Method and system for predicting and evaluating life of electric energy meter based on multi-source data fusion

Abstract

The application discloses a kind of based on multi-source data fusion's electric energy meter life prediction and evaluation method and system, collect the static attribute data of electric energy meter, dynamic time series data and event record data, after fusion processing, degradation feature extraction and supervision label construction, establish the mixed supervision learning model including degradation state identification submodel, residual life prediction submodel and self-adapting integrated module, and according to this realization electric energy meter degradation state evaluation, residual service life prediction, reliability index calculation and graded early warning output.The application can improve electric energy meter life prediction accuracy and reliability evaluation level, provide basis for predictive maintenance and operation decision.

Description

Technical Field

[0001] This invention relates to the field of electricity meters, and specifically to a method and system for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion. Background Technology

[0002] As a crucial infrastructure device in the power system for electricity metering, trade settlement, and electricity consumption information collection, the reliability and accuracy of electricity meters directly impact the economic interests of both power suppliers and consumers, as well as the authenticity of power grid operation data. With the continuous advancement of smart grid construction, the number of electricity meters in use is constantly increasing, their coverage is expanding, and their operating environment is becoming increasingly complex. Different models, batches, and installation environments of electricity meters exhibit significant differences in performance degradation rates and failure risks over long-term operation. Therefore, how to scientifically assess the lifespan and reliability level of electricity meters has become a key issue in power equipment operation and maintenance and asset management.

[0003] In existing technologies, the operation and maintenance management of electricity meters typically employs two main modes: periodic rotation and replacement after failure. The periodic rotation mode usually involves replacing meters in batches that have reached their specified service life, based on relevant verification procedures or service life requirements, without considering individual differences in actual operation. While this method facilitates unified management, it can easily lead to the premature replacement of a large number of meters still in normal working condition and whose performance has not yet significantly degraded, resulting in wasted equipment resources and increased operation and maintenance costs. The replacement after failure mode addresses issues only after a meter experiences communication abnormalities, metering inaccuracies, complete damage, or fails random inspections. While this mode avoids some of the problems associated with premature replacement, it has a significant lag, making it difficult to detect the slow degradation process of meters before failure, which may lead to long-term metering deviations, distorted line loss statistics, and trade settlement disputes.

[0004] To improve the scientific nature of electricity meter operation and maintenance management, some condition monitoring-based analysis methods have been proposed in existing technologies. For example, some solutions set fixed thresholds for single indicators such as battery voltage, metering error, and communication success rate, issuing an alarm signal when the detected value exceeds the preset threshold. While these methods are simple to implement, they typically rely on only one or a few indicators, making it difficult to comprehensively reflect the overall health status of the electricity meter. Furthermore, they focus more on whether static values ​​exceed limits at a specific moment, failing to fully reflect the changing trends of the meter's performance over long-term operation, thus hindering the effective identification of early signs of degradation. In addition, setting thresholds too strictly can easily lead to false alarms, while setting them too leniently may result in missed alarms, making it difficult to balance sensitivity and accuracy in practical applications.

[0005] Other technical solutions are based on traditional reliability statistical models, such as fitting lifetime models like the Weibull distribution to historical failure data to calculate the mean time to failure (MTBF) or group failure patterns. While these methods can reflect the overall lifetime characteristics of a type of equipment at the group level, they are essentially statistical analyses focused on the equipment group and are difficult to perform individual lifetime predictions for the specific operating conditions of a single operating electricity meter. Especially given the high overall reliability of modern electricity meters and the long accumulation period of complete failure samples, relying solely on historical failure data for modeling often fails to produce lifetime assessment results with sufficient guiding value for individual equipment. Furthermore, these methods typically fail to fully utilize the metering, electrical, environmental, and event information continuously generated during electricity meter operation, resulting in low data utilization and limited precision in the prediction results.

[0006] With the development of data-driven methods, some technologies have attempted to use machine learning algorithms such as neural networks and support vector machines to predict the operating status or remaining lifespan of electricity meters at a future point in time. However, existing solutions generally suffer from the following shortcomings: First, the input features are mostly simple splicing of raw monitoring data or limited manually selected features, lacking deep feature construction that can reveal the mechanism of equipment degradation, making it difficult for the model to fully extract effective information reflecting the performance degradation process; Second, many solutions only model for a single task, such as only determining whether a fault exists, or only outputting a remaining lifespan estimate, failing to organically combine degradation status identification, lifespan prediction, and reliability assessment, resulting in a lack of synergy between results and hindering the formation of a complete basis for operation and maintenance decisions; Third, most existing models only output a single point prediction result, lacking a quantitative description of prediction uncertainty, making it difficult for operation and maintenance personnel to assess risk levels and formulate differentiated maintenance strategies; Fourth, for a large number of in-service electricity meters that have not yet failed but have accumulated a large amount of operating data, existing technologies are insufficient in utilizing this type of censored data, which can easily cause training bias and affect the model's generalization ability and actual prediction effect.

