A TMR data quality intelligent diagnosis and self-healing system
By leveraging a cloud-edge-device collaborative architecture and federated learning optimization, combined with the power business rules and lightweight models of edge nodes, the real-time, accuracy, and security issues of TMR data quality diagnosis were resolved, achieving efficient data quality governance and self-healing processing.
Patent Information
- Application Number
- CN202611354546.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-17
AI Technical Summary
Existing TMR data quality diagnostic technologies cannot simultaneously meet the requirements of real-time performance, accuracy, and data security. The self-healing strategies lack refinement and cross-regional adaptability, resulting in low efficiency in data quality governance.
A cloud-edge-device collaborative architecture is adopted, which combines the power business rules and lightweight models of edge nodes to perform data quality diagnosis. The model is aggregated and optimized through federated learning to achieve hierarchical self-healing processing.
It improves the efficiency of rapid detection and processing of TMR data quality anomalies, reduces diagnostic response time, improves the accuracy of anomaly identification and cross-regional model adaptability, and enhances data security and system adaptability.
Smart Images

Figure CN123115117A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system automation technology, specifically to a TMR data quality intelligent diagnosis and self-healing system. Background Technology
[0002] In the large-scale application of current electricity consumption data acquisition (TMR) systems in power systems, inconsistent data quality has become a core issue restricting electricity metering, line loss management, and load monitoring. Existing data quality diagnostic technologies are mainly divided into two categories: centralized diagnostic architecture and single edge model diagnostic architecture. Both of these structures have obvious defects and significant shortcomings in self-healing processing and model optimization.
[0003] In a centralized diagnostic architecture, all terminal data is uploaded to the main station for quality diagnosis and processing. The massive influx of terminal data into the main station at the same time causes network congestion and generally results in long diagnostic response times, which cannot meet the real-time monitoring needs of power sites. At the same time, cross-domain transmission of raw power and user data increases the risk of data leakage.
[0004] In a single edge model diagnostic architecture, although the industry has tried to deploy diagnostic models at edge nodes to improve real-time performance, the model design is divorced from the specific business rules and domain knowledge of the power system. For anomalies such as meter stagnation that can be quickly identified by deterministic rules, complex model reasoning is still relied upon, which leads to increased diagnostic latency and high false alarm rate. Moreover, the heterogeneity of electricity consumption characteristics in different distribution areas is not considered, and the diagnostic performance of the unified model varies greatly in residential / industrial / commercial / rural distribution areas.
[0005] In terms of handling anomalous data self-healing, existing methods typically perform interpolation compensation or data backtracking uniformly on all diagnosed anomalies, which cannot distinguish the confidence level of the anomalies. This results in a large number of low-confidence anomalies being incorrectly repaired, which in turn introduces new data quality problems.
[0006] In terms of collaborative optimization, existing methods lack an effective data interaction channel between the main station and edge nodes, making it impossible to continuously optimize the diagnostic model using the field data from each edge node. Furthermore, they fail to incorporate the results of manual on-site verification into the model / rule optimization system, resulting in a failure to continuously improve the system's intelligence level and making it difficult to adapt to the dynamic changes in power field business scenarios.
[0007] In summary, existing technologies cannot simultaneously meet the requirements of real-time performance, accuracy, and data security for TMR data diagnosis. Furthermore, the level of sophistication of self-healing strategies and the cross-regional adaptability of models are insufficient, which have become technical bottlenecks for the large-scale deployment of TMR systems. Summary of the Invention
[0008] The present invention aims to solve the problems existing in the prior art and provide a TMR data quality intelligent diagnosis and self-healing system. Under the existing power consumption information collection system architecture, it realizes intelligent diagnosis and self-healing of TMR data quality, and improves the efficiency of data quality governance and the reliability of system operation without adding a large number of hardware devices.
[0009] To achieve the above objectives, the present invention adopts the following solution:
[0010] A TMR data quality intelligent diagnosis and self-healing system, the system comprising:
[0011] The terminal acquisition layer collects electricity meter data and electricity parameter data;
[0012] The edge diagnostic layer includes multiple transformer area edge nodes. Each transformer area edge node is configured with a rule diagnostic module, a lightweight model inference module, and a self-healing module. The rule diagnostic module performs data quality diagnosis based on preset business rules, and the lightweight model inference module calls locally stored lightweight models to perform data quality diagnosis. The self-healing module performs graded self-healing based on the anomaly identification results of the diagnostic module and the business level of the anomaly type.
[0013] The main site aggregation layer includes a federated learning coordination module and a dynamic aggregation module deployed on the main site server. The federated learning coordination module receives local model parameters and diagnostic performance statistics uploaded by each edge node, calls the dynamic aggregation module to determine the aggregation weight of each edge node based on the diagnostic performance statistics, aggregates the local model parameters based on the aggregation weight to obtain global model parameters, and sends them to each edge node for updates.
[0014] The business application layer is used to receive abnormal task information sent by the edge nodes and / or the main station aggregation layer, and to provide a manual confirmation interface for the abnormal task information.
[0015] In some embodiments of the present invention, the edge diagnostic layer is configured with a power business rule base, and the rule diagnostic module calls the power business rule base to perform rule diagnostics based on the preset business rules in the power business rule base. The power business rule base includes at least one of the following data diagnostic rules:
[0016] Electricity meter stagnation detection rule: When the electricity meter readings at three consecutive collection points remain unchanged, and the synchronous current is greater than the preset effective current threshold TH1, the first type of rule abnormal alarm is triggered.
[0017] Line loss sudden change detection rule: Calculate the line loss rate based on the power supply of the main meter of the transformer area and the power consumption of all sub-meters. When the line loss rate increases by more than a% within a preset time period t1 and lasts for more than a preset time period t2, trigger a second-class rule abnormal alarm.
[0018] Voltage and current over-limit detection rules: When the voltage parameter exceeds the preset upper or lower voltage threshold and the duration exceeds the preset time t3, the current parameter exceeds the preset current threshold and the duration exceeds the preset time t4, or the three-phase imbalance exceeds the preset threshold TH2 and the duration exceeds the preset time t5, three types of rule abnormal alarms are triggered.
[0019] Power balance detection rules: The total power of the transformer area is compared with the sum of the power of each sub-meter in real time. When the deviation between the total power of the transformer area and the sum of the power of each sub-meter exceeds the preset ratio b% and the duration exceeds the preset time t6, four types of rule abnormal alarms are triggered.
[0020] In some embodiments of the present invention, the line loss change detection rule further includes household-transformer relationship verification: when detecting the line loss rate of the transformer area, the household-transformer relationship information in the transformer area archive data is obtained simultaneously, the target users whose household-transformer relationship has changed are marked, and the electricity consumption data of the marked users is corrected, so as to redetermine the line loss status of the transformer area with the corrected user electricity consumption.
[0021] In some embodiments of the present invention, the lightweight model is constructed using a one-dimensional convolutional neural network. The input features of the lightweight model include daily load curve features, historical anomaly labels, and meteorological features. The output results include at least probability values corresponding to normal operation, meter flyaway anomaly, meter stagnation anomaly, line loss anomaly, voltage over-limit anomaly, current over-limit anomaly, and power imbalance anomaly, respectively.
