Anti-creeping protection system based on embedded electric energy metering box

Abstract

The invention relates to the technical field of electric energy metering protection, and discloses an anti-creeping protection system based on an embedded electric energy metering box. A data acquisition module of the system acquires real-time electric energy parameter data and historical electric leakage event data of an electric energy metering box; the model training module trains an artificial intelligence detection model based on the collected data; the electric leakage detection module performs real-time electric leakage detection by using the trained model; the event analysis module performs feature extraction and constraint analysis on the detected electric leakage event; the risk prediction module predicts the future electric leakage risk according to the analysis result; the optimization module continuously optimizes and iterates the artificial intelligence detection model according to the prediction result; and the protection execution module executes a leakage protection action based on the optimization model after model optimization. According to the system, through cooperative work of multiple modules, accurate detection, deep analysis, risk prediction and intelligent protection of electric energy metering box electric leakage faults are realized, the reliability and the intelligent level of electric leakage protection are effectively improved, and potential safety hazards are reduced.

Description

Technical Field

[0001] This invention relates to the field of power metering and protection technology, specifically to a leakage protection system based on an embedded power metering box. Background Technology

[0002] During the operation of power systems, electricity metering boxes, as crucial equipment for electricity metering and management, are widely distributed in various scenarios such as residential communities, commercial buildings, and industrial plants. With the continuous increase in electricity load and the growing complexity of the electricity environment, the probability of leakage faults in electricity metering boxes is gradually rising. Once a leakage occurs, it may not only lead to inaccurate metering data, affecting the reasonable billing and allocation of electricity resources, but may also cause electric shock accidents, electrical fires, and other safety hazards, posing a serious threat to the safety of people and property. Most commercially available leakage current protection solutions for electricity metering boxes rely on traditional electrical components such as residual current devices (RCDs) and circuit breakers. These solutions typically use fixed current or voltage thresholds for judgment, triggering protection when the detected electrical parameters exceed the preset threshold. However, in actual electricity usage scenarios, the operating status of electricity metering boxes is affected by various factors, such as the type and number of electrical devices, fluctuations in grid voltage, and changes in ambient temperature and humidity, resulting in diverse and complex manifestations of leakage faults. Traditional protection solutions based on fixed thresholds are ill-suited to this complex operating environment, often leading to false or missed trips. For example, during peak electricity consumption periods, grid voltage may experience brief fluctuations, causing a sudden increase in current within the electricity metering box. In such cases, traditional protection solutions may misinterpret this as a leakage fault, triggering unnecessary power outages and disrupting normal electricity use. Conversely, for weak, intermittent leakage faults, where electrical parameter changes do not reach the preset threshold, traditional protection solutions may fail to detect them in time, leaving potential safety hazards. Traditional leakage current protection (RCD) solutions lack effective analysis and risk prediction capabilities for leakage events. When a leakage fault occurs, staff often need to conduct on-site investigations to determine the cause and location of the fault, a time-consuming and labor-intensive process that hinders rapid response and handling. Furthermore, the inability to deeply mine and analyze historical leakage event data, and the inability to predict future leakage risks based on past operating conditions, leaves RCD protection in a reactive state, failing to fundamentally reduce the probability of leakage faults. With the increasing application of artificial intelligence (AI) technology in power systems, using AI models to analyze and process operational data from electricity metering boxes to achieve more accurate and intelligent RCD protection has become an industry trend. However, current related technical solutions are still in the exploratory stage, and a complete RCD protection system integrating data acquisition, model training, real-time detection, event analysis, risk prediction, model optimization, and protection execution has not yet been formed. This system cannot fully meet the high-precision, high-reliability, and intelligent requirements of electricity metering box RCD protection in practical applications. Summary of the Invention

[0003] The purpose of this invention is to provide a leakage protection system based on an embedded power metering box to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a leakage protection system based on an embedded power metering box, the system comprising: The data acquisition module is used to collect real-time power parameter data and historical leakage event data from the power metering box. The model training module trains an artificial intelligence detection model based on the data collected by the data acquisition module. The leakage current detection module performs real-time leakage current detection on the power metering box based on a trained artificial intelligence detection model. The event analysis module performs feature extraction and constraint analysis on leakage events detected by the leakage detection module in real time; The risk prediction module predicts future leakage risks based on the analysis results of the event analysis module. The optimization module continuously optimizes and iterates the artificial intelligence detection model based on the prediction results of the risk prediction module. The protection execution module performs leakage protection actions based on the optimized artificial intelligence detection model after the optimization module completes the optimization iteration.

[0005] Preferably, the data acquisition module includes: Real-time current fluctuation data and voltage anomaly data are obtained from the embedded sensors in the power metering box; Extract the event occurrence time and event duration from historical leakage current event data; Real-time current fluctuation data and voltage anomaly data are integrated with historical leakage event data to form a training dataset.

[0006] Preferably, the model training module includes: The leakage current range is selected based on the training dataset integrated by the data acquisition module. Calculate the first and second feature parameters of each historical leakage event in the training dataset; The first feature parameter is used as the input and the second feature parameter is used as the output to form a training set. The training set is used to train the artificial intelligence detection model.

