A method and system for quantifying the root cause of electric meter failure by fusing multi-dimensional risk features

Abstract

The application discloses a kind of fusion multi-dimensional risk characteristics's meter fault root cause quantification method and system, method includes: obtaining the historical fault data of intelligent electric meter, the severity level of each fault type is determined by expert scoring;Combining attribution feedback self-adapting training fault prediction model;Through the attribution analysis in the prediction process of fault prediction model by explainable AI technology, the SHAP value of each fault element is obtained, the SHAP value of each fault element, fault occurrence time and severity level form attribution triple;Through comprehensive risk weight quantitative calculation, the comprehensive risk weight of each fault element is saved;The comprehensive risk weight of the fault element of the meter in the network is found and accumulated, and the total risk score corresponding to each meter in the network is obtained, so as to obtain the static risk image of the meter in the network.The application can generate more accurate equipment risk image, which significantly improves the intelligent level of power grid asset active operation and data management.

Description

Technical Field

[0001] This invention relates to power big data analysis technology, specifically to a method and system for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics. Background Technology

[0002] With the widespread adoption of Advanced Metering Architectures (AMIs) in smart grids, massive amounts of smart meter operation and fault data are being collected, providing valuable data resources for refined operation and maintenance and asset management of the power grid. Utilizing artificial intelligence and machine learning technologies to automate the analysis of this data for rapid fault diagnosis and root cause analysis has become the mainstream technological development direction in this field.

[0003] Currently, some advanced machine learning methods, such as gradient boosting decision trees (e.g., XGBoost), combined with interpretable AI technologies (e.g., SHAP), have been applied to identify key features related to faults. These technologies can extract potential factors leading to specific faults from complex feature relationships and, to some extent, quantify the contribution of each factor, significantly improving efficiency and accuracy compared to traditional troubleshooting methods that rely on human experience.

[0004] However, existing technologies still have the following deep-seated limitations when performing root cause quantification and risk assessment for failures: 1. Risk assessment lacks business severity awareness: Existing attribution analysis methods typically output business-neutral contribution values, failing to distinguish whether a root cause leads to a minor issue like "display malfunction" or a serious consequence like "measuring inaccuracies." This results in a discrepancy between the model's risk assessment results and the actual business priorities of the operations department, making it difficult to directly guide the governance of key risks.

[0005] 2. Lack of Timely Insight in Risk Assessment: Risk assessment models are typically trained statically once using all historical data, and the analysis results are a comprehensive reflection of all historical patterns. This "time-insensitive" assessment method is slow to react to new failure modes that emerge due to equipment batch aging, firmware version updates, or changes in environmental factors, where the risk is rapidly escalating, resulting in a lag in risk profiling.

[0006] 3. Lack of Pattern Differentiation in Risk Profiling: Traditional attribution analysis often provides an average or cumulative contribution value, resulting in a diagnosis with ambiguous patterns. It struggles to effectively distinguish between two distinct risk patterns: a "high-frequency, low-risk" pattern that occurs frequently but has manageable impact, and a "low-frequency, high-risk" pattern that is sporadic, unpredictable, but can have significant consequences once it occurs. This confusion between the two risk patterns hinders the development of differentiated and proactive risk management strategies. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method and system for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics, which can go beyond the traditional, static contribution calculation. It can quantify the root causes of failures in a deep and three-dimensional way from multiple dimensions such as business impact, time dynamics and risk volatility, thereby generating a more insightful equipment risk profile and providing more accurate decision support for the intelligent and proactive management of power grid assets.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics includes the following steps: S1: Obtain historical fault data of smart meters. Each sample in the historical fault data includes a multi-source heterogeneous feature vector and fault type label of a fault event. The severity level of each fault type is determined by expert scoring. S2: Using historical fault data, a fault prediction model is trained by combining an adaptive optimization mechanism based on attribution stability feedback, so that the fault prediction model can predict the corresponding fault type based on the input feature vector. S3: Through interpretable AI technology, attribution analysis is performed on each sample in the historical fault data during the prediction process of the fault prediction model to obtain the SHAP value of each fault element. The SHAP value of each fault element, the corresponding fault occurrence time and the corresponding severity level are combined into an attribution triplet. Each fault element is the specific value of the corresponding feature of each sample in the historical fault data. S4: Based on the attribution triplet corresponding to each fault element, calculate and save the comprehensive risk weight of each fault element through comprehensive risk weight quantification; S5: Obtain the fault elements of all on-grid operating meters, find the comprehensive risk weight of the fault elements of the same on-grid operating meter and sum them up to obtain the total risk score corresponding to each on-grid operating meter, thereby obtaining the static risk profile of the on-grid operating meters.