[0022] In some embodiments of the present invention, the edge diagnostic layer performs hierarchical diagnostics on the collected data, including:
[0023] The first level of diagnosis involves performing a diagnosis on the collected data through the rule diagnosis module. If the identified anomaly type matches a preset deterministic rule, the rule diagnosis result is directly determined as the anomaly identification result without calling the lightweight model inference module for model inference.
[0024] The second level of diagnosis involves calling the lightweight model inference module to verify the rule diagnosis results when the anomaly type identified by the rule diagnosis module matches a preset fuzzy rule. If the rule diagnosis result is consistent with the model diagnosis result and the overall confidence level reaches a preset threshold, a confirmed anomaly is output; otherwise, it is marked as a suspected anomaly and added to the observation list. The overall confidence level is obtained by weighting the model diagnosis confidence level and the rule diagnosis confidence level.
[0025] Encrypted data collection and / or remote recall are initiated for the data in the observation list. Rule diagnosis and model inference are re-executed in the subsequent N diagnostic windows. When the comprehensive confidence level continuously reaches the corresponding judgment condition, the suspected anomaly is upgraded to a confirmed anomaly and sent to the self-healing processing module. When the rule diagnosis result returns to normal and the model confidence level is lower than the corresponding judgment threshold, the suspected anomaly is canceled. If the anomaly status is not determined after the preset observation period, a corresponding manual verification task is generated for subsequent push to the business application layer.
[0026] The third level of diagnosis is determined when none of the business rules in the power business rule base meet the abnormal triggering conditions, and the confidence level of the lightweight model's output for at least one preset abnormal type reaches the second judgment threshold. The corresponding abnormal identification result is then determined based on the output result of the lightweight model.
[0027] In some embodiments of the present invention, each edge node periodically uses local TMR data and anomaly labels to train a local model, and calculates performance indicators and statistical local data distribution characteristics based on local validation data to form diagnostic performance statistics.
[0028] The federated learning coordination module receives and verifies the model parameters and diagnostic performance statistics uploaded by each edge node, and sends the verified data to the dynamic aggregation module for weight calculation. The dynamic aggregation module calculates the aggregation weights based on the following formula:
[0029]
[0030]
[0031]
[0032]
[0033] In the formula, is the aggregation weight of the k-th edge node; K is the total number of edge nodes; , These are the performance index values of the k-th and j-th edge nodes, respectively; , These are the data heterogeneity adjustment factors for the k-th edge node; Let the data distribution entropy of the k-th edge node be denoted as . The proportion of outlier samples at the k-th edge node; The average proportion of outlier samples across all nodes participating in the aggregation. For indicator functions; Encode the station type for the k-th edge node; Encode the mainstream station types for all participating aggregation nodes; , , is a coefficient.
[0034] In some embodiments of the present invention, the confidence level of the diagnostic result is obtained based on the following method:
[0035] Obtain the confidence level of the rule diagnosis results output by the rule diagnosis module, and use it as the confidence level of the first-level diagnosis;
[0036] Obtain the confidence score of the model diagnosis results output by the lightweight model inference module, and use it as the confidence score of the third-level diagnosis;
[0037] The confidence score of the second-level diagnosis is based on the confidence score of the rule-based diagnostic results and the comprehensive confidence score calculated by the lightweight model inference module.
[0038] Based on the confidence levels of the first-level diagnosis, the third-level diagnosis, and the second-level diagnosis, the confidence levels are divided according to the preset confidence thresholds to obtain the initial confidence interval.
[0039] Obtain verified historical diagnostic samples, calculate the positive predictive value corresponding to different candidate confidence thresholds, and determine the boundary threshold of each confidence level based on the preset target positive predictive value of each interval, which is used to correct the boundary threshold of the initial confidence interval.
[0040] The boundary threshold is dynamically adjusted based on the updated status of historical diagnostic samples.
[0041] In some embodiments of the present invention, the confidence level of the rule diagnosis result is determined based on the rule matching strength and taken from a pre-constructed rule matching strength-confidence mapping table; the values in the mapping table are assigned based on the statistical results of on-site samples.
[0042] The confidence level of the model's diagnostic results is determined based on the statistical characteristics of multiple random forward inference results.
[0043] In some embodiments of the present invention, the self-healing processing module determines the self-healing level based on a preset mapping relationship according to the confidence level of the diagnostic results and the business level corresponding to the anomaly type, and executes the self-healing strategy corresponding to the self-healing level.
[0044] The self-healing strategy includes data interpolation compensation or data backtracking, remote recall and data repair, generating abnormal work orders and pushing them to the business application layer, and reporting to the main station for in-depth diagnosis and generating troubleshooting plans.
[0045] In some embodiments of the present invention, the self-healing processing module determines the selection priority of the self-healing strategy corresponding to different anomaly types based on historical self-healing operation data and through a multi-armed slot machine algorithm based on an ε-greedy strategy, and optimizes the selection priority according to the historical success rate of each healing action.
[0046] The present invention has the following beneficial effects:
[0047] (1) By combining edge-side business rules with lightweight models, the diagnostic method can achieve rapid detection and localized processing of TMR data quality anomalies, reduce diagnostic response time, and improve the accuracy of anomaly identification.
[0048] (2) The present invention constructs a dynamic aggregation mechanism based on federated learning, which adjusts the model aggregation weight according to the diagnostic performance and data characteristics of different transformer substations, reduces the fluctuation of model performance caused by the difference in data distribution in transformer substations, and improves the adaptability and diagnostic stability of the model in cross-transformer substation scenarios.
[0049] (3) The present invention establishes a graded self-healing mechanism based on the confidence level of anomaly diagnosis and the level of business impact. It matches the corresponding self-healing strategy for different anomaly types and confidence levels, avoids ineffective repair caused by using a single disposal method, improves the efficiency of anomaly handling, and reduces manual intervention.
[0050] (4) The invention adopts a federated learning architecture for collaborative model optimization, exchanging only model parameters and necessary statistical information, avoiding the centralized transmission of original business data, and improving data security; at the same time, it combines self-healing result feedback to realize continuous optimization of diagnostic models and strategy libraries, and improves the adaptability of the system in the long-term operation process. Attached Figure Description
[0051] Figure 1 This is a schematic diagram of the system architecture of the present invention.
[0052] Figure 2 This is the judgment process of the three-level diagnostic mechanism of the present invention.
[0053] Figure 3 This is a sequence diagram of federated learning. Detailed Implementation
[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0055] Example 1
[0056] The TMR data quality diagnostic system of this invention adopts a four-layer architecture of cloud-edge-device collaboration, such as... Figure 1 As shown, from bottom to top, the layers are: terminal acquisition layer, edge diagnostic layer, main station aggregation layer, and business application layer.
[0057] Terminal acquisition layer: consists of smart meters and acquisition terminals. The smart meters include the main meter for the distribution area and the user sub-meters. The acquisition terminals include various power monitoring terminals, each responsible for collecting raw power, voltage, current, power and other data. The default acquisition frequency is 15 minutes / time (set according to relevant specifications), which can be dynamically adjusted to 5 minutes / time based on the abnormal diagnosis results of the edge nodes (encrypted acquisition in abnormal scenarios).
[0058] Edge diagnostic layer: This layer includes multiple station area edge nodes. In this embodiment, the concentrator of each station area is used as the edge node of that station area. The concentrator (edge node) includes an ARM Cortex-A7 processor, 512MB of memory, local memory, and a communication interface. An edge diagnostic program runs on the processor, which includes a data access module, a local database, a rule diagnostic module, a lightweight model inference module, and a self-healing processing module.