[0007] Furthermore, existing technologies often struggle to integrate and analyze the static attributes, dynamic time-series data, and historical event records of electricity meters. For instance, different production batches, installation locations, load levels, and environmental temperature and humidity conditions all affect the aging rate of electricity meters. However, traditional methods often fail to collaboratively model these data from different sources, with varying structures and time series, and struggle to establish a correlation between multi-dimensional data and the device's degradation state. Consequently, existing technologies often only provide rough risk assessments in practical applications, failing to achieve accurate lifespan predictions and dynamic reliability evaluations for individual electricity meters.

[0008] Therefore, existing technologies still lack a lifespan prediction and reliability assessment method and system that can be applied to individual electricity meters, integrate multi-source heterogeneous data, take into account both degradation state identification and remaining lifespan prediction, and further output reliability indicators and early warning decision suggestions. How to build a unified assessment framework that can adapt to the complex operating scenarios of electricity meters, based on fully utilizing historical operating data, status data, and event data, and realize the transformation of electricity meters from periodic maintenance to predictive maintenance, remains a pressing technical problem to be solved in this field. Summary of the Invention

[0009] To address the aforementioned problems, this invention provides a method and system for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion, which can effectively overcome the shortcomings of existing technologies.

[0010] This invention is achieved through the following technical solution: a method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion, comprising the following steps: S1. Collect multi-source heterogeneous data from the target electricity meter and fuse the multi-source heterogeneous data to obtain a raw dataset for lifetime prediction and reliability assessment; the multi-source heterogeneous data includes at least static attribute data, dynamic time series data and event record data. S2. The original dataset is preprocessed, degradation features are extracted, and supervised labels are constructed to obtain the model input feature set and the training label set; the training label set includes at least a classification label representing the current degradation stage of the target energy meter and a label representing the continuous value of the remaining service life of the target energy meter from the current time to the time of failure. S3. Construct and train a hybrid supervised learning model based on the model input feature set and training label set. The hybrid supervised learning model includes at least a degradation state identification sub-model, a remaining life prediction sub-model, and an adaptive fusion module. The degradation state identification sub-model is used to output the probability that the target energy meter belongs to each predefined degradation state. The remaining life prediction sub-model is used to output the remaining life point estimation result and corresponding probability distribution of the target energy meter. The adaptive fusion module is used to dynamically weight and fuse the prediction results based on the uncertainty of the output results of the two sub-models to obtain the degradation state assessment result and the remaining life prediction result of the target energy meter. S4. Based on the remaining service life prediction results and their probability distribution, perform online dynamic reliability assessment on the target energy meter to obtain reliability indicators that change over time, and output graded early warning information and maintenance priority suggestions when the remaining service life is lower than a preset threshold and / or the target energy meter enters a preset high-risk degradation stage. S5. Collect newly added running data and newly added failure data, and perform incremental learning and / or periodic retraining on the hybrid supervised learning model to achieve continuous optimization and updating of the model.

[0011] As a preferred technical solution, the static attribute data includes one or more of the following: production model, production batch, initial calibration value, and installation environment classification; the dynamic time series data includes one or more of the following: measurement error value, voltage, current, temperature, battery voltage, and communication success rate; and the event record data includes one or more of the following: on-site inspection records, fault records, and maintenance and replacement records.

[0012] As a preferred technical solution, when extracting degradation features from the original dataset in step S2, at least the following should be included: Statistical features are extracted based on a sliding time window. Extract trend features from key performance indicators; Construct a composite degradation index based on the degradation mechanism of electricity meters and domain knowledge; The classification labels and the continuous value labels of the remaining service life are constructed based on historical failure time points and / or preset performance thresholds.

[0013] As a preferred technical solution, the degradation state identification sub-model is constructed using a gradient boosting decision tree model or a deep learning classification model, and is used to output the probability distribution of the target energy meter currently belonging to multiple predefined degradation states based on the input feature set of the model.