[0009] Furthermore, when training the fault prediction model using an adaptive optimization mechanism based on attribution stability feedback, the specific steps include: The historical fault data is divided into a training set and a validation set. In each iteration of hyperparameter optimization, a temporary prediction model is trained on a sub-training set based on the current hyperparameter combination. A temporary prediction model is used to predict and calculate the evaluation index of the validation set, and the SHAP value of each feature in the validation set is extracted through interpretable AI technology. Calculate the standard deviation and mean of the SHAP values ​​of each feature in the validation set, and calculate the attribution volatility based on the standard deviation and mean; Construct a joint fitness function, which is the weighted difference between the prediction evaluation index and the attribution volatility; With the goal of maximizing the joint fitness function, the hyperparameter combination is iteratively updated through an optimization algorithm, the optimal hyperparameter combination is output, and the final fault prediction model is trained.

[0010] Furthermore, when calculating the comprehensive risk weight of each fault element through comprehensive risk weight quantification, it includes: The weighted average contribution of dynamic severity perception is calculated based on all attribution triples of the current fault element. Calculate the mean and standard deviation of all SHAP values ​​for the current faulty element, and calculate the risk volatility amplifier based on the mean and standard deviation; The weighted average contribution of the dynamic severity perception of the current fault element is multiplied by the result of the risk volatility amplifier calculation to obtain the comprehensive risk weight of the current fault element.

[0011] Furthermore, the mathematical expression for the weighted average contribution calculation result of the dynamic severity perception is as follows:

[0012] in, This indicates the first fault in the historical fault data. The absolute value of the SHAP value of the current fault element in each fault sample. This indicates the first fault in the historical fault data. The severity level corresponding to the current fault element in each fault sample. Based on the time interval of the fault occurrence Calculated time decay factor, This indicates the first fault in the historical fault data. The interval between the time of occurrence of the current fault element and the current time in each fault sample. This is the attenuation coefficient.

[0013] Furthermore, the mathematical expression for the calculation result of the risk volatility amplifier is as follows:

[0014] in, This represents the standard deviation of the SHAP value of the current faulty element across all historical fault samples. This represents the mean SHAP value of the current faulty element across all historical fault samples. It is the volatility sensitivity coefficient.

[0015] Furthermore, before training the fault prediction model using historical fault data, a data preprocessing step is also included. Specifically, the categorical features in the feature vector of each sample in the historical fault data are converted into numerical format using one-hot encoding.

[0016] Furthermore, after obtaining the static risk profile of the electricity meters in operation, the process also includes: sorting the total risk score of each electricity meter in operation in descending order, and specifying the corresponding inspection or maintenance plan based on the sorting results.

[0017] Furthermore, after obtaining the static risk profile of the on-grid operating meters, the method also includes: if there are on-grid operating meters with a total risk score greater than the risk threshold, increasing the data collection frequency of the on-grid operating meters.

[0018] Furthermore, in step S1, when obtaining historical fault data of the smart meter, the historical fault data of the smart meter for different time periods is obtained according to a specified time interval. After calculating the comprehensive risk weight of each fault element by quantifying the comprehensive risk weight, the method further includes: if there are fault elements whose comprehensive risk weight continues to increase over time, the smart meter is replaced, upgraded, or modified.

[0019] The present invention also proposes a meter fault root cause quantification system that integrates multi-dimensional risk features, including a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the meter fault root cause quantification method that integrates multi-dimensional risk features.

[0020] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for quantifying the root causes of meter failures by incorporating multidimensional risk features.

[0021] Compared with the prior art, the advantages of the present invention are as follows: The accuracy and depth of fault tracing are significantly improved. This invention achieves a leap from simply locating the root cause to deeply characterizing the risk profile of the root cause. Its output risk weights are rich in content, not only quantifying the magnitude of risk, but also revealing the business impact, timeliness trend, and potential patterns of the risk, resulting in more accurate and comprehensive diagnostic results.

[0022] Model accuracy and attribution stability achieve closed-loop synergy. Traditional machine learning model parameter tuning focuses solely on improving prediction accuracy, which easily leads to drastic fluctuations in the interpretation of underlying features across different samples, reducing the reliability of root cause analysis. This invention pioneers a hyperparameter adaptive optimization mechanism with a closed-loop feedback between prediction and attribution. It introduces "attribution volatility" as a penalty term during the model training phase, ensuring that the final fault prediction model is not only accurate in prediction but also stable in attribution, laying a solid and high-quality data foundation for subsequent comprehensive risk weight quantification.