[0059] Edge nodes acquire data collected by smart meters and data acquisition terminals, and perform rule-based local real-time diagnosis on the collected data by calling the power business rule base stored in the local database, and perform model-based local real-time diagnosis on the collected data by loading a lightweight diagnostic model stored in the local memory. The diagnostic latency is ≤5 seconds / data entry and the startup time is ≤30 seconds.
[0060] Main station aggregation layer: Deployed on the power company's main station server, running the federated learning coordination module and the dynamic aggregation module. The federated learning coordination module is used to create and manage training rounds, select participating nodes, distribute the global model and training configuration, control local training cycles or trigger conditions, receive and verify model parameters and statistical parameters uploaded by each node, handle timeouts and failed retransmissions, call the dynamic aggregation module for aggregation, and manage and redistribute the aggregated global model. The dynamic aggregation module is responsible for calculating node weights and performing parameter aggregation.
[0061] In addition, the main station includes a meteorological service interface, which obtains temperature and humidity data according to the geographical location of the station area. After being timestamped, the data is distributed and cached to edge nodes for training of lightweight models.
[0062] Business Application Layer: Provides a mobile APP (manual confirmation interface) for regional managers to receive abnormal work orders pushed by edge nodes / main site, and feeds back the confirmation results of on-site verification, the cause of the abnormality, and the handling measures to the edge nodes for continuous optimization of the rule base and model.
[0063] The core operating logic of the entire system is as follows: The terminal acquisition layer uploads raw data to the edge diagnosis layer → The edge diagnosis layer identifies anomalies through a composite diagnosis method (rule diagnosis + model diagnosis) and performs graded self-healing based on confidence level and business level → Anomalies that cannot be automatically repaired are pushed to the business application layer for manual verification → Edge nodes upload local model parameters, diagnostic performance, and manual feedback data to the main site aggregation layer → The main site dynamically aggregates and generates a global model through federated learning, and distributes the global model and training configuration to the edge nodes for updates → Edge nodes use manual feedback data to optimize the rule base and local model, achieving a closed-loop self-evolution throughout the entire process.
[0064] Example 2
[0065] This embodiment provides a specific implementation method for the edge diagnostic layer.
[0066] The edge diagnostic layer implements edge diagnostics by embedding a power business rule base and a lightweight diagnostic model in the edge nodes, and performs self-healing processing based on the diagnostic results. Both the power business rule base and the lightweight diagnostic model perform diagnostic functions, thus achieving dual diagnostics through both.
[0067] The power business rule base is responsible for the rapid identification of deterministic anomalies (no model inference required, response time ≤ 1 second), while the lightweight model is responsible for judging complex anomaly patterns. Both output diagnostic results such as anomaly type, confidence level, and anomaly duration, which are used to formulate self-healing strategies.
[0068] The edge diagnostic layer data input includes real-time collected data, historical load data, and transformer area archive data. The historical load data is the collected data of the most recent few days that are continuously cached in the local database of the edge node. In this embodiment, the data of the same period of the past 7 days is taken. The transformer area archive data is regularly synchronized to the local database of the edge node by the electricity information collection master station and the marketing archive system, and includes basic information such as household-transformer relationship, transformer parameters, and user type.
[0069] The diagnostic program is compatible with existing area concentrator hardware configurations, requiring no additional hardware investment. It can run stably on devices with an ARM Cortex-A7 processor and 512MB of memory, with a startup time of ≤30 seconds and a diagnostic latency of ≤5 seconds per diagnostic.
[0070] The power business rule base is solidified based on years of experience in power field anomaly diagnosis. It mainly includes four types of rules, each of which has been verified by more than a thousand sets of anomaly samples in the field and can be directly adapted to the application scenarios of different types of transformer substations.
[0071] Category 1: Rules for detecting electricity meter stagnation
[0072] Edge nodes monitor changes in electricity meter readings in real time. When the electricity meter readings of three consecutive collection points remain unchanged, they are marked as candidates for meter stagnation. If the synchronous current is greater than a preset effective current threshold, it is determined to be an abnormal meter stagnation. If the synchronous current of a user's sub-meter is not greater than the effective current threshold, it is not directly determined to be abnormal. If the synchronous current of the main meter in the distribution area is not greater than the effective current threshold, it is further distinguished as power outage, abnormal data collection and communication, and main meter failure.
[0073] The rule was verified in 12 pilot areas, and a total of 47 meter malfunctions were detected, with only 1 misjudgment.
[0074] Category 2: Rules for detecting sudden changes in line loss
[0075] Edge nodes periodically calculate the synchronous line loss rate of the transformer area. The calculation formula is: Line loss rate = (Total power supply of the transformer area - Sum of power consumption of all sub-meters) / Total power supply of the transformer area × 100%. If the line loss rate suddenly increases by more than 5 percentage points within 15 minutes and lasts for more than 30 minutes, an abnormal line loss alarm is triggered.
[0076] To eliminate misjudgments caused by changes in user records, the system simultaneously acquires the user-transformer relationship data from the transformer substation records during line loss anomaly monitoring. Users whose user-transformer relationship has recently changed are marked, and their electricity consumption is recalculated to rule out line loss anomalies. In the pilot application, this rule successfully identified 23 line loss anomalies caused by incorrect user-transformer relationships.
[0077] Category 3: Voltage and Current Exceedance Detection Rules
[0078] Based on State Grid standards, normal voltage and current ranges are set: 198V-242V for residential users and 342V-418V for industrial and commercial users. An over-limit alarm is triggered if the current exceeds 120% of the rated value for more than 5 minutes. Special attention is paid to scenarios where the three-phase voltage imbalance exceeds 2%, as such scenarios often indicate latent faults in the transformer substation. During the pilot period, this rule identified 31 latent faults, including insufficient transformer capacity and poor line contact, effectively reducing the fault rate in the transformer substation.
[0079]
[0080] Category 4: Power Balance Detection Rules
[0081] Based on real-time comparison of the total power meter reading for the distribution area and the sum of the power readings of each sub-meter, a power balance anomaly alarm is triggered if the deviation exceeds 10% and persists for 15 minutes. Time alignment processing is introduced, employing a sliding window matching mechanism to reduce time-section deviations caused by acquisition delays, ensuring that the data in the total meter and sub-meters are at the same time point. Specifically, the time t of the total meter reading for the distribution area is used as the alignment benchmark. Within the window, the power sample with the smallest difference from time t is selected for each user's sub-table. If samples exist on both sides of t, the power at time t is obtained through linear interpolation. If a sub-table sample is missing from the window, that time window is marked as incomplete, and no power imbalance alarm is triggered. The edge node compares the time-aligned total power of the transformer area with the power of each user's sub-table, and calculates the power deviation rate at time t using the following formula:
[0082]
[0083] In the formula, Let be the power deviation rate at time t. Let be the total active power of the transformer substation at time t. M represents the active power of the i-th user sub-table at time t after time alignment, and M is the number of valid user sub-tables participating in this calculation. This is a preset positive number used to prevent the denominator from being zero.
[0084] Here This refers to the valid user sub-table data collected within a preset time window before and after the timestamp of the overall transformer area table. If a user sub-table has no valid data within this time window, then that user sub-table will not participate in this power deviation rate calculation and will be recorded as missing data.
