A sensor data anomaly pattern processing method and system

Abstract

The application provides a sensor data abnormal mode processing method and system, and relates to the technical field of sensor data processing. The method comprises the following steps: acquiring multi-channel sensor data, identifying the correlation deviation mode between different channel sensor data based on a preset correlation evaluation rule, and calculating the corresponding value evaluation score according to the identification result; when the value evaluation score is greater than or equal to the value judgment threshold, a local data slice is intercepted and a refined feature set is obtained through multi-dimensional feature extraction; and abnormal description information containing the refined feature set is reported. The method of the application aims to effectively solve the problem of insufficient identification ability of existing edge monitoring systems for new and weak faults, realize early and accurate early warning of the operating state of high-voltage cable joints, and significantly improve the stability and safety of the power system operation.

Description

Technical Field

[0001] This invention relates to the field of sensor data processing technology, and more specifically, to a method and system for processing abnormal patterns in sensor data. Background Technology

[0002] In modern power transmission systems, high-voltage cable joints play a crucial role, serving as hubs connecting different cable segments. Their operational status directly impacts the stability and security of the entire power grid. To ensure the healthy operation of these critical nodes, a specialized protective box is typically installed outside the high-voltage cable joint, integrating various sensors such as those for temperature, partial discharge, and humidity monitoring, aiming to prevent potential faults through real-time monitoring. The massive amounts of data collected by these sensors are initially processed and analyzed by an edge computing unit within the protective box, designed to detect anomalies and issue early warnings as soon as possible. However, with the increasing operating time of power systems and the increasing complexity of the environment, a new type of cable joint defect, characterized by extremely slow development, has emerged. Early signs of this defect often manifest as extremely weak and constantly changing correlation patterns between data from different sensors, rather than dramatic fluctuations in data from a single sensor. This presents unprecedented challenges to existing edge computing-based monitoring systems, as these systems are often limited by computing power and storage space, making it difficult for them to autonomously and in real-time perceive and learn these subtle correlation changes. Consequently, fault information is "hidden," failing to provide timely early warnings.

[0003] For example, in high-voltage transmission networks, cable joints are critical connection points in the entire system, and their long-term stable operation is crucial to power supply reliability. To ensure the health of these critical components, a dedicated protective box is typically installed outside each high-voltage cable joint. This protective box integrates various sensors, such as temperature sensors to monitor temperature changes, partial discharge detection units to detect minute discharge phenomena within the insulation, and humidity sensors to sense ambient humidity. These sensors continuously collect operational data on the cable joint and its surrounding environment. All collected data is transmitted in real-time to an edge computing unit within the protective box. This edge computing unit is pre-loaded with an analysis method for common cable joint fault modes, such as insulation aging and increased contact resistance. Under normal operating conditions, the edge computing unit continuously analyzes the sensor data locally. Once it detects a data pattern matching preset fault characteristics, it immediately identifies an anomaly and sends a warning to a remote monitoring center via its built-in wireless communication module, in the form of a brief alarm message and a summary of key data. This design, which moves data processing tasks closer to the data source, significantly reduces the consumption of communication bandwidth and enables faster fault warning response, thereby improving the operating efficiency and safety of the entire power transmission system.

[0004] In its initial stages of operation, the monitoring system demonstrated excellent performance, accurately identifying some known and easily identifiable early signs of faults, thus fully validating its effectiveness. However, as the system operated for an extended period, a new and more subtle challenge gradually emerged.

[0005] During long-term operation, a new type of defect, developing extremely slowly, begins to appear inside the cable joint. This defect is not a traditional fault that causes drastic changes in individual sensor data, such as a sudden rise in temperature or a sudden increase in partial discharge intensity. Instead, it manifests as a very weak and gradual change, mainly reflected in subtle shifts in the relationships between different sensor data. For example, a tiny, slowly expanding void may form inside the cable joint insulation material. The initial expansion of this void does not immediately trigger a significant partial discharge signal; the resulting discharge energy is extremely low, almost blending into the background noise. Crucially, these extremely weak discharge signals may exhibit a previously unobserved, extremely weak periodicity under specific AC voltage phases, or show a very subtle, intermittent correlation with minute fluctuations in the ambient temperature within the protection box. These subtle changes in the correlation between sensor information are the earliest precursors to this new type of fault.

