A 0.4kV power customer internal load sudden drop determination method and system

Abstract

The present application relates to the technical field of load sudden drop determination, and discloses a 0.4kV power customer internal load sudden drop determination method and system, which comprises the following steps: collecting real-time current and 24-hour historical data from a power utilization information collection system at a 15-minute cycle, and synchronously collecting fault text data; determining suspected sudden drop based on current zeroing and 24-hour backtracking, outputting rule confidence based on automatic power failure record by a rule engine, analyzing and matching historical features by a text intelligent analysis engine to output text analysis confidence, and obtaining comprehensive confidence by weighted fusion to determine the type of sudden drop; automatically generating a differentiated service response strategy according to the comprehensive confidence; and verifying and counting the accuracy rate by using hierarchical random sampling. The present application significantly improves the determination accuracy, reduces misjudgment and missed judgment by deep fusion of multi-source data and fusion of double-engine confidence, and realizes low-cost fine power supply service under zero hardware addition by combining closed-loop optimization and hierarchical response.

Description

Technical Field

[0001] This invention relates to the field of load drop determination technology, and more specifically, to a method and system for determining load drops within a 0.4kV power customer premises. Background Technology

[0002] In 0.4kV distribution network power supply services, customer-side internal faults are the main cause of power outages, and load surges are the core characteristic of customer-side faults. Traditional load surge detection mainly relies on manual inspections and passive verification after customer reports, which suffers from problems such as delayed judgment, inability to achieve batch real-time identification, and untimely service response. At the same time, traditional methods lack standardized judgment rules, easily misjudging abnormal data collection and automatic customer power outages as load surges, resulting in low accuracy and increased ineffective inspections and service costs. In addition, traditional methods lack the ability to subdivide surge types, cannot effectively distinguish between customer-side internal faults and active power outages, leading to service resource misallocation, and lacking dynamic optimization mechanisms, making it difficult to adapt to the load monitoring needs of large-scale power supply areas.

[0003] Existing power information systems already possess high-frequency current data acquisition capabilities. Some systems attempt to construct judgment rules based on current data, setting simple zero-current judgment rules to identify load surges. While this approach achieves batch judgment, it suffers from the following drawbacks: relying solely on single current data points fails to differentiate between various zero-current scenarios, such as internal customer faults, automatic power outages by the power supply company, abnormal data acquisition, and customer-initiated power disconnection, resulting in a generally high false-judgment rate. Therefore, there is an urgent need to provide a method and system for determining internal load surges in 0.4kV power customers to address these issues. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a method and system for determining sudden load drops within a 0.4kV power customer's system. It integrates current data with work order text to improve accuracy; employs tiered response to reduce ineffective costs; utilizes closed-loop optimization for self-iteration; and requires zero hardware investment, making it highly economical.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for determining sudden load drops within a 0.4kV power customer, comprising:

[0007] The system collects real-time current data and 24-hour historical current data of 0.4kV power customers every 15 minutes from the electricity information collection system. It also collects unstructured fault text data from the 95598 work order system and the grid service management system and performs standardized preprocessing.

[0008] The judgment is based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the current value recovers to non-zero in the future, the judgment result is removed. If it remains at 0, it is retained.

[0009] The system performs initial classification based on customer automatic power outage records and outputs rule confidence scores. Natural language parsing is then performed on the unstructured fault text data to extract fault features and match them with historical fault features, outputting text analysis confidence scores. The rule confidence scores and text analysis confidence scores are then weighted and fused according to a preset mixed judgment weight to generate a comprehensive confidence score. Finally, the system generates a load drop type determination result based on the comprehensive confidence score.

[0010] Based on the comprehensive confidence level and judgment results, a differentiated service response strategy is automatically generated;

[0011] The judgment results are verified by stratified random sampling. The text intelligent parsing engine monitors the 95598 work order system and grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

[0012] As a preferred embodiment of the present invention, the natural language parsing is performed by a text intelligent parsing engine, which is an AI large model engine based on deep learning training. It uses a lightweight and finely tuned BERT-Chinese-base model as the core of text parsing and is deployed on a cloud inference server. Its pre-trained corpus contains no less than 100,000 historical 95598 work orders, grid service records and on-site emergency repair reports. The inference latency is no more than 200 milliseconds per message, and it supports a batch text processing rate of no less than 100 messages per second.

[0013] As a preferred embodiment of the present invention, the standardization preprocessing includes:

[0014] The structured current data is cleaned, with a timed trigger cycle of 15 minutes ± 10 seconds. The current data verification threshold is set to 1.2 times the rated current from zero. Data exceeding this range is marked as a metering anomaly, and the timestamp synchronization error does not exceed 30 seconds.

[0015] After cleaning and removing special characters, garbled text, and duplicate content from unstructured fault text data and unifying it into simplified Chinese, the text intelligent parsing engine performs entity recognition to extract core entities such as fault phenomenon, fault location, fault cause, and processing result, and automatically classifies them into internal faults, power grid faults, automatic power outages, or other categories, outputting a classification probability value between zero and one; at the same time, the fault text is converted into a high-dimensional feature vector and stored in the fault feature library.

[0016] As a preferred embodiment of the present invention, the determination based on real-time current data includes:

[0017] For single-phase customers, the current of the corresponding phase at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample; for three-phase customers, the current of all three phases at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample.

[0018] Define a set of valid sampling points within 24 hours, and count the total number of sampling points with a current value of zero. If the total number is zero, proceed to the next step; otherwise, remove the candidate samples.

[0019] Monitor the next 3 sampling cycles, i.e., 45 minutes. If the current at 3 consecutive sampling points is not greater than the preset minimum metering resolution threshold, it is confirmed as a suspected load drop, and the drop time is recorded.

[0020] The confidence level of the rule is calculated, which decreases as the number of valid sampling points decreases.