[0014] The remaining life prediction sub-model is constructed using a hybrid architecture that combines a survival analysis model and a time series neural network. The survival analysis model is used to characterize the basic shape of the life distribution, while the time series neural network is used to learn degradation patterns from dynamic time series data. Together, they output point estimates and probability distributions of the remaining life.

[0015] As a preferred technical solution, the time series neural network is a long short-term memory network, a long short-term memory network with attention mechanism, or other recurrent neural networks capable of handling time series dependencies; the survival analysis model is an accelerated failure time model, whose distribution parameters are dynamically modulated by the hidden layer output of the time series neural network.

[0016] As a preferred technical solution, the adaptive integration module dynamically weights and fuses the output results of the two sub-models based on the uncertainty of the state probability output by the degradation state identification sub-model and the uncertainty of the lifetime probability distribution output by the remaining lifetime prediction sub-model, so as to form the final degradation state assessment result and the remaining lifetime prediction result.

[0017] As a preferred technical solution, the reliability indicators in step S4 include at least a reliability function, a failure rate function, and a remaining useful life confidence interval at a specified confidence level; the graded early warning information includes at least risk level information generated based on the remaining useful life prediction results and the degradation stage assessment results.

[0018] As a preferred technical solution, the method is deployed in a cloud-edge collaborative execution mode, wherein data acquisition and preliminary fusion processing are performed on intelligent fusion terminals or edge terminals, model training, model management and service deployment are performed on the cloud platform, and online prediction and reliability assessment are performed on the cloud platform and / or edge terminals.

[0019] The present invention provides a system for predicting the lifespan and assessing the reliability of electricity meters based on multi-source data fusion and supervised learning algorithms, comprising: The data acquisition and fusion module is used to acquire multi-source heterogeneous data from the target electricity meter and perform data cleaning, format conversion, time synchronization and fusion processing. The data communication and storage module is used to transmit and store the merged data to a database or big data platform. The model training and service module is used to train a hybrid supervised learning model based on the training dataset and encapsulate the trained model as a prediction service. An online prediction and evaluation engine is used to call the prediction service to identify the degradation status of the target energy meter, predict its remaining service life, and evaluate its reliability. The visualization and decision support module is used to output individual lifespan prediction results, group reliability assessment results, graded early warning information, and maintenance decision recommendations.

[0020] The beneficial effects of this invention are as follows: By constructing a multi-source data fusion mechanism, this invention unifies the collection, cleaning, alignment, and fusion processing of static attribute data, dynamic time series data, and event record data of electricity meters. This breaks through the limitations of existing technologies that rely solely on a single indicator or a small number of monitoring parameters for analysis. It enables both internal and external factors affecting the lifespan and reliability of electricity meters to be included in the evaluation system, thereby significantly improving the comprehensiveness and accuracy of lifespan prediction and reliability assessment.

[0021] This invention, by setting up degradation feature engineering and status label construction steps, can not only extract sliding window statistical features, trend features, and composite degradation indicators from raw operating data, but also simultaneously construct degradation stage classification labels and continuous value labels for remaining service life. This enables the model to both identify the current degradation state of the electricity meter and predict its future remaining life, achieving a continuous characterization of the electricity meter degradation process and overcoming the shortcomings of existing technologies that can only perform single fault judgment or single life estimation.

[0022] This invention constructs a hybrid supervised learning model framework that organically combines a degradation state identification sub-model with a remaining life prediction sub-model. It also dynamically fuses the two types of prediction results through an adaptive fusion module, creating a complementary and supportive relationship between degradation state identification and life prediction. This not only improves the model's ability to identify complex degradation patterns but also enhances its ability to predict the life evolution trend of individual electricity meters, thereby realizing the transformation from traditional group statistical analysis to individualized and accurate assessment.

[0023] This invention introduces a lifetime probability distribution and uncertainty quantification mechanism into the remaining lifetime prediction process, so that the output results include not only the estimated remaining lifetime point value, but also the corresponding probability distribution, confidence interval, reliability function, failure rate function and other reliability indicators. Compared with the existing technology that only outputs a single predicted value, it can provide maintenance personnel with richer and more robust decision-making basis, facilitate the formulation of differentiated maintenance strategies according to risk level, and improve the engineering availability of prediction results.

[0024] This invention can effectively utilize a large amount of operational data from in-service but not yet failed electricity meters. By combining survival analysis with a supervised learning model, it makes reasonable use of censored data, thereby avoiding the training bias problem caused by insufficient utilization of non-failed samples in traditional lifetime prediction methods. Even with relatively limited failed samples, a relatively robust lifetime prediction model can still be established, which is beneficial to improving the generalization performance and actual deployment effect of the model.