[0006] However, the fault analysis methods pre-programmed within edge computing units are primarily trained on fault patterns with clear characteristics and significant changes in historical data. These methods excel at identifying situations where single sensor data exceeds thresholds, or where strong, predefined correlation patterns exist between data from several sensors. Their design philosophy is to find "obvious" anomalies. Therefore, existing methods fall short when faced with this new type of fault characterized by weak, evolving changes in cross-sensor information correlation. They cannot identify these extremely subtle, non-linear, and constantly evolving shifts in data correlation because these changes do not trigger any pre-defined alarm thresholds or match any known strong fault patterns. The edge computing unit continuously classifies these sensor data containing potential fault information as "normal" operating conditions and processes them according to standard procedures.

[0007] Because the edge computing unit fails to recognize these subtle early signs of failure, it does not flag these raw, high-frequency sensor data streams as anomalous and upload them. According to its design principles, the edge unit only periodically sends brief health reports or status summaries, processed and aggregated locally, to the monitoring center. These summaries typically only contain averages, maximum values, or whether preset alarms have been triggered for key indicators. The raw, high-resolution, multi-channel data containing crucial information about new types of failures, due to its massive volume, cannot be stored locally by the edge unit for extended periods, nor can it be continuously uploaded to the remote monitoring center for deeper analysis. The edge computing unit's computing power and storage space are relatively limited, insufficient to support algorithms that require substantial computing resources and can adaptively learn and identify complex, nonlinear, and evolving data correlation patterns. Therefore, although early information about failures is actually collected by sensors, this information is "trapped" at the edge and cannot be effectively identified and utilized.

[0008] The result of this situation is that this slowly developing, novel cable joint fault remains "hidden" under the edge monitoring system. The monitoring center, receiving only "normal" status reports from the edge units, remains unaware of the impending risk. The fault continues to develop inside the cable joint until its characteristics become sufficiently apparent—for example, the gap widens enough to trigger a stronger partial discharge, or causes a significant increase in local temperature—ultimately triggering the alarm thresholds in the edge computing unit for traditional, obvious fault modes. Only then does the system issue an alarm. However, by this time, the fault has already progressed to a relatively severe stage, missing the optimal window for early intervention and handling.

[0009] This delayed alerting has serious operational implications. When a fault is finally detected, it is often sudden and without warning, catching the operations team off guard. They are forced to respond urgently to a problem that has already progressed to the middle or late stages, which typically means more complex repairs, longer power outages (if emergency repairs are needed), and higher maintenance costs. Because the system fails to provide true early warnings, the trust of operations personnel in this advanced monitoring system is also undermined. They may begin to question its "early warning" capabilities, viewing it as a "late-stage alarm" that only alerts when the problem has become severe. This lack of trust may lead to less timely responses or reduced attention to system alerts in the future, further increasing the risk of major incidents—a direct contradiction to the original intention of deploying edge computing architecture for proactive, preventative maintenance.

[0010] There is currently no effective technical solution to the above problems. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for processing abnormal sensor data patterns, which aims to effectively solve the problem of insufficient ability of existing edge monitoring systems to identify new and weak faults, realize early and accurate early warning of the operating status of high-voltage cable joints, and significantly improve the stability and safety of power system operation.

[0012] In a first aspect, the present invention provides a sensor data anomaly mode processing method, used to process sensor data collected by sensors inside a protection box through an edge computing unit to monitor the high-voltage cable joints inside the protection box; the sensor data anomaly mode processing method includes the following steps: S1. Acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint, and perform time synchronization processing on the sensor data of each channel; S2. Based on preset correlation evaluation rules, the correlation offset patterns between different channels of sensor data are identified by real-time analysis of the multi-channel sensor data, and the corresponding value evaluation scores are calculated based on the identification results. S3. Obtain the real-time resource occupancy status of the edge computing unit, and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status; S4. When the value assessment score is greater than or equal to the value judgment threshold, a micro-slice storage mechanism is triggered. The micro-slice storage mechanism includes extracting the multi-channel sensor data before and after the triggering time for a preset duration as a local data slice for temporary caching. S5. By performing multidimensional feature extraction on the local data slices, a refined feature set representing the abnormal pattern is obtained; S6. Report the anomaly description information containing the refined feature set, and release the cache space occupied by the local data slice.