[0007] Preferably, the leakage current detection module includes: The artificial intelligence detection model trained by the model training module monitors the real-time power parameter data of the power metering box. The leakage event detection is triggered when the real-time current fluctuation data exceeds a preset threshold. The generated leakage current event detection report includes the event type and the event severity level.

[0008] Preferably, the event analysis module includes: Select a baseline anomaly point from the leakage event detection report generated by the leakage current detection module; Tracing the historical work sequence within the target detection period corresponding to the baseline anomaly point; Based on the traversal order corresponding to the anomaly type of the baseline anomaly point, the sequence features in the historical working sequence are traversed sequentially. When the traversed sequence features match the trigger sequence features, the corresponding replay mechanism and verification rules are obtained; The replay mechanism controls the replay of historical working sequences and executes verification rules to generate feature analysis results.

[0009] Preferably, the event analysis module selects benchmark anomaly points from the leakage event detection report, including: Determine the primary and secondary objectives; The earlier occurrence of the first objective and the second objective is taken as the baseline outlier. The determination of the first objective is based on the fact that the sum of the cost values ​​of the first few events in the leakage current event detection report is close to the cost threshold; The second objective is determined based on the fact that the first anomaly type in the leakage event detection report is the same as the standard anomaly type.

[0010] Preferably, the risk prediction module includes: The feature analysis results generated by the event analysis module are used to obtain feature description vectors; Determine the prediction strategy corresponding to the feature description vector from the prediction strategy library; The probability of future leakage risks and the scope of their impact are predicted based on the predictive strategy. The generated risk forecast report includes the predicted time of occurrence and the predicted risk level.

[0011] Preferably, the optimization module includes: The optimized parameter set is obtained by parsing the risk prediction report generated by the risk prediction module. Determine the optimization iteration strategy corresponding to the optimization parameter set from the optimization strategy library; The model structure of the artificial intelligence detection model is adjusted and the parameters are updated based on an optimization iteration strategy. Output the optimized version of the artificial intelligence detection model.

[0012] Preferably, the protection execution module includes: The optimized AI detection model version output by the optimization module is used to monitor the real-time status of the power metering box. The protection mechanism is triggered when the detected leakage risk exceeds the execution threshold. The protection mechanism includes current cut-off commands or alarm activation commands; The protection execution log includes the execution time and execution result.

[0013] Preferably, the system further includes: a visualization monitoring module, which generates a visualization monitoring model of the optimized artificial intelligence detection model to assist in work monitoring; The visualization monitoring module includes: multimodal work information of an artificial intelligence detection model integrated and optimized based on the visualization template corresponding to the personnel profile of the management personnel; detection of multiple monitoring timing events occurring in the visualization monitoring model; generation of a monitoring task quick selection table based on each monitoring timing event; and assistance to management personnel in quickly selecting monitoring tasks from the monitoring task quick selection table to perform monitoring operations.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This embedded energy metering box leakage protection system provides a more comprehensive and intelligent solution for leakage protection of energy metering boxes through the collaborative work of multiple modules. From the data acquisition perspective, the data acquisition module can not only collect real-time energy parameter data but also collect historical leakage event data. Compared to traditional solutions that rely solely on real-time data for protection judgment, the rich data sources provide more comprehensive information support for subsequent model training and analysis. Real-time energy parameter data reflects the current operating status of the energy metering box, ensuring timely monitoring of potential leakage risks. Historical leakage event data contains characteristic information of leakage faults in different scenarios, helping the artificial intelligence detection model better learn and identify various leakage fault modes, reducing misjudgments or missed judgments due to insufficient understanding of fault modes. The model training module trains an AI detection model based on collected multi-dimensional data, giving the model stronger adaptability and generalization capabilities. Traditional protection schemes rely on fixed thresholds and cannot be adjusted according to different operating environments and fault types. However, the AI ​​detection model trained through a data-driven approach can continuously learn the operating patterns of the power metering box under different power loads, grid fluctuations, and environmental conditions, as well as the characteristics of different types of leakage faults, thereby more accurately identifying leakage faults. Even in complex and changing operating scenarios, such as current fluctuations caused by frequent switching of electrical equipment and unstable grid voltage, the model can accurately distinguish between normal operating fluctuations and leakage faults based on its learned knowledge, effectively reducing the incidence of malfunctions and leakage trips. The leakage current detection module performs real-time leakage current detection based on a trained artificial intelligence detection model, enabling rapid response to leakage current faults. Compared to traditional solutions that require waiting for electrical parameters to exceed a fixed threshold before triggering detection, the artificial intelligence model can continuously analyze and calculate real-time data to capture abnormal signals in the early stages of a leakage current fault, identifying potential leakage risks in advance. This allows more time for subsequent protective actions, preventing the leakage current fault from escalating and reducing safety accidents and property damage caused by leakage current. The event analysis module performs real-time feature extraction and constraint analysis on detected leakage events, enabling in-depth mining of key information. In traditional solutions, after a leakage fault occurs, staff struggle to quickly obtain specific fault characteristics, such as the timing of the leakage, the trend of leakage current changes, and its correlation with other electrical parameters, leading to low efficiency in fault diagnosis and handling. However, the event analysis module, through feature extraction of leakage events, clearly presents the various attributes of the fault. Combined with constraint analysis, it eliminates interference from irrelevant factors, accurately pinpointing the possible causes and scope of impact of the fault. This provides significant assistance to staff in quickly locating the fault and developing targeted solutions, greatly shortening fault handling time and reducing the impact of the fault on normal power supply. The risk prediction module predicts future leakage risks based on event analysis results, changing the traditional reactive approach to leakage protection. Through comprehensive analysis of historical leakage event characteristics and current operating status, the risk prediction module can identify potential risk points in the operation of the electricity metering box, such as the high incidence trend of leakage faults during specific electricity consumption periods and under specific environmental conditions, and issue early warnings. This enables staff to proactively take preventive measures, such as conducting early maintenance on high-risk equipment and adjusting power consumption plans, reducing the probability of leakage faults at the source and improving the safety and stability of the electricity metering box operation. The optimization module continuously optimizes and iterates the AI ​​detection model based on risk prediction results, ensuring that the model maintains high detection accuracy and adaptability. As the operating environment of the electricity metering box changes, electrical equipment is updated, and leakage fault types evolve, the initially trained model may experience a decrease in detection accuracy. The optimization module continuously adjusts and improves the model using new operational data and risk prediction information, updating the model's parameters and fault identification rules. This enables the model to adapt to new operating scenarios and fault modes, preventing performance degradation due to model aging and ensuring the long-term stable and reliable operation of the leakage protection system. After model optimization and iteration, the protection execution module executes leakage current protection actions based on the optimized model, further improving the accuracy and effectiveness of the protection actions. Because the optimized model can more accurately identify leakage faults, the protection execution module can more accurately determine whether protection operations are necessary and what level of protection operation to execute when triggering protection actions. For example, if a serious leakage fault is confirmed, the power supply can be cut off in time to prevent the accident from escalating; while if a minor, controllable leakage risk is detected, measures such as adjusting operating parameters can be taken to reduce unnecessary power outages and ensure the continuity of power supply. Simultaneously, the protection execution module forms a closed-loop workflow with the preceding modules, enabling the entire leakage current protection system to continuously improve itself based on actual operating conditions. This achieves intelligent management of the entire process from data acquisition, model training, detection and analysis, risk prediction, model optimization to protection execution, comprehensively improving the overall level of leakage current protection in the power metering box and providing strong support for the safe and stable operation of the power system. Attached Figure Description