[0025] This invention establishes an online reliability dynamic assessment and graded early warning mechanism, which can output corresponding risk level information and maintenance priority suggestions in a timely manner based on the remaining life prediction results, degradation status assessment results and reliability indicators of the target electricity meter. This transforms the operation and maintenance work from the traditional periodic inspection and post-fault handling mode to a predictive maintenance mode. It can not only identify high-risk equipment in advance, but also reduce metering inaccuracies, passive operation and maintenance and economic losses caused by sudden failures.

[0026] This invention can be deployed by combining intelligent fusion terminals, cloud platforms and edge computing terminals. It can make full use of the data access capabilities and computing resources of the existing power system. It can be implemented without large-scale hardware modifications to the on-site electricity meters. It has the advantages of low deployment cost, strong adaptability and good real-time performance, and is easy to promote and apply in large-scale electricity meter asset operation and maintenance scenarios. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0029] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0030] Any feature disclosed in this specification (including any appended claims, abstract, and drawings) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0031] like Figure 1 As shown, this invention presents a method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion. This method is designed to meet the individualized lifespan management and reliability analysis needs of in-operation electricity meters. It is applicable to smart grids, electricity consumption information collection systems, and electricity metering equipment status monitoring scenarios, enabling continuous tracking, degradation identification, lifespan prediction, and maintenance decision support for electricity meter operation. By uniformly collecting, fusing, modeling, and evaluating multi-dimensional data generated by the electricity meter during operation, this method achieves a comprehensive judgment of the target electricity meter's current health status, remaining lifespan, and risk level. This transforms the traditional maintenance model, which relies on periodic rotation or replacement after failure, into a predictive maintenance model based on individual status.

[0032] Specifically, in this embodiment, multi-source heterogeneous data acquisition and fusion processing is first performed. The system establishes a multi-source data access link around the target energy meter, and uniformly collects data from different sources, structures, and sampling frequencies. The collected multi-source heterogeneous data includes at least static attribute data, dynamic time series data, and event log data. Among them, static attribute data is mainly used to characterize the inherent attributes of the target energy meter and installation background information, such as production model, production batch, initial calibration value, and installation environment classification. The production model and production batch can reflect the differences in equipment design generation, manufacturing process, and component consistency. The initial calibration value can reflect the baseline performance state of the equipment at the time of leaving the factory. The installation environment classification can reflect the external stress conditions of the energy meter's long-term service scenarios, such as high temperature, high humidity, enclosed distribution box, outdoor exposure, or well-ventilated environment. Dynamic time series data is used to characterize the continuous state change information of the target energy meter during operation, preferably including periodically acquired parameters such as metering error value, voltage, current, temperature, battery voltage, and communication success rate. Among them, the metering error value directly reflects the stability of the metering performance; voltage and current parameters reflect load fluctuations and power supply conditions; temperature parameters reflect the level of thermal stress on the equipment; battery voltage reflects the attenuation of auxiliary power supply capacity; and communication success rate reflects the stability of the meter's communication link and internal operation. Event log data is used to characterize important operational and maintenance events that occur during the target meter's lifecycle, preferably including on-site inspection records, fault records, and maintenance and replacement records. On-site inspection records can reflect the results of previous verifications or spot checks; fault records can reflect events such as abnormal codes, communication interruptions, abnormal metering, power outages, and system crashes; and maintenance and replacement records are used to characterize historical actions such as equipment repair, parts replacement, and complete meter rotation.

[0033] In actual implementation, to ensure the accuracy of subsequent model training and predictive analysis, the system performs fusion processing on the collected multi-source heterogeneous data. This fusion processing is not a simple splicing, but rather an integrated preprocessing approach addressing the format differences, time granularity differences, and semantic differences between different data sources. Specifically, a data mapping relationship is first established based on the unique identifier of the electricity meter, associating corresponding data from different systems; then, dynamic time series data is aligned according to a unified time benchmark, and data with inconsistent sampling periods are resampled, interpolated, or aggregated to form a unified time step; for missing values, methods such as mean imputation, median imputation, nearest-time interpolation, model estimation completion, or missing value encoding are used depending on the data characteristics; for outliers, threshold rules, box plotting methods, or time series mutation detection methods are used for identification and correction; for categorical static attributes, one-hot encoding, embedding encoding, or hierarchical mapping are performed; for event record data, event occurrence time, event type, event frequency, and event severity are transformed into structured input features. After the above processing, a raw dataset for lifetime prediction and reliability assessment is formed, thus providing a unified data foundation for subsequent degradation feature construction and supervised modeling.