[0013] The sensor data anomaly pattern processing method provided by this invention can effectively identify new types of faults characterized by weak and evolving cross-sensor information correlation changes. By dynamically adjusting the value judgment threshold and micro-slice storage mechanism, it can accurately capture and report key anomaly information when edge unit resources are limited. This solves the problem that fault information is "trapped" at the edge in the prior art and realizes early warning and effective monitoring of the operating status of high-voltage cable joints.

[0014] Secondly, the present invention provides a sensor data anomaly mode processing system, used to process sensor data collected by sensors inside a protection box through an edge computing unit to monitor the high-voltage cable joints inside the protection box; the sensor data anomaly mode processing system includes: The processing module is used to acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint and to perform time synchronization processing on the sensor data of each channel. The analysis module is used to identify the correlation offset patterns between different channels of sensor data by performing real-time analysis on the multi-channel sensor data based on preset correlation evaluation rules, and to calculate the corresponding value evaluation score based on the identification results. The adjustment module is used to obtain the real-time resource occupancy status of the edge computing unit and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status. The interception module is used to trigger a micro-slice storage mechanism when the value assessment score is greater than or equal to the value judgment threshold. The micro-slice storage mechanism includes intercepting the multi-channel sensor data before and after the triggering time for a preset duration as local data slices for temporary caching. The extraction module is used to obtain a refined feature set representing abnormal patterns by performing multi-dimensional feature extraction on the local data slices; The reporting module is used to report anomaly description information containing the refined feature set and release the cache space occupied by the local data slice.

[0015] As can be seen from the above, the sensor data anomaly pattern processing method provided by the present invention solves the problem in the prior art where edge computing units cannot effectively identify new, weak, and evolving cross-sensor information correlation faults, resulting in fault information being "trapped" at the edge and failing to be effectively identified and utilized, through the following means: First, this application lays the foundation for subsequent correlation analysis by acquiring and time-synchronizing multi-channel sensor data. Based on this, and using pre-defined correlation evaluation rules, it performs real-time analysis of the multi-channel sensor data, identifies correlation offset patterns between different channels, and calculates corresponding value assessment scores. This method can capture subtle, non-linear, and evolving changes in data correlation that are difficult to detect with traditional threshold alarms or single-sensor analysis, thereby effectively identifying early warning signs of new types of faults.

[0016] Secondly, this application introduces the acquisition of real-time resource occupancy status of edge computing units and dynamically adjusts the preset value judgment threshold based on this status. This mechanism enables the system to flexibly adjust the sensitivity of anomaly identification according to the actual operating load of the edge units, avoiding excessive storage when resources are scarce and improving identification accuracy when resources are abundant, thereby optimizing the resource utilization efficiency of the edge units.

[0017] Furthermore, when the value assessment score reaches a threshold, this application triggers a micro-slice storage mechanism, extracting multi-channel sensor data for a preset duration before and after the triggering time as local data slices for temporary caching. This "micro-slice" strategy avoids the long-term storage and uploading of massive amounts of raw data, effectively addressing the challenges of limited storage space and bandwidth in edge units.

[0018] Finally, by performing multi-dimensional feature extraction on local data slices, a refined feature set representing abnormal patterns is obtained, and anomaly description information containing this refined feature set is reported. Simultaneously, the cache space occupied by the local data slices is released. The refined feature set occupies minimal storage space but contains key information for identifying novel faults, ensuring that critical anomaly information can be effectively extracted and uploaded to the monitoring center for deeper analysis. This overcomes the limitation of existing technologies where fault information is "trapped" at the edge.

[0019] In summary, this application effectively solves the problem of insufficient ability of existing edge monitoring systems to identify new and weak faults through innovative mechanisms such as multi-channel data correlation analysis, dynamic threshold adjustment, micro-slice storage and refined feature extraction. It achieves early and accurate early warning of the operating status of high-voltage cable joints, and significantly improves the stability and safety of power system operation.

[0020] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0021] Figure 1 This is a flowchart of a sensor data anomaly mode processing method provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a sensor data anomaly mode processing system provided in an embodiment of the present invention.