[0023] Risk identification has shifted from passive to proactive. By generating accurate and dynamic risk profiles for each networked device, this invention can effectively identify potential "high-risk entities," enabling operation and maintenance management to transform from traditional reactive response to abnormal data to proactive control of high-risk sources, fundamentally improving the reliability of power grid operation.

[0024] The decision support capability is greatly enhanced. The quantitative risk results output by this invention are strongly correlated with business operations, and can directly provide "surgical" level precision decision support for the entire life cycle management of assets. For example, it can guide the formulation of differentiated proactive maintenance strategies, optimize equipment replacement and decommissioning plans, and accurately allocate operation and maintenance resources, thereby significantly improving the scientific nature of management and the return on investment.

[0025] The system is highly intelligent and adaptable. The method framework of this invention has good adaptability, and can automatically learn and adapt to newly emerging fault modes through a time decay mechanism. Its risk profiling results can also be used as input for upper-level intelligent applications, such as enabling the establishment of risk-adaptive dynamic data acquisition strategies, optimizing resource allocation through intelligent means, and improving the intelligence level of the entire power grid data governance system. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0028] To address the shortcomings of existing machine learning-based fault tracing technologies in risk assessment, such as insufficient perception of the severity of business impact, lack of insight into the timeliness of risks, and inability to effectively distinguish risk patterns, this embodiment proposes a multi-dimensional risk feature-integrated root cause quantification method for meter faults. This method constructs a complete quantification path from instance-level attribution to multi-dimensional risk fusion. First, a pre-trained fault attribution model and interpretable AI technology are used to obtain the instance-level basic contribution value of each fault element. Then, an innovative Comprehensive Risk Weight (CRW) quantification model is used to nonlinearly fuse the basic contribution value with fault severity level, time decay factor, and risk volatility index to calculate the final root cause risk weight.

[0029] like Figure 1 As shown, the method includes the following steps: S1: Data preparation and business knowledge integration: Obtain historical fault data of smart meters. Each sample in the historical fault data includes a multi-source heterogeneous feature vector and fault type label of a fault event. The severity level of each fault type is determined by expert scoring. S2: Construct a high-precision fault prediction model: Use historical fault data and combine it with an adaptive optimization mechanism based on attribution stability feedback to train the fault prediction model, so that the fault prediction model can predict the corresponding fault type based on the input feature vector. S3: Instance-level fault attribution and information extraction: Through interpretable AI technology, attribution analysis is performed on each sample in the historical fault data during the prediction process of the fault prediction model to obtain the SHAP value of each fault element. The SHAP value of each fault element, the corresponding fault occurrence time and the corresponding severity level are combined into an attribution triplet. Each fault element is the specific value of the corresponding feature of each sample in the historical fault data. S4: Quantitative calculation of Comprehensive Risk Weight (CRW): Based on the attribution triplet corresponding to each fault element, the comprehensive risk weight of each fault element is quantitatively calculated and saved through comprehensive risk weight; S5: Risk Profile Generation and Application: Obtain the fault elements of all on-grid operating meters, find the comprehensive risk weight of the fault elements of the same on-grid operating meter and sum them up to obtain the total risk score corresponding to each on-grid operating meter, thereby obtaining the static risk profile of the on-grid operating meter.

[0030] The following provides a detailed explanation of each step.

[0031] This embodiment aims to collect and process the required data from a multi-source, heterogeneous power business system by acquiring meter data containing device attributes and historical fault records in step S1. It also innovatively parameterizes the business knowledge of domain experts, providing rich and high-quality input for subsequent models. The specific process is as follows: S11) Through a preset data interface, obtain historical data sets of smart meters within a specified range from data sources such as the electricity consumption information collection system (AMI), marketing business system, and geographic information system (GIS) of the power company, denoted as... Each sample in this dataset represents a confirmed failure event, and its data structure can be represented as a feature vector. A fault type label The combination of .

[0032] A specific fault sample eigenvectors Including but not limited to the following characteristics: Static device attributes: {Hardware version (e.g., 'V2020', 'V2021'), Rated voltage (e.g., '220V', '380V'), Connection method (e.g., 'Direct connection', 'Through current transformer')}.