[0021] As a preferred embodiment of the present invention, the initial classification based on customer automatic power outage records includes:

[0022] If the customer has no record of automatic power outage, it is directly judged as a suspected internal fault, with a rule confidence level of 0.9;

[0023] If the customer has an automatic power outage record, compare the load drop time with the trip success time: when the load drop time is earlier than the trip success time, it is judged as a suspected internal fault with a rule confidence level of 0.85; when the load drop time is later than or equal to the trip success time, it is judged as a suspected automatic power outage with a rule confidence level of 0.92.

[0024] As a preferred embodiment of the present invention, the step of performing natural language parsing on unstructured fault text data to extract fault features and matching them with historical fault features includes:

[0025] The text intelligent parsing engine extracts the feature vectors of all fault texts from the past 12 months of the customer's history, as well as the feature vector of the current sudden drop event;

[0026] Calculate the average similarity between the current sudden drop event and the feature vectors of historical faults. This similarity is measured by the cosine of the angle between the vectors.

[0027] Based on the similarity and the rule confidence, the text analysis confidence is calculated using a nonlinear mapping function.

[0028] As a preferred embodiment of the present invention, the step of weighting and fusing the rule confidence and the text analysis confidence according to a preset mixed judgment weight includes:

[0029] Set a mixed judgment weight, with the sum of the two weights being one. The default weight for the traditional rule engine is 0.6, and the weight for the text intelligent parsing engine is 0.4.

[0030] The rule confidence score and the text analysis confidence score are weighted and summed according to a preset mixed judgment weight to obtain the comprehensive confidence score;

[0031] If the overall confidence level is not lower than 0.7, output the final judgment result and the overall confidence level; if the overall confidence level is between 0.5 and 0.7, mark it as pending manual review and push it to the power supply service personnel; if the overall confidence level is lower than 0.5, judge it as data abnormal and remove this result.

[0032] As a preferred embodiment of the present invention, the automatic generation of differentiated service response strategy based on comprehensive confidence level and judgment result includes:

[0033] When a suspected internal fault is detected, a personalized warning SMS is generated and automatically pushed to the customer. At the same time, the customer information and the time of the sudden incident are pushed to the mobile terminal of the area manager. If there is no customer call within 1 hour and the overall confidence level is not lower than 0.85, a voice callback is automatically triggered.

[0034] When a suspected automatic power outage is detected, an overdue payment reminder SMS will be automatically generated and a manual consultation channel will be maintained, but on-site inspections will not be arranged.

[0035] When marked as pending manual review, the request is sent to the power supply service operator for manual verification before a decision is made on whether to dispatch the order.

[0036] As a preferred embodiment of the present invention, it also includes model closed-loop optimization:

[0037] Invalid judgment data, misjudged samples and missed judgment samples found during the verification process are included in the sample library. The text intelligent parsing engine automatically analyzes their current data features and text features, extracts misjudgment reason tags and generates error feature vectors.

[0038] New judgment rule suggestions are automatically generated based on the error sample library. The format is to output the corresponding judgment result and attach the confidence level if a specific condition is met.

[0039] The rule review mechanism is as follows: if the rule confidence score is not lower than 0.9, it is automatically deployed to the rule engine; if the rule confidence score is between 0.7 and 0.9, it is pushed to the algorithm engineer for review; if the rule confidence score is lower than 0.7, it is marked as pending observation.

[0040] Each week, new validation samples and error samples are added to the training set. The text intelligent parsing engine is incrementally fine-tuned and the mixed judgment weights are recalculated to maximize the overall accuracy. The fault feature library is updated monthly.

[0041] A load surge detection system for 0.4kV power customers, comprising:

[0042] The data acquisition and preprocessing module is configured to collect real-time current data and 24-hour historical current data of 0.4kV power customers from the electricity information acquisition system at a 15-minute interval, and simultaneously collect unstructured fault text data from the 95598 work order system and the grid service management system, and perform standardized preprocessing.

[0043] The load drop initial judgment module is configured to make judgments based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the current value returns to non-zero in the future, the judgment result is removed; if it remains at 0, it is retained.

[0044] The load drop type fusion determination module includes a rule engine initial classification unit, a text intelligent parsing engine, and a weighted fusion unit. The rule engine initial classification unit is configured to perform initial classification based on customer automatic power outage records and output rule confidence scores. The text intelligent parsing engine is configured to perform natural language parsing on the unstructured fault text data to extract fault features and match them with historical fault features, outputting text analysis confidence scores. The weighted fusion unit is configured to weight and fuse the rule confidence scores and text analysis confidence scores according to preset mixed determination weights to generate a comprehensive confidence score, and generate the final load drop type determination result based on the comprehensive confidence score. The service response module is configured to automatically generate differentiated service response strategies based on the comprehensive confidence score and the determination result.

[0045] The result verification module is configured to verify the judgment results using a stratified random sampling method. The text intelligent parsing engine monitors the 95598 work order system and the grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments, and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

[0046] The beneficial technical effects of this invention are:

[0047] This invention not only performs 24-hour historical backtracking judgment based on 15-minute high-frequency current sampling data, but also simultaneously collects unstructured fault text data from the 95598 work order system and the grid service management system. Through AI large model analysis of fault descriptions, extraction of fault features and matching with historical features, it achieves deep integration of structured current data and unstructured text data, effectively making up for the information blind spots of a single data source and significantly improving the accuracy of load drop judgment.

[0048] This invention uses a traditional rule engine to process deterministic logic judgments and output rule confidence scores; it uses an AI large-scale model engine to process fuzzy reasoning and feature matching of unstructured text and output text analysis confidence scores; then, the two are weighted and fused according to mixed judgment weights to generate a comprehensive confidence score. This approach retains the reliability of rule engines in high-deterministic scenarios while leveraging the text understanding capabilities of large-scale models to capture hidden fault features that are difficult for manual rules to cover, thus reducing the false positive and false negative rates.