[0023] Label Explanation: 100. Processing module; 200. Analysis module; 300. Adjustment module; 400. Capture module; 500. Extraction module; 600. Reporting module. Detailed Implementation

[0024] 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] In modern power transmission systems, high-voltage cable joints are critical hubs connecting different cable segments, and their operational status directly affects the stability and security of the entire power grid. To ensure the healthy operation of these critical nodes, protective boxes are typically installed outside the high-voltage cable joints, integrating temperature sensors, partial discharge sensors, and humidity sensors to prevent potential faults through real-time monitoring. The massive amounts of data collected by these sensors are first processed and analyzed by an edge computing unit inside the protective box, aiming to detect anomalies and issue early warnings as soon as possible. However, existing edge computing-based monitoring systems primarily train their fault analysis methods on fault patterns with clear characteristics and significant changes in historical data. These methods excel at identifying situations where single sensor data exceeds thresholds or where there are strong, predefined correlation patterns between several sensor data points, aiming to find those "obvious" anomalies. Therefore, existing methods fall short when faced with new faults characterized by weak, evolving changes in cross-sensor information correlation. They cannot identify these extremely subtle, nonlinear, and constantly evolving data correlation shifts because these changes do not reach any preset alarm thresholds or match any known strong fault patterns. The edge computing unit continuously classifies sensor data containing potential fault information as "normal" and processes it according to standard procedures. Because the edge computing unit fails to identify these subtle early signs of faults, it does not mark these raw, high-frequency sensor data streams as anomalous data for uploading. According to its design principles, the edge unit only periodically sends brief health reports or status summaries, processed and aggregated locally, to the monitoring center. These summaries typically only contain averages, maximum values, or whether preset alarms have been triggered for key indicators. The raw, high-resolution, multi-channel data containing novel fault information, due to its massive volume, cannot be stored locally by the edge unit for extended periods, nor can it be continuously uploaded to the remote monitoring center for deeper analysis. The edge computing unit's computing power and storage space are relatively limited, insufficient to support algorithms that require substantial computing resources and can adaptively learn and identify complex, nonlinear, and evolving data association patterns. Therefore, although early fault information has actually been collected by sensors, this information is "trapped" at the edge and cannot be effectively identified and utilized. As a result of this situation, the slowly developing, new type of cable joint failure continues to "hide" from the edge monitoring system. Because the monitoring center only receives "normal" status reports from the edge units, it is completely unaware of the impending risks and misses the best opportunity for early intervention and handling.

[0027] For reference, see the appendix. Figure 1This invention provides a method for processing abnormal sensor data patterns, used to process sensor data collected by sensors inside a protection box through an edge computing unit to monitor high-voltage cable joints inside the protection box; the method includes the following steps: S1. Acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint, and perform time synchronization processing on the sensor data of each channel; the multi-channel sensor data includes sensor data from temperature sensor, partial discharge sensor and humidity sensor respectively; S2. Based on the preset correlation evaluation rules, the correlation offset patterns between different channel sensor data are identified by real-time analysis of multi-channel sensor data, and the corresponding value evaluation scores are calculated based on the identification results. The correlation evaluation rules include the co-occurrence rules of partial discharge signals and temperature fluctuation signals, the non-random jitter rules of single-channel sensor data, and the phase or frequency correlation rules between sensor data of each channel. S3. Obtain the real-time resource occupancy status of the edge computing unit and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status; the real-time resource occupancy status includes the CPU utilization rate and memory utilization rate of the edge computing unit; S4. When the value assessment score is greater than or equal to the value judgment threshold, the micro-slice storage mechanism is triggered. The micro-slice storage mechanism includes extracting multi-channel sensor data before and after the trigger time for a preset duration as local data slices for temporary caching. S5. By performing multidimensional feature extraction on local data slices, a refined feature set representing abnormal patterns is obtained; S6. Report anomaly description information containing a refined feature set and release the cache space occupied by the local data slice.

[0028] This application aims to address the problem in the prior art that edge computing units have difficulty identifying new, weak, and evolving cross-sensor data correlation anomalies, thereby effectively improving the early warning capability of high-voltage cable joint faults.