[0033] User profile information: {User category (e.g., 'Residential', 'Commercial', 'High-voltage transformer'), Industry classification (e.g., 'Industrial', 'Service')}.

[0034] Environmental spatiotemporal information: {season in which the fault occurred (e.g., 'spring', 'summer'), geographical region (e.g., 'area A', 'area B')}.

[0035] The fault type label of this sample The fault type for this fault event, for example: ∈ {'Metering failure', 'Power failure', 'Communication failure', 'Display failure', ...}.

[0036] S12) This step performs a key business knowledge fusion operation, namely defining the Severity Coefficient, denoted as... This coefficient is quantitatively assessed by senior operation and maintenance experts based on the comprehensive impact of different fault types on power grid safety, metering and billing, and user complaints. In this embodiment, the coefficient is stored in a preset mapping table, as shown in Table 1.

[0037] Table 1 Fault Severity Level Mapping Table

[0038] After this step is completed, a structured training dataset is obtained. and quantified business knowledge parameters This has laid a solid foundation for subsequent steps.

[0039] Step S2 in this embodiment aims to train a machine learning model capable of deep learning and understanding the complex nonlinear relationships in meter fault data, serving as an intelligent engine for subsequent accurate attribution analysis.

[0040] In this embodiment, XGBoost (eXtreme Gradient Boosting) is selected as the core fault prediction model. This model was chosen because of its efficiency in processing tabular data, its powerful ability to capture nonlinear relationships, and its inherent feature interaction learning mechanism, making it very suitable for the application scenario of this embodiment. The specific process is as follows: S21) Data Preprocessing: Before model training, the feature vectors obtained in step S1 are processed... Perform necessary preprocessing. Specifically, for all categorical features (such as "hardware version", "user category", etc.), use one-hot encoding to convert them into a numerical format that the model can process.

[0041] S22) Model training: The preprocessed training dataset... Used to train the XGBoost model, denoted as The training objective of the model is a multi-class classification task, that is, to learn a mapping function. This enables the model to adapt to the input feature vector. It can accurately predict the corresponding fault type. .

[0042] S23) Adaptive Hyperparameter Configuration Based on Multidimensional Risk Feedback: To overcome the shortcomings of traditional static hyperparameter configuration in balancing model prediction accuracy and attribution stability, this embodiment abandons the method of manually setting static hyperparameters (such as manually specifying scale_pos_weight, max_depth, etc.) and innovatively introduces a closed-loop hyperparameter adaptive optimization mechanism based on attribution stability feedback. The specific process is as follows: First, the preprocessed training dataset The dataset is divided into a training set and a validation set. Then, a Bayesian optimization algorithm (such as the TPE algorithm) is used for dynamic iterative optimization of the hyperparameters. In each iteration: The optimization algorithm generates a set of current hyperparameter combinations (covering max_depth, n_estimators, scale_pos_weight, etc.); A temporary prediction model is obtained by training on a sub-training set based on this combination of hyperparameters. The temporary prediction model is applied to the validation set, and its prediction evaluation index (such as F1-Score) is calculated to take into account the sample imbalance problem. Extracting attributable volatility: Simultaneously apply SHAP technology on the validation set to extract the SHAP values ​​of each fault element feature in the validation set samples, and calculate the coefficient of variation (i.e., standard deviation) of the SHAP values ​​of key features. with the mean The ratio of the two values ​​is defined as the "attributable volatility" for that iteration. Constructing a joint fitness function: in, This is a pre-defined balance coefficient, specifically designed to penalize drastic fluctuations in attribution results caused by excessive tree depth or weight imbalance.

[0043] To maximize the joint fitness function The goal is to drive the Bayesian optimization algorithm to continuously update the hyperparameter search space. After the iteration, the output makes... The largest optimal combination of hyperparameters, and in the full set Retraining yields a final prediction model with high accuracy and high interpretability stability. .

[0044] Through this adaptive mechanism, the underlying model training in step S2 and the risk volatility quantification in the subsequent step S4 achieve deep synergy, ensuring that the model captures complex failure modes without generating false high-risk volatility due to overfitting, thus providing a highly confident basic contribution value for the calculation of CRW.

[0045] Step S3 in this embodiment aims to make the "black box" model trained in step S2 transparent. Through interpretable AI technology, it deeply analyzes the model's judgment basis for each specific fault and extracts structured attribution information for subsequent quantitative analysis.