[0085] when An exception is triggered when the alignment rate exceeds 10% and N consecutive alignment windows last for at least 15 minutes. It should be determined according to the actual maximum acquisition delay, for example, 5 minutes; the window step size can be one acquisition point.
[0086] The rule has proven highly effective in industrial and commercial transformer substations, successfully identifying 18 cases of power imbalance caused by sub-meter failures, with an accuracy rate of 94.7%.
[0087] To address the one-dimensional time-series characteristics of the load curve data, a lightweight diagnostic model is designed using a one-dimensional convolutional neural network (1D-CNN) as the base model. To accommodate the limited computing power and storage of the edge nodes (concentrators), knowledge distillation is employed to transfer knowledge from the complex teacher model to the lightweight student model. A conventional model structure can be used here, ensuring that the final model parameter size is kept below 300KB, the inference time is ≤200ms / database, and it is compatible with the existing concentrator's ARM Cortex-A7 processor and 512MB of memory.
[0088] Model input features: 96-point daily load curves (15 minutes / collection point, continuous 24 hours), 7-day historical anomaly labels, 2D meteorological features (temperature, humidity, optional).
[0089] The model structure consists of the following connected components:
[0090] Convolutional Layer 1: Kernel size 3, number of channels 16, stride 1, activation function ReLU
[0091] Pooling layer 1: Max pooling, pooling window 2
[0092] Convolutional layer 2: Kernel size 3, number of channels 32, stride 1, activation function ReLU
[0093] Pooling layer 2: Average pooling, pooling window 2
[0094] Flattening layer: flattens the feature map into a one-dimensional vector.
[0095] Fully connected layer 1: 64 neurons, ReLU activation function, dropout rate 0.2
[0096] Fully connected layer 2: 32 neurons, activation function ReLU
[0097] Output layer: Initially 7 neurons, with the activation function Softmax, outputting the probability values of six types of abnormalities: meter flying away, meter stopping, abnormal line loss, voltage exceeding the limit, current exceeding the limit, and power imbalance, as well as normal categories; after new abnormalities are discovered by subsequent manual inspection, corresponding output neurons are added and the model is retrained.
[0098] Model training: Using 3 months of historical data from the pilot transformer area, totaling 1.2 million samples, including 180,000 abnormal samples, the training was iterated for 50 training rounds, and the convergence accuracy reached 98.5%.
[0099] like Figure 2 As shown, based on the above business rules and model's anomaly diagnosis results, this invention sets up the following three-level diagnostic mechanism:
[0100] Level 1: Deterministic rules, rule-first.
[0101] The acquired data is first quickly identified through the power business rule base. If the identified anomaly type matches a deterministic rule (such as meter malfunction, which is an anomaly with deterministic judgment rules), the rule output result is directly used without model reasoning, reducing diagnostic delay. The response time for this type of anomaly diagnosis is ≤1 second.
[0102] The deterministic rule refers to a rule that sets triggering and exclusion conditions based on a clear power business mechanism. After the corresponding triggering conditions are met and the normal business state is excluded, the rule diagnosis module can independently confirm the anomaly without calling the lightweight model inference module for verification.
[0103] Level 2: Fuzzy rules, rule-model mutual verification (Rule-triggered, Model-verified)
[0104] If the anomaly type identified by the power business rule base does not match the deterministic rule, then the fuzzy rule judgment is entered. If the anomaly category belongs to anomalies such as sudden line loss or voltage / current exceeding limits, which are supported by rules and require verification, the model is called for secondary confirmation after the rule is triggered. When the rule diagnosis result is consistent with the model diagnosis result and the final diagnosis confidence reaches the preset threshold, the confirmed anomaly is output; otherwise, it is marked as a suspected anomaly and added to the observation list.
[0105] The data in the observation list is reassessed: after entering the list, encrypted data collection or remote testing is initiated for 5 minutes; the rule conditions and model confidence are recalculated in the subsequent N windows; if the threshold is reached continuously, it is upgraded to confirmed anomaly and enters self-healing; if it returns to normal, it is automatically cancelled; if it is still uncertain after the observation period, a three-level manual verification work order is generated.
[0106] This mechanism increases the diagnostic accuracy of such anomalies to over 97%.
[0107] Level 3: Model-dominant, Rule-assisted
[0108] When none of the business rules have reached the trigger threshold, but the model's confidence level for one of the six types of anomalies reaches the model's judgment threshold, the model will perform supplementary identification.
[0109] The accuracy rate for diagnosing this type of abnormality is ≥92%.
[0110] The diagnostic output includes five pieces of information: anomaly type, anomaly occurrence time, anomaly duration, rule trigger identifier, and confidence score. In one specific embodiment, the rule trigger identifier can be configured to include:
[0111] rule_triggered: Whether the rule was triggered; rule_id: The specific rule number that was triggered; match_strength: The degree of rule matching or the extent to which the match exceeded the threshold.
[0112] In some embodiments of the present invention, the rule diagnosis module further includes a data integrity and transmission status diagnosis unit. The data integrity and transmission status diagnosis unit identifies abnormal states such as single-point data loss, acquisition delay and communication data interruption based on the predetermined acquisition time, actual acquisition time, data arrival time and data reporting sequence of the acquisition point, and outputs the corresponding abnormality type and rule confidence level.
[0113] Example 3
[0114] This embodiment provides a specific implementation method for the main site aggregation layer.
[0115] 1. Federated Learning Architecture
[0116] Confederate learning sequence as follows Figure 3 As shown, each edge node completes a local training task once a week. The training data consists of local TMR data collected by the edge node in the past week. The training process uses local historical anomaly labels as supervision signals and employs the cross-entropy loss function to optimize model parameters. After training, the node uses local validation data to calculate three performance metrics: recall (the proportion of actual anomalies that are correctly identified), precision (the proportion of real anomalies among the identified anomalies), and F1 score (the harmonic mean of recall and precision). At the same time, it statistically analyzes local data distribution characteristics, including data distribution entropy (reflecting data diversity), anomaly sample ratio (the percentage of anomalies in the total number of samples), and transformer area type coding (residential transformer area 0, industrial and commercial transformer area 1, rural transformer area 2).
[0117] The main site aggregator's federated learning coordination module receives and verifies the model parameters and statistical parameters uploaded by each edge node, and then sends the verified data to the dynamic aggregation module; the dynamic aggregation module calculates the weights and performs model parameter aggregation.
[0118] 2. Dynamic Aggregation Module
[0119] The dynamic aggregation module performs dynamic aggregation weight calculations based on the acquired model parameters, performance indicators, and distribution characteristics to generate global model parameters. These dynamic aggregation weights comprehensively consider node diagnostic performance and data heterogeneity. Compared to the traditional federated averaging algorithm which aggregates according to equal weights, this algorithm can dynamically adjust the aggregation weights based on the model performance and data distribution differences of each node, thereby reducing the impact of cross-regional data heterogeneity on model adaptability.
[0120] The formula for calculating aggregate weights is as follows:
[0121]
[0122] in, The aggregation weight of the k-th edge node satisfies And the weight of a single node ≤0.5 (Constraint: Avoid a single node dominating the global model and ensure model diversity). , These are the performance index values of the kth and jth edge nodes, respectively, reflecting the diagnostic performance. In this embodiment, the performance index value is the F1 score. , , respectively, are the data heterogeneity adjustment factors for the k-th and j-th edge nodes, reflecting the impact of the node's heterogeneity on aggregation. The larger the value, the higher the weight of the edge node in the aggregation; K is the total number of edge nodes participating in federated aggregation.