[0029] This application utilizes an edge computing unit to efficiently process sensor data collected by sensors inside the protection box, enabling precise monitoring of the operating status of high-voltage cable joints. The "edge computing unit" refers to a device deployed near the data source, possessing certain computing, storage, and networking capabilities, such as an industrial PC, embedded system, or dedicated gateway. Its primary function is to perform local data processing and preliminary analysis to alleviate cloud-based pressure and shorten response time. The "protection box" typically refers to the housing used to encapsulate and protect the high-voltage cable joint and its internal sensors, providing physical protection and environmental isolation. The "high-voltage cable joint" is a critical component connecting high-voltage cables; its internal insulation condition and connection quality directly affect the reliability of power transmission.

[0030] "Multi-channel sensor data" refers to data streams collected from different types of sensors (such as temperature sensors, partial discharge sensors, and humidity sensors), which collectively characterize the operating status of the high-voltage cable joint. "Time synchronization processing" refers to time calibration of data from different sensors to ensure all data points are aligned on the time axis for accurate correlation analysis. "Correlation evaluation rules" are a pre-defined set of logic or algorithms used to identify abnormal correlation patterns between different sensor data, such as co-occurrence rules for partial discharge signals and temperature fluctuation signals, non-random jitter rules for single-channel sensor data, and phase or frequency correlation rules between sensor data from different channels.

[0031] The co-occurrence rule of partial discharge signals and temperature fluctuation signals means that when a partial discharge sensor collects a partial discharge signal, if a temperature sensor simultaneously detects a significant temperature fluctuation or increase, then these two phenomena are considered to co-occur, potentially indicating anomalies such as overheating or insulation degradation inside the high-voltage cable joint. This rule can effectively capture the coupling relationship between partial discharge and thermal effects, thereby improving the ability to identify specific fault modes.

[0032] The non-random jitter rule for single-channel sensor data refers to analyzing sensor data from a single sensor to identify whether there are continuous, non-random fluctuations, drifts, or specific patterns of jitter, rather than occasional noise interference. For example, if temperature data shows a slow rise or periodic fluctuation over a period of time, and this fluctuation does not conform to normal operating patterns, it may indicate a fault in the sensor itself or a slowly developing anomaly inside the high-voltage cable joint. This rule helps identify potential problems caused by abnormal changes in a single physical quantity.

[0033] It should be noted that to determine or identify whether sensor data meets the non-random jitter rule, continuous monitoring and analysis of the sensor data are required, along with signal processing and statistical methods to distinguish between random noise and non-random patterns. Specific methods include: 1. Conduct trend analysis: For example, analysis based on moving averages: by calculating the average of data points over a certain time window, short-term random noise can be smoothed out, making it easier to reveal potential long-term upward, downward, or stable trends in the data. If the moving average continues to drift upward or downward, it may indicate the presence of non-random trend fluctuations.

[0034] For example, analysis based on linear regression or polynomial fitting: data is fitted over a period of time. If the fitted curve shows a significant slope or a specific curve shape, and this trend persists, it indicates that there is a non-random drift or pattern in the data.

[0035] 2. Perform fluctuation pattern identification: For example, based on periodicity analysis: if the data exhibits repetitive, regular rising and falling patterns, i.e., periodic fluctuations, its period and amplitude can be identified using frequency domain analysis methods such as Fourier Transform or Wavelet Analysis. If the identified periodic fluctuations do not conform to the known period of normal system operation, it may be a type of non-random jitter.

[0036] Another example is statistical feature analysis: calculating the statistical characteristics of data over different time periods, such as mean, variance, and standard deviation. If these statistical characteristics change significantly and continuously over a period of time (for example, a continuously increasing standard deviation indicates increased volatility, and a continuous shift in the mean indicates drift), it may indicate the presence of non-random fluctuations.

[0037] 3. Perform anomaly detection: For example, threshold settings: In addition to setting absolute value thresholds, you can also set rate of change thresholds. For instance, if the rate at which temperature data rises or falls within a unit of time exceeds a preset normal range, it may indicate an anomaly.