[0049] Invalid, misjudged, and missed judgments discovered during the verification process are incorporated into a sample library. The AI ​​model automatically analyzes the error characteristics and generates new judgment rule suggestions, supporting automatic deployment or deployment after manual review. Simultaneously, the AI ​​model undergoes incremental fine-tuning weekly, and the fault feature library is updated monthly, with the mixed judgment weights recalculated. This closed-loop mechanism enables the model to continuously adapt to the electricity consumption characteristics and data collection conditions of customers in the power supply area, eliminating the need for frequent manual rule adjustments and reducing long-term operation and maintenance costs.

[0050] Based on a comprehensive confidence level, a differentiated service strategy is automatically generated. High-confidence internal faults automatically send personalized warning SMS messages and work orders to the distribution center manager; suspected automatic power outages only send overdue payment reminders; and medium-confidence events are marked for manual review. This tiered response mechanism effectively avoids service resource misallocation, reduces ineffective inspections and on-site work orders, and improves the precision and economy of power supply services.

[0051] Relying on the existing electricity consumption information collection system, 95598 work order system, and grid service management system of power supply enterprises, no new hardware collection equipment is required. Meanwhile, using a lightweight, finely tuned BERT-Chinese-base model as the core of text parsing, it has low server configuration requirements and can be quickly deployed and promoted in power supply enterprises at all levels nationwide. This transforms existing high-frequency current collection data from simple collection to service support, improving the utilization efficiency of power information system data resources and maximizing data value. Furthermore, grid-based service push based on model judgment results can improve the connection rate and problem resolution rate of power supply service calls, enhancing the utilization efficiency of service resources. Attached Figure Description

[0052] Figure 1 This is a flowchart of the processing method of the present invention. Detailed Implementation

[0053] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0054] Combination Figure 1 The present invention provides the following embodiments:

[0055] Example 1:

[0056] A method for determining sudden load drops within a 0.4kV power customer, comprising:

[0057] The system collects real-time current data and 24-hour historical current data of 0.4kV power customers every 15 minutes from the electricity information collection system. It also collects unstructured fault text data from the 95598 work order system and the grid service management system and performs standardized preprocessing.

[0058] The judgment is based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the current value recovers to non-zero in the future, the judgment result is removed. If it remains at 0, it is retained.

[0059] The system performs initial classification based on customer automatic power outage records and outputs rule confidence scores. Natural language parsing is then performed on the unstructured fault text data to extract fault features and match them with historical fault features, outputting text analysis confidence scores. The rule confidence scores and text analysis confidence scores are then weighted and fused according to a preset mixed judgment weight to generate a comprehensive confidence score. Finally, the system generates a load drop type determination result based on the comprehensive confidence score.

[0060] Based on the comprehensive confidence level and judgment results, a differentiated service response strategy is automatically generated;

[0061] The judgment results are verified by stratified random sampling. The text intelligent parsing engine monitors the 95598 work order system and grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

[0062] Furthermore, the natural language parsing is performed by a text intelligent parsing engine, which is an AI large model engine trained based on deep learning. It uses a lightweight, finely tuned BERT-Chinese-base model as the core of text parsing and is deployed on a cloud inference server. Its pre-trained corpus contains no less than 100,000 historical 95598 work orders, grid service records, and on-site emergency repair reports. The inference latency is no more than 200 milliseconds per message, and it supports a batch text processing rate of no less than 100 messages per second.

[0063] Power fault texts possess strong domain characteristics, containing a large amount of colloquialisms, dialects, and technical terms, making them difficult for general natural language processing models to accurately identify. This paper employs a BERT-Chinese-base model, pre-trained on general Chinese corpora and lightweightly fine-tuned on power domain corpora. This model can learn the terminology distribution and contextual relationships in power fault texts while retaining general language understanding capabilities, thereby improving the accuracy of entity recognition and fault classification. Deploying the model on a cloud-based inference server enables centralized scheduling and elastic scaling of computing resources, avoiding redundant deployment of computing power at the edge. By controlling inference latency to within 200 milliseconds and a batch processing rate of no less than 100 messages per second, it ensures the completion of text parsing for hundreds of thousands of customers within a 15-minute sampling period, without hindering the real-time requirements of the main business process.

[0064] Furthermore, the standardization preprocessing includes:

[0065] The structured current data is cleaned, with a timed trigger cycle of 15 minutes ± 10 seconds. The current data verification threshold is set to 1.2 times the rated current from zero. Data exceeding this range is marked as a metering anomaly, and the timestamp synchronization error does not exceed 30 seconds.

[0066] After cleaning and removing special characters, garbled text, and duplicate content from unstructured fault text data and unifying it into simplified Chinese, the text intelligent parsing engine performs entity recognition to extract core entities such as fault phenomenon, fault location, fault cause, and processing result, and automatically classifies them into internal faults, power grid faults, automatic power outages, or other categories, outputting a classification probability value between zero and one; at the same time, the fault text is converted into a high-dimensional feature vector and stored in the fault feature library.

[0067] If structured current data has timestamp discrepancies or exceeds the measurement range, it will directly lead to distortion in the 24-hour historical backtracking verification. Constraining the timing trigger cycle to 15 minutes ± 10 seconds and the timestamp synchronization error to ≤30 seconds ensures that the 96 sampling points are evenly distributed and strictly aligned within 24 hours. A current verification threshold of 0-1.2 times the rated current can filter out outliers caused by metering device drift or electromagnetic interference. Special characters and garbled text in unstructured text can interfere with the word segmentation and attention weight calculation of the BERT model; cleaning and unifying them to simplified Chinese can reduce character set noise. Transforming unstructured text into structured semantic tags through entity recognition provides a clear vector dimension for subsequent similarity calculations. After mapping the text to high-dimensional feature vectors, semantically similar fault descriptions are closer in the vector space, facilitating the quantification of the correlation between historical faults and current events using cosine similarity.