[0123] by The following example illustrates how the data heterogeneity adjustment factor is calculated:
[0124]
[0125] in, Let the data distribution entropy of the k-th edge node be . Characterizing the concentration of data distribution, The smaller, The larger the value, the more typical the data characteristics, and the higher the weight percentage. The proportion of outlier samples at the k-th edge node; This represents the average proportion of outlier samples across all nodes participating in the aggregation. , The item represents the deviation of the anomaly ratio. The larger the value, the more likely the node may have a special anomaly pattern and needs to be closely monitored. This is an indicator function; when the condition inside the parentheses is true, =1; otherwise =0; Encode the station type for the k-th edge node; Encode the mainstream transformer area type for all participating aggregation nodes (mainstream transformer area refers to the transformer area type with the highest percentage of samples; if there is no mainstream transformer area type, then...). (No possible value)
[0126] , , The coefficients, determined using the Analytic Hierarchy Process (AHP) combined with expert scoring from power field sites and test data from pilot transformer substations, reflect the weight of each component's impact on the overall model performance. In this embodiment, the values determined using test data from 100 pilot transformer substations are... , , .
[0127] (1) The coefficient 0.5 was obtained by the analytic hierarchy process (the influence weight of data distribution entropy on model aggregation is 0.52, and the approximate value is 0.5) and the on-site expert scores (80% of power experts believe that "typical scenario data" has the greatest supporting role for the global model). The more concentrated (typical) the data distribution, the higher the reference value for the global model, and the higher the weight should be given.
[0128] (2) The coefficient 0.3 was obtained by the analytic hierarchy process (the influence weight of the deviation of the abnormal proportion is 0.28, and the approximate value is 0.3) and on-site verification (the data of the transformer area with special abnormal patterns can improve the global model’s ability to identify niche abnormalities). The larger the deviation of the abnormal proportion from the mean, the more likely the node may have a special abnormal pattern, which should be given priority in the aggregation to avoid the model losing the characteristics of niche abnormalities.
[0129] (3) The coefficient 0.2 is obtained by the analytic hierarchy process (the influence weight of the transformer area type is 0.20, and the exact value is 0.2) and business needs (to ensure that niche transformer areas such as rural power grids have a voice in the aggregation and to avoid the global model from being overly biased towards residential / commercial transformer areas). When the transformer area type is different from the mainstream type, it is given additional weight to ensure that the global model is adapted to all transformer area types.
[0130] For ease of understanding, this embodiment uses a federated aggregation cycle with 3 edge nodes as an example to provide complete steps for indicator reporting, mean calculation, adjustment factor calculation, and aggregation weight calculation:
[0131] Step 1: Obtain node-reported metrics
[0132]
[0133] Step 2: Calculate the global statistical mean
[0134] Mean proportion of outlier samples: =(3.2%+5.8%+4.5%) / 3=4.5%;
[0135] Mainstream TV zone types: There is no absolutely mainstream type (1 of each), take =None, indicator function (All nodes are assigned additional weights based on their area type).
[0136] Step 3: Calculate the adjustment factor for each node.
[0137] α1=0.5×(1 / 2.1)+0.3×∣3.2%−4.5%∣+0.2×1=0.2381+0.0039+0.2=0.4420
[0138] α2=0.5×(1 / 3.5)+0.3×∣5.8%−4.5%∣+0.2×1=0.1429+0.0039+0.2=0.3468
[0139] α3=0.5×(1 / 1.8)+0.3×∣4.5%−4.5%∣+0.2×1=0.2778+0+0.2=0.4778
[0140] Step 4: Calculate the numerator
[0141] F 11 ⋅α1=0.92×0.4420=0.4066
[0142] F 12 α² = 0.88 × 0.3468 = 0.3052
[0143] F 13 α3 = 0.85 × 0.4778 = 0.4061
[0144] Step 5: Calculate the denominator term
[0145] Step 6: Calculate the aggregate weights
[0146] =0.4066 / 1.1179≈0.3637
[0147] =0.3052 / 1.1179≈0.2730
[0148] =0.4061 / 1.1179≈0.3633
[0149] Verification: 0.3637 + 0.2730 + 0.3633 = 1, which satisfies the weight constraint condition.
[0150] Results analysis: Although node 3 has the lowest F1 score, it has the lowest data distribution entropy (the most typical data), so the adjustment factor is the largest and the aggregation weight is basically the same as that of node 1; node 2 is an industrial and commercial transformer area. Although it has special abnormal patterns, its data distribution entropy is relatively large, so its weight is the lowest. This is in line with the design idea of prioritizing typical data, taking into account special patterns, and fairly adapting to different transformer area types.
[0151] 3. Secure communication channel
[0152] After generating global model parameters, the dynamic aggregation module distributes them to each edge node through a secure channel.
[0153] To meet the data security and related graded protection requirements of the power industry, this invention adopts a multi-layered secure communication mechanism of "restricted transmission of model parameters, secure multi-party computation, TLS 1.3 encryption, and data anonymization" to ensure the confidentiality, integrity, and immutability of data during communication. The specific design is as follows:
[0154] Transmission content limitations: Edge nodes only transmit about 300KB of model parameters and statistical parameters required for dynamic aggregation to the main station, without transmitting raw power data and user data, so that the raw data is always kept locally;
[0155] Secure Multi-Party Computation (SMPC): Based on the secret-sharing SMPC algorithm, the model parameters of each edge node are split into multiple secret shares and transmitted to the master station respectively. The master station can only restore the parameters after collecting all shares. A single share has no practical meaning. The secret-sharing threshold and the number of participants are determined according to the actual scale of the deployed edge nodes and security requirements. In this embodiment, the SMPC parameters are set as follows: secret-sharing threshold k=2, and the participants are edge nodes + master station.
[0156] TLS 1.3 Encryption: The communication protocol adopts the latest TLS 1.3 standard, using the AES-256-GCM symmetric encryption algorithm to encrypt transmitted data, and employing the ECDHE-P256 key exchange algorithm to complete key negotiation. The key is periodically updated and rotated (weekly in this embodiment, synchronized with the federated learning cycle) to prevent security issues caused by key leakage;
[0157] Data anonymization: Sensitive information in model parameters is anonymized: ① Substation identifiers are encrypted using SHA-256 hash, making reverse parsing impossible; ② Numerical model parameters are normalized to eliminate their actual physical meaning.
[0158] The secure communication mechanism was piloted in three cities and operated continuously for three months, transmitting model parameters a total of 720 times without any data leakage or tampering incidents. The communication delay was ≤1 second, meeting the real-time requirements.
[0159] Example 4
[0160] This embodiment specifically illustrates the implementation method of the hierarchical self-healing strategy based on confidence level and business level in this invention.
[0161] The self-healing strategy is executed by the self-healing processing module of the edge diagnostic layer. The self-healing processing module selects the self-healing strategy based on the anomaly type, overall confidence level, security risk, and action rollbackability. Level 1 and Level 2 self-healing are executed locally by the corresponding edge nodes and the results are recorded; Level 3 anomalies that require manual verification generate work orders; Level 4 actions involving cross-regional or unified configuration of the main station are approved or executed by the main station.