[0038] In the above embodiments, defining "whether it conforms to normal operating rules" is key to identifying non-random jitter, which requires combining system design specifications, historical data, physical principles, and domain knowledge. Specific definition methods include: 1. Based on system design specifications and operation manual: Normal operating range: Under normal operating conditions, the output parameters (such as temperature, pressure, voltage, etc.) of a system or sensor should be within a clearly defined range. Any fluctuations that continuously exceed this range can be considered as not conforming to normal operating conditions.

[0039] Response characteristics: The system has design specifications for response time, response amplitude, etc., to specific inputs or operations. For example, if a heating system should reach a stable temperature within a specific time after startup, but actual data shows that its temperature rises slowly and cannot stabilize, this does not conform to normal behavior.

[0040] 2. Based on historical data and baseline analysis: Establish a normal operating baseline: Collect sensor data of the system under known normal operating conditions and establish a "normal" data model or statistical distribution. This includes the average value, fluctuation range, and trend under normal conditions.

[0041] Baseline Comparison: Compare the current real-time sensor data with the established normal operating baseline. If the current data consistently deviates from the statistical characteristics of the baseline (e.g., the mean continuously drifts out of the normal range, the fluctuation range continues to increase, or a new periodic pattern appears), it can be considered inconsistent with normal operating patterns. For example, if historical data shows that the temperature should remain at 20℃±1℃ during normal operation, while current data shows that the temperature is slowly rising between 22℃ and 25℃, this is abnormal.

[0042] 3. Based on physical principles and domain knowledge: Understanding the physical processes: A deep understanding of the physical processes of the monitored object. For example, regarding the temperature data inside a high-voltage cable joint, under normal circumstances, its temperature should maintain a reasonable correlation with factors such as load current and ambient temperature. If the temperature continues to rise when the load is stable and the ambient temperature does not change significantly, it may indicate abnormal internal heating, which does not conform to normal energy conservation and heat dissipation laws.

[0043] Causal analysis: This involves analyzing the causal relationship between sensor data fluctuations and other system parameters or external events. If a fluctuation occurs but no reasonable external cause or internal system operation can be found to correspond to it, this fluctuation is likely an anomaly that does not conform to normal operating patterns. For example, if temperature data exhibits periodic fluctuations without changes in load or ambient temperature, it may indicate a problem with the sensor itself or its connections.

[0044] 4. Statistical Process Control (SPC) Method: Control charts: Control charts such as X-bar charts, R charts, EWMA charts, or CUSUM charts are used to monitor the mean and variability of sensor data. These charts use statistical methods to set upper and lower control limits. If data points consistently fall outside the control limits, or if multiple consecutive points appear on the same side of the center line, or if there are continuous increases or decreases—indicating an "out-of-control" pattern—it means that the process does not conform to statistically normal operating conditions.

[0045] The phase or frequency correlation rules between sensor data from different channels refer to analyzing the phase relationship or frequency coupling characteristics of sensor data from different channels (such as partial discharge signals and structural vibration data, temperature data and humidity data, etc.) in the frequency domain. For example, if a specific frequency component of a partial discharge signal has a fixed phase difference or frequency synchronization with a specific frequency component of the vibration data of a high-voltage cable joint structure, it may indicate a specific mechanical or electrical fault mode. This rule can reveal deep interactions between different physical quantities, thus providing a more comprehensive dimension for anomaly pattern recognition.

[0046] "Association offset pattern" refers to anomaly patterns where sensor data deviate from normal correlation. "Value assessment score" is a quantitative indicator calculated based on the identified association offset patterns and their strength, used to measure the potential risk or importance of the anomaly pattern. "Real-time resource occupancy status" refers to the computing and storage resource usage of the edge computing unit at a given moment, including CPU and memory utilization. "Value judgment threshold" is a critical value used to determine whether to trigger the anomaly handling mechanism; it is dynamically adjusted based on the resource status of the edge computing unit. "Micro-slice storage mechanism" is an efficient data storage strategy. When a potential anomaly is detected, only multi-channel sensor data for a preset duration before and after the trigger moment is temporarily cached as "local data slices" to minimize storage overhead. "Refined feature set" consists of key features extracted from the local data slices that concisely and accurately characterize the anomaly pattern. They occupy minimal storage space but contain crucial information for identifying novel faults, effectively addressing the challenge of limited storage space in edge units.