[0068] Furthermore, the determination based on real-time current data includes:

[0069] For single-phase customers, the current of the corresponding phase at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample; for three-phase customers, the current of all three phases at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample.

[0070] Define a set of valid sampling points within 24 hours, and count the total number of sampling points with a current value of zero. If the total number is zero, proceed to the next step; otherwise, remove the candidate samples.

[0071] Monitor the next 3 sampling cycles, i.e., 45 minutes. If the current at 3 consecutive sampling points is not greater than the preset minimum metering resolution threshold, it is confirmed as a suspected load drop, and the drop time is recorded.

[0072] The confidence level of the rule is calculated, which decreases as the number of valid sampling points decreases.

[0073] 0.01A is the minimum metering resolution of common smart meters. Values ​​below this are considered no-load conditions, and using this as a candidate screening threshold can exclude minor leakage current interference. Defining a set of valid 24-hour sampling points and counting the total number of zero-current sampling points is because if the current was zero within 24 hours, this zeroing might be due to normal customer activity or periodic power outages, not meeting the sudden drop characteristic. For points not sampled, the default current value is ≠ 0, avoiding misjudging communication interruptions as power outages. A subsequent 45-minute validity period is used to exclude momentary interruptions or brief communication packet loss. If the current recovers within 45 minutes, it indicates that the customer side did not experience continuous power loss, avoiding false triggering of subsequent service responses. The initial confidence level decreases as the number of valid sampling points decreases because fewer sampling points result in weaker statistical significance for the 24-hour backtesting, reducing the reliability of evidence of no historical zero current. Therefore, the weight of this evidence needs to be reduced in subsequent fusion stages to encourage the system to accept it cautiously.

[0074] Furthermore, the initial classification based on customer automatic power outage records includes:

[0075] If the customer has no record of automatic power outage, it is directly judged as a suspected internal fault, with a rule confidence level of 0.9;

[0076] If the customer has an automatic power outage record, compare the load drop time with the trip success time: when the load drop time is earlier than the trip success time, it is judged as a suspected internal fault with a rule confidence level of 0.85; when the load drop time is later than or equal to the trip success time, it is judged as a suspected automatic power outage with a rule confidence level of 0.92.

[0077] Automatic power outage records are remote power outage operations initiated by the power supply company's marketing system, and are considered external deterministic events. If a customer has no such record but the current suddenly drops to 0, the cause of the power outage is most likely internal to the customer, and is therefore directly classified as a suspected internal fault with a high confidence level of 0.9. If an automatic power outage record exists, causal inference is made through time-based comparison: if the load drop time is earlier than the trip time, it indicates that the customer had already lost power before the system's remote power outage, and the current drop to zero was caused by an internal fault; if the drop time is later than or equal to the trip time, it indicates that the current drop to zero was directly caused by the system's remote power outage.

[0078] Furthermore, the step of performing natural language parsing on unstructured fault text data to extract fault features and matching them with historical fault features includes:

[0079] The text intelligent parsing engine extracts the feature vectors of all fault texts from the past 12 months of the customer's history, as well as the feature vector of the current sudden drop event;

[0080] Calculate the average similarity between the current sudden drop event and the feature vectors of historical faults. This similarity is measured by the cosine of the angle between the vectors.

[0081] Based on the similarity and the rule confidence, the text analysis confidence is calculated using a nonlinear mapping function.

[0082] Historical fault texts from the same customer contain personalized information such as their electricity usage habits, equipment aging patterns, and common fault modes. Extracting fault texts from the past 12 months strikes a balance between sample sufficiency and timeliness. Using cosine similarity to measure the vector angle between the current event and historical faults eliminates differences in vector magnitudes, focuses on semantic consistency, and accurately characterizes the degree of correlation between fault descriptions in the semantic space. Since cosine similarity and rule confidence differ in their dimensions, value ranges, and physical meanings, direct linear fusion can lead to dimensional conflicts. Normalizing multidimensional evidence to the same probability scale through a nonlinear mapping function makes textual semantic evidence and rule-based judgment evidence comparable, laying a unified mathematical foundation for subsequent weighted fusion.

[0083] Furthermore, the step of weighting and fusing the rule confidence score and the text analysis confidence score according to a preset mixed judgment weight includes:

[0084] Set a mixed judgment weight, with the sum of the two weights being one. The default weight for the traditional rule engine is 0.6, and the weight for the text intelligent parsing engine is 0.4.

[0085] The rule confidence score and the text analysis confidence score are weighted and summed according to a preset mixed judgment weight to obtain the comprehensive confidence score;

[0086] If the overall confidence level is not lower than 0.7, output the final judgment result and the overall confidence level; if the overall confidence level is between 0.5 and 0.7, mark it as pending manual review and push it to the power supply service personnel; if the overall confidence level is lower than 0.5, judge it as data abnormal and remove this result.

[0087] Traditional rule engines rely on physical measurements (current, timestamps) to make deterministic logical judgments, which are objective and interpretable when the data is complete and the records are clear. Text intelligent parsing engines rely on semantic reasoning and provide supplementary evidence when the rule boundaries are ambiguous (such as missing timestamps, abnormal collection, or new types of load interference).

[0088] The default weight of the traditional rule engine is 0.6, and the weight of the text intelligent parsing engine is 0.4, because the comparison between the current returning to zero and time is hard physical evidence, and its reliability has a higher priority than semantic inference.

[0089] Setting two threshold levels of 0.7 and 0.5 to form a three-level output is to achieve a balance between automation efficiency and judgment security: a comprehensive confidence level ≥ 0.7 indicates that the rules and textual evidence corroborate each other, which can directly drive automatic order dispatch; the range of 0.5-0.7 indicates that there is discrepancy or uncertainty in the evidence, and the risk of mis-dispatch if orders are dispatched directly is high, so it is marked for manual review; a value below 0.5 indicates that the rules and textual evidence contradict each other or the data quality is extremely poor, and direct rejection can prevent noisy samples from contaminating subsequent verification statistics and model training.