[0046] In this embodiment, SHAP (SHapley Additive exPlanations) technology is selected as the core attribution analysis tool. Based on the Shapley value theory in game theory, SHAP technology has a solid theoretical foundation. Its "additivity" property can accurately decompose the model's prediction results into the sum of the contributions of each input feature, making it very suitable for the refined attribution requirements of this embodiment. The specific process is as follows: S31) Attribution Model Principles: For the model For any fault sample In the prediction process, SHAP constructs a linear interpretable model. For a specific fault type (e.g., "metering fault"), the output of the model can be represented as:

[0047] in: It is the model's final prediction output (e.g., probability value) that the sample belongs to this fault type. It is a simplified feature vector, where The binary value (1 or 0) represents the first... Does the feature exist? It represents the total number of input features. It is the baseline value of the model, that is, the mean of the predicted values ​​of all training samples. It is the first The SHAP value of each feature precisely quantifies how that feature affects the model's predictions from the baseline. Pushing towards the final value The contributions made during the process will be included in this embodiment. Defined as "basic contribution value".

[0048] S32) Attribution Analysis Example: Suppose a sample has experienced a "metering failure". Attribution analysis was performed. The characteristics of this sample include {hardware version='V2020', access method='via instrument transformer', user category='high voltage dedicated transformer'}. The SHAP analysis process is as follows: First, determine the model's prediction baseline for "metering failure". For example, 0.05.

[0049] Then, the SHAP value of each feature's specific value (referred to as "fault element" in this embodiment) for this prediction is calculated: The SHAP value of the fault element "Access Method = via Current Transformer" This means that the presence of this element increases the probability of a "metering failure" prediction by 0.30.

[0050] The SHAP value of the fault element "Hardware Version=V2020" .

[0051] The SHAP value of the fault element "User Category = High Voltage Transformer" This means that the presence of this element actually lowers the predicted probability by 0.05.

[0052] The sum of the SHAP values ​​of all other features is -0.02.

[0053] Final model prediction output:

[0054] S33) Structured information extraction: The core output of this step is to structurally integrate the discrete information obtained from the above analysis. For each "fault element" participating in the fault sample attribution process, an "attribution triple" containing rich contextual information is extracted and constructed. For example, for the fault element "connection method = via transformer" in the above sample, the generated attribution triple is:

[0055] The specific content is as follows: Basic contribution value ( ): Obtained directly from SHAP analysis, here it is +0.30.

[0056] Fault occurrence timestamp ( From this fault sample Extracted from the original record, such as '2024-10-25 14:30:00'. This information is the basis for subsequent calculation of the time decay factor.

[0057] Fault severity level ( Based on the fault type of this sample: = 'Metering fault', obtained by querying the mapping table established in step S1, is 10.0 here. This information is the basis for subsequent weighted calculation of business impact.

[0058] By performing this step on all fault samples in the training set, a large set of attribution triples organized around each "fault element" will be obtained. This set is no longer a simple list of SHAP values, but a carefully designed, structured dataset rich in multi-dimensional information.

[0059] Therefore, through steps S2 and S3, the method of this embodiment analyzes historical fault records based on a pre-trained fault attribution model and interpretable AI technology, and obtains instance-level basic contribution values ​​corresponding to each fault element in the fault records, providing complete and necessary input for the next step of performing quantitative calculation of Comprehensive Risk Weight (CRW).

[0060] Step S4 of this embodiment aims to integrate the discrete, multi-dimensional attribution information extracted in step S3 into a single comprehensive indicator that can fully reflect the nature of risk by using an original quantitative model—the comprehensive risk weight model.

[0061] In this embodiment, the external business features include: The fault severity level coefficient is used to weight the basic contribution value when calculating the comprehensive risk weight, so as to reflect the differentiated impact of different fault types on the business. The fault severity level coefficient is a quantitative value pre-set for each fault type based on historical operation and maintenance cost data or domain expert knowledge, which is directly obtained through the attribution triplet. The time decay factor, calculated based on the timestamps of historical fault records in the attribution triplet, is used to assign higher weight to the basic contribution values ​​generated in recent fault records when calculating the comprehensive risk weight.

[0062] In this embodiment, the integrated risk weight model also incorporates a statistical distribution characteristic of the basic contribution value itself. This characteristic is used to amplify the risk weight of fault elements whose contribution values ​​exhibit high volatility. This statistical distribution characteristic acts as a risk volatility amplifier, and its value is calculated based on the mean and standard deviation of all basic contribution values ​​of the fault element.