[0047] The core technological concept of this solution lies in its transformation of how traditional edge monitoring systems perceive anomalies. Instead of relying solely on whether data from a single sensor exceeds a preset absolute threshold, it shifts its focus to the subtle, non-linear, and constantly changing relationships between data from different sensors. Within the edge computing unit, the "value density" of the data stream is evaluated in real time—that is, whether it contains subtle correlation patterns that might lead to new types of faults. Once such high-value information is detected, the system initiates a lightweight "micro-slicing" mechanism to perform deep analysis on high-resolution raw data within a specific time period, extracting "high-value features" that reveal the essence of the fault, and reporting them as a highly compressed "anomaly summary." This method enables edge units to autonomously and in real-time capture and utilize early fault precursors hidden deep within the data, which are difficult to detect using traditional methods, even in resource-constrained environments, thereby achieving timely early warning of new, slowly evolving faults.

[0048] In the embodiments of this application, it is first necessary to acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint and perform time synchronization processing on the sensor data of each channel. The multi-channel sensor data may include sensor data from temperature sensors, partial discharge sensors, and humidity sensors, respectively. For example, the temperature sensor may be a thermistor, thermocouple, or infrared temperature sensor, used to monitor temperature changes on the joint surface in real time. The partial discharge sensor may be an ultra-high frequency (UHF) sensor, acoustic sensor, or extra-high frequency (TEV) sensor, used to detect partial discharge signals in the insulating medium. The humidity sensor may be a capacitive humidity sensor or a resistive humidity sensor, used to monitor the humidity level inside the protection box. These sensors can continuously acquire data at a fixed sampling frequency (e.g., 1000 times per second). To ensure the accuracy of subsequent analysis, time synchronization processing of these multi-channel sensor data is required. One implementation is that all sensors sample using a unified clock source, or that post-processing is performed after data acquisition using a timestamp alignment algorithm. For example, a high-precision timestamp can be attached to each sensing data point, and then the data from different channels can be aligned based on these timestamps to eliminate time deviations caused by transmission delays or sampling asynchrony.

[0049] Next, based on preset correlation evaluation rules, real-time analysis of multi-channel sensor data is performed to identify correlation shift patterns between different channel sensor data, and corresponding value evaluation scores are calculated based on the identification results. Correlation evaluation rules can include co-occurrence rules for partial discharge signals and temperature fluctuation signals, non-random jitter rules for single-channel sensor data, and phase or frequency correlation rules between sensor data from different channels. For example, a rule can be set that when a partial discharge signal appears within a certain time period and is accompanied by a small but continuous upward trend in temperature sensor data, this may be identified as a correlation shift pattern. Another example is that when sensor data from a certain channel (e.g., humidity data) exhibits non-random, small-amplitude repetitive jitter within a short period, rather than stable changes or random noise, this is also considered a potential abnormal pattern. Furthermore, abnormal phase or frequency correlation patterns can be identified by analyzing the phase difference or frequency characteristics between different channel sensor data (e.g., partial discharge signals and AC voltage signals).

[0050] Specifically, the steps for identifying anomalous phase correlation patterns by analyzing phase differences include: 1. Establish a baseline for normal phase correlation modes: For healthy insulation systems or specific types of known defects, partial discharge (PD) activity typically occurs within a specific phase window of the AC voltage cycle. For example, some internal cavity discharges may exhibit higher discharge intensity near voltage peaks (e.g., 90° and 270°), while surface discharges or corona discharges may be more active near voltage zero crossings (e.g., 0° and 180°). Through long-term monitoring and the accumulation of historical data, a normal phase difference distribution pattern between these signals can be established. The most common visualization method is to plot a phase-resolved partial discharge (PRPD) map, which visually displays the amplitude, number, and phase distribution of partial discharge pulses within the AC voltage cycle. This baseline pattern represents the "fingerprint" of the equipment under normal operating conditions.

[0051] 2. Identify anomalous phase correlation patterns: When the insulation condition of a device changes or new defects appear, the phase correlation pattern between the partial discharge signal and the AC voltage signal will deviate from the normal reference. The anomaly can be identified through the following aspects: Phase shift: If the partial discharge activity shifts significantly relative to the typical phase window of the AC voltage (e.g., the discharge that originally occurred near the peak of the positive half-cycle shifts to near the peak of the negative half-cycle, or the overall discharge phase is advanced or delayed), this may indicate a change in the nature, location, or local electric field distribution of the defect. Such a shift often foreshadows a deterioration in the insulation condition or a change in the defect type.