[0015] Figure 1 This is a timing diagram of the leakage protection system based on an embedded power metering box described in this invention. Figure 2 This is a flowchart illustrating the working principle of the data acquisition module. Figure 3 This is a flowchart illustrating the working principle of the model training module. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a leakage current protection system based on an embedded power metering box. The system acquires real-time power parameters and historical leakage event data from the power metering box through a data acquisition module. A model training module trains an artificial intelligence detection model based on the acquired data. A leakage current detection module monitors leakage current in real time using the trained model. An event analysis module extracts features and performs constraint analysis on detected leakage events. A risk prediction module predicts future leakage risks based on the analysis results. An optimization module continuously optimizes the artificial intelligence detection model based on the prediction results. A protection execution module executes leakage current protection actions after model optimization. The system is deployed embedded in the power metering box to achieve real-time, intelligent leakage current protection.

[0018] Example 1: See Figure 2 and Figure 3 The data acquisition module is directly coupled to the electrical circuit of the power metering box through an embedded sensor network. These sensors capture instantaneous changes in current and voltage using high-frequency sampling. Current fluctuation data is acquired through a Hall effect sensor with a sampling frequency of 10kHz, which can accurately record leakage current changes at the microampere level. Voltage anomaly data is monitored using a differential measurement circuit to monitor the potential difference between the live and neutral wires, simultaneously recording waveform distortion and instantaneous drops. Historical leakage event data is extracted from the metering box's non-volatile memory and stored in a time-series log format. Each record contains a precise timestamp of the event occurrence and millisecond-level precision data of the event duration. The data acquisition module preprocesses the real-time and historical data. The real-time data undergoes filtering, noise reduction, and baseline correction, while the historical data is deduplicated and time-aligned. Finally, the two types of data are integrated into a structured training dataset, which is stored in a time-series format and includes current waveform segments, voltage anomaly flags, and event duration labels.

[0019] After loading the training dataset, the model training module first standardizes the data, converting current fluctuation data into per-unit values ​​and normalizing voltage anomaly data to the 0-1 range. Then, a clustering algorithm is used to divide the standard leakage range, which is determined based on the statistical quantiles of current surges and voltage drops in historical leakage events. For example, the region where the current surge exceeds 15% of the rated value and the voltage drop exceeds 10% is defined as the leakage anomaly zone. When calculating the first characteristic parameter of each historical leakage event, a sliding window is used to analyze the rising and falling slopes of the current waveform, and the root mean square error within the window is calculated as a fluctuation intensity index. The second characteristic parameter is obtained by integral calculation to obtain the energy loss value during the duration of the leakage event, and the number of abnormal pulses exceeding the threshold during the event duration is counted. The training set is constructed by organizing the first feature parameter into a multi-dimensional input vector and the second feature parameter as the target output vector. The artificial intelligence detection model adopts a multilayer perceptron structure, with the number of input layer nodes corresponding to the dimension of the first feature parameter. The hidden layer uses the ReLU activation function, and the output layer is designed as a regression layer to predict the second feature parameter. The training process adopts the stochastic gradient descent algorithm, with the loss function being the mean squared error. Iterative training is conducted until the prediction error of the model on the validation set tends to stabilize.

[0020] The embedded sensors in the data acquisition module have self-calibration capabilities, periodically performing zero-point drift correction using a reference signal source to ensure the long-term accuracy of current fluctuation data. Voltage anomaly data acquisition is synchronized using a hardware watchdog circuit to prevent sampling interruptions or data loss. The extraction process for historical leakage event data includes integrity verification, ensuring log files are undamaged through cyclic redundancy checks. Event occurrence time parsing is compatible with multiple time formats, including Unix timestamps and ISO standard formats. Event duration calculations are accurate to milliseconds, and interference signals with durations less than 1ms are filtered out. The integration of the training dataset employs a dynamic buffering mechanism, concatenating real-time data streams and historical data blocks in memory, adding data source markers and quality identifiers. Inferior data, such as signal saturation segments or sampling loss segments, are automatically removed.

[0021] The standard leakage current range screening in the model training module employs an adaptive threshold algorithm. The threshold is dynamically adjusted based on the distribution of historical data to avoid misjudgments caused by fixed thresholds due to environmental changes. Frequency domain analysis is incorporated into the calculation of the first feature parameter, extracting harmonic components of the current waveform through Fast Fourier Transform as additional feature input. The energy loss value calculation for the second feature parameter considers the impact of voltage fluctuations, using dynamic power integration instead of simple current integration. The construction of the training set includes data augmentation steps, expanding sample diversity through time series interpolation and noise injection. The training of the AI ​​detection model employs an early stopping strategy to prevent overfitting. The validation set loss is monitored in real time during training; training is terminated and the optimal model parameters are retained when the loss does not decrease for several consecutive epochs.

[0022] Example 2: The core function of the leakage current detection module is to continuously monitor the power metering box based on a trained artificial intelligence detection model. This module receives current fluctuation data and voltage anomaly data from the data acquisition module in real time through the embedded system's data interface. This data is continuously input at a sampling frequency of 1000 times per second, forming a continuous data stream for model analysis. The artificial intelligence detection model adopts a lightweight neural network structure, with its weight parameters stored in the module's flash memory. The model inference process is accelerated by a dedicated digital signal processor, ensuring that the processing and analysis of the input data is completed within milliseconds. During real-time monitoring, the model outputs a continuous leakage current risk index, which comprehensively reflects the degree of current anomaly and voltage distortion characteristics.

[0023] When real-time current fluctuations exceed a preset threshold, the system immediately triggers a leakage event detection process. The preset threshold is dynamically adjusted based on the rated operating parameters of the power metering box. For example, for a metering box with a rated current of 60A, the threshold is set to three standard deviations of the normal fluctuation range. After the detection process is initiated, the module first performs windowing processing on the current data window, extracting current waveform segments and voltage sampling sequences from the last 500 milliseconds. This data is then input into an artificial intelligence detection model for in-depth analysis. The model output includes three main indicators: anomaly probability value, fault type classification, and severity score. The leakage event detection process includes a multi-stage verification mechanism. The first stage performs waveform feature matching, calculating the similarity between the real-time current waveform and typical leakage patterns in the model library. The second stage performs anomaly point localization, identifying the precise start and peak points of current surges using a sliding window difference algorithm. The third stage performs event classification, distinguishing different leakage types such as insulation aging, grounding faults, and equipment breakdown based on voltage anomaly characteristics and current surge patterns. The entire detection process is completed within 50 milliseconds, meeting the requirements of real-time protection.

[0024] When generating leakage current event detection reports, the module uses a structured data format. The report includes fields such as event number, detection timestamp, event type code, severity rating, current surge amplitude, and voltage anomaly index. Event types are classified and coded according to International Electrotechnical Commission (IEC) standards, and severity levels use a four-level classification system, from Level 1 (minor anomaly) to Level 4 (critical fault). After generation, the report is immediately sent to the event analysis module via a communication interface, and a copy is simultaneously stored in local non-volatile memory for subsequent auditing. The module's real-time monitoring algorithm employs a multi-threaded architecture. One thread is dedicated to data acquisition and preprocessing, while another thread handles model inference and event detection. This design ensures continuous monitoring even under high load conditions. Preprocessing includes data filtering, baseline correction, and outlier removal, using moving average and median filters to eliminate random noise. The model inference stage employs quantization techniques to optimize the computational efficiency of the neural network, converting floating-point operations to fixed-point operations to improve processing speed.

[0025] The dynamic adjustment mechanism for preset thresholds is based on historical operating data. The system automatically recalculates the threshold baseline weekly, taking into account factors such as load fluctuations due to seasonal changes, equipment aging trends, and recent leakage event statistical characteristics. This adaptive threshold method effectively reduces false alarm rates while ensuring sensitive response to real leakage events. The threshold update process is fully automated and requires no manual intervention. Leakage event detection reports are transmitted using a reliable communication protocol, with CRC checksums ensuring data integrity. The report format is compatible with the IEEE C37.118 standard, facilitating integration with other power monitoring systems. Locally stored reports are retained for at least 90 days and support multi-dimensional queries and retrieval by time range, event type, severity level, etc., providing a data foundation for subsequent event analysis and system optimization. The module has self-diagnostic capabilities, periodically checking the health status of the artificial intelligence model and verifying its detection accuracy by injecting test signals. When model performance degradation is detected, a maintenance alarm is automatically issued. Simultaneously, the module records operating logs, including data processing volume, event detection statistics, resource usage, and other indicators. These logs are used for system performance monitoring and troubleshooting.

[0026] Example 3: The core task of the event analysis module is to perform deep feature extraction and constraint analysis on detected leakage events. After receiving the detection report from the leakage detection module, this module first selects a baseline anomaly point from the report. The selection process adopts a dual-objective decision-making mechanism: the first objective is based on the cumulative calculation of the event cost value, and the cost value calculation formula is as follows:

[0027] in: Represents the accumulated cost value. For the first The unit time cost weight of each event. For each event duration, when the cumulative value approaches a preset cost threshold, the event is marked as the first target candidate. The second target is achieved through pattern matching, calculating the similarity between the feature vector of the first abnormal event in the report and a standard anomaly type library. A cosine similarity greater than 0.9 is selected as the second target candidate. Finally, the earlier timestamp of the two targets is chosen as the baseline anomaly point. After determining the baseline anomaly point, the module traces the historical working sequence within the target detection period corresponding to that point. The historical working sequence is extracted from the operation log of the energy metering box and includes time-series data of current, voltage, and power factor. The sequence length is dynamically adjusted according to the severity of the event, generally covering the period from 5 minutes before to 2 minutes after the anomaly point. The sequence data undergoes standardization to eliminate the influence of dimensions and retain relative change characteristics.

[0028] Based on the anomaly type of the baseline anomaly point, the module determines the corresponding traversal order. Anomaly types are divided into transient and persistent types. Transient anomalies are traversed in reverse time, tracing back from the anomaly point; persistent anomalies are traversed in forward time, extending forward from the anomaly point. During the traversal, sequence features are extracted from the historical working sequence, including indicators such as fluctuation frequency, peak interval, and slope change rate. These features are matched and compared with a predefined trigger sequence feature library. When a traversed sequence feature matches a trigger sequence feature, the module immediately invokes the corresponding replay mechanism and verification rules. The replay mechanism reconstructs the running state of the historical working sequence in a simulation environment, simulating the dynamic changes of electrical parameters through differential equations. The verification rules perform logical judgments on the replay results, including consistency checks (evaluation of the deviation between replay data and actual recorded data) and constraint verification (whether electrical parameters violate physical constraints). The feature analysis results are generated by integrating the replay output and validation results. The results are represented in multi-dimensional vector form, containing information such as anomaly root cause localization, propagation path mapping, and influencing factor weight allocation. This result is transmitted to the risk prediction module as input data and stored in the analysis database for subsequent model optimization. The entire analysis process emphasizes temporal correlation and causal reasoning, and the physical credibility of the analysis results is ensured through replay validation.

[0029] Historical working sequence extraction employs multi-source data fusion technology, incorporating auxiliary sensor data such as ambient temperature and humidity in addition to electrical energy parameters to improve the integrity of sequence features. Sequence standardization uses Z-score normalization to preserve the statistical distribution characteristics of the data. The trigger sequence feature library is built based on historical leakage event analysis, with each feature template containing a feature vector, matching tolerance, and weighting coefficients. The simulation environment for the replay mechanism is built based on an electrical model of the electricity metering box. Model parameters are calibrated in real-time according to equipment specifications, and the replay process uses a variable-step numerical integration algorithm to balance computational accuracy and efficiency. The constraints of the verification rules are set based on circuit theory, including Kirchhoff's laws, energy conservation constraints, and other physical principles. The data structure of the feature analysis results adopts a standardized architecture, including fields such as timestamps, anomaly type codes, root cause device identifiers, and confidence scores, supporting multi-dimensional queries and analysis. Intermediate data and final results generated during the analysis process are digitally signed to ensure data authenticity and integrity.

[0030] Taking a leakage current event in the main power metering box of a power distribution room in an industrial park during the early morning hours as an example, the leakage current detection module generates a detection report with the number E-20230615-112. This report records a continuous grounding fault in phase B, with an event severity level of Level 4. After receiving the report, the event analysis module starts the analysis process. First, it selects a baseline anomaly point: calculates the cost value of the first 5 events in the report (based on a weighted average of fault energy loss and maintenance cost). The cumulative value reaches 85% of the preset cost threshold of 480 cost units, meeting the first objective condition. At the same time, the first anomaly event type code IEC-607 in the detection report completely matches the standard grounding fault type, satisfying the second objective condition. The event corresponding to the first objective with the earliest timestamp is selected as the baseline anomaly point. The time corresponding to this point is 2023-06-15 02:15:33.216.

[0031] The module traces the historical working sequence of the three minutes before and after the benchmark anomaly point, extracting 1256 sampling records from the metering box operation database for that period, including parameters such as three-phase current, effective voltage value, zero-sequence current, and insulation resistance. After data standardization, a time series matrix is ​​formed. Sequence feature extraction shows a combination of characteristics: a gradual increase in current (from 85A to 112A) and intermittent voltage flicker (fluctuation of 220V±8V). Based on the anomaly type of ground fault, the module uses a forward time traversal method to analyze the sequence features from the benchmark anomaly point backward. When traversing the sequence features at 02:16:01.500, it is found that the abrupt change in zero-sequence current (from 0.8A to 5.6A) matches the "insulation breakdown" mode in the trigger sequence feature library with a 91% match. The system immediately calls the corresponding replay mechanism, which is based on a simulation environment constructed from circuit differential equations, with parameter settings including actual measured values ​​such as line distributed capacitance of 0.3μF and ground resistance of 1.2kΩ. The replay process simulates the electrical behavior of historical operating sequences, calculating the current and voltage changes at each time step through numerical integration to reproduce the process of thermal collapse of the insulation material leading to breakdown. Verification rules are applied to physical constraint checks, confirming that the deviation rate between the replay data and the actual recorded data is less than 3%, and that the energy conservation verification error is within the allowable range. Feature analysis results pinpoint the root cause of the fault as insulation aging at the B-phase cable joint, with the propagation path showing the fault current flowing through the grounding wire to the metal casing of the distribution cabinet. The analysis results generate a multi-dimensional vector containing: fault location coordinates (distribution cabinet X3 position), radius of influence (2.5 meters), and weights of key influencing factors (insulation aging 0.7%, ambient humidity 0.2%, load fluctuation 0.1%). This result is transmitted to the risk prediction module and simultaneously stored in the analysis database and marked as a typical grounding fault case.

[0032] The entire analysis process took 1.8 seconds, consumed 12.3MB of system memory, and had a peak CPU load of 42%. During the analysis, the module simultaneously monitored environmental data, detecting an ambient humidity of 85% RH and a temperature of 42℃ during the fault period. These data were incorporated into the auxiliary parameters of the feature analysis results. The electrical model parameters used in the replay mechanism were calibrated according to the actual model of the metering box, including equipment-specific parameters such as line impedance characteristics and protection device operating characteristics. In addition to basic physical constraints, the verification rules also included industry standard constraints: checking whether the protection action time requirements specified in GB / T14285 and the contact voltage safety limits specified in IEC60364 were violated during the replay process. The data structure of the feature analysis results includes time dimension information, displaying the time stamps for each stage of fault development: insulation degradation start time 02:13:22, breakdown occurrence time 02:15:33, and protection device sensing time 02:15:33.250. The module's runtime log records the execution status of each step during the analysis process, including data extraction (0.2 seconds), feature matching (0.6 seconds), replay calculation (0.9 seconds), and verification check (0.1 seconds). Upon completion of the analysis, the system automatically updates the feature patterns from this analysis to the trigger sequence feature library for reference in pattern recognition of subsequent events.

[0033] Example 4: After receiving the feature analysis results from the event analysis module, the risk prediction module first performs feature description to generate a feature description vector. This vector encapsulates the essential attributes of the abnormal event using a structured data format. For example, for a detected insulation aging event, the feature description vector includes the following dimensions: anomaly intensity coefficient (0.87), persistence trend index (1.23), associated equipment identifier (metering box number X-7B3), environmental humidity influence factor (0.62), and frequency of historical similar events (3 times). These dimensional values ​​are calculated using the anomaly root cause location and propagation path mapping data in the feature analysis results. The anomaly intensity coefficient is based on the weighted integral of the current mutation amplitude and duration, and the persistence trend index obtains the slope value through time series regression analysis.

[0034] When determining the prediction strategy corresponding to the feature description vector from the prediction strategy library, the module employs a multi-level matching mechanism. The first level matches a basic strategy template based on the associated device type. The second level refines the strategy parameters by considering the numerical range of the anomaly intensity coefficient and the persistence trend index. The third level fine-tunes the strategy by incorporating auxiliary parameters such as the environmental humidity influence factor. The prediction strategy library is stored in the embedded system's non-volatile memory and includes various prediction algorithms, such as ARIMA model strategies based on time series analysis, Markov model strategies based on state transitions, and risk propagation model strategies based on neural networks.

[0035] When predicting future leakage risks based on a predictive strategy, the module performs a multi-step calculation process. First, the feature description vector is converted into the input format of the prediction model. Then, the model inference is run to calculate the risk probability distribution within the future time window. Finally, a spatial mapping algorithm is used to determine the scope of risk impact. For example, for an insulation aging event, the prediction model outputs a curve showing the leakage probability changing over time in the next 24 hours, while also identifying the range of potentially affected equipment, including adjacent circuit modules and associated protection devices.

[0036] The risk prediction report is generated using a standardized document structure. The report includes data fields such as predicted event number, baseline timestamp, predicted time range, probability distribution of predicted occurrence time, list of affected devices, and predicted risk level assessment. The predicted risk level uses a five-level classification system, from R1 (low risk) to R5 (emergency risk). The level assessment is a weighted calculation based on both the probability of occurrence and the scope of impact.

[0037] After receiving the risk prediction report, the optimization module obtains an optimization parameter set through parameter parsing. The parsing process extracts key performance indicators from the report, including parameters such as model prediction bias rate, feature recognition error coefficient, and response latency. These parameters constitute the optimization parameter set, which guides the iterative optimization of the artificial intelligence detection model.

[0038] When determining the optimization iteration strategy corresponding to the optimization parameter set from the optimization strategy library, the module uses a parameter matching algorithm to perform similarity matching between the numerical features of the optimization parameter set and the strategy templates in the strategy library, selecting the strategy with the highest matching degree as the optimization basis. The optimization strategy library contains a variety of optimization methods, including model structure adjustment strategies, parameter update strategies, feature weight adjustment strategies, and algorithm replacement strategies.

[0039] When adjusting the model structure and updating parameters of an AI detection model based on an optimization iterative strategy, specific optimization operations are performed. For example, when the prediction bias rate exceeds a threshold, the strategy instructs to increase the number of hidden layer nodes in the neural network; when the feature recognition error coefficient is too high, the strategy requires adjusting the convolutional kernel size of the feature extraction layer; parameter updates employ incremental learning, fine-tuning the model weight matrix based on the latest event data. The entire optimization process runs in real time on an embedded system, ensuring that the model continuously adapts to changes in the operating environment.

[0040] When outputting the optimized AI detection model version, the module generates metadata such as a model version identifier, optimization timestamp, and change log. The new model version is deployed to the leakage current detection module after passing security verification, while the old version model is backed up for rollback emergencies. Version management adopts an incremental update mechanism, transmitting only changed parameters to reduce communication overhead (see Table 1).

[0041] Table 1: Feature Description Vector Dimension Table

[0042] The risk prediction module operates in sync with the leakage current detection module. Each new leakage current event detection report triggers the prediction process, and the prediction time range can be adjusted according to configuration, with a default setting of 24 hours. The optimization module operates at a relatively low frequency, typically performing batch optimization once daily, or triggering immediate optimization after a certain number of risk prediction reports have been received. Both modules employ fault-tolerant design, automatically retrying when a single prediction or optimization task fails and recording detailed logs for fault analysis.

[0043] Example 5: The protection execution module implements real-time monitoring of the electricity metering box based on an optimized version of the artificial intelligence detection model. This model version is identified as V2.3.5, and its neural network structure contains 128 hidden nodes and an improved leakage current identification algorithm. The module continuously acquires real-time status data of the metering box through an RS-485 communication interface, including three-phase current sampling values ​​(one group per millisecond), ground insulation resistance readings, box temperature monitoring values, and humidity sensor data. After preprocessing, the real-time status data is input into the optimized model for calculation. The model outputs a comprehensive risk score (range 0-100) and a specific risk classification identifier. When the risk score exceeds the execution threshold of 75 points for 5 consecutive sampling cycles, the module immediately triggers the protection mechanism. The triggering process of the protection mechanism includes multiple layers of verification. First, the similarity between the current risk pattern and the historical event database is compared to confirm whether it belongs to a known leakage type. Then, the operating status of related equipment is checked to eliminate misjudgments caused by temporary overload or equipment startup. Finally, instructions are sent to the execution agency through a safety authentication protocol. When implementing the protection mechanism, the module selects the operation type according to the risk level: for risks with a score of 75-85, it sends an alarm activation command to the monitoring center platform and the on-site audible and visual alarm; for emergency risks with a score higher than 85, it simultaneously sends a current cut-off command to the control unit of the smart circuit breaker. The command uses encrypted transmission and a dual verification mechanism to ensure reliability.

[0044] The protection execution log is recorded in a structured format. Each log entry includes an event sequence number, timestamp (accurate to milliseconds), execution action type, target device identifier, execution result status, and a risk score snapshot. For example, in an actual operation, the log record shows: event sequence number E-20231027-083, timestamp 2023-10-27 08:45:32.456, execution action type "alarm activation + current cutoff", target device identifier CB-7B3-12, execution result status "success", and risk score 87. The log file is stored locally and simultaneously synchronized to the central management system. The visualization monitoring module configures the display interface based on the administrator profile. For the operations engineer role, the visualization template highlights the real-time risk heatmap and historical event timeline; for the administrator role, the template focuses on system availability indicators and risk trend statistics. The module integration and optimization of the AI ​​detection model provides multimodal working information, including real-time inference confidence (displayed as a percentage dashboard), risk hotspot geographic distribution map (drawn based on GIS coordinates), historical event backtracking timeline (interactively zoomable), and model version performance comparison charts.

[0045] The detection of monitoring events is achieved through a rule engine. Rules include: sudden changes in risk scores (change rate exceeding 20% ​​per minute), continuous decline in model confidence (below 80% for 10 consecutive samples), and communication latency exceeding limits (greater than 500 milliseconds). When multiple monitoring events are detected, the module generates a quick-selection table of monitoring tasks. This table is prioritized and lists tasks to be processed, recommended operations, and estimated processing times. For example, when both a sudden change in risk scores and a communication latency event are detected simultaneously, the quick-selection table prioritizes the "Check Communication Link" task, followed by the "Verify Sensor Calibration" task. Administrators select tasks from the quick-selection table via a touch interface or an external mouse. After selection, the module automatically brings up the corresponding operation interface pre-filled with relevant parameters. For example, when the "Verify Sensor Calibration" task is selected, the interface automatically displays a graph showing the deviation between the current sensor reading and the standard value, and provides two options: "Start Automatic Calibration" and "Manual Adjustment." All monitoring operations are recorded in the audit log, including operator identity, operation time, and execution results.

[0046] The real-time data display of the module adopts a multi-window linkage design. The main window shows the current risk score curve, the sidebar scrolls to update the monitoring task list, and the bottom status bar continuously outputs the system health indicators. The display refresh frequency is dynamically adjusted according to the data criticality. Under normal conditions, it is refreshed once per second, and under high-risk conditions, the speed is increased to 5 times per second. The historical data query supports multi-dimensional filtering by time range, event type, device number, etc., and the query results are presented in two forms: visual charts and data tables. The communication interface adopts a redundant design. The main channel is the Ethernet TCP / IP protocol, and the backup channel is 4G wireless transmission. When the main channel delay is detected to exceed the threshold, the backup channel is automatically switched. All transmitted data is encrypted with AES-256 and verified with HMAC signatures to ensure the security of instruction transmission. The system self-diagnosis function regularly checks the status of each component and generates a system health report every week, including key indicators such as processor load rate, memory usage, and remaining storage space.

[0047] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A leakage protection system based on an embedded energy metering box, characterized in that, include: The data acquisition module is used to collect real-time power parameter data and historical leakage event data from the power metering box. The model training module trains an artificial intelligence detection model based on the data collected by the data acquisition module. The leakage current detection module performs real-time leakage current detection on the power metering box based on a trained artificial intelligence detection model. The event analysis module performs feature extraction and constraint analysis on leakage events detected by the leakage detection module in real time; The risk prediction module predicts future leakage risks based on the analysis results of the event analysis module. The optimization module continuously optimizes and iterates the artificial intelligence detection model based on the prediction results of the risk prediction module. The protection execution module performs leakage protection actions based on the optimized artificial intelligence detection model after the optimization module completes the optimization iteration.

2. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The data acquisition module includes: Real-time current fluctuation data and voltage anomaly data are obtained from the embedded sensors in the power metering box; Extract the event occurrence time and event duration from historical leakage current event data; Real-time current fluctuation data and voltage anomaly data are integrated with historical leakage event data to form a training dataset.

3. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The model training module includes: The leakage current range is selected based on the training dataset integrated by the data acquisition module. Calculate the first and second feature parameters of each historical leakage event in the training dataset; The first feature parameter is used as the input and the second feature parameter is used as the output to form a training set. The training set is used to train the artificial intelligence detection model.

4. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The leakage current detection module includes: The artificial intelligence detection model trained by the model training module monitors the real-time power parameter data of the power metering box. The leakage event detection is triggered when the real-time current fluctuation data exceeds a preset threshold. The generated leakage current event detection report includes the event type and the event severity level.

5. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The event analysis module includes: Select a baseline anomaly point from the leakage event detection report generated by the leakage current detection module; Tracing the historical work sequence within the target detection period corresponding to the baseline anomaly point; Based on the traversal order corresponding to the anomaly type of the baseline anomaly point, the sequence features in the historical working sequence are traversed sequentially. When the traversed sequence features match the trigger sequence features, the corresponding replay mechanism and verification rules are obtained; The replay mechanism controls the replay of historical working sequences and executes verification rules to generate feature analysis results.

6. The leakage protection system based on an embedded power metering box as described in claim 5, characterized in that, The event analysis module selects baseline anomalies from the leakage current event detection report, including: Determine the primary and secondary objectives; The earlier of the first and second objectives is used as the baseline outlier. The determination of the first objective is based on the fact that the sum of the cost values ​​of the first few events in the leakage current event detection report is close to the cost threshold; The second objective is determined based on the fact that the first anomaly type in the leakage event detection report is the same as the standard anomaly type.

7. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The risk prediction module includes: The feature analysis results generated by the event analysis module are used to obtain feature description vectors; Determine the prediction strategy corresponding to the feature description vector from the prediction strategy library; The probability of future leakage risks and the scope of their impact are predicted based on the predictive strategy. The generated risk forecast report includes the predicted time of occurrence and the predicted risk level.

8. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The optimization module includes: The optimized parameter set is obtained by parsing the risk prediction report generated by the risk prediction module. Determine the optimization iteration strategy corresponding to the optimization parameter set from the optimization strategy library; The model structure of the artificial intelligence detection model is adjusted and the parameters are updated based on an optimization iteration strategy. Output the optimized version of the artificial intelligence detection model.

9. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The protection execution module includes: The optimized AI detection model version output by the optimization module is used to monitor the real-time status of the power metering box. The protection mechanism is triggered when the detected leakage risk exceeds the execution threshold. The protection mechanism includes current cut-off commands or alarm activation commands; The protection execution log includes the execution time and execution result.

10. The leakage protection system based on an embedded power metering box as described in claim 1, characterized in that, The system also includes: a visualization monitoring module, which generates a visualization monitoring model of the optimized artificial intelligence detection model to assist in work monitoring; The visualization monitoring module includes: multimodal work information of an artificial intelligence detection model integrated and optimized based on the visualization template corresponding to the personnel profile of the management personnel; detection of multiple monitoring timing events occurring in the visualization monitoring model; generation of a monitoring task quick selection table based on each monitoring timing event; and assistance to management personnel in quickly selecting monitoring tasks from the monitoring task quick selection table to perform monitoring operations.

Images