[0162] 1. Calculation of overall confidence level
[0163] Confidence level is the core quantitative basis for graded self-healing. This invention uses a weighted average of rule confidence and model confidence to calculate the comprehensive confidence level, thus balancing the certainty of the rules with the accuracy of the model, providing a quantitative basis for graded self-healing. The calculation formula is as follows:
[0164]
[0165] In the formula To assess the overall confidence level, , These are the weights for rule confidence and model confidence, respectively, and they satisfy the constraints. .
[0166] (1) Rule confidence
[0167] Rule confidence is calculated based on rule matching strength. Different matching degrees correspond to different confidence values, and the assignment criteria are determined based on the statistical results of on-site anomaly samples to ensure that the assignment is highly correlated with the anomaly accuracy rate. In this embodiment, the confidence assignment criteria for the core rules determined based on the statistical results of 1200 sets of on-site anomaly samples are shown in the table below. Rules not listed are calibrated according to the principle of "the higher the rule matching strength, the higher the rule confidence." After a rule is triggered, the magnitude, duration, and fulfillment of combined conditions of the anomaly index exceeding the threshold are used to represent the rule matching strength. For continuous numerical rules, the rule matching strength is determined based on at least one of the threshold exceedance magnitude, duration, and number of consecutive triggers; for combined rules consisting of multiple judgment conditions, the rule matching strength is determined based on the fulfillment of each judgment condition and the number of consecutive fulfillments. A mapping relationship between rule matching strength and rule confidence is established for different rules.
[0168] The table below only lists some examples of rule matching strength and confidence values. Specific thresholds and confidence levels can be adjusted based on the operational characteristics of different transformer substations and historical sample statistics:
[0169]
[0170] When the rule diagnosis module performs rule judgment, it first determines whether the data to be diagnosed meets the triggering conditions of the corresponding rule. After the rule is triggered, the rule matching strength is determined based on the extent to which the diagnostic feature exceeds the rule threshold, the duration of the abnormality, and the satisfaction of the combination conditions. The rule confidence is determined based on the rule matching strength. The output rule judgment result includes the rule type and confidence.
[0171] (2) Model confidence
[0172] Model confidence scores were calculated using the Monte Carlo Dropout method to estimate uncertainty. Ten forward inferences were performed on the same input (with Dropout enabled each time), and the mean μ and variance σ² of the probabilities for each class were calculated. The confidence score was... The larger the variance, the lower the confidence level. This method can effectively reduce the uncertainty error of model prediction. In the pilot verification, the model confidence error was ≤3%.
[0173] (3) Rule weight / Allocation
[0174] Rule weight With model weights The allocation is based on rule type, and the criteria are the rule's accuracy and determinism. The higher the accuracy and the stronger the determinism, the greater the rule's weight. The specific allocation criteria are shown in the table below:
[0175]
[0176] 2. Classification of Business Impact Levels
[0177] Based on the actual impact of anomalies on the accuracy of power metering, the safe operation of the power grid, and on-site business processing, and in accordance with the "Operation and Management Specifications for Power Consumption Information Collection System," anomalies are classified into four levels of business impact. The processing time requirements and typical anomaly types for each level are clearly defined. The classification criteria have been verified in pilot distribution areas and are consistent with actual on-site business practices, as shown in the table below:
[0178]
[0179] 3. Development of a graded self-healing strategy
[0180] Based on the combination of confidence interval and business impact level, the self-healing strategy is divided into four levels, and corresponding self-healing actions are set for each level to achieve refined repair.
[0181] Specifically, historical diagnostic results are first grouped according to the confidence intervals given by the comprehensive confidence level calculation, and then the cutoff thresholds are calibrated based on on-site verification. The on-site verification labels are continuously updated as the sample size increases; therefore, the correction of the confidence intervals is dynamic. The specific process is as follows:
[0182] Periodically retrieve historical diagnostic samples that have undergone manual verification. Let the final diagnostic confidence level of the j-th sample be . The manual verification result is ;Real anomaly retrieval =1, non-real exceptions are taken =0. For the candidate confidence threshold... The positive predictive value (PPV) corresponding to a diagnostic result with a confidence level not lower than this threshold is calculated according to the following formula:
[0183]
[0184] Where n represents the total number of historical labeled samples, This represents the indicator function, taking the value 1 if the condition is true and 0 otherwise. PPV represents the proportion of samples deemed anomalous by the system that are confirmed as genuine anomalies by manual verification; it is not equivalent to the overall accuracy across all samples. Simultaneously, the confidence level must not be lower than the candidate threshold. The sample size is:
[0185]
[0186] The calibrated confidence threshold is determined according to the following formula:
[0187]
[0188] in, Represents the set of candidate thresholds. Indicates the target PPV. This represents the minimum sample size required to guarantee the validity of the statistical results. This indicates the confidence threshold used after calibration.
[0189] Analyze the historical PPV corresponding to different confidence thresholds, and then determine the confidence interval boundary based on the target PPV. For example, when the historical sample statistics show that the PPV corresponding to samples with a confidence level of not less than 0.9 is not less than 98%, and the sample size meets the minimum requirement, 0.9 is used as the lower limit of the high confidence interval.
[0190] In this embodiment, four confidence levels are defined, including >C1, C2-C1, C3-C2, and <C3. C1-C4 are confidence thresholds, which are calibrated based on the PPV statistics of historical labeled samples and can be configured according to different substation areas and anomaly types.
[0191] The detailed content, pilot implementation examples, and pilot success rates of the Level 4 self-healing strategy are shown in the table below. All self-healing actions require no additional hardware investment and can be directly implemented in existing systems. Since the confidence interval threshold is a dynamic value, the confidence boundary threshold in the table differs in different demonstration scenarios. Therefore, only symbols are used in the table as examples.
[0192]
[0193] Examples of different levels of self-healing involved in the pilot program are as follows:
[0194] Level 1 Self-Healing Process: Data loss occurred at a residential transformer substation collection point, diagnosed as a momentary communication outage (confidence level 0.95 > C1, Level 1 service impact). The system automatically executed historical average interpolation, retrieving data from the same time period over the past four weeks (2.35kWh, 2.41kWh, 2.28kWh, 2.33kWh), calculating an average of 2.34kWh to fill the missing data. The repaired data trended consistent with subsequent data collection, reflecting actual on-site electricity consumption. During the pilot phase, Level 1 self-healing was executed 1245 times, with 1109 successful attempts, a success rate of 89.1%.
[0195] Level 2 Self-Healing Process: For a commercial or industrial user's electricity meter, the readings at three collection points remained unchanged, and the current value was normal (confidence level 0.82 ∈ [C1, C2], Level 2 business impact). The system triggered a remote recall mechanism, conducting three consecutive recalls (1 minute intervals). The results were 156.2, 156.3, and 156.2 respectively. The three results were largely consistent and aligned with historical trends. The last recall value of 156.3 was used to overwrite the original data, and the meter was added to the key monitoring list. Enhanced data collection and observation were conducted for 24 hours, and no further anomalies were observed. During the pilot period, Level 2 self-healing was executed a total of 876 times, with 688 successful executions, a success rate of 78.5%.