[0034] After constructing the original dataset, this implementation further preprocesses, extracts degradation features, and constructs supervised labels to generate a model input feature set and a training label set. The core of this step lies in extracting effective features from multi-source raw data that reflect the performance degradation patterns of electricity meters, and establishing classification labels and continuous value labels that support the training of a hybrid supervised learning model. Specifically, for dynamic time series data, a sliding time window approach is preferred for segmentation. For example, data from several consecutive months can be extracted in a window format, or model input samples can be constructed according to fixed-length time slices. For each sliding time window, statistical features can be extracted, including mean, maximum, minimum, standard deviation, variance, range, skewness, kurtosis, quantiles, and fluctuation coefficient, to reflect the overall distribution and dispersion of key operating indicators within the window. Furthermore, trend features are extracted from key performance indicators, such as the slope, acceleration, year-on-year change rate, and consistency of trend direction of changes in sequences of metering error values, battery voltage, communication success rate, and temperature, to characterize whether the performance changes of the target energy meter tend to stabilize, continuously deteriorate, or fluctuate in the short term. To enhance the model's understanding of degradation mechanisms, composite degradation indicators can be constructed based on domain knowledge. For example, the error drift rate can be coupled with ambient temperature to form a temperature drift-related degradation indicator, the communication success rate can be coupled with the frequency of fault records to form a communication health indicator, and voltage and current fluctuations can be coupled with temperature rise characteristics to form a load stress indicator, thereby improving the model's ability to characterize degradation behavior under the coupled effects of multiple factors.

[0035] Meanwhile, in this embodiment, supervisory labels are constructed based on historical data. The training label set includes at least a classification label representing the current degradation stage of the target energy meter and a continuous value label representing the remaining service life of the target energy meter from the current time to the failure time. For the construction of classification labels, the equipment status can be divided into several predefined degradation stages such as healthy, mild degradation, moderate degradation, severe degradation, and high-risk criticality, based on the historical failure time points, verification results, error thresholds, abnormal event density, or expert annotation rules of the target energy meter. This allows the classification model to learn the feature patterns corresponding to different degradation states. For the construction of continuous value labels for remaining service life, the time difference between the current sample time of the equipment and the actual failure time, retirement time, or verification failure time can be defined to obtain the remaining service life label in days, weeks, or months. For in-service equipment that has not yet failed but has accumulated a long period of operating data, it can be introduced into the modeling process as censored samples and processed in the subsequent life prediction model using survival analysis, thereby avoiding the waste of information caused by simply removing such samples.

[0036] After obtaining the model input feature set and training label set, this implementation constructs and trains a hybrid supervised learning model. The hybrid supervised learning model includes at least a degradation state identification sub-model, a remaining life prediction sub-model, and an adaptive ensemble module. The degradation state identification sub-model is mainly used to identify the current degradation stage of the target energy meter. Its input is the aforementioned model input feature set, and its output is the probability distribution of the target energy meter belonging to each predefined degradation state. This degradation state identification sub-model can be constructed using a gradient boosting decision tree model or a deep learning classification model. When using a gradient boosting decision tree model, its strong tolerance for nonlinear feature relationships, missing values, and feature importance analysis capabilities can be utilized to achieve efficient classification of static features, statistical features, and composite indicators. When using a deep learning classification model, multi-layer fully connected networks, sequence coding networks, or attention mechanisms can be further introduced to enhance the extraction capability of complex time-series features and implicit degradation patterns. Through the degradation state identification sub-model, the system can output the probability that the target energy meter is in a healthy, early warning, or high-risk degradation stage, providing a status basis for subsequent operation and maintenance strategies.