[0063] Therefore, in step S4, when calculating the comprehensive risk weight of each fault element through comprehensive risk weight quantification, it includes: S41) Calculation of the weighted average contribution of dynamic severity perception: The weighted average contribution of dynamic severity perception is calculated based on all attribution triples of the current fault element. The weighted average contribution of dynamic severity perception is an average base contribution value after contextual intelligent weighting. The mathematical expression of the calculation result is as follows:

[0064] in: This indicates the first fault in the historical fault data. The absolute value of the SHAP value of the current fault element in the fault sample, i.e., the th fault sample. The absolute value of the basic contribution of an element in a particular fault represents its basic contribution intensity in that fault. This indicates the first fault in the historical fault data. The severity level corresponding to the current fault element in each fault sample is used as a "severity level" modulator. It uses the severity level coefficient of this fault (derived from step S1) as a weight to directly modulate the base contribution value. For example, in a "metering fault"... An element that contributes +0.2 SHAP value in a given context will have a far greater impact than an element that contributes in a "display failure" context. The element that contributed +0.2 SHAP value also appeared in the previous one. Based on the time interval of the fault occurrence The calculated time decay factor serves as the "time decay" modulator. Wherein, This indicates the first fault in the historical fault data. The interval between the time of occurrence of the current fault element and the current time in each fault sample. This is a preset decay coefficient (e.g., 0.01). This exponential term gives higher weight to recent failures, while the influence of older failures decreases exponentially, allowing the model to dynamically focus on more recent and relevant risk patterns.

[0065] S42) Risk volatility amplifier calculation: Calculate the mean and standard deviation of all SHAP values ​​for the current faulty element. Then, calculate a risk volatility amplifier based on these mean and standard deviation to obtain a multiplier term used to quantify and amplify the "unpredictability" of risk. The mathematical expression for the calculation result is as follows:

[0066] in: This represents the standard deviation of the SHAP values ​​of the current faulty element across all historical fault samples. It directly measures the dispersion or instability of the element's risk contribution. An element with a high value means that its contribution is sometimes very low and sometimes it can produce extremely high values. This represents the mean SHAP value of the current faulty element across all historical fault samples. This is the volatility sensitivity coefficient. The amplifier term is passed through a preset volatility sensitivity coefficient. (For example, 0.5) is used to adjust for the effect of standard deviation. Its overall value is always greater than or equal to 1. For an element with a stable risk contribution ( When the value approaches 0, the amplifier value is close to 1, which basically does not change its risk assessment; however, for a "high-risk, occasional" element whose risk contribution is extremely unstable ( Much larger This amplifier value will be significantly greater than 1, thus penalizingly amplifying its final CRW weight. This design is specifically designed to accurately identify and highlight potentially high-risk low-frequency (HRLF) elements.

[0067] S43) Calculation of comprehensive risk weights: The weighted average contribution of the dynamic severity perception of the current fault element is multiplied by the result of the risk volatility amplifier calculation to obtain the comprehensive risk weight of the current fault element.

[0068] By combining the above two parts, the final Comprehensive Risk Weight (CRW) is a comprehensive risk indicator that deeply integrates risk intensity, business impact, time dynamics, and pattern volatility, possessing extremely rich connotations. For any "fault element" in the dataset (e.g., the specific value of the feature "access method" "via transformer"), firstly, the "attribution triplet" information set of all associated historical fault samples is collected from the output of step S3. Then, its final risk weight is calculated using the following Comprehensive Risk Weight (CRW) calculation formula:

[0069] Wherein, CRW represents the final calculated comprehensive risk weight; To iterate through all historical fault samples associated with this fault element; For the first The absolute value of the basic contribution value of each fault sample; For the first The severity level coefficient corresponding to each fault sample; Based on the time interval between failures Calculate the time decay factor; and These are the mean and standard deviation of all basic contribution values ​​for this element, respectively; and These are the preset hyperparameters.

[0070] In this embodiment, step S5, based on the calculated comprehensive risk weights of each fault element, sums the weights of all feature elements contained in any given meter to generate the meter's total risk score, thus forming an equipment-level risk profile. This aims to transform the quantitative risk results calculated in step S4 into precise insights that can directly guide actual business operations. The specific process is as follows: S51) Generation of Equipment-Level Risk Profiles: For any electricity meter operating on the network, its characteristic elements (such as hardware version, access method, etc.) have all had their corresponding CRW calculated in step S4. The total risk score of this electricity meter can be obtained by summing the CRWs of all its characteristic elements:

[0071] For example, the characteristics of an electricity meter A are {Hardware Version='V2020', Connection Method='Direct Connection', User Category='Residential'}. Assuming that the table lookup yields: CRW(Hardware Version='V2020') = 1.85; CRW(Connection Method='Direct Connection') = 0.23; CRW(User Category='Residential') = 0.11, then the total risk score for electricity meter A is 1.85 + 0.23 + 0.11 = 2.19. This score and its components together constitute a precise and quantifiable static risk profile for this device.