[0196] Level 3 Self-Healing Process: In a certain transformer substation, the line loss rate suddenly increased from 3.2% to 8.7% (confidence level 0.65 ∈ [C2, C3], Level 3 business impact). The system generated an exception work order and pushed it to the substation manager's APP. The work order included the exception type, occurrence time, confidence level, list of associated meters, and troubleshooting suggestions. The substation manager's on-site verification revealed that the transformer-household relationship for newly installed users had not been updated, causing a discrepancy between the master meter and sub-meter statistics. After correcting the records, the line loss rate returned to normal. During the pilot period, the effectiveness rate of Level 3 self-healing work orders increased from 18.3% before the pilot to 82.8%.
[0197] Level 4 self-healing process: When multiple transformer areas experience power imbalance simultaneously (confidence level 0.42 < C4, Level 4 business), the main station conducts in-depth diagnosis to detect transformer faults, promptly arranges maintenance, and avoids power outages.
[0198] 4. The self-healing strategy library is dynamically updated.
[0199] To adapt to the dynamic changes in power field business scenarios, an ε-greedy multi-armed slot machine algorithm is adopted to dynamically adjust the selection priority of the self-healing strategy library. The core design idea is to balance strategy exploration and utilization, that is, to prioritize the self-healing action with the highest success rate (utilization), while trying other actions with a small probability (exploration), and continuously optimize the strategy priority based on historical self-healing data to improve the self-healing success rate. Specifically:
[0200] Algorithm type: ε-greedy multi-armed slot machine algorithm, ε=0.1, selected based on the following: tests found that when ε=0.1, the balance between exploration and exploitation is optimal, neither reducing the success rate due to excessive exploration nor causing the strategy to become rigid due to excessive exploitation;
[0201] Data logging: The system records all data for each self-healing operation, including: anomaly type, self-healing action, success or failure of repair, manual confirmation result, repair time, etc.
[0202] Statistical cycle: Statistics are compiled weekly (in sync with the federated learning cycle) to calculate the success rate of different self-healing actions under each anomaly type;
[0203] Priority Update: When the success rate of a certain self-healing action is more than 5% higher than the current preferred action for two consecutive weeks, the system will automatically update the strategy library and set the action as the preferred self-healing action for the corresponding exception type.
[0204] Pilot verification example:
[0205] Taking communication interruption anomalies as an example, the statistical data of the three self-healing actions during the pilot period are shown in the table below. The system dynamically adjusts the strategy priority based on the success rate, increasing the self-healing success rate of this type of anomaly from 89.1% to 91.2%.
[0206]
[0207] Example 5
[0208] This embodiment specifically illustrates the implementation method of the business application layer.
[0209] The business application layer includes the area manager application (APP) and the manual confirmation feedback module. The manual confirmation results and system operation feedback are used to continuously optimize the diagnostic model and business rules, realize the self-evolution of the system, and improve long-term operating performance.
[0210] 1. Manual confirmation of return mechanism
[0211] After the regional manager confirms the pushed abnormal work order on the mobile application, the confirmation result is fed back to the edge node through a secure channel. The fed-back data includes core information such as work order ID, abnormality type, occurrence time, on-site confirmation result, cause of abnormality, handling measures, whether it is a false alarm, and suggested weight. The data is in a standardized format and can be directly used for model training and rule optimization.
[0212] Example of reflux data format:
[0213] {
[0214] "Work Order ID": "GZ20250315001",
[0215] "Abnormal Type": "Linear Loss Sudden Change"
[0216] "Time of Occurrence": "2025-03-15 10:00:00",
[0217] On-site confirmation result: "True anomaly"
[0218] Reason for the anomaly: "Household registration information not updated".
[0219] "Handling Measures": "Revision of Archives"
[0220] Confirmed by: Zhang San
[0221] Confirmation Time: 2025-03-15 14:23:00
[0222] "False alarm": false,
[0223] Recommended weight: 1.5
[0224] }
[0225] Edge nodes use a lightweight SQLite database to store the backflow data locally, while the main site uses a MySQL cluster for centralized storage. The backflow data is stored as new label data in the local training set for the next round of model training. Samples identified as false alarms are marked as negative samples, and their weights are set to 2.0 during training (based on the fact that tests showed that when the weight is 2.0, the false alarm rate decreases by an average of 1.2% per month, and too high a weight will lead to model overfitting), reducing the probability of similar false alarms occurring again.
[0226] 2. Rule base self-optimization
[0227] The system periodically analyzes manually confirmed data and uses the FP-Growth association rule mining algorithm to extract new rule patterns and enrich the rule base. The mining process is as follows: collect samples that have been manually confirmed as true anomalies in the past three months, extract multi-dimensional features such as time, weather, load, and archives, mine frequently occurring feature combinations, and select rules with support >5% and confidence >70% as candidates. After manual review, these rules are added to the rule base.
[0228] During the pilot phase, three new rules were discovered through this mechanism, as follows:
[0229]
[0230] 3. Federal Retraining Triggering Conditions
[0231] The main station continuously monitors the diagnostic performance of each edge node. When any trigger condition is met, a new round of federated training is automatically started. The incremental learning method is used to continue training on the basis of the existing model, avoiding catastrophic forgetting and saving computing resources.
[0232]
[0233] Incremental learning parameter settings: Continue training for 5 more training epochs based on the existing model, and adjust the learning rate to 0.1 times the initial learning rate. The training time is shortened by 60% compared to retraining, and the model performance does not decrease significantly. In the pilot verification, the F1 score of the model improved by an average of 0.05 after incremental training.
[0234] The solution of this invention has been piloted in the TMR systems of several municipal power supply companies and has achieved good application results.
[0235] In terms of diagnostic accuracy, the composite diagnostic mechanism combining business rules and lightweight models reduced the false alarm rate from 18.3% to 4.2%, while maintaining a recall rate of over 93%, effectively reducing the workload of manual review.
[0236] In terms of real-time diagnostics, by conducting local diagnostics at edge nodes, the average system response time has been reduced from 12 minutes under a centralized architecture to less than 45 seconds, which can meet the business needs of real-time monitoring.
[0237] In terms of self-healing effectiveness, after applying the hierarchical self-healing strategy based on the comprehensive confidence interval, the proportion of invalid repairs decreased by 67%, the overall repair success rate increased to 89%, and the number of work orders pushed to the area manager decreased by 67%, enabling maintenance personnel to concentrate their limited human resources on abnormal events that truly require manual handling.
[0238] In terms of model adaptability, after adopting the heterogeneous federated aggregation method oriented towards data quality diagnosis, the standard deviation of the F1 score of the global model ranged from 0.15 to 0.06 across different transformer substation types, maintaining stable performance across residential, industrial and commercial, and rural transformer substations.
[0239] In terms of privacy and security, the system adopts a federated learning architecture that uploads model parameters instead of raw data, and meets the Level 3 protection requirements of the power data security classification guidelines through secure multi-party computation and encrypted communication, achieving data usability without visibility.
Claims
1. A TMR data quality intelligent diagnosis and self-healing system, characterized in that, The system includes: The terminal acquisition layer collects electricity meter data and electricity parameter data; The edge diagnostic layer includes multiple transformer area edge nodes. Each transformer area edge node is configured with a rule diagnostic module, a lightweight model inference module, and a self-healing module. The rule diagnostic module performs data quality diagnosis based on preset business rules, and the lightweight model inference module calls locally stored lightweight models to perform data quality diagnosis. The self-healing module performs graded self-healing based on the anomaly identification results of the diagnostic module and the business level of the anomaly type. The main site aggregation layer includes a federated learning coordination module and a dynamic aggregation module deployed on the main site server. The federated learning coordination module receives local model parameters and diagnostic performance statistics uploaded by each edge node, calls the dynamic aggregation module to determine the aggregation weight of each edge node based on the diagnostic performance statistics, aggregates the local model parameters based on the aggregation weight to obtain global model parameters, and sends them to each edge node for updates. The business application layer is used to receive abnormal task information sent by the edge nodes and / or the main station aggregation layer, and to provide a manual confirmation interface for the abnormal task information.