[0090] Furthermore, the strategy for automatically generating differentiated service responses based on comprehensive confidence levels and judgment results includes:

[0091] When a suspected internal fault is detected, a personalized warning SMS is generated and automatically pushed to the customer. At the same time, the customer information and the time of the sudden incident are pushed to the mobile terminal of the area manager. If there is no customer call within 1 hour and the overall confidence level is not lower than 0.85, a voice callback is automatically triggered.

[0092] When a suspected automatic power outage is detected, an overdue payment reminder SMS will be automatically generated and a manual consultation channel will be maintained, but on-site inspections will not be arranged.

[0093] When marked as pending manual review, the request is sent to the power supply service operator for manual verification before a decision is made on whether to dispatch the order.

[0094] Different judgment results correspond to differentiated service resource allocation logic. Suspected internal faults directly affect customer electricity safety and power supply reliability, requiring rapid customer outreach via SMS, while simultaneously enabling rapid grid-side response from the district manager, forming a dual-channel system of customer self-inspection and grid-based emergency repair to shorten fault recovery time. When there is no call within one hour and the confidence level is high, a voice callback is triggered because high-confidence internal faults, if not proactively reported by the customer, may pose hidden service risks such as elderly people living alone, communication difficulties, or customers being unaware of the issue; proactive confirmation is necessary to avoid service blind spots. Suspected automatic power outages fall under electricity billing management, not equipment failure. Dispatching the district manager for on-site inspection would incur unnecessary manpower and vehicle costs; simply sending a payment reminder is sufficient. In scenarios awaiting manual review, directly assigning orders automatically has a significantly higher probability of misassignment than in high-confidence scenarios. Pushing orders to a human agent for secondary verification can effectively reduce service error rates and complaint risks while controlling manpower costs.

[0095] Furthermore, this also includes model closed-loop optimization:

[0096] Invalid judgment data, misjudged samples and missed judgment samples found during the verification process are included in the sample library. The text intelligent parsing engine automatically analyzes their current data features and text features, extracts misjudgment reason tags and generates error feature vectors.

[0097] New judgment rule suggestions are automatically generated based on the error sample library. The format is to output the corresponding judgment result and attach the confidence level if a specific condition is met.

[0098] The rule review mechanism is as follows: if the rule confidence score is not lower than 0.9, it is automatically deployed to the rule engine; if the rule confidence score is between 0.7 and 0.9, it is pushed to the algorithm engineer for review; if the rule confidence score is lower than 0.7, it is marked as pending observation.

[0099] Each week, new validation samples and error samples are added to the training set. The text intelligent parsing engine is incrementally fine-tuned and the mixed judgment weights are recalculated to maximize the overall accuracy. The fault feature library is updated monthly.

[0100] Misjudged and missed cases expose blind spots in the rules; vectorizing and storing these samples in an error sample library helps pinpoint weaknesses in the model. Rule suggestions are automatically generated based on this error sample library, and undergo a three-level review process to balance iterative efficiency and system stability. Weekly incremental fine-tuning and monthly feature library updates ensure the model continuously adapts to evolving customer electricity consumption behavior, preventing concept drift that could distort similarity calculations and maintaining long-term stability in judgment accuracy.

[0101] Example 2:

[0102] A load surge detection system for 0.4kV power customers, comprising:

[0103] The data acquisition and preprocessing module is configured to collect real-time current data and 24-hour historical current data of 0.4kV power customers from the electricity information acquisition system at a 15-minute interval, and simultaneously collect unstructured fault text data from the 95598 work order system and the grid service management system, and perform standardized preprocessing.

[0104] The load drop initial judgment module is configured to make judgments based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the current value returns to non-zero in the future, the judgment result is removed; if it remains at 0, it is retained.

[0105] The load drop type fusion determination module includes a rule engine initial classification unit, a text intelligent parsing engine, and a weighted fusion unit. The rule engine initial classification unit is configured to perform initial classification based on customer automatic power outage records and output rule confidence scores. The text intelligent parsing engine is configured to perform natural language parsing on the unstructured fault text data to extract fault features and match them with historical fault features, outputting text analysis confidence scores. The weighted fusion unit is configured to weight and fuse the rule confidence scores and text analysis confidence scores according to preset mixed determination weights to generate a comprehensive confidence score, and generate the final load drop type determination result based on the comprehensive confidence score. The service response module is configured to automatically generate differentiated service response strategies based on the comprehensive confidence score and the determination result.

[0106] The result verification module is configured to verify the judgment results using a stratified random sampling method. The text intelligent parsing engine monitors the 95598 work order system and the grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments, and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

[0107] Application example:

[0108] A method for determining a sudden load drop within a 0.4kV power customer's premises includes the following steps:

[0109] Step S1: Automatic acquisition and standardized preprocessing of multi-source data

[0110] 1.1 Structured Current Data Acquisition

[0111] Using the existing electricity consumption information collection system, real-time current data and historical current data from the previous 24 hours are automatically extracted from 0.4kV power customers at 15-minute intervals. The timed trigger cycle is controlled within 15 minutes ± 10 seconds, and the current data verification threshold is set to 0 to 1.2 times the rated current. Data points exceeding this range are marked as metering anomalies and removed. The timestamp synchronization error of each data source does not exceed 30 seconds to ensure time alignment of multi-source data.

[0112] The data fields collected include: customer number, collection time, and three-phase current. , , ), neutral current, power outage indicator, etc.

[0113] 1.2 Unstructured Fault Text Data Acquisition

[0114] It collects fault repair work orders and consultation work orders from the 95598 work order system in real time; and synchronizes daily with the grid service management system's follow-up records, on-site repair reports, and customer feedback records. Collected fields include: work order number, customer number, acceptance time, fault description, processing procedure, processing result, and handler.