[0072] S52) Application of risk profiling: After obtaining the static risk profile of the electricity meters in operation, the generated risk profile can be directly applied to various advanced operation and maintenance and management application scenarios, including but not limited to: Proactive maintenance prioritization: All meters in operation are sorted in descending order of their total risk score, and corresponding inspection or maintenance plans are assigned based on the ranking results. Specifically, all meters within the jurisdiction are sorted in descending order of their total risk score. The maintenance department can then develop proactive inspection or maintenance plans based on this, precisely allocating limited human and material resources to the equipment with the highest risk, achieving "prevention before problems arise."

[0073] Risk-adaptive data acquisition strategy: If any on-grid meter has a total risk score greater than a risk threshold, the data acquisition frequency of that meter is increased. Specifically, a risk threshold can be set. When the total risk score of a meter exceeds this threshold, the system can automatically trigger an instruction to increase the data acquisition frequency of that meter (e.g., from once per hour to once every 15 minutes), thereby intelligently acquiring high-density data from key equipment at extremely low resource cost, providing support for subsequent refined diagnostics.

[0074] Asset Management Decision Support: In step S1 of this embodiment, when acquiring historical fault data of smart meters, specifically, historical fault data of smart meters for different time periods is acquired at specified time intervals. Correspondingly, after quantifying and calculating the comprehensive risk weight of each fault element through comprehensive risk weight, the following steps are also included: if there are fault elements whose comprehensive risk weight continuously increases over time, the smart meters are replaced, upgraded, or modified. For example, with a 7-day time interval, the historical fault data of the previous week is acquired each time, and steps S1 to S4 are executed on the acquired week's historical fault data to obtain the comprehensive risk weight of each fault element in the historical fault data of each week. When the CRW of a certain fault element (such as the "hardware version" of a certain batch) shows a continuous upward trend over time, it can provide strong data support for the asset management department to initiate special technical transformation, firmware upgrade, or batch replacement plans for the batch of equipment.

[0075] Furthermore, this embodiment also proposes a meter fault root cause quantification system that integrates multi-dimensional risk features, including a processor and a computer-readable storage medium. The computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the meter fault root cause quantification method that integrates multi-dimensional risk features as described in this embodiment.

[0076] Furthermore, this embodiment also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the meter fault root cause quantification method that integrates multi-dimensional risk features as described in this embodiment.

[0077] In summary, this invention proposes a method and system for quantifying the root causes of electricity meter faults by integrating multi-dimensional risk features. First, a pre-trained gradient boosting decision tree model optimized for the characteristics of power fault data is used to perform deep relationship learning on historical fault data. Next, interpretable AI technology is used to perform instance-level attribution analysis on specific fault samples to accurately obtain the basic contribution values ​​of each element constituting the fault. Subsequently, an original Comprehensive Risk Weight (CRW) quantification model is used to further process and enhance the obtained basic contribution values. This CRW model organically combines the basic contribution values ​​with the following three key external features through a non-linear fusion formula: Business Severity Dimension: Introduce a fault severity level coefficient that is pre-quantified based on operation and maintenance costs or expert experience, so that the risk assessment results are directly linked to the actual business impact; Time dynamic dimension: Introduce a time decay factor based on the fault occurrence timestamp to give higher analysis weight to recent fault modes, enabling risk assessment to have dynamic evolution capabilities; Risk volatility dimension: Introduce a risk volatility amplifier based on the statistical distribution characteristics of the underlying contribution value (such as mean and standard deviation) to quantify and amplify the occasional but highly impactful "low-frequency, high-risk" pattern.

[0078] By integrating the above-mentioned multi-dimensional features, the present invention finally calculates a comprehensive risk weight that can fully and three-dimensionally reflect the true risk level of each fault element, and can further generate a device-level risk profile. It realizes a complete automated process from data access and intelligent attribution to multi-dimensional risk quantification and application, effectively solving the problems raised in the background technology.