2. The system according to claim 1, characterized in that, The edge diagnostic layer is configured with a power business rule base. The rule diagnostic module calls the power business rule base and performs rule diagnostics based on the preset business rules in the power business rule base. The power business rule base includes at least one of the following data diagnostic rules: Electricity meter stagnation detection rule: When the electricity meter readings at three consecutive collection points remain unchanged, and the synchronous current is greater than the preset effective current threshold TH1, the first type of rule abnormal alarm is triggered. Line loss sudden change detection rule: Calculate the line loss rate based on the power supply of the main meter of the transformer area and the power consumption of all sub-meters. When the line loss rate increases by more than a% within a preset time period t1 and lasts for more than a preset time period t2, trigger a second-class rule abnormal alarm. Voltage and current over-limit detection rules: When the voltage parameter exceeds the preset upper or lower voltage threshold and the duration exceeds the preset time t3, the current parameter exceeds the preset current threshold and the duration exceeds the preset time t4, or the three-phase imbalance exceeds the preset threshold TH2 and the duration exceeds the preset time t5, three types of rule abnormal alarms are triggered. Power balance detection rules: The total power of the transformer area is compared with the sum of the power of each sub-meter in real time. When the deviation between the total power of the transformer area and the sum of the power of each sub-meter exceeds the preset ratio b% and the duration exceeds the preset time t6, four types of rule abnormal alarms are triggered.
3. The system according to claim 2, characterized in that, The line loss mutation detection rule also includes household-transformer relationship verification: when detecting the line loss rate of the transformer area, the household-transformer relationship information in the transformer area archive data is obtained simultaneously, the target users whose household-transformer relationship has changed are marked, and the electricity consumption data of the marked users are corrected, so as to redetermine the line loss status of the transformer area with the corrected user electricity consumption.
4. The system according to claim 1, characterized in that, The lightweight model is constructed using a one-dimensional convolutional neural network. The input features of the lightweight model include daily load curve features, historical anomaly labels, and meteorological features. The output results include at least the probability values corresponding to normal operation, meter flyaway anomaly, meter stagnation anomaly, line loss anomaly, voltage over-limit anomaly, current over-limit anomaly, and power imbalance anomaly, respectively.
5. The system according to claim 1, characterized in that, The edge diagnostic layer performs hierarchical diagnostics on the collected data, including: The first level of diagnosis involves performing a diagnosis on the collected data through the rule diagnosis module. If the identified anomaly type matches a preset deterministic rule, the rule diagnosis result is directly determined as the anomaly identification result without calling the lightweight model inference module for model inference. The second level of diagnosis involves calling the lightweight model inference module to verify the rule diagnosis results when the anomaly type identified by the rule diagnosis module matches a preset fuzzy rule. If the rule diagnosis result is consistent with the model diagnosis result and the overall confidence level reaches a preset threshold, a confirmed anomaly is output; otherwise, it is marked as a suspected anomaly and added to the observation list. The overall confidence level is obtained by weighting the model diagnosis confidence level and the rule diagnosis confidence level. Encrypted data collection and / or remote recall are initiated for the data in the observation list. Rule diagnosis and model inference are re-executed in the subsequent N diagnostic windows. When the comprehensive confidence level continuously reaches the corresponding judgment condition, the suspected anomaly is upgraded to a confirmed anomaly and sent to the self-healing processing module. When the rule diagnosis result returns to normal and the model confidence level is lower than the corresponding judgment threshold, the suspected anomaly is canceled. If the anomaly status is not determined after the preset observation period, a corresponding manual verification task is generated for subsequent push to the business application layer. The third level of diagnosis is determined when none of the business rules in the power business rule base meet the abnormal triggering conditions, and the confidence level of the lightweight model's output for at least one preset abnormal type reaches the second judgment threshold. The corresponding abnormal identification result is then determined based on the output result of the lightweight model.
6. The system according to claim 1, characterized in that, Each edge node periodically uses local TMR data and anomaly labels to train a local model, and calculates performance indicators and statistical local data distribution characteristics based on local validation data to form diagnostic performance statistics. The federated learning coordination module receives and verifies the model parameters and diagnostic performance statistics uploaded by each edge node, and sends the verified data to the dynamic aggregation module for weight calculation. The dynamic aggregation module calculates the aggregation weights based on the following formula: ; ; ; ; In the formula, is the aggregation weight of the k-th edge node; K is the total number of edge nodes; , These are the performance index values of the k-th and j-th edge nodes, respectively; , These are the data heterogeneity adjustment factors for the k-th edge node; Let the data distribution entropy of the k-th edge node be denoted as . The proportion of outlier samples at the k-th edge node; The average proportion of outlier samples across all nodes participating in the aggregation. For indicator functions; Encode the station type for the k-th edge node; Encode the mainstream station types for all participating aggregation nodes; , , is a coefficient.
7. The system according to claim 5, characterized in that, The confidence level of the diagnostic results is obtained based on the following method: Obtain the confidence level of the rule diagnosis results output by the rule diagnosis module, and use it as the confidence level of the first-level diagnosis; Obtain the confidence score of the model diagnosis results output by the lightweight model inference module, and use it as the confidence score of the third-level diagnosis; The confidence score of the second-level diagnosis is based on the confidence score of the rule-based diagnostic results and the comprehensive confidence score calculated by the lightweight model inference module. Based on the confidence levels of the first-level diagnosis, the third-level diagnosis, and the second-level diagnosis, the confidence levels are divided according to the preset confidence thresholds to obtain the initial confidence interval. Obtain verified historical diagnostic samples, calculate the positive predictive value corresponding to different candidate confidence thresholds, and determine the boundary threshold of each confidence level based on the preset target positive predictive value of each interval, which is used to correct the boundary threshold of the initial confidence interval. The boundary threshold is dynamically adjusted based on the updated status of historical diagnostic samples.
8. The system according to claim 7, characterized in that, The confidence level of the rule diagnosis result is determined based on the rule matching strength, according to a pre-constructed rule matching strength-confidence mapping table; the values in the mapping table are assigned based on the statistical results of on-site samples. The confidence level of the model's diagnostic results is determined based on the statistical characteristics of multiple random forward inference results.
9. The system according to claim 1 or 8, characterized in that, The self-healing processing module determines the self-healing level based on the confidence level of the diagnostic results and the business level corresponding to the anomaly type, and executes the self-healing strategy corresponding to the self-healing level. The self-healing strategy includes data interpolation compensation or data backtracking, remote recall and data repair, generating abnormal work orders and pushing them to the business application layer, and reporting to the main station for in-depth diagnosis and generating troubleshooting plans.
10. The system according to claim 1, characterized in that, The self-healing processing module determines the selection priority of the self-healing strategy corresponding to different anomaly types based on historical self-healing operation data and uses a multi-armed slot machine algorithm based on an ε-greedy strategy. It also optimizes the selection priority based on the historical success rate of each healing action.