[0037] A residual lifetime prediction sub-model, set up in parallel with the degradation state identification sub-model, outputs the point estimate of the residual lifetime of the target energy meter and its corresponding probability distribution. This sub-model is constructed using a hybrid architecture combining a survival analysis model and a time-series neural network. The survival analysis model characterizes the basic shape of the lifetime distribution, handles censored data issues in lifetime prediction, and is suitable for learning the time distribution patterns of equipment from normal operation to failure. The time-series neural network automatically learns complex degradation patterns and long-term dependencies from dynamic time-series data, thereby extracting temporal evolution information that is difficult to fully express using traditional manual features. Specifically, the time-series neural network is preferably a long short-term memory network, a long short-term memory network with an attention mechanism, or other recurrent neural networks capable of handling time-series dependencies. The long short-term memory unit effectively preserves long-term degradation trend information, avoiding the gradient vanishing problem in long-sequence learning of ordinary neural networks. By introducing an attention mechanism, higher weights can be assigned to time segments and feature dimensions that have a greater impact on lifetime prediction, thereby improving the model's response to key degradation signals. The survival analysis model is preferably an accelerated failure time model, whose parameters are dynamically modulated by the hidden layer output of a time-series neural network. That is, the neural network automatically learns and outputs lifetime distribution parameters based on the input sequence, making the distribution parameters no longer fixed but adaptively adjustable for different individual devices, different time periods, and different degradation modes. In this way, the remaining lifetime prediction sub-model can not only output a point estimate of the remaining lifetime at a given moment but also the corresponding probability distribution or confidence interval, thereby giving the lifetime prediction results a stronger risk interpretation capability.

[0038] In this embodiment, the outputs of the degradation state identification sub-model and the remaining useful life prediction sub-model are not used independently, but are further input into the adaptive integration module for fusion. The adaptive integration module dynamically weights and fuses the outputs of the two sub-models based on the uncertainty of the state probability output from the degradation state identification sub-model and the uncertainty of the lifetime probability distribution output from the remaining useful life prediction sub-model, to form the final degradation state assessment result and the remaining useful life prediction result. Specifically, when the degradation state identification sub-model has a highly concentrated probability of determining a certain state and a clear classification boundary, its weight in the final assessment can be increased; when the remaining useful life prediction sub-model has a small lifetime distribution variance and a narrow confidence interval, it indicates high prediction stability, and its output weight in the integrated result can be increased. Conversely, when the uncertainty of a sub-model is large, its weight is reduced accordingly to reduce the impact of a single model error on the final result. Through this uncertainty-based dynamic fusion method, the system can organically combine discrete state identification capabilities with continuous lifetime prediction capabilities, making the final output both interpretable of the current state and quantifiable of future lifetime, thereby improving the comprehensive assessment accuracy and the reliability of engineering applications.

[0039] After completing model training, this implementation uses the hybrid supervised learning model for online dynamic reliability assessment. Specifically, after the system obtains the latest operating data of the target electricity meter, it first generates online input features according to the aforementioned fusion processing and feature construction process, and then calls the trained hybrid supervised learning model for inference to obtain the degradation state assessment result and remaining service life prediction result of the target electricity meter at the current moment. Based on this, and based on the remaining service life prediction result and its probability distribution, reliability indicators that change over time are further calculated. The reliability indicators include at least a reliability function, a failure rate function, and a remaining service life confidence interval at a specified confidence level. The reliability function describes the probability that the equipment will maintain normal operation before a certain point in the future, the failure rate function describes the intensity of the failure risk of the equipment at a certain moment, and the remaining service life confidence interval characterizes the interval credibility of the prediction result. By comprehensively analyzing these reliability indicators, the system can not only determine whether the target electricity meter is close to a high-risk state, but also quantify its risk level and time urgency.

[0040] Furthermore, when the predicted remaining useful life is lower than a preset threshold and / or the target energy meter enters a preset high-risk degradation stage, the system outputs tiered early warning information and maintenance priority recommendations. The tiered early warning information includes at least risk level information generated based on the remaining useful life prediction results and degradation stage assessment results. For example, the early warning levels can be divided into multiple levels such as normal, attention, yellow warning, orange warning, and red warning. Normal indicates low equipment operating risk; attention indicates a slight degradation trend; yellow warning indicates the need to increase inspection frequency; orange warning suggests inclusion in the recent maintenance plan; and red warning indicates that on-site verification, calibration, or replacement should be arranged within a limited time. Maintenance priority recommendations can be generated by combining equipment importance, installation location, load level, historical failure frequency, and group asset configuration requirements, enabling the operation and maintenance department to make prioritization decisions according to risk level and resource constraints. Thus, this implementation method realizes the transformation from model output to actual maintenance strategy recommendations, no longer merely remaining at the data analysis level, but directly supporting the execution of on-site operation and maintenance work.