[0115] 1.3 Standardized Preprocessing

[0116] The structured current data is cleaned to remove duplicates, missing values, and outliers exceeding the threshold.

[0117] Unstructured faulty text data is cleaned to remove special characters, garbled text, and duplicate content, and then uniformly converted to Simplified Chinese. Subsequently, an intelligent text parsing engine performs entity recognition and feature extraction.

[0118] Entity recognition: Extracts core entities such as fault phenomena (e.g., tripping, no power, smoke), fault location (e.g., meter box, circuit breaker, socket), fault cause, and handling results (e.g., replacing the circuit breaker, restoring power). The output format is as follows:

[0119]

[0120] Fault Classification: Automatically classify text into four categories: "Internal Fault," "Power Grid Fault," "Automatic Power Outage," and "Other," and output the classification probability:

[0121]

[0122] Feature vectorization: Converting faulty text into 768-dimensional feature vectors.

[0123]

[0124] Stored in the fault feature database for subsequent historical feature matching.

[0125] Step S2: Initial Judgment of Sudden Load Drop

[0126] 2.1 Candidate Sample Screening

[0127] Read the current value at the current sampling time and filter it according to the following threshold:

[0128] Single-phase customers: If the corresponding phase current is not greater than the preset minimum metering resolution threshold, i.e. If so, it will be included in the candidate sample;

[0129] Three-phase customers: If the three-phase current simultaneously meets the requirements If it is, then it enters the candidate sample.

[0130] 2.2 24-hour historical backtracking verification

[0131] Using the moment when the current is 0 as the time interface, go back 24 hours and define the set of valid sampling points within 24 hours:

[0132]

[0133] Count the total number of sampling points in this set where the current value is zero:

[0134]

[0135] in This is an indicator function. If... If the current returns to zero, it indicates that the customer has not experienced a zero-current situation in the past 24 hours, and we proceed to the next step; otherwise, the candidate sample is removed.

[0136] For sampling points where no data was collected within 24 hours, the current value is assumed to be non-zero to avoid misjudging communication interruption or missing data collection as a power outage.

[0137] 2.3 Verification of Sudden Drop Survival and Calculation of Initial Confidence Level

[0138] For candidate samples that pass the historical backtracking verification, continue monitoring for the next 3 sampling periods (i.e., 45 minutes). If the current at 3 consecutive sampling points is not greater than the preset minimum metering resolution threshold, it is confirmed as a suspected load drop, and the time of the first drop is recorded. .

[0139] Simultaneously calculate the initial confidence level. The confidence level increases with the number of valid sampling points within 24 hours. The decrease is linear, and the calculation formula is:

[0140]

[0141] Step S3: Determine the type of load drop

[0142] This step is completed collaboratively by the rule engine's initial classification unit, the text intelligent parsing engine, and the weighted fusion unit.

[0143] 3.1 Preliminary Classification of Traditional Rule Engines

[0144] Retrieve the customer's daily automatic power outage (power outage due to unpaid bills) operation records:

[0145] If the customer has no record of automatic power outages, it is directly determined as a suspected internal fault, and the rule confidence level is output: ;

[0146] If the customer has records of automatic power outages, compare the time of the load drop. Time of successful trip :

[0147] when At that time, it was determined to be a suspected internal fault:

[0148]

[0149] when At that time, it was determined to be a suspected automatic power outage:

[0150]

[0151] 3.2 Text Intelligent Parsing Engine Feature Matching

[0152] The text intelligent parsing engine extracts feature vectors of all fault texts from the past 12 months for this customer: and the feature vector of the current sudden drop event Calculate the average similarity between the current sudden drop event and the feature vectors of historical faults. This similarity is measured by the cosine of the angle between the vectors.

[0153]

[0154] Based on this similarity and the aforementioned initial confidence level Rule confidence The confidence level of text analysis is calculated using a nonlinear mapping function. :

[0155]

[0156] in For the sigmoid function, This is the weight matrix. These are bias vectors, all learned from the training set.

[0157] 3.3 Weighted Fusion and Comprehensive Confidence Determination

[0158] Set mixed decision weights and ,satisfy:

[0159]

[0160] Default traditional rule engine weights Text intelligent parsing engine weight .

[0161] The rule confidence score and the text analysis confidence score are weighted and summed according to a preset mixed decision weight to obtain the comprehensive confidence score:

[0162]

[0163] Classification based on overall confidence level:

[0164] like Output the final judgment result and overall confidence level;

[0165] like Marked as "Pending manual review," it is sent to the power supply service personnel;

[0166] like The data was deemed abnormal, and the result was removed.

[0167] 3.4 Refinement of Fault Subtypes

[0168] When it is determined to be a suspected internal fault and At that time, the text intelligent parsing engine automatically predicts the fault subtype based on historical fault characteristics, including circuit breaker tripping, residual current device operation, fault in the line below the meter, and short circuit in internal equipment, and outputs the two subtypes with the highest probability as a reference for the area manager to troubleshoot on site.

[0169] Step S4: Hierarchical and Classified Intelligent Service Response

[0170] Based on the overall confidence level and the final judgment result, a differentiated service response strategy is automatically generated:

[0171] Suspected internal fault: Generate a personalized alert SMS and automatically push it to the customer; simultaneously, customer information, sudden occurrence time, and predicted fault subtype are pushed to the area manager's mobile terminal, automatically dispatching an inspection work order. If no customer calls within 1 hour and It automatically triggers a voice callback.

[0172] Suspected automatic power outage: Automatically generate an overdue payment reminder SMS, including the amount owed, payment method, power restoration process, and retain a manual consultation channel; no on-site inspection will be arranged.

[0173] Pending manual review: The request is sent to the power supply service operator, who will verify the nature of the customer's power outage before deciding whether to dispatch the order.