[0079] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create an implementation for the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics, characterized in that, Includes the following steps: S1: Obtain historical fault data of smart meters. Each sample in the historical fault data includes a multi-source heterogeneous feature vector and fault type label of a fault event. The severity level of each fault type is determined by expert scoring. S2: Using historical fault data, a fault prediction model is trained using an adaptive optimization mechanism based on attribution stability feedback. This model predicts the corresponding fault type based on the input feature vector. The training of the fault prediction model using the adaptive optimization mechanism based on attribution stability feedback specifically includes: The historical fault data is divided into a training set and a validation set. In each iteration of hyperparameter optimization, a temporary prediction model is trained on a sub-training set based on the current hyperparameter combination. A temporary prediction model is used to predict and calculate the evaluation index of the validation set, and the SHAP value of each feature in the validation set is extracted through interpretable AI technology. Calculate the standard deviation and mean of the SHAP values ​​of each feature in the validation set, and calculate the attribution volatility based on the standard deviation and mean; Construct a joint fitness function, which is the weighted difference between the prediction evaluation index and the attribution volatility; With the goal of maximizing the joint fitness function, the hyperparameter combination is iteratively updated through an optimization algorithm, the optimal hyperparameter combination is output, and the final fault prediction model is trained. S3: Through interpretable AI technology, attribution analysis is performed on each sample in the historical fault data during the prediction process of the fault prediction model to obtain the SHAP value of each fault element. The SHAP value of each fault element, the corresponding fault occurrence time and the corresponding severity level are combined into an attribution triplet. Each fault element is the specific value of the corresponding feature of each sample in the historical fault data. S4: Based on the attribution triplet corresponding to each fault element, calculate and save the comprehensive risk weight of each fault element through comprehensive risk weight quantification, including: The weighted average contribution of dynamic severity perception is calculated based on all attribution triples of the current fault element. Calculate the mean and standard deviation of all SHAP values ​​for the current faulty element, and calculate the risk volatility amplifier based on the mean and standard deviation; The weighted average contribution of the dynamic severity perception of the current fault element is multiplied by the result of the risk volatility amplifier calculation to obtain the comprehensive risk weight of the current fault element. S5: Obtain the fault elements of all on-grid operating meters, find the comprehensive risk weight of the fault elements of the same on-grid operating meter and sum them up to obtain the total risk score corresponding to each on-grid operating meter, thereby obtaining the static risk profile of the on-grid operating meters.

2. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, The mathematical expression for the weighted average contribution calculation result of the dynamic severity perception is as follows: in, This indicates the first fault in the historical fault data. The absolute value of the SHAP value of the current fault element in each fault sample. This indicates the first fault in the historical fault data. The severity level corresponding to the current fault element in each fault sample. Based on the time interval of the fault occurrence Calculated time decay factor, This indicates the first fault in the historical fault data. The interval between the time of occurrence of the current fault element and the current time in each fault sample. This is the attenuation coefficient.

3. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, The mathematical expression for the calculation result of the risk volatility amplifier is as follows: in, This represents the standard deviation of the SHAP value of the current faulty element across all historical fault samples. This represents the mean SHAP value of the current faulty element across all historical fault samples. It is the volatility sensitivity coefficient.

4. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, Before training a fault prediction model using historical fault data, a data preprocessing step is also included. Specifically, the categorical features in the feature vector of each sample in the historical fault data are converted into numerical format using one-hot encoding.

5. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, After obtaining the static risk profile of the electricity meters in operation, the process also includes: sorting the total risk score of each electricity meter in operation in descending order, and specifying the corresponding inspection or maintenance plan based on the sorting results.

6. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, After obtaining the static risk profile of the on-grid electricity meters, the method further includes: if there are on-grid electricity meters with a total risk score greater than the risk threshold, increasing the data collection frequency of the on-grid electricity meters.

7. The method for quantifying the root causes of meter failures by integrating multi-dimensional risk characteristics according to claim 1, characterized in that, In step S1, when obtaining historical fault data of the smart meter, the specific steps are to obtain historical fault data of the smart meter for different time periods according to a specified time interval; after calculating the comprehensive risk weight of each fault element by quantifying the comprehensive risk weight, the steps also include: if there are fault elements whose comprehensive risk weight continues to increase over time, the smart meter is replaced, upgraded or modified.

8. A root cause quantification system for electricity meter failures that integrates multi-dimensional risk characteristics, characterized in that, The device includes a processor and a computer-readable storage medium storing a computer program, which is executed by the processor to implement the steps of the meter fault root cause quantification method according to any one of claims 1 to 7, which integrates multi-dimensional risk features.

Images