[0041] To adapt to the actual deployment scenarios of large-scale electricity meter assets, this implementation method preferably adopts a cloud-edge collaborative execution approach for system implementation. Specifically, data acquisition and preliminary fusion processing are performed on intelligent fusion terminals or edge terminals, model training, model management, and service deployment are performed on the cloud platform, while online prediction and reliability assessment are performed on the cloud platform and / or edge terminals. Specifically, the data acquisition and fusion module deployed on the electricity meter field side or distribution area side can directly obtain raw data from the electricity meter itself, sensors, distribution automation systems, or electricity consumption information collection systems, and complete data cleaning, format conversion, time synchronization, and preliminary fusion locally to reduce the data processing pressure on the central platform and improve real-time performance. The data communication and storage module is responsible for securely transmitting the fused data to a database or big data platform and completing data indexing, historical archiving, and query support. The model training and service deployment module is deployed on the cloud platform to train a hybrid supervised learning model based on the training dataset, and performs parameter optimization, model validation, version management, and service encapsulation, enabling the trained model to provide external calling interfaces in the form of standardized prediction services. The online prediction and assessment engine is used to call the prediction service to identify the degradation status of target electricity meters, predict their remaining service life, and assess their reliability. For scenarios with high real-time requirements, some lightweight inference logic can be offloaded to edge terminals for execution to achieve low-latency local early warning. The visualization and decision support module is used to output individual life prediction results, group reliability assessment results, tiered early warning information, and maintenance decision suggestions. These are displayed to maintenance personnel through dashboards, lists, charts, risk maps, or reports, enabling them to intuitively grasp the health status of equipment, lifespan distribution, and a list of high-risk equipment.

[0042] In the actual operation of this implementation, the system can continuously receive new operational data and new failure data according to a predetermined cycle, and perform incremental learning and / or periodic retraining on the hybrid supervised learning model to achieve continuous model optimization and updates. Specifically, when the scale of new samples is small but representative, incremental learning can be performed, allowing the model parameters to quickly absorb new data features without complete retraining. When the cumulative number of new samples reaches a certain scale, or when the equipment degradation mode drifts significantly due to seasonal changes, environmental changes, or load structure changes, periodic retraining can be initiated to retrain the model on the entire training set or an expanded training set to avoid model aging and prediction bias. Through this continuous optimization mechanism, the model can be continuously updated as the operating status of the electricity meter group evolves, thereby maintaining long-term prediction performance and applicability.

[0043] For example, in a typical application scenario, a smart fusion terminal can continuously collect operational data, installation environment data, grid load data, and previous verification conclusions of smart meters in operation in a certain region. The final determination of each target meter as either continuing to be used or failing to meet the standards for retirement can be used as a source of supervisory labels. For each meter, multi-feature sequences from the past several consecutive months can be selected as input samples, and these sample sequences can be input into the hybrid supervised learning model for training or prediction. After receiving the data from the latest time window, the model outputs, on the one hand, the probability of the meter currently being in each degradation stage, and on the other hand, a point estimate of its remaining service life in months and the corresponding confidence interval. If the model determines that it has entered a high-risk degradation stage and the predicted remaining service life is lower than a set threshold, the system automatically generates a high-level warning and includes the meter in the priority replacement list; if the model determines that it is still in a healthy or mildly degraded state, even if its service life is close to the normal rotation period, it can be recommended to extend the service life, thereby reducing unnecessary replacements. In this way, the system can identify potentially failing equipment in advance, reduce the risk of sudden failures and inaccurate measurement, improve equipment utilization efficiency, and reduce overall operation and maintenance costs.

[0044] The method and system for predicting and assessing the lifespan of electricity meters based on multi-source data fusion provided in this embodiment do not simply make threshold judgments on a single state indicator, nor do they perform isolated regression on lifespan. Instead, they establish an integrated closed-loop architecture that encompasses multi-source data access, degradation feature extraction, joint classification and regression modeling, uncertainty integration, reliability function calculation, hierarchical early warning output, and continuous model updates. This enables dynamic, continuous, and quantifiable assessment of the individual lifespan and reliability status of electricity meters. This solution can more realistically reflect the individual degradation differences of electricity meters under different batches, environments, and load conditions. It has advantages such as high prediction accuracy, strong reliability, good interpretability, strong deployment adaptability, and outstanding support capabilities for operation and maintenance decisions. It can be widely applied to scenarios such as smart meter lifespan management, equipment condition evaluation, and lean operation and maintenance of power grid assets.