[0052] PRPD pattern distortion: Changes in the shape, density, or distribution of the PRPD pattern are an important indicator of anomalies. For example, broadening (expansion of the discharge phase range), narrowing (shrinkage of the discharge phase range), the appearance of new discharge clusters (indicating the possible existence of new discharge sources or discharge mechanisms), or the disappearance of existing discharge clusters may all indicate the development of defects or the generation of new defects.

[0053] The appearance or disappearance of characteristic patterns: Certain types of insulation defects (such as internal cavity discharge, surface discharge, corona discharge, etc.) have their own unique PRPD spectral characteristics. If a characteristic pattern associated with a severe defect is detected, or a pattern associated with a known benign defect disappears, it should be considered an anomaly and further investigation is required.

[0054] Multi-channel phase difference analysis: When using multiple sensors for partial discharge detection, the phase difference between the partial discharge signals received by different sensors can be used to locate defects. Abnormal changes in these phase differences may indicate a shift in the defect location or a change in the signal propagation path, thus indicating structural problems within the equipment.

[0055] The steps for identifying anomalous frequency correlation patterns by analyzing frequency characteristics include: 1. Establish a benchmark for normal frequency characteristics: Partial discharge signals are broadband signals, and their frequency components are affected by various factors such as discharge type, defect size, discharge medium, and signal propagation path. For stable insulation systems or known defects, the frequency spectrum of partial discharge signals usually remains relatively stable. By performing spectral analysis on partial discharge signals (such as Fast Fourier Transform, FFT), their frequency components and corresponding amplitude distributions can be obtained, thereby establishing a normal frequency characteristic pattern. This pattern reflects the "frequency fingerprint" of the partial discharge signal under normal conditions.

[0056] 2. Identify abnormal frequency correlation patterns: When the insulation condition of equipment changes, the frequency characteristics of the partial discharge signal also change. Anomalies can be identified through the following aspects: The emergence of new frequency components: If previously unobserved frequency components appear in the spectrum of a partial discharge signal, this may indicate the emergence of a new type of defect, a change in the discharge mechanism, or an increase in defect size. For example, resonances at certain specific frequencies may be associated with specific types of defects or equipment structures.

[0057] Changes in the amplitude of existing frequency components: A significant increase or decrease in the amplitude of a specific frequency band may indicate a change in the intensity of partial discharge or an increase in the severity of defects. For example, a significant enhancement of high-frequency components may be associated with more intense discharge activity, while changes in low-frequency components may be related to an increase in discharge energy.

[0058] Spectral broadening or narrowing: The broadening or narrowing of the overall spectrum of a partial discharge signal may reflect changes in the complexity or simplification of the discharge process. For example, spectral broadening may indicate the coexistence of multiple discharge mechanisms or an increase in the instability of the discharge process.

[0059] Frequency correlation patterns between different signals: In addition to the frequency characteristics of the partial discharge signal itself, the frequency correlation between the partial discharge signal and AC voltage signals or other environmental signals (such as noise) can also be analyzed. For example, if harmonics are present in the AC voltage signal, and these harmonics exhibit abnormal coupling or correlation with specific frequency components in the partial discharge signal, this may indicate abnormal interactions or resonance phenomena within the system, thus revealing potential faults.

[0060] For each identified associated offset pattern, a value assessment score can be calculated. For example, different scores can be assigned based on the pattern's duration, intensity, or similarity to historical anomalous patterns. A simple calculation method is to pre-set a base score for each identified associated offset pattern and then weight it according to its duration or intensity.

[0061] Simultaneously, it's necessary to acquire the real-time resource utilization status of the edge computing unit and dynamically adjust the preset value judgment threshold based on this status. Real-time resource utilization status can include the CPU and memory utilization rates of the edge computing unit. For example, the edge computing unit can have a built-in resource monitoring module that periodically (e.g., every 5 seconds) collects CPU and memory usage data. Based on these real-time resource utilization statuses, the value judgment threshold can be dynamically adjusