A multi-source sensing-based edge terminal health degree self-diagnosis method and system

Abstract

The application provides a multi-source sensing-based edge terminal health degree self-diagnosis method and system, and belongs to the technical field of power distribution. The method comprises the following steps: performing key feature extraction and data fusion processing on multi-source sensing data collected by an edge distribution transformer terminal; performing terminal sampling performance prediction to generate a predicted sampling accuracy reliability score according to a key sensing data sequence set; dynamically evaluating a current noise level coefficient and a real-time false report risk coefficient, and configuring an adaptive accuracy reliability score threshold; if the predicted sampling accuracy reliability score is less than the adaptive accuracy reliability score threshold, it is determined that the edge distribution transformer terminal is in sampling accuracy drift anomaly, and an automatic calibration program is triggered; if the calibration fails, a maintenance request is generated and reported to a master station management platform. The application realizes real-time prediction and evaluation of the sampling accuracy of the edge distribution transformer terminal and adaptive threshold adjustment of the environment, and improves the detection accuracy of the sampling accuracy drift anomaly and the automatic calibration success rate.

Description

Technical Field

[0001] This invention relates to the field of power distribution technology, and in particular to a method and system for self-diagnosing the health of edge terminals based on multi-source sensing. Background Technology

[0002] As crucial data acquisition nodes at the end of the power system, edge distribution transformer terminals undertake key functions such as power quality monitoring, load analysis, and fault detection. These terminal devices are typically deployed in harsh outdoor environments and require long-term stable operation to provide accurate data acquisition services.

[0003] In existing technologies, health status monitoring of edge distribution transformer terminals mainly relies on periodic manual inspections or simple alarm mechanisms based on fixed thresholds. However, due to the complex and variable operating environment of distribution transformer terminals, including factors such as temperature fluctuations, electromagnetic interference, and equipment aging, their sampling accuracy exhibits dynamic drift. Traditional fixed threshold monitoring methods cannot adapt to this dynamic change, easily generating false alarms or missed alarms, making it difficult to accurately determine the true health status of the terminal. Simultaneously, existing fault diagnosis methods often employ single sensor data or simple data comparison methods, lacking effective fusion of multi-source sensor data. They cannot dynamically adjust diagnostic thresholds according to the real-time operating environment, resulting in difficulty in timely detection and automatic calibration of sampling accuracy drift. Summary of the Invention

[0004] This invention addresses the technical problem in existing technologies where edge distribution transformer terminals lack an adaptive sampling accuracy monitoring mechanism, making it impossible to dynamically adjust diagnostic thresholds according to the real-time operating environment, resulting in difficulty in timely detection and automatic calibration of sampling accuracy drift. It provides a self-diagnosis method and system for the health of edge terminals based on multi-source sensing to solve this problem.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a self-diagnosis method for the health of an edge terminal based on multi-source sensing, comprising: extracting key features and performing data fusion processing on multi-source sensing data collected by the edge distribution transformer terminal; generating a predicted sampling accuracy reliability score based on the terminal sampling performance prediction based on the key sensing data sequence set; dynamically evaluating the current noise level coefficient and the real-time missed alarm risk coefficient based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, and configuring an adaptation accuracy reliability score threshold; if the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, the edge distribution transformer terminal is determined to have an abnormal sampling accuracy drift, and an automatic calibration procedure is triggered; if the calibration fails, a maintenance request is generated and reported to the main station management platform.

[0006] Secondly, this invention provides a self-diagnostic system for the health of an edge terminal based on multi-source sensing, comprising: a multi-source data fusion module, used to extract key features and perform data fusion processing on multi-source sensing data collected by the edge distribution transformer terminal, and generate a predicted sampling accuracy reliability score based on the key sensing data sequence set; a dynamic threshold configuration module, used to dynamically evaluate the current noise level coefficient and the real-time missed alarm risk coefficient based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, and configure an appropriate accuracy reliability score threshold; and a health determination module, used to determine that the edge distribution transformer terminal has an abnormal sampling accuracy drift if the predicted sampling accuracy reliability score is less than the appropriate accuracy reliability score threshold, and trigger an automatic calibration procedure, and generate a maintenance request to be reported to the main station management platform if the calibration fails.

[0007] The beneficial effects of this invention are: Key features are extracted and data fusion is performed on multi-source sensor data collected from edge distribution transformer terminals. Based on the key sensor data sequence set, terminal sampling performance is predicted to generate a predicted sampling accuracy reliability score. Through deep fusion and predictive analysis of multi-source sensor data, the current sampling performance status of the terminal can be accurately assessed, providing a reliable data foundation for subsequent anomaly detection. Based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, the current noise level coefficient and real-time missed detection risk coefficient are dynamically evaluated, and an adaptive accuracy reliability score threshold is configured. Through real-time environmental perception and risk assessment, the diagnostic threshold is dynamically and adaptively adjusted to ensure accurate anomaly detection capability under different operating environments. If the predicted sampling accuracy reliability score is less than the adaptive accuracy reliability score threshold, the edge distribution transformer terminal is determined to have a sampling accuracy drift anomaly, and an automatic calibration procedure is triggered. If calibration fails, a maintenance request is generated and reported to the main station management platform. Through intelligent anomaly detection and automatic calibration mechanisms, closed-loop management of terminal health status is achieved, reducing manual intervention and improving fault handling efficiency.

[0008] The above technical solution enables real-time prediction and evaluation of sampling accuracy of edge distribution transformer terminals and environmental adaptive threshold adjustment, effectively solving the problems of lack of adaptive monitoring mechanism and inability to dynamically adjust diagnostic threshold in existing technologies, and improving the detection accuracy of abnormal sampling accuracy drift and the success rate of automatic calibration. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a self-diagnosis method for the health of an edge terminal based on multi-source sensing provided by the present invention; Figure 2 This is a schematic diagram of the structure of an edge terminal health self-diagnosis system based on multi-source sensing provided by the present invention.

[0010] In the attached diagram, the components represented by each number are as follows: Multi-source data fusion module 11, dynamic threshold configuration module 12, and health determination module 13. Detailed Implementation

[0011] 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.

[0012] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0013] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0014] Example 1, as Figure 1 As shown, this embodiment of the invention provides a self-diagnosis method for the health of an edge terminal based on multi-source sensing, including: S1. Extract key features and perform data fusion processing on the multi-source sensor data collected by the edge distribution transformer terminal, and generate a prediction sampling accuracy reliability score based on the key sensor data sequence set to predict the terminal sampling performance.

[0015] Specifically, firstly, multi-source sensor data is collected from the edge distribution transformer terminal. This multi-source sensor data covers multiple dimensions of information during the terminal's operation, including basic physical state data sequences and communication state data sequences. The basic physical state data sequences include at least key indicators reflecting the terminal's physical operating status, such as temperature, voltage, current, power, clock, vibration, and ambient temperature and humidity. The communication state data sequences include at least key indicators reflecting the terminal's communication performance, such as signal strength, data transmission rate, link quality indication, bit error rate, packet loss rate, communication latency, and physical port connection status.

[0016] Key feature extraction and data fusion processing are performed on the collected multi-source sensor data. Specifically, data alignment, outlier removal, and dimensional normalization are performed on the basic physical state data sequence and the communication state data sequence to obtain standardized data sequences. Based on this, an initial feature space containing time-domain features, statistical features, and state features is calculated and constructed to meet the needs of sampling performance analysis. The initial feature space includes, but is not limited to, multi-dimensional feature parameters such as the mean voltage sampling, the variance voltage sampling, the mean current sampling, the variance current sampling, the fluctuation period of the power curve, the rise slope of the core chip temperature, the cumulative jitter of the clock signal, the main frequency amplitude of the vibration energy spectrum, and the burst frequency of the communication bit error rate. Subsequently, a feature importance evaluation method based on a tree model is used to analyze the correlation between each feature in the initial feature space and the historical sampling accuracy deviation, and the top K features with the highest correlation are selected to construct a key sensor data sequence set.

[0017] When predicting terminal sampling performance based on key sensor data sequence sets, the first step is to obtain an appropriate accuracy and reliability evaluation channel by matching the key sensor data types. This evaluation channel includes L accuracy and reliability evaluation branches, which are obtained through deep learning training on historical sample data and can evaluate the terminal's sampling performance from different perspectives. Volatility analysis is performed on each key sensor data sequence in the key sensor data sequence set to calculate the coefficient of variation of each key sensor data, and a weighted average is calculated to obtain the comprehensive data volatility. Based on the ratio of the comprehensive data volatility to the preset benchmark comprehensive data volatility, the number Q of appropriate evaluation branches is dynamically determined, and Q branches are randomly selected from the L accuracy and reliability evaluation branches for terminal sampling performance prediction. The mean of the Q prediction results is used as the predicted sampling accuracy and reliability score, which comprehensively reflects the current sampling accuracy health status of the terminal.

[0018] Through the aforementioned key feature extraction, data fusion processing, and multi-branch prediction mechanism, the system can accurately extract key features strongly correlated with sampling performance from massive multi-source sensor data, avoiding interference from redundant features in the prediction results. Simultaneously, a strategy based on dynamically adjusting the number of evaluation branches according to the comprehensive data volatility allows the prediction process to adaptively balance prediction accuracy and computational resource consumption based on the stability of the terminal's real-time operating status. When the terminal's operating status fluctuates significantly, the system automatically increases the number of evaluation branches to improve prediction robustness; when the operating status is relatively stable, it reduces the number of evaluation branches to lower the computational load. The mean fusion mechanism of multi-branch prediction results further reduces the random error of single-model predictions, enabling the prediction sampling accuracy reliability score to more objectively and accurately reflect the actual sampling performance health status of the terminal, providing a reliable quantitative basis for subsequent anomaly detection and automatic calibration.

[0019] By generating a predictive sampling accuracy reliability score, the complex multi-source sensing status of the terminal can be quantified into a unified reliability evaluation index, providing a clear quantitative basis for subsequent anomaly detection and automatic calibration.

[0020] S2. Based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, dynamically evaluate the current noise level coefficient and the real-time missed reporting risk coefficient, and configure the adaptation accuracy reliability score threshold.

[0021] Specifically, firstly, real-time acquisition of operational status and environmental data of the edge distribution transformer terminal within a preset time zone is performed. The operational status data includes active power load rate sequences, core chip temperature sequences, power supply voltage ripple RMS value sequences, and communication channel signal-to-noise ratio sequences. These data reflect key operational indicators such as load level, thermal management status, power supply quality, and communication quality during actual operation. The environmental data includes ambient temperature and humidity sequences, reflecting the external environmental conditions of the terminal.

[0022] Then, the coefficients of variation for the active power load rate sequence, core chip temperature sequence, voltage ripple RMS value sequence, signal-to-noise ratio sequence, ambient temperature sequence, and ambient humidity sequence were calculated respectively. These coefficients of variation quantify the fluctuation degree of each operating state parameter and environmental parameter. By weighted summing of the above coefficients of variation, the current noise level coefficient is obtained. The current noise level coefficient comprehensively reflects the complexity and interference level of the terminal's current operating environment. The higher the noise level coefficient, the more interference factors the terminal faces and the harsher the operating environment.

[0023] Simultaneously, a real-time missed detection risk coefficient is calculated based on the active power load rate sequence. Specifically, the mean of the active power load rate sequence is calculated, and the ratio of the mean active power load rate to a preset benchmark active power load rate is used as the first risk compensation coefficient. The coefficient of variation of the active power load rate sequence is used as the second risk compensation coefficient. A comprehensive risk compensation coefficient is obtained by weighting and fusing the first and second risk compensation coefficients. The product of the comprehensive risk compensation coefficient and the preset benchmark missed detection risk coefficient is used as the real-time missed detection risk coefficient. The real-time missed detection risk coefficient reflects the potential risk of terminal sampling accuracy anomalies that are not detected in time under the current load conditions. When the load rate is high or the load fluctuation is large, the missed detection risk coefficient increases accordingly, indicating that more stringent diagnostic criteria are needed to avoid missed detections.

[0024] After obtaining the current noise level coefficient and the real-time missed detection risk coefficient, a pre-constructed two-dimensional dynamic mapping table is consulted to determine the adaptation accuracy reliability score threshold based on the matching of the current noise level coefficient and the real-time missed detection risk coefficient. This two-dimensional dynamic mapping table uses the current noise level coefficient as the horizontal axis and the real-time missed detection risk coefficient as the vertical axis, discretizing the two continuous coefficient value ranges into regular discretized grids. For each intersection node in the grid, a unique corresponding adaptation accuracy reliability score threshold is pre-stored. The thresholds stored in the two-dimensional dynamic mapping table follow a clear configuration rule: the adaptation accuracy reliability score threshold is negatively correlated with the current noise level coefficient, i.e., the higher the noise level, the lower the threshold, and the relatively relaxed diagnostic criteria to avoid excessive false alarms due to environmental interference; the adaptation accuracy reliability score threshold is positively correlated with the real-time missed detection risk coefficient, i.e., the higher the missed detection risk, the higher the threshold, and the relatively tighter the diagnostic criteria to reduce the probability of missed detection and ensure that anomalies under critical operating conditions can be detected in a timely manner.

[0025] By dynamically evaluating the current noise level coefficient and the real-time missed detection risk coefficient, and configuring an appropriate accuracy and reliability scoring threshold, the diagnostic criteria are adaptively adjusted. This ensures that the most suitable judgment criteria can be used under different operating environments and load conditions, effectively balancing the false alarm rate and the missed detection rate, and improving the accuracy and reliability of anomaly judgment.

[0026] S3. If the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, the edge distribution transformer terminal is determined to have an abnormal sampling accuracy drift, and an automatic calibration procedure is triggered. If the calibration fails, a maintenance request is generated and reported to the main station management platform.

[0027] Specifically, the predicted sampling accuracy reliability score is compared with the adaptation accuracy reliability score threshold. When the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, it indicates that the current sampling accuracy health status of the terminal is below an acceptable safety level, and the edge distribution transformer terminal is determined to have experienced abnormal sampling accuracy drift.

[0028] Once an abnormal drift in sampling accuracy is detected, an automatic calibration procedure is immediately triggered. This procedure attempts to restore the terminal's sampling accuracy to the normal range by adjusting the terminal's sampling parameter configuration, recalibrating the sensor reference value, or executing a built-in error compensation algorithm. During calibration, the terminal's sampling performance indicators are continuously monitored to verify whether the calibration effect meets expectations.

[0029] If the sampling accuracy of the edge distribution transformer terminal fails to return to normal after the automatic calibration procedure, indicating a calibration failure, it suggests that the anomaly may stem from hardware failure, sensor aging, or other deeper issues that automatic calibration cannot resolve. In this case, a maintenance request is automatically generated, containing detailed information such as the anomaly type, occurrence time, terminal identifier, current operating status parameters, and calibration attempt records. This request is then reported to the main station management platform via the communication link, allowing maintenance personnel to manually intervene or arrange on-site maintenance.

[0030] By comparing the predicted sampling accuracy reliability score with the adaptive accuracy reliability score threshold, and combining the closed-loop processing mechanism of automatic calibration and maintenance request reporting, a complete self-diagnosis process from anomaly detection and autonomous repair to manual intervention is realized. It can detect and automatically repair in the early stage of sampling accuracy degradation, avoiding data distortion and misjudgment caused by the accumulation of anomalies. For faults that cannot be automatically repaired, they are reported in a timely manner. This realizes real-time prediction and evaluation of the sampling accuracy of edge distribution transformer terminals and environmental adaptive threshold adjustment. It effectively solves the technical problems in the existing technology of edge distribution transformer terminals lacking an adaptive sampling accuracy monitoring mechanism, unable to dynamically adjust diagnostic thresholds according to the real-time operating environment, and making it difficult to detect and automatically calibrate sampling accuracy drift in a timely manner. It improves the detection accuracy of sampling accuracy drift anomalies and the success rate of automatic calibration.

[0031] Furthermore, the multi-source sensing data includes a basic physical state data sequence and a communication state data sequence. The basic physical state data includes at least temperature parameters, voltage parameters, current parameters, power parameters, clock parameters, vibration parameters, and ambient temperature and humidity. The communication state data includes at least signal strength, data transmission rate, link quality indication, bit error rate, packet loss rate, communication delay, and physical port connection status.

[0032] In one feasible implementation, the multi-source sensor data includes a basic physical state data sequence and a communication state data sequence. The basic physical state data sequence reflects the hardware operational health and physical environment conditions of the edge distribution transformer terminal. The communication state data sequence reflects the communication quality and network connectivity health between the edge distribution transformer terminal and the main station management platform or other devices.

[0033] Basic physical state data includes at least temperature, voltage, current, power, clock, vibration, and ambient temperature and humidity parameters. Temperature parameters include the core processor chip junction temperature, power management module surface temperature, communication module operating temperature, internal chassis ambient temperature, critical power device temperatures, and battery temperature. This data reflects the thermal management status and potential overheating risks of key components. Voltage, current, and power parameters include the main power input voltage and current, battery output voltage and current and remaining capacity, core chip operating voltage, power supply voltage for each functional module, instantaneous and average power, and power supply ripple noise. These electrical parameters reflect the terminal's power supply quality, energy consumption, and power stability. Clock parameters include the system master clock frequency stability, real-time clock accuracy, and timer interrupt response interval deviation. These clock parameters directly affect the terminal's data sampling timing accuracy and system synchronization. Vibration parameters reflect the vibration intensity and spectral characteristics of the terminal's mechanical environment; excessive vibration may lead to loose hardware connections or sensor performance degradation. Ambient temperature and humidity parameters reflect the terminal's external environmental conditions and have a significant impact on the operational stability and lifespan of terminal components.

[0034] Communication status data includes at least signal strength, data transmission rate, link quality indicator, bit error rate, packet loss rate, communication latency, and physical port connection status. Signal strength reflects the signal reception power of the wireless communication link; low signal strength leads to communication instability. Data transmission rate reflects the actual data transmission speed; a decrease in transmission rate may indicate deteriorating communication performance. Link quality indicator comprehensively reflects the overall quality level of the communication link. Bit error rate represents the proportion of errors occurring during data transmission; an increased bit error rate indicates decreased communication quality. Packet loss rate reflects the proportion of data packets lost during transmission; a high packet loss rate leads to incomplete data. Communication latency reflects the time required for data to travel from transmission to reception; excessive latency affects applications with high real-time requirements. Physical port connection status reflects the connection status of the wired communication interface; port abnormalities may cause communication interruptions.

[0035] By collecting multi-source sensor data covering basic physical and communication states, the operating status of edge distribution transformer terminals can be comprehensively perceived from multiple dimensions such as hardware health, electrical performance, timing accuracy, mechanical environment, and communication quality. This provides a rich source of feature information for subsequent sampling accuracy prediction, thereby enabling more accurate identification of various potential factors affecting sampling accuracy and improving the sensitivity of anomaly detection and the generalization ability of prediction models.

[0036] Furthermore, key feature extraction and data fusion processing are performed on the multi-source sensor data collected by the edge distribution transformer terminal, including: S11. Perform data alignment, outlier removal and dimension normalization on the basic physical state data sequence and the communication state data sequence respectively to obtain the standard physical state data sequence and the standard communication state data sequence. S12. Based on the standard physical state data sequence and the standard communication state data sequence, for sampling performance analysis, calculate and construct an initial feature space containing time-domain features, statistical features and state features. The initial feature space includes, but is not limited to, the voltage sampling mean, voltage sampling variance, current sampling mean, current sampling variance, the fluctuation period of the power curve, the rise slope of the core chip temperature, the cumulative jitter of the clock signal, the main frequency amplitude of the vibration energy spectrum and the burst frequency of the communication bit error rate. S13. Using a tree-based feature importance evaluation method, analyze the correlation between each feature in the initial feature space and the historical sampling accuracy deviation, and select the top K features with the highest correlation to construct a key sensor data sequence set.

[0037] In a preferred embodiment, the basic physical state data sequence and communication state data sequence are first processed by data alignment, outlier removal, and dimension normalization to obtain standard physical state data sequences and standard communication state data sequences. Specifically, data alignment is performed first. Since the sampling frequencies and sampling times of different sensors in the edge distribution transformer terminal differ—for example, a temperature sensor may sample once per second, while a voltage sensor may sample once every 100 milliseconds—it is necessary to align each data sequence to a unified time reference. A unified reference sampling time sequence is determined. For data sequences with sampling frequencies higher than the reference frequency, the average value or the closest sampling value is taken within a time window at each reference time point. For data sequences with sampling frequencies lower than the reference frequency, linear interpolation or spline interpolation methods are used to estimate the corresponding data values ​​at the reference time points, ensuring that all sensor data sequences are completely synchronized in the time dimension. Subsequently, outlier removal is performed. During data acquisition, outliers that significantly deviate from the normal range may occur due to factors such as sensor malfunctions, electromagnetic interference, or communication errors. Outliers are identified using the 3σ criterion or an outlier detection method based on interquartile range, marking data points outside the normal range as outliers. Detected outliers are handled based on their location, either by interpolating preceding and following data or by marking entire segments as invalid, to avoid interference with subsequent feature calculations. Next, dimensional normalization is performed. Different types of sensor data exhibit significant differences in physical dimensions and numerical ranges; for example, temperature ranges may be -40℃ to +85℃, and voltage ranges may be 0V to 500V. This difference can lead to features with large numerical ranges dominating subsequent analysis. Min-max normalization is used to linearly map the data to a unified interval, or Z-score normalization is used to standardize the data to a distribution with a mean of zero and a standard deviation of one. Normalization eliminates the influence of dimensional differences, making different types of features comparable. After the above processing, standard physical state data sequences and standard communication state data sequences are obtained.

[0038] Next, based on standard physical state data sequences and standard communication state data sequences, an initial feature space including time-domain features, statistical features, and state features is calculated and constructed to meet the requirements of sampling performance analysis. The time-domain features include the fluctuation period of the power curve, the rise slope of the core chip temperature, and the cumulative jitter of the clock signal, reflecting the dynamic characteristics and evolution of the data sequence over time. Specifically, the fluctuation period of the power curve is identified by spectral analysis of the power time series to identify the main periodic components, reflecting the periodic variation of the load. The rise slope of the core chip temperature is calculated by slope analysis of the temperature sequence to obtain the rate of temperature change, reflecting the rate of heat accumulation in the chip; rapid temperature rise may lead to chip performance degradation and decreased sampling accuracy. The cumulative jitter of the clock signal is the cumulative deviation of the clock signal from the ideal reference, reflecting the long-term stability of the clock; increased jitter can cause sampling time offset, affecting sampling synchronization accuracy. Statistical features include the voltage sampling mean, voltage sampling variance, current sampling mean, and current sampling variance, reflecting the probability distribution characteristics and the degree of centralization and dispersion of the data sequence. Specifically, the voltage sampling mean is calculated by averaging the voltage sequence within a preset time window, reflecting the average voltage level. Deviations in this value may indicate power supply anomalies or sampling reference drift. The voltage sampling variance calculates the degree of fluctuation in the voltage sequence; increased variance indicates voltage instability, which may affect the sampling accuracy of the analog-to-digital converter. The current sampling mean reflects the current central tendency, while the current sampling variance reflects the current fluctuation characteristics. Abnormal current fluctuations can also affect the terminal's sampling stability. State characteristics include the dominant frequency amplitude of the vibration energy spectrum and the burst frequency of the communication bit error rate, reflecting typical characteristics under specific operating conditions. Specifically, the dominant frequency amplitude of the vibration energy spectrum is obtained by performing spectral analysis on the vibration signal to identify the frequency components with the most concentrated energy and their amplitudes, reflecting the main frequency characteristics and intensity of the vibration. An excessively large dominant frequency amplitude may lead to loosening of internal components or mechanical fatigue of the sensor, thus affecting sampling stability. The burst frequency of the communication bit error rate (BER) is the number of times the BER exceeds a set threshold within a preset time window. This reflects the frequency of sudden deterioration in communication quality. An increase in the burst frequency of the BER may indicate performance degradation of the communication module or a deterioration of the electromagnetic environment, affecting the integrity and timeliness of data transmission, and consequently affecting the reliable uploading of sampled data. The initial feature space, by comprehensively extracting the aforementioned temporal, statistical, and state features, forms a high-dimensional feature vector set containing multiple feature dimensions. This comprehensively characterizes the terminal's operating state and performance, providing a rich information foundation for subsequent feature selection and sampling performance prediction.

[0039] Next, a tree-based feature importance evaluation method is used to filter features in the initial feature space. Tree models, such as random forests or gradient boosting trees, can quantify the predictive importance of each feature to the target variable by analyzing the contribution of features during the tree structure splitting process. Specifically, using historical sampling accuracy deviation as the target variable and each feature in the initial feature space as the input variable, the tree model is trained and the importance score of each feature is calculated. The feature importance score reflects the correlation strength between the feature and the historical sampling accuracy deviation; a higher score indicates a more significant impact of the feature on sampling accuracy. Features are sorted in descending order according to their importance scores, and the top K features with the highest correlation are selected to construct a key sensor data sequence set. Through feature filtering, the most valuable key features for predicting sampling performance are retained, while the dimensionality of the feature space is reduced, thus decreasing the computational complexity of subsequent model training and prediction.

[0040] By preprocessing data alignment, outlier removal, and dimensional normalization, combined with multi-dimensional feature extraction from the time domain, statistics, and state, and feature importance evaluation and screening based on tree models, the most valuable key features for predicting sampling accuracy can be accurately extracted from the original multi-source sensor data. This effectively reduces the dimensionality of the feature space while ensuring prediction accuracy, thereby improving the training efficiency and generalization ability of the subsequent prediction model.

[0041] Furthermore, methods for determining the value of K include: S131. Calculate the cumulative contribution rate of the importance score of the features after evaluation and ranking. When the cumulative contribution rate reaches the preset threshold for the first time, record the number of features included at this time as the base K value. S132. Real-time acquisition of terminal CPU utilization and remaining memory capacity, determination of the maximum number of features currently allowed to be processed according to a preset resource load lookup table, comparison of the basic K value with the maximum number of features, and taking the smaller of the two as the K value after capacity constraint, wherein the resource load lookup table is a matrix mapping table based on a two-dimensional query of terminal CPU utilization and remaining memory capacity. S133. Based on the operating stage of the edge distribution transformer terminal within a preset sliding time window, the K value is optimized and adjusted. If the terminal is in the initial operation stage, after software upgrade, or in the adaptation period after drastic changes in the external environment, the K value is multiplied by an expansion coefficient greater than 1 to enhance the feature exploration capability. If the terminal is in a stable operating period, the K value remains unchanged. If the terminal performance is detected to show a continuous deterioration trend, the K value is multiplied by a focusing coefficient less than 1 to strengthen the monitoring of core features. The continuous deterioration trend refers to the fact that the predicted sampling accuracy deviation exceeds 1.5 times the standard deviation of the historical baseline value in 5 consecutive diagnostic cycles.

[0042] In a preferred embodiment, after the feature importance evaluation based on the tree model in step S13, the features are sorted in descending order of importance score. The importance scores of the sorted features are accumulated one by one to calculate the cumulative contribution rate. The cumulative contribution rate reflects the proportion of the sum of the importance scores of the top N features to the total sum of the importance scores of all features. A higher proportion indicates a stronger explanatory power of these N features for predicting sampling accuracy. When the cumulative contribution rate first reaches a preset threshold, for example, 85%, it indicates that the selected features can explain most of the sampling accuracy variation, and the number of features included at this point is recorded as the base K value. This method based on the cumulative contribution rate ensures the predictive power of the feature set while avoiding the inclusion of too many redundant features.

[0043] Subsequently, a capacity constraint is applied to the K value based on the terminal's computing resource status. Edge transformer terminals have limited computing resources; too many features would increase the computational load and potentially affect the terminal's real-time response performance. The terminal's CPU utilization and remaining memory capacity are acquired in real time, reflecting the terminal's current computing resource usage. The maximum allowed number of features to be processed is determined based on a pre-defined resource load lookup table. This table is a matrix mapping table based on a two-dimensional lookup of CPU utilization and remaining memory capacity. It discretizes the ranges of CPU utilization and remaining memory capacity into several levels, forming a two-dimensional grid. Each grid node corresponds to a pre-defined maximum feature quantity limit. When CPU utilization is low and memory is sufficient, the maximum feature quantity limit is larger; when CPU utilization is high or memory is insufficient, the maximum feature quantity limit is correspondingly lower. The base K value is compared with the maximum feature quantity limit, and the smaller of the two is taken as the capacity-constrained K value, ensuring that feature processing does not exceed the terminal's computing resource capacity.

[0044] Next, the K value is optimized and adjusted according to the operating stage of the edge distribution transformer terminal within a preset sliding time window. The terminal's feature set requirements differ at different operating stages. If the terminal is in the initial operation phase, after a software upgrade, or in the adaptation period after a drastic change in the external environment, the terminal's operating mode and fault characteristics are not yet fully understood. A wider range of features is needed to establish an accurate performance baseline. Therefore, the K value is multiplied by an expansion coefficient greater than 1, such as 1.2 or 1.5, to increase the number of features and enhance feature exploration capabilities. If the terminal is in a stable operating period, with a relatively fixed operating mode and sufficient historical data, keeping the K value unchanged is sufficient for diagnostic needs. If a continuous deterioration trend in terminal performance is detected, i.e., the prediction sampling accuracy deviation exceeds 1.5 times the standard deviation of the historical baseline value for five consecutive diagnostic cycles, it indicates that a specific potential fault may be developing in the terminal. In this case, it is necessary to focus on the most critical features for in-depth monitoring. The K value is multiplied by a focus coefficient less than 1, such as 0.7 or 0.8, to reduce the number of features and strengthen the monitoring of core features, thereby improving the sensitivity and computational efficiency of anomaly detection.

[0045] By determining the base K value based on the cumulative contribution rate, constraining the computing resource capacity, and adaptively adjusting the operation stage, a dynamic balance configuration of the number of features is achieved. This ensures that the selected key features have sufficient predictive and interpretive capabilities, avoids the impact of computing resource overload on the terminal's real-time performance, and can flexibly adjust the feature exploration breadth and monitoring focus depth according to different operation stages, thereby improving the adaptability and robustness of self-diagnosis in different application scenarios.

[0046] Furthermore, based on the key sensor data sequence set, a prediction sampling accuracy reliability score is generated to predict the terminal sampling performance, including: S14. Obtain the adaptation accuracy reliability evaluation channel by matching the key sensor data types in the key sensor data sequence set, wherein the adaptation accuracy reliability evaluation channel includes L accuracy reliability evaluation branches. S15. Perform key sensor data volatility analysis on several key sensor data sequences in the key sensor data sequence set to obtain several key sensor data variation coefficients, and calculate the comprehensive data volatility by weighting them. S16. Multiply the ratio of the comprehensive data volatility to the preset benchmark comprehensive data volatility by the number of benchmark evaluation branches selected K and round down to obtain the number of adaptation evaluation branches selected Q, where K is one-third of L. If Q is less than 1, Q is equal to 1; if Q is greater than L, Q is equal to L. S17. Randomly select Q accuracy reliability evaluation branches from the L accuracy reliability evaluation branches, perform terminal sampling performance prediction based on the key sensor data sequence set, and use the average of the Q prediction results as the prediction sampling accuracy reliability score.

[0047] In a preferred embodiment, firstly, an appropriate accuracy and reliability evaluation channel is obtained by matching the key sensor data types in the key sensor data sequence set. Since different types of key sensor data correspond to different physical characteristics and data distribution patterns, targeted evaluation models are required for prediction. The appropriate accuracy and reliability evaluation channel includes L accuracy and reliability evaluation branches. Each accuracy and reliability evaluation branch is an independently trained deep learning model capable of predicting and evaluating the sampling accuracy and reliability of the terminal based on the key sensor data sequence set. Based on the data types contained in the current key sensor data sequence set, the corresponding appropriate accuracy and reliability evaluation channel is selected to ensure that the evaluation model used in subsequent prediction processes matches the data characteristics.

[0048] Then, volatility analysis is performed on each key sensor data sequence in the key sensor data sequence set. Volatility reflects the stability of the data sequence and is an important indicator for assessing the complexity of the current operating state. The coefficient of variation (COP) of each key sensor data sequence is calculated. The COP is the ratio of the standard deviation to the mean of the data sequence. This indicator can eliminate the influence of dimensions and objectively reflect the relative volatility of the data. The larger the COP, the more drastic the fluctuation of the sensor data and the more unstable the operating state. The COPs of all key sensor data sequences are weighted and summed to obtain the comprehensive data volatility. Specifically, according to the degree of influence of each key sensor data on sampling performance, a corresponding weight coefficient is assigned to the COP of each key sensor data sequence. Sensor data with a greater impact, such as voltage, current, and clock signals, which are directly related to sampling accuracy, are assigned higher weights, such as 0.3 or 0.25; sensor data with a relatively smaller impact, such as ambient temperature and humidity, vibration, and other indirectly affecting parameters, are assigned lower weights, such as 0.1 or 0.05. The overall data volatility is obtained by multiplying the coefficient of variation of each key sensor data sequence by its corresponding weighting coefficient and then summing the results. The overall data volatility comprehensively reflects the overall stability of the terminal's current operating state; higher volatility indicates that the terminal is in a complex and ever-changing operating environment.

[0049] Next, the number of suitable evaluation branches, Q, is dynamically determined based on the overall data volatility. The ratio of the current overall data volatility to the preset benchmark overall data volatility is calculated; this ratio reflects the degree of deviation of the current data volatility from the benchmark state. This ratio is multiplied by the benchmark evaluation branch selection number K and rounded down to obtain the initial number of suitable evaluation branches. The benchmark evaluation branch selection number K is set to one-third of L, representing the recommended number of evaluation branches to use under standard operating conditions. When the overall data volatility is high, the ratio is large, and the calculated Q value is also large, meaning more evaluation branches are needed to participate in the prediction to improve robustness; when the overall data volatility is low, the ratio is small, and the Q value is correspondingly smaller, reducing computational overhead. To ensure that the Q value is within a reasonable range, boundary constraints are set: if the calculated Q is less than 1, Q is set to 1 to ensure that at least one evaluation branch participates in the prediction; if Q is greater than L, Q is set to L to avoid selecting more than the total number of available evaluation branches.

[0050] Subsequently, Q accuracy and reliability evaluation branches are randomly selected from L branches for prediction. This random selection mechanism covers different combinations of evaluation branches in multiple predictions, reducing the impact of bias in specific branches on the results. The selected Q evaluation branches predict terminal sampling performance based on key sensor data sequence sets, with each branch outputting a sampling accuracy and reliability score. The mean of the Q prediction results is used as the final prediction sampling accuracy and reliability score. This mean fusion mechanism effectively suppresses prediction bias and random errors from a single model, making the final score more stable and reliable, objectively reflecting the terminal's true sampling accuracy and health status.

[0051] By adopting a prediction mechanism based on data type matching to adapt the evaluation channel, dynamically determining the number of evaluation branches based on comprehensive data volatility, and random selection and mean fusion of multiple branches, adaptive prediction of sampling accuracy reliability score is achieved. When the data volatility is large, the number of evaluation branches is automatically increased to improve prediction robustness, and when the data volatility is small, the number of evaluation branches is reduced to reduce computational overhead. Multi-branch mean fusion effectively suppresses the prediction bias of a single model, so that the predicted sampling accuracy reliability score can accurately and stably reflect the real sampling performance health status of the terminal, providing a reliable quantitative basis for subsequent anomaly judgment.

[0052] Furthermore, the training method for the adaptation accuracy reliability evaluation channel includes: S141. Using the key sensor data type as a constraint, collect several sample key sensor data sequence sets, calculate the statistical deviation of each sample key sensor data sequence set from the preset health baseline state, map the statistical deviation to the preset scoring interval, automatically generate the corresponding sample sampling accuracy reliability score, and obtain several sample sampling accuracy reliability scores. S142. The key sensor data sequence set of several samples and the sampling accuracy reliability score of several samples are used as training data, and L-fold cross-partitioning with replacement is performed to obtain L sample training sets. S143. Train the deep learning model using the L sample training sets until convergence, and generate L accuracy and reliability evaluation branches.

[0053] In a preferred embodiment, firstly, several key sensor data sequence sets are collected as training samples, constrained by the type of key sensor data. These sample data originate from historical data of edge distribution transformer terminals under different operating scenarios and health states, ensuring that the samples cover various typical operating states that the terminals may experience. For each key sensor data sequence set, its statistical deviation from a preset health baseline state is calculated. The preset health baseline state refers to the sensor data characteristic distribution when the terminal operates under ideal conditions and the sampling accuracy is completely normal. The statistical deviation quantifies the degree of difference between the two by comparing the statistical characteristics of the sample data (such as mean, variance, distribution pattern, etc.) with the statistical characteristics of the health baseline state. The larger the deviation, the further the terminal's operating state deviates from the healthy state, and the lower the sampling accuracy reliability. The calculated statistical deviation is mapped to a preset scoring interval, such as a scoring interval from 0 to 100; the larger the deviation, the lower the corresponding score, automatically generating the corresponding sample sampling accuracy reliability score. In this way, each key sensor data sequence set is automatically labeled with a sampling accuracy reliability score, resulting in several sample sampling accuracy reliability scores and constructing a complete training dataset.

[0054] Next, several sets of key sensor data sequences and corresponding sample sampling accuracy and reliability scores are used as training data, and an L-fold cross-split with replacement is performed. L-fold cross-split with replacement involves randomly sampling all training data L times, constructing a new training set each time. Due to sampling with replacement, the same sample may appear in multiple training sets, resulting in sample overlap between different training sets, but with random differences in the overall distribution. This partitioning method increases the diversity of each training set, which helps train evaluation branches with different generalization abilities. Through L-fold cross-split, L training sets are obtained, each containing a certain number of key sensor data sequences and their corresponding sample sampling accuracy and reliability scores.

[0055] Subsequently, deep learning models are trained to convergence using L training sets of samples, generating L accuracy and reliability evaluation branches. For each training set of samples, a deep learning model is constructed. This model takes the key sensor data sequence set as input and the sampling accuracy and reliability score as the output target. The model parameters are continuously adjusted using backpropagation and gradient descent optimization methods, so that the model's predicted output gradually approaches the true sampling accuracy and reliability score. Training is complete when the prediction error of the model on the training set converges to a stable level. Due to the random differences between the L training sets of samples, the L deep learning models trained also have certain differences in parameters and prediction characteristics, forming L independent accuracy and reliability evaluation branches. These evaluation branches together constitute the adaptive accuracy and reliability evaluation channel, which can predict and evaluate the sampling accuracy and reliability of the terminal from different perspectives. Here, L is an integer greater than or equal to 10. A larger value of L can provide richer evaluation branches and improve the robustness and accuracy of the prediction. For example, when L is set to 15, 15 multilayer perceptron network models are constructed for 15 training samples. Each model contains 3 fully connected layers. Each model has a learning rate of 0.001, a batch size of 64, and a loss function of mean squared error. After 300 training iterations, convergence is determined when the loss function of each model decreases by less than 0.001 within 10 consecutive training iterations, ultimately generating 15 accuracy and reliability evaluation branches. Although each branch uses the same network structure and training parameter configuration, due to the random differences in the training data, the model parameters learned by the 15 branches are different, resulting in differentiated evaluation capabilities. In another implementation, different types of deep learning models can be used to construct L evaluation branches to further enhance the differences and complementarity between branches. For example, when L is set to 12, four multilayer perceptron models, four convolutional neural network models, and four long short-term memory network models can be constructed. Multilayer perceptron models excel at handling nonlinear mapping relationships between features, convolutional neural network models excel at extracting local pattern features of data, and long short-term memory network models excel at capturing long-term dependencies in time-series data. Different types of models learn the correlation pattern between sampling accuracy and sensor data from different perspectives. The 12 evaluation branches generated have greater diversity and can more comprehensively evaluate the sampling accuracy reliability of the terminal when performing multi-branch prediction fusion in the subsequent process.

[0056] Training samples were generated using an automatic labeling method based on statistical deviation. Combined with L-fold cross-validation with replacement and independent model training, an adaptation accuracy and reliability evaluation channel with multiple evaluation branches was constructed. Each evaluation branch, based on a different training set sample distribution, forms a complementary prediction perspective. This ensures the predictive ability of individual branches while improving the overall prediction robustness through the differences between branches, laying the model foundation for subsequent multi-branch random selection and mean fusion prediction.

[0057] Furthermore, based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, the current noise level coefficient and real-time missed alarm risk coefficient are dynamically assessed, including: S21. Real-time acquisition of the operating status data and environmental data of the edge distribution transformer terminal within a preset time zone, wherein the operating status data includes an active power load rate sequence, a core chip temperature sequence, a power supply voltage ripple RMS value sequence, and a communication channel signal-to-noise ratio sequence, and the environmental data includes an ambient temperature sequence and an ambient humidity sequence. S22. Calculate the coefficients of variation of the active power load rate sequence, core chip temperature sequence, voltage ripple RMS value sequence, signal-to-noise ratio sequence, ambient temperature sequence, and ambient humidity sequence respectively, and perform weighted summation to obtain the current noise level coefficient. S23. Calculate the real-time underreporting risk coefficient based on the active power load rate sequence.

[0058] In a preferred embodiment, firstly, real-time acquisition of operational status data and environmental data of the edge distribution transformer terminal within a preset time zone is performed. The preset time zone refers to a time window used to assess noise levels and the risk of missed alarms, such as data from the most recent 1 hour or 2 hours. Operational status data reflects the terminal's current workload and operational quality, including an active power load rate sequence, a core chip temperature sequence, a power supply voltage ripple RMS value sequence, and a communication channel signal-to-noise ratio sequence. The active power load rate sequence reflects the ratio of the terminal's actual load level to its rated load; a higher load rate indicates a greater workload for the terminal. The core chip temperature sequence reflects the processor's thermal state; excessively high temperatures may affect chip stability. The voltage ripple RMS value sequence reflects the fluctuation level of the power supply; a larger ripple indicates poorer power supply quality. The signal-to-noise ratio sequence reflects the signal quality of the communication channel; a lower signal-to-noise ratio indicates more severe communication interference. Environmental data reflects the external environmental conditions of the terminal, including an ambient temperature sequence and an ambient humidity sequence; extreme temperature and humidity conditions can affect the terminal's operational stability and sensor accuracy.

[0059] Next, the coefficients of variation (COPs) for the active power load rate sequence, core chip temperature sequence, voltage ripple RMS value sequence, signal-to-noise ratio (SNR) sequence, ambient temperature sequence, and ambient humidity sequence are calculated separately. The COP is defined as the ratio of the standard deviation to the mean of the data sequence, reflecting the relative fluctuation of the data. After calculating the COPs for each of the six data sequences, a weighted sum is performed to obtain the current noise level coefficient. Specifically, based on the degree of influence of each operating state parameter and environmental parameter on sampling accuracy, corresponding weights are assigned to each COP. For example, the voltage ripple RMS value and SNR directly affect sampling quality and can be assigned higher weights, such as 0.25 and 0.2; the active power load rate and core chip temperature affect system stability and can be assigned medium weights, such as 0.2 and 0.15; ambient temperature and ambient humidity are indirect influencing factors and can be assigned lower weights, such as 0.1 and 0.1. Each COP is multiplied by its corresponding weight and then summed to obtain the current noise level coefficient. The current noise level coefficient comprehensively reflects the complexity and interference intensity of the terminal's current operating environment. A higher coefficient indicates more severe noise interference faced by the terminal, and the more easily the sampling accuracy is affected.

[0060] Simultaneously, a real-time missed detection risk coefficient is calculated based on the active power load rate sequence. The active power load rate reflects the current workload intensity of the terminal. Under high load conditions, the terminal's computing and communication resources tend to be saturated, and self-diagnosis may reduce detection sensitivity due to resource contention, leading to an increased risk of missed detection due to abnormal sampling accuracy. Based on the statistical characteristics of the active power load rate sequence, the current missed detection risk level is quantitatively assessed, generating a real-time missed detection risk coefficient, providing a basis for subsequent threshold configuration.

[0061] By collecting real-time operational status data and environmental data, and calculating the current noise level coefficient based on the weighted fusion of multi-parameter variation coefficients and the real-time missed detection risk coefficient based on load characteristics, the complexity of the terminal's current operating environment and the risk of abnormal missed detection can be accurately quantified. This provides a reliable decision-making basis for the subsequent dynamic configuration of the accuracy and reliability scoring threshold, ensuring that the diagnostic criteria can be adaptively adjusted according to real-time operating conditions.

[0062] Furthermore, a real-time underreporting risk coefficient is calculated based on the active power load rate sequence, including: S231. The mean value of the active power load rate is obtained by calculating the mean value of the active power load rate sequence. S232. The ratio of the average active power load rate to the preset active power load rate benchmark value is used as the first risk compensation coefficient. S233. The coefficient of variation of the active power load rate sequence is used as the second risk compensation coefficient. S234. A comprehensive risk compensation coefficient is obtained by weighted fusion of the first risk compensation coefficient and the second risk compensation coefficient. S235. The product of the comprehensive risk compensation coefficient and the preset benchmark underreporting risk coefficient is used as the real-time underreporting risk coefficient.

[0063] In a preferred embodiment, firstly, the average active power load rate is calculated by averaging the active power load rate sequence. The active power load rate sequence includes load rate data from multiple sampling times within a preset time zone. By calculating the arithmetic mean, the average load level within that time zone is obtained, reflecting the overall workload of the terminal during that period.

[0064] Then, the ratio of the average active power load rate to the preset active power load rate benchmark is used as the first risk compensation coefficient. The preset active power load rate benchmark refers to the typical load rate level of the terminal under normal and stable operating conditions, such as 60%. When the average active power load rate is higher than the benchmark, the ratio is greater than 1, indicating that the terminal is currently under high load, resources are scarce, and the risk of missed alarms increases accordingly. When the average active power load rate is lower than the benchmark, the ratio is less than 1, indicating that the terminal load is relatively light, resources are sufficient, and the risk of missed alarms is relatively low. The first risk compensation coefficient reflects the impact of the average load level on the risk of missed alarms.

[0065] Subsequently, the coefficient of variation of the active power load rate sequence was used as the second risk compensation coefficient. The coefficient of variation is defined as the ratio of the standard deviation to the mean of the active power load rate sequence, reflecting the degree of load fluctuation. A larger coefficient of variation indicates more frequent and severe load fluctuations, potentially leading to momentary resource shortages or state switching delays at the terminal during rapid load changes, increasing the risk of missed anomalies. The second risk compensation coefficient reflects the impact of load volatility on the risk of missed detections.

[0066] Next, a comprehensive risk compensation coefficient is obtained by weighting and fusing the first and second risk compensation coefficients. Specifically, weights are assigned to the first and second risk compensation coefficients, for example, the first risk compensation coefficient has a weight of 0.6 and the second risk compensation coefficient has a weight of 0.4. This weighting is determined based on statistical analysis of a large amount of historical operating data and the experience of domain experts. The analysis shows that the average load level has a more direct and significant impact on system resource consumption and false negative risk, while load volatility, although it also increases false negative risk, has a relatively minor impact. Therefore, the average load level is assigned a higher weight. The comprehensive risk compensation coefficient is obtained by multiplying the two risk compensation coefficients by their corresponding weights and then summing them. The comprehensive risk compensation coefficient takes into account both the average load level and the volatility characteristics, comprehensively reflecting the degree of false negative risk under the current load conditions.

[0067] Subsequently, the product of the comprehensive risk compensation coefficient and the preset benchmark underreporting risk coefficient is used as the real-time underreporting risk coefficient. The preset benchmark underreporting risk coefficient is a benchmark value for underreporting risk pre-set under standard operating conditions based on historical statistical data and expert experience. By using the comprehensive risk compensation coefficient as an adjustment factor, the benchmark underreporting risk coefficient is dynamically adjusted to obtain the real-time underreporting risk coefficient. When the comprehensive risk compensation coefficient is greater than 1, the real-time underreporting risk coefficient is higher than the benchmark value, indicating that the current underreporting risk is high and the diagnostic threshold needs to be tightened; when the comprehensive risk compensation coefficient is less than 1, the real-time underreporting risk coefficient is lower than the benchmark value, indicating that the current underreporting risk is low and the diagnostic threshold can be appropriately relaxed.

[0068] By analyzing the mean and coefficient of variation of the active power load rate sequence and performing weighted fusion calculations based on preset benchmark values, the real-time missed detection risk can be comprehensively assessed from two dimensions: load level and load fluctuation. This allows the missed detection risk coefficient to accurately reflect the potential risk of abnormal missed detections under different load conditions, providing a precise risk quantification indicator for subsequent threshold configuration. This ensures improved detection sensitivity and effectively reduces the probability of missed detections under high load or drastic load fluctuations.

[0069] Furthermore, configure the accuracy and reliability scoring thresholds, including: S24. By querying the pre-built two-dimensional dynamic mapping table, the adaptation accuracy reliability score threshold is determined according to the matching of the current noise level coefficient and the real-time missed detection risk coefficient, wherein the adaptation accuracy reliability score threshold is negatively correlated with the current noise level coefficient and positively correlated with the real-time missed detection risk coefficient.

[0070] In a preferred embodiment, the adaptation accuracy reliability score threshold is determined by querying a pre-built two-dimensional dynamic mapping table and matching the current noise level coefficient and the real-time missed detection risk coefficient.

[0071] The two-dimensional dynamic mapping table is pre-built before formal deployment and operation. This table uses the current noise level coefficient as the horizontal axis and the real-time missed detection risk coefficient as the vertical axis, discretizing the two continuous coefficient ranges to form a regular discretized grid. Specifically, the noise level coefficient and the missed detection risk coefficient are divided into several discrete levels, and the intersection of these two dimensions forms the two-dimensional grid structure. For example, the noise level coefficient range of 0 to 1 is divided into 10 levels, each with a width of 0.1; the missed detection risk coefficient range of 0 to 2 is divided into 10 levels, each with a width of 0.2. The intersection of these two dimensions forms 100 grid nodes.

[0072] For each intersection node in the grid, a unique, corresponding accuracy and reliability score threshold is pre-calculated and stored by comprehensively utilizing historical operational data statistical analysis and domain expert experience rules. Historical data statistical analysis involves mining a large amount of historical terminal operational data to statistically determine which threshold achieves the optimal diagnostic accuracy under different noise levels and varying false negative risks. Domain expert experience rules refer to the suggestions and requirements for diagnostic standards under different operational scenarios proposed by power system operation and maintenance experts based on practical engineering experience. By integrating these two aspects of information, the most suitable threshold is configured for each grid node, forming a complete two-dimensional dynamic mapping table.

[0073] The internal data of this two-dimensional dynamic mapping table follows a clear configuration rule, reflecting the correlation between the threshold and the two coefficients. The adaptation accuracy reliability score threshold is negatively correlated with the current noise level coefficient. That is, when a certain missed detection risk coefficient is fixed and the vertical axis index remains constant, the stored threshold decreases in a step-like manner as the noise level coefficient index increases along the horizontal axis. This is because the higher the noise level, the more severe the interference faced by the terminal. Fluctuations and deviations in the sampling data are more due to environmental noise than actual sampling accuracy degradation. If strict diagnostic standards are still used in this situation, it will lead to a large number of false alarms. Therefore, in high-noise environments, the threshold is lowered, and the diagnostic standards are relatively relaxed, improving the system's tolerance to noise and avoiding excessive false alarms that affect operational efficiency. The adaptation accuracy reliability score threshold is positively correlated with the real-time missed detection risk coefficient. That is, when a certain noise level coefficient is fixed and the horizontal axis index remains constant, the stored threshold increases in a step-like manner as the missed detection risk coefficient index increases along the vertical axis. This is because a higher risk of missed detection indicates that the terminal is under high load or experiencing severe load fluctuations. If an anomaly in sampling accuracy goes undetected, it could lead to incorrect power dispatching decisions or equipment protection failures, resulting in serious consequences. Therefore, under high risk of missed detection, the threshold is increased and the diagnostic criteria are tightened to improve the system's detection sensitivity, ensuring that anomalies under critical operating conditions can be detected in a timely manner, reducing the probability of missed detection and protecting system safety.

[0074] By using a two-dimensional mapping table structure, the complex dynamic threshold determination process is transformed into a simple and efficient table query operation, significantly reducing the computational overhead of real-time diagnosis and improving response speed. During real-time diagnosis, the calculated continuous numerical form of the current noise level coefficient and the real-time missed detection risk coefficient are first discretized according to preset intervals, converting them into corresponding horizontal axis index i and vertical axis index j, respectively. Specifically, based on the magnitude of the current noise level coefficient, it is determined which preset noise level interval it falls into; this interval number is the horizontal axis index i. Similarly, based on the magnitude of the real-time missed detection risk coefficient, it is determined which preset missed detection risk interval it falls into; this interval number is the vertical axis index j. For example, if the current noise level coefficient is 0.35, assuming the 4th interval is 0.3 to 0.4, then the coefficient falls into the 4th interval, and the horizontal axis index i is 4. If the real-time missed detection risk coefficient is 1.2, assuming the 7th interval is 1.2 to 1.4, then the coefficient falls into the 7th interval, and the vertical axis index j is 7. Using the discretized index pairs as query keys, the corresponding grid node is directly located in the two-dimensional dynamic mapping table. The pre-stored adaptation accuracy reliability score threshold of that node is retrieved and output as the adaptation accuracy reliability score threshold for this diagnosis. The entire query process only requires index matching and data reading, with extremely low computational complexity, which can meet the real-time requirements of edge terminals.

[0075] In a preferred embodiment, an online update mechanism for the mapping table can also be implemented to further enhance the adaptive capability of threshold configuration. Feedback on the subsequent results of diagnostics performed by terminals using various thresholds in actual operation is collected periodically, including multi-dimensional performance indicators such as diagnostic accuracy, false alarm rate, false negative rate, and diagnostic response time. Statistical analysis of the feedback data is used to evaluate the actual performance of the currently configured thresholds for each grid node. Based on the feedback data, the thresholds for the corresponding nodes in the mapping table are fine-tuned and optimized. For example, if the threshold of a certain grid node is found to cause a persistently high false alarm rate in actual application, exceeding the preset acceptable false alarm rate upper limit, the threshold of that node is automatically lowered appropriately to relax the diagnostic criteria and reduce false alarms. If the threshold of a certain grid node is found to cause a high false negative rate, indicating that anomalies are not being detected in a timely manner, the threshold of that node is automatically raised appropriately to tighten the diagnostic criteria and reduce the risk of false negatives. Through continuous online learning and optimization, the threshold configuration in the mapping table can continuously adapt to the characteristics and needs of actual operation and maintenance scenarios, making the threshold settings more in line with the accuracy and security requirements of actual operation and maintenance scenarios, further improving the adaptability and reliability of self-diagnosis.

[0076] Example 2, as Figure 2As shown, based on the same inventive concept as the edge terminal health self-diagnosis method based on multi-source sensing provided in Embodiment 1, this embodiment of the invention also provides an edge terminal health self-diagnosis system based on multi-source sensing, comprising: The multi-source data fusion module 11 is used to extract key features and perform data fusion processing on the multi-source sensor data collected by the edge distribution transformer terminal, and generate a prediction sampling accuracy reliability score based on the key sensor data sequence set to predict the terminal sampling performance. The dynamic threshold configuration module 12 is used to dynamically evaluate the current noise level coefficient and the real-time missed alarm risk coefficient based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, and configure the adaptation accuracy reliability score threshold. The health determination module 13 is used to determine that the edge distribution transformer terminal has an abnormal sampling accuracy drift if the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, and to trigger an automatic calibration program. If the calibration fails, a maintenance request is generated and reported to the main station management platform.

[0077] Furthermore, the multi-source sensing data includes a basic physical state data sequence and a communication state data sequence. The basic physical state data includes at least temperature parameters, voltage parameters, current parameters, power parameters, clock parameters, vibration parameters, and ambient temperature and humidity. The communication state data includes at least signal strength, data transmission rate, link quality indication, bit error rate, packet loss rate, communication delay, and physical port connection status.

[0078] Furthermore, the multi-source data fusion module 11 is also used for: The basic physical state data sequence and the communication state data sequence are respectively processed by data alignment, outlier removal and dimension normalization to obtain standard physical state data sequence and standard communication state data sequence; Based on the standard physical state data sequence and the standard communication state data sequence, for sampling performance analysis, an initial feature space containing time-domain features, statistical features and state features is calculated and constructed. The initial feature space includes, but is not limited to, the voltage sampling mean, voltage sampling variance, current sampling mean, current sampling variance, the fluctuation period of the power curve, the rise slope of the core chip temperature, the cumulative jitter of the clock signal, the main frequency amplitude of the vibration energy spectrum and the burst frequency of the communication bit error rate. A tree-based feature importance assessment method is used to analyze the correlation between each feature in the initial feature space and the historical sampling accuracy deviation, and the top K features with the highest correlation are selected to construct a key sensor data sequence set.

[0079] Furthermore, methods for determining the value of K include: Calculate the cumulative contribution rate of the importance score for the features after evaluation and ranking. When the cumulative contribution rate reaches the preset threshold for the first time, record the number of features included at this time as the base K value. The system obtains the terminal's CPU utilization and remaining memory capacity in real time, determines the maximum number of features that can be processed at present based on a preset resource load lookup table, and compares the basic K value with the maximum number of features, taking the smaller value as the K value after capacity constraint. The resource load lookup table is a matrix mapping table based on a two-dimensional query of the terminal's CPU utilization and remaining memory capacity. The K value is optimized and adjusted according to the operating stage of the edge distribution transformer terminal within a preset sliding time window. If the terminal is in the initial operation stage, after software upgrade, or in the adaptation period after drastic changes in the external environment, the K value is multiplied by an expansion coefficient greater than 1 to enhance the feature exploration capability. If the terminal is in a stable operating period, the K value remains unchanged. If the terminal performance is detected to show a continuous deterioration trend, the K value is multiplied by a focusing coefficient less than 1 to strengthen the monitoring of core features. The continuous deterioration trend refers to the fact that the predicted sampling accuracy deviation exceeds 1.5 times the standard deviation of the historical baseline value in 5 consecutive diagnostic cycles.

[0080] Furthermore, the multi-source data fusion module 11 is also used for: The adaptation accuracy reliability evaluation channel is obtained by matching the key sensor data types in the key sensor data sequence set, wherein the adaptation accuracy reliability evaluation channel includes L accuracy reliability evaluation branches. The key sensor data volatility analysis is performed on several key sensor data sequences in the key sensor data sequence set to obtain several key sensor data variation coefficients, and the weighted calculation is used to obtain the comprehensive data volatility. Multiply the ratio of the comprehensive data volatility to the preset benchmark comprehensive data volatility by the number of benchmark evaluation branches selected, K, and round down to obtain the number of adaptation evaluation branches selected, Q. Here, K is one-third of L. If Q is less than 1, Q is equal to 1. If Q is greater than L, Q is equal to L. Q accuracy reliability evaluation branches are randomly selected from the L accuracy reliability evaluation branches. The terminal sampling performance is predicted based on the key sensor data sequence set, and the mean of the Q prediction results is used as the prediction sampling accuracy reliability score.

[0081] Furthermore, the training method for the adaptation accuracy reliability evaluation channel includes: Using the key sensor data type as a constraint, several sample key sensor data sequence sets are collected, and the statistical deviation of each sample key sensor data sequence set from the preset health baseline state is calculated. The statistical deviation is mapped to the preset scoring interval, and the corresponding sample sampling accuracy reliability score is automatically generated to obtain several sample sampling accuracy reliability scores. The key sensor data sequence set of several samples and the sampling accuracy reliability score of several samples are used as training data, and L-fold cross-partitioning with replacement is performed to obtain L sample training sets. The deep learning model is trained using the L sample training sets until convergence, generating L accuracy and reliability evaluation branches.

[0082] Furthermore, the dynamic threshold configuration module 12 is also used for: The system collects and acquires the operating status data and environmental data of the edge distribution transformer terminal in a preset time zone in real time. The operating status data includes the active power load rate sequence, the core chip temperature sequence, the voltage ripple RMS value sequence of the power supply, and the signal-to-noise ratio sequence of the communication channel. The environmental data includes the ambient temperature sequence and the ambient humidity sequence. The coefficients of variation of the active power load rate sequence, core chip temperature sequence, voltage ripple RMS value sequence, signal-to-noise ratio sequence, ambient temperature sequence, and ambient humidity sequence are calculated respectively, and then weighted and summed to obtain the current noise level coefficient. The real-time underreporting risk coefficient is calculated based on the active power load rate sequence.

[0083] Furthermore, the dynamic threshold configuration module 12 is also used for: The mean of the active power load rate is obtained by averaging the active power load rate sequence. The ratio of the average active power load rate to the preset active power load rate benchmark value is used as the first risk compensation coefficient. The coefficient of variation of the active power load rate sequence is used as the second risk compensation coefficient. The comprehensive risk compensation coefficient is obtained by weighting and fusing the first risk compensation coefficient and the second risk compensation coefficient. The product of the comprehensive risk compensation coefficient and the preset benchmark underreporting risk coefficient is used as the real-time underreporting risk coefficient.

[0084] Furthermore, the dynamic threshold configuration module 12 is also used for: By querying a pre-built two-dimensional dynamic mapping table, an adaptation accuracy reliability score threshold is determined based on the matching of the current noise level coefficient and the real-time missed detection risk coefficient. The adaptation accuracy reliability score threshold is negatively correlated with the current noise level coefficient and positively correlated with the real-time missed detection risk coefficient.

[0085] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0091] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A self-diagnosis method for the health of an edge terminal based on multi-source sensing, characterized in that, The methods include: Key features are extracted and data fusion is performed on the multi-source sensor data collected by the edge distribution transformer terminal. Based on the key sensor data sequence set, the terminal sampling performance is predicted and a prediction sampling accuracy reliability score is generated. Based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, the current noise level coefficient and real-time missed alarm risk coefficient are dynamically evaluated, and the adaptation accuracy reliability score threshold is configured. If the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, the edge distribution transformer terminal is determined to have an abnormal sampling accuracy drift, and an automatic calibration procedure is triggered. If the calibration fails, a maintenance request is generated and reported to the main station management platform.

2. The self-diagnosis method for the health of an edge terminal based on multi-source sensing according to claim 1, characterized in that, The multi-source sensing data includes a basic physical state data sequence and a communication state data sequence. The basic physical state data includes at least temperature parameters, voltage parameters, current parameters, power parameters, clock parameters, vibration parameters, and ambient temperature and humidity. The communication state data includes at least signal strength, data transmission rate, link quality indication, bit error rate, packet loss rate, communication delay, and physical port connection status.

3. The self-diagnosis method for the health of an edge terminal based on multi-source sensing according to claim 2, characterized in that, Key feature extraction and data fusion processing are performed on multi-source sensor data collected by edge distribution transformer terminals, including: The basic physical state data sequence and the communication state data sequence are respectively processed by data alignment, outlier removal and dimension normalization to obtain standard physical state data sequence and standard communication state data sequence; Based on the standard physical state data sequence and the standard communication state data sequence, for sampling performance analysis, an initial feature space containing time-domain features, statistical features and state features is calculated and constructed. The initial feature space includes, but is not limited to, the voltage sampling mean, voltage sampling variance, current sampling mean, current sampling variance, the fluctuation period of the power curve, the rise slope of the core chip temperature, the cumulative jitter of the clock signal, the main frequency amplitude of the vibration energy spectrum and the burst frequency of the communication bit error rate. A tree-based feature importance assessment method is used to analyze the correlation between each feature in the initial feature space and the historical sampling accuracy deviation, and the top K features with the highest correlation are selected to construct a key sensor data sequence set.

4. The edge terminal health self-diagnosis method based on multi-source sensing according to claim 3, characterized in that, Methods for determining the K value include: Calculate the cumulative contribution rate of the importance score for the features after evaluation and ranking. When the cumulative contribution rate reaches the preset threshold for the first time, record the number of features included at this time as the base K value. The system obtains the terminal's CPU utilization and remaining memory capacity in real time, determines the maximum number of features that can be processed at present based on a preset resource load lookup table, and compares the basic K value with the maximum number of features, taking the smaller value as the K value after capacity constraint. The resource load lookup table is a matrix mapping table based on a two-dimensional query of the terminal's CPU utilization and remaining memory capacity. The K value is optimized and adjusted according to the operating stage of the edge distribution transformer terminal within a preset sliding time window. If the terminal is in the initial operation stage, after software upgrade, or in the adaptation period after drastic changes in the external environment, the K value is multiplied by an expansion coefficient greater than 1 to enhance the feature exploration capability. If the terminal is in a stable operating period, the K value remains unchanged. If the terminal performance is detected to show a continuous deterioration trend, the K value is multiplied by a focusing coefficient less than 1 to strengthen the monitoring of core features. The continuous deterioration trend refers to the fact that the predicted sampling accuracy deviation exceeds 1.5 times the standard deviation of the historical baseline value in 5 consecutive diagnostic cycles.

5. The edge terminal health self-diagnosis method based on multi-source sensing according to claim 1, characterized in that, Based on the key sensor data sequence set, the terminal sampling performance is predicted to generate a predicted sampling accuracy reliability score, including: The adaptation accuracy reliability evaluation channel is obtained by matching the key sensor data types in the key sensor data sequence set, wherein the adaptation accuracy reliability evaluation channel includes L accuracy reliability evaluation branches. The key sensor data volatility analysis is performed on several key sensor data sequences in the key sensor data sequence set to obtain several key sensor data variation coefficients, and the weighted calculation is used to obtain the comprehensive data volatility. Multiply the ratio of the comprehensive data volatility to the preset benchmark comprehensive data volatility by the number of benchmark evaluation branches selected, K, and round down to obtain the number of adaptation evaluation branches selected, Q. Here, K is one-third of L. If Q is less than 1, Q is equal to 1. If Q is greater than L, Q is equal to L. Q accuracy reliability evaluation branches are randomly selected from the L accuracy reliability evaluation branches. The terminal sampling performance is predicted based on the key sensor data sequence set, and the mean of the Q prediction results is used as the prediction sampling accuracy reliability score.

6. The edge terminal health self-diagnosis method based on multi-source sensing according to claim 5, characterized in that, The training method for the adaptation accuracy reliability evaluation channel includes: Using the key sensor data type as a constraint, several sample key sensor data sequence sets are collected, and the statistical deviation of each sample key sensor data sequence set from the preset health baseline state is calculated. The statistical deviation is mapped to the preset scoring interval, and the corresponding sample sampling accuracy reliability score is automatically generated to obtain several sample sampling accuracy reliability scores. The key sensor data sequence set of several samples and the sampling accuracy reliability score of several samples are used as training data, and L-fold cross-partitioning with replacement is performed to obtain L sample training sets. The deep learning model is trained using the L sample training sets until convergence, generating L accuracy and reliability evaluation branches.

7. The self-diagnosis method for the health of an edge terminal based on multi-source sensing according to claim 1, characterized in that, Based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, the current noise level coefficient and real-time missed alarm risk coefficient are dynamically assessed, including: The system collects and acquires the operating status data and environmental data of the edge distribution transformer terminal in a preset time zone in real time. The operating status data includes the active power load rate sequence, the core chip temperature sequence, the voltage ripple RMS value sequence of the power supply, and the signal-to-noise ratio sequence of the communication channel. The environmental data includes the ambient temperature sequence and the ambient humidity sequence. The coefficients of variation of the active power load rate sequence, core chip temperature sequence, voltage ripple RMS value sequence, signal-to-noise ratio sequence, ambient temperature sequence, and ambient humidity sequence are calculated respectively, and then weighted and summed to obtain the current noise level coefficient. The real-time underreporting risk coefficient is calculated based on the active power load rate sequence.

8. The edge terminal health self-diagnosis method based on multi-source sensing according to claim 7, characterized in that, The real-time underreporting risk coefficient is calculated based on the active power load rate sequence, including: The mean of the active power load rate is obtained by averaging the active power load rate sequence. The ratio of the average active power load rate to the preset active power load rate benchmark value is used as the first risk compensation coefficient. The coefficient of variation of the active power load rate sequence is used as the second risk compensation coefficient. The comprehensive risk compensation coefficient is obtained by weighting and fusing the first risk compensation coefficient and the second risk compensation coefficient. The product of the comprehensive risk compensation coefficient and the preset benchmark underreporting risk coefficient is used as the real-time underreporting risk coefficient.

9. The self-diagnosis method for the health of an edge terminal based on multi-source sensing according to claim 1, characterized in that, Configure the accuracy and reliability scoring thresholds, including: By querying a pre-built two-dimensional dynamic mapping table, an adaptation accuracy reliability score threshold is determined based on the matching of the current noise level coefficient and the real-time missed detection risk coefficient. The adaptation accuracy reliability score threshold is negatively correlated with the current noise level coefficient and positively correlated with the real-time missed detection risk coefficient.

10. A self-diagnostic system for the health of an edge terminal based on multi-source sensing, characterized in that, For implementing the edge terminal health self-diagnosis method based on multi-source sensing as described in any one of claims 1 to 9, the system comprises: The multi-source data fusion module is used to extract key features and fuse data from multi-source sensor data collected by edge distribution transformer terminals, and to generate a prediction sampling accuracy reliability score based on the key sensor data sequence set to predict terminal sampling performance. The dynamic threshold configuration module is used to dynamically evaluate the current noise level coefficient and the real-time missed alarm risk coefficient based on the real-time operating status and environmental conditions of the edge distribution transformer terminal, and configure the adaptation accuracy reliability score threshold. The health assessment module is used to determine that the edge distribution transformer terminal has an abnormal sampling accuracy drift if the predicted sampling accuracy reliability score is less than the adaptation accuracy reliability score threshold, and to trigger an automatic calibration procedure. If the calibration fails, a maintenance request is generated and reported to the main station management platform.

Images