[0045] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions conceived without inventive effort should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data from the target electricity meter and fuse the multi-source heterogeneous data to obtain a raw dataset for lifetime prediction and reliability assessment; the multi-source heterogeneous data includes at least static attribute data, dynamic time series data and event record data. S2. The original dataset is preprocessed, degradation features are extracted, and supervised labels are constructed to obtain the model input feature set and the training label set; the training label set includes at least a classification label representing the current degradation stage of the target energy meter and a label representing the continuous value of the remaining service life of the target energy meter from the current time to the failure time. S3. Construct and train a hybrid supervised learning model based on the model input feature set and training label set. The hybrid supervised learning model includes at least a degradation state identification sub-model, a remaining life prediction sub-model, and an adaptive fusion module. The degradation state identification sub-model is used to output the probability that the target energy meter belongs to each predefined degradation state. The remaining life prediction sub-model is used to output the remaining life point estimation result and corresponding probability distribution of the target energy meter. The adaptive fusion module is used to dynamically weight and fuse the prediction results based on the uncertainty of the output results of the two sub-models to obtain the degradation state assessment result and the remaining life prediction result of the target energy meter. S4. Based on the remaining service life prediction results and their probability distribution, perform online dynamic reliability assessment on the target energy meter to obtain reliability indicators that change over time, and output graded early warning information and maintenance priority suggestions when the remaining service life is lower than a preset threshold and / or the target energy meter enters a preset high-risk degradation stage. S5. Collect newly added running data and newly added failure data, and perform incremental learning and / or periodic retraining on the hybrid supervised learning model to achieve continuous optimization and updating of the model.

2. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The static attribute data includes one or more of the following: production model, production batch, initial calibration value, and installation environment classification; the dynamic time series data includes one or more of the following: measurement error value, voltage, current, temperature, battery voltage, and communication success rate; the event record data includes one or more of the following: on-site inspection records, fault records, and maintenance and replacement records.

3. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: Step S2, when extracting degenerate features from the original dataset, includes at least the following: Statistical features are extracted based on a sliding time window. Extract trend features from key performance indicators; Construct a composite degradation index based on the degradation mechanism of electricity meters and domain knowledge; The classification labels and the continuous value labels of the remaining service life are constructed based on historical failure time points and / or preset performance thresholds.

4. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The degradation state identification sub-model is constructed using a gradient boosting decision tree model or a deep learning classification model, and is used to output the probability distribution of the target energy meter currently belonging to multiple predefined degradation states based on the input feature set of the model.

5. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The remaining life prediction sub-model is constructed using a hybrid architecture that combines a survival analysis model and a time series neural network. The survival analysis model is used to characterize the basic shape of the life distribution, while the time series neural network is used to learn degradation patterns from dynamic time series data. Together, they output point estimates and probability distributions of the remaining life.

6. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 5, characterized in that: The time-series neural network is a long short-term memory network, a long short-term memory network with attention mechanism, or other recurrent neural networks capable of handling time-series dependencies; the survival analysis model is an accelerated failure time model, whose distribution parameters are dynamically modulated by the hidden layer output of the time-series neural network.

7. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The adaptive integration module dynamically weights and fuses the outputs of the two sub-models based on the uncertainty of the state probability output by the degradation state identification sub-model and the uncertainty of the lifetime probability distribution output by the remaining lifetime prediction sub-model, in order to form the final degradation state assessment result and the remaining lifetime prediction result.

8. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The reliability indicators mentioned in step S4 include at least a reliability function, a failure rate function, and a remaining useful life confidence interval at a specified confidence level; the graded early warning information includes at least risk level information generated based on the remaining useful life prediction results and the degradation stage assessment results.

9. The method for predicting and evaluating the lifespan of electricity meters based on multi-source data fusion according to claim 1, characterized in that: The method is deployed in a cloud-edge collaborative execution mode, wherein data acquisition and preliminary fusion processing are performed on intelligent fusion terminals or edge terminals, model training, model management and service deployment are performed on the cloud platform, and online prediction and reliability assessment are performed on the cloud platform and / or edge terminals.

10. A system for predicting the lifespan and assessing the reliability of electricity meters based on multi-source data fusion and supervised learning algorithms, characterized in that, include: The data acquisition and fusion module is used to acquire multi-source heterogeneous data from the target electricity meter and perform data cleaning, format conversion, time synchronization and fusion processing. The data communication and storage module is used to transmit and store the merged data to a database or big data platform. The model training and service module is used to train a hybrid supervised learning model based on the training dataset and encapsulate the trained model as a prediction service. An online prediction and evaluation engine is used to call the prediction service to identify the degradation status of the target energy meter, predict its remaining service life, and evaluate its reliability. The visualization and decision support module is used to output individual lifespan prediction results, group reliability assessment results, graded early warning information, and maintenance decision recommendations. The online prediction and evaluation engine performs the method as described in any one of claims 1 to 9.

Images