[0174] Step S5: Multi-dimensional result verification

[0175] The judgment results are systematically verified using stratified random sampling. This mainly includes sample selection, automatic verification, and accuracy statistics.

[0176] The sample selection process is as follows: 10% of the results of the day's judgment are selected each day, of which 60% are high-confidence samples, 30% are medium-confidence samples, and 10% are samples pending review.

[0177] The automatic verification process involves a text intelligent parsing engine that monitors the 95598 work order system and the grid service system in real time, automatically matching subsequent new work orders and follow-up records corresponding to the sudden event. It then parses the fault descriptions and processing results in the work orders to extract the actual fault type. , and the model's judgment results Perform comparisons and generate a verification report:

[0178] Correct judgment: ;

[0179] Misjudgment: Record the reasons for misjudgment;

[0180] Missed detection: A fault actually occurs but is not detected by the model; record the missed detection feature.

[0181] The accuracy statistics are as follows:

[0182] Internal fault identification accuracy calculation formula:

[0183]

[0184] The false negative rate is calculated using the following formula:

[0185]

[0186] in The number of correctly identified internal fault samples. This represents the number of samples that were misclassified as internal faults. This represents the number of samples that actually experienced internal faults but were not identified by the model.

[0187] Step S6: Model closed-loop optimization

[0188] Invalid judgment data, misjudged samples, and missed judgment samples discovered during the verification process are added to the sample library, and dynamic closed-loop optimization is performed:

[0189] Error Feature Analysis: The intelligent text parsing engine automatically analyzes the current data features and text features of misjudged and missed samples, extracts the labels of reasons for misjudgment (such as no occupancy, abnormal data collection, incorrect basic information, new types of load interference, etc.), and generates error feature vectors. Store in the error sample library.

[0190] Automatic rule generation and review: New decision rule suggestions are automatically generated based on the error sample library, in the format "IF [condition] THEN [decision result], confidence level [C]". The rule review mechanism is as follows: if the rule confidence level... Then it will be automatically deployed to the rules engine; if Then it will be pushed to the algorithm engineer for review; if Then mark it as to be observed.

[0191] Incremental tuning of the engine: New validation samples and error samples are added to the training set every week to incrementally fine-tune the text intelligent parsing engine and update the internal weights of entity recognition and fault classification.

[0192] Weight and feature library update: Recalculate the hybrid decision weights and To maximize overall accuracy; update the fault feature database monthly, adding newly discovered fault types and features, and cleaning up outdated features.

[0193] For certain special cases, the following measures are taken: Large-scale power outage correlation determination: When more than 30% of customers in the same transformer area are simultaneously determined to have experienced a load surge, the text intelligent parsing engine automatically analyzes recent grid maintenance records and fault repair work orders for that transformer area. If the keywords "transformer area power outage" or "line maintenance" are matched, the load surge type for all customers is automatically corrected to a public grid fault. New load interference handling: For customers equipped with distributed photovoltaic, residential energy storage, or charging piles, the text intelligent parsing engine automatically identifies new load current characteristics and dynamically adjusts the determination threshold and mixed determination weight, ensuring that the weights of both the traditional rule engine and the text intelligent parsing engine are equal. To adapt to the characteristics of bidirectional power flow and intermittent load.

[0194] Extreme Weather Emergency Response: When issuing warnings for extreme weather such as heavy rain or typhoons, the system automatically increases the priority of internal fault detection and the frequency of work order pushes to the area manager, prioritizing the handling of fault warnings in high-risk areas.

[0195] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for determining sudden load drops within a 0.4kV power customer, characterized in that, include: The system collects real-time current data and 24-hour historical current data of 0.4kV power customers every 15 minutes from the electricity information collection system. It also collects unstructured fault text data from the 95598 work order system and the grid service management system and performs standardized preprocessing. The judgment is based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the subsequent current value returns to non-zero, the judgment result is discarded; if it remains zero, it is retained. The system performs initial classification based on customer automatic power outage records and outputs rule confidence scores. Natural language parsing is then performed on the unstructured fault text data to extract fault features and match them with historical fault features, outputting text analysis confidence scores. The rule confidence scores and text analysis confidence scores are then weighted and fused according to a preset mixed judgment weight to generate a comprehensive confidence score. Finally, the system generates a load drop type determination result based on the comprehensive confidence score. Based on the comprehensive confidence level and judgment results, a differentiated service response strategy is automatically generated; The judgment results are verified by stratified random sampling. The text intelligent parsing engine monitors the 95598 work order system and grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

2. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The natural language parsing is performed by a text intelligent parsing engine, which is an AI large model engine trained based on deep learning. It uses a lightweight, finely tuned BERT-Chinese-base model as the core of text parsing and is deployed on a cloud inference server. Its pre-trained corpus contains no less than 100,000 historical 95598 work orders, grid service records, and on-site emergency repair reports. The inference latency is no more than 200 milliseconds per message, and it supports a batch text processing rate of no less than 100 messages per second.

3. The method for determining sudden load drops within a 0.4kV power customer as described in claim 2, characterized in that, The standardized preprocessing includes: The structured current data is cleaned, with a timed trigger cycle of 15 minutes ± 10 seconds. The current data verification threshold is set to 1.2 times the rated current from zero. Data exceeding this range is marked as a metering anomaly, and the timestamp synchronization error does not exceed 30 seconds. After cleaning and removing special characters, garbled text, and duplicate content from unstructured fault text data and unifying it into simplified Chinese, the text intelligent parsing engine performs entity recognition to extract core entities such as fault phenomenon, fault location, fault cause, and processing result, and automatically classifies them into internal faults, power grid faults, automatic power outages, or other categories, outputting a classification probability value between zero and one; at the same time, the fault text is converted into a high-dimensional feature vector and stored in the fault feature library.

4. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The determination based on real-time current data includes: For single-phase customers, the current of the corresponding phase at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample; for three-phase customers, the current of all three phases at the current sampling time is not greater than the preset minimum metering resolution threshold when the current is included in the candidate sample. Define a set of valid sampling points within 24 hours, and count the total number of sampling points with a current value of zero. If the total number is zero, proceed to the next step; otherwise, remove the candidate samples. Monitor the next 3 sampling cycles, i.e., 45 minutes. If the current at 3 consecutive sampling points is not greater than the preset minimum metering resolution threshold, it is confirmed as a suspected load drop, and the drop time is recorded. The confidence level of the rule is calculated, which decreases as the number of valid sampling points decreases.

5. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The initial classification based on customer automatic power outage records includes: If the customer has no record of automatic power outage, it is directly judged as a suspected internal fault, with a rule confidence level of 0.9; If the customer has an automatic power outage record, compare the load drop time with the trip success time: when the load drop time is earlier than the trip success time, it is judged as a suspected internal fault with a rule confidence level of 0.85; when the load drop time is later than or equal to the trip success time, it is judged as a suspected automatic power outage with a rule confidence level of 0.

92.

6. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The step of performing natural language parsing on unstructured fault text data to extract fault features and matching them with historical fault features includes: The text intelligent parsing engine extracts the feature vectors of all fault texts from the past 12 months of the customer's history, as well as the feature vector of the current sudden drop event; Calculate the average similarity between the current sudden drop event and the feature vectors of historical faults. This similarity is measured by the cosine of the angle between the vectors. Based on the similarity and the rule confidence, the text analysis confidence is calculated using a nonlinear mapping function.

7. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The step of weighting and fusing rule confidence and text analysis confidence according to a preset mixed judgment weight includes: Set a mixed judgment weight, with the sum of the two weights being one. The default weight for the traditional rule engine is 0.6, and the weight for the text intelligent parsing engine is 0.

4. The rule confidence score and the text analysis confidence score are weighted and summed according to a preset mixed judgment weight to obtain the comprehensive confidence score; If the overall confidence level is not lower than 0.7, output the final judgment result and the overall confidence level; if the overall confidence level is between 0.5 and 0.7, mark it as pending manual review and push it to the power supply service personnel; if the overall confidence level is lower than 0.5, judge it as data abnormal and remove this result.

8. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, The strategy for automatically generating differentiated service responses based on comprehensive confidence level and judgment results includes: When a suspected internal fault is detected, a personalized warning SMS is generated and automatically pushed to the customer. At the same time, the customer information and the time of the sudden incident are pushed to the mobile terminal of the area manager. If there is no customer call within 1 hour and the overall confidence level is not lower than 0.85, a voice callback is automatically triggered. When a suspected automatic power outage is detected, an overdue payment reminder SMS will be automatically generated and a manual consultation channel will be maintained, but on-site inspections will not be arranged. When marked as pending manual review, the request is sent to the power supply service operator for manual verification before a decision is made on whether to dispatch the order.

9. The method for determining sudden load drops within a 0.4kV power customer as described in claim 1, characterized in that, It also includes model closed-loop optimization: Invalid judgment data, misjudged samples and missed judgment samples found during the verification process are included in the sample library. The text intelligent parsing engine automatically analyzes their current data features and text features, extracts misjudgment reason tags and generates error feature vectors. New judgment rule suggestions are automatically generated based on the error sample library. The format is to output the corresponding judgment result and attach the confidence level if a specific condition is met. The rule review mechanism is as follows: if the rule confidence score is not lower than 0.9, it is automatically deployed to the rule engine; if the rule confidence score is between 0.7 and 0.9, it is pushed to the algorithm engineer for review; if the rule confidence score is lower than 0.7, it is marked as pending observation. Each week, new validation samples and error samples are added to the training set. The text intelligent parsing engine is incrementally fine-tuned and the mixed judgment weights are recalculated to maximize the overall accuracy. The fault characteristic database is updated monthly.

10. A system for determining sudden load drops within a 0.4kV power customer's premises, characterized in that, include: The data acquisition and preprocessing module is configured to collect real-time current data and 24-hour historical current data of 0.4kV power customers from the electricity information acquisition system at a 15-minute interval, and simultaneously collect unstructured fault text data from the 95598 work order system and the grid service management system, and perform standardized preprocessing. The load drop initial judgment module is configured to make judgments based on real-time current data. When the real-time current value is 0, the historical current data of the previous 24 hours is traced back to that time as the time interface. If there are no sampling points with a current value of 0 within 24 hours and the current value of the sampling points without data collection is not equal to 0 by default, it is judged as a suspected load drop and the time of the first drop is recorded. If the subsequent current value returns to non-zero, the judgment result is discarded; if it remains zero, it is retained. The load drop type fusion determination module includes a rule engine initial classification unit, a text intelligent parsing engine, and a weighted fusion unit; the rule engine initial classification unit is configured to perform initial classification based on customer automatic power outage records and output rule confidence scores; The text intelligent parsing engine is configured to perform natural language parsing on the unstructured fault text data to extract fault features and match them with historical fault features, and output text analysis confidence scores; the weighted fusion unit is configured to perform weighted fusion of the rule confidence scores and text analysis confidence scores according to preset mixed judgment weights to generate a comprehensive confidence score, and generate the final load drop type judgment result based on the comprehensive confidence score. The service response module is configured to automatically generate differentiated service response strategies based on the comprehensive confidence level and the judgment result. The result verification module is configured to verify the judgment results using a stratified random sampling method. The text intelligent parsing engine monitors the 95598 work order system and the grid service system in real time, automatically matches the subsequent work orders and follow-up records corresponding to the sudden drop event, parses the fault description and processing results in the work orders to extract the actual fault type, compares it with the model judgment results, generates a verification report containing correct judgments, misjudgments, and missed judgments, and calculates the internal fault identification accuracy rate and missed judgment rate.

Images