the total number of transactions.

5. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 3, characterized in that, The range of the support threshold is as follows: ； Where, ω max This represents the maximum value of the contribution weights of all features.

6. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The formulas for calculating the scores of the four evaluation indicators are as follows: Corr = 0.6 × MIC + 0.4 × KCC; Where Corr represents the score of the correlation index; MIC represents the maximum information coefficient; and KCC represents the kernel correlation coefficient. Stab = 0.7 × (1 - CV) + 0.3 × TSI; Where Stab represents the stability index score; CV represents the coefficient of variation within the sliding window, CV=σ / μ, where σ is the standard deviation of the data within the sliding window and μ is the mean of the data within the sliding window; TSI is the trend stability index, TSI =(1 - |k|), where k is the slope obtained by linear regression on the data within the sliding window; Red = 0.5 × I1 + 0.5 × I2; Where Red represents the redundancy score; I1 and I2 are indicator functions, I1=1 when r≥0.8, otherwise I1=0, and r is the Pearson correlation coefficient; I2=1 when MI≥0.6, otherwise I2=0, and MI is the mutual information entropy. Timelinesss=Δt÷T duration； Where Timelinesss represents the score of the timeliness index; Δt is the time difference between the time when the feature shows abnormal signs and the actual time when the energy meter shows abnormality; T duration The duration of the abnormality.

7. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The formula for calculating the real-time anomaly occurrence rate is as follows: ρ t =0.3×ρ t-1 +0.7× Ns' / Ns; Where, ρ t ρ represents the real-time anomaly occurrence rate for the current period. t-1 N represents the real-time anomaly occurrence rate of the previous period. s 'N represents the number of samples with abnormal labels in the current period.' s This represents the total number of samples in the current period.

8. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The step of dynamically adjusting the weights of the four evaluation indicators based on the real-time anomaly occurrence rate includes: When the real-time anomaly occurrence rate is greater than or equal to the preset rate, the ratio between the weight of the correlation index and the weight of the stability index is increased. When the real-time anomaly occurrence rate is less than the preset rate, the ratio between the weight of the correlation index and the weight of the stability index is reduced.

9. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, The formula for calculating the dynamic threshold is as follows: T=α×S 综合max +(1-a)×S 综合mean ; Where T is the dynamic threshold; S 综合max S is the maximum value of the combined score of all features; 综合mean The average score of all features is α, which is a threshold parameter with a value range of [0.3, 0.7]. The threshold parameter is optimized using the Q-learning algorithm.

10. The multi-dimensional electrical feature fusion screening method for electricity meters according to claim 1, characterized in that, After determining a dynamic threshold based on the maximum and average of the comprehensive scores of all features, filtering out features whose comprehensive scores are greater than the dynamic threshold, and obtaining a set of key features, the method further includes: Difficult example samples are generated using adversarial generation techniques based on datasets from different user scenarios; these user scenarios include urban residents, industrial users, and users in remote areas. The set of key features is validated using the difficult examples and standard examples. If the set of key features passes verification, the set of key features is input into the abnormal state detection to obtain the state of the electricity meter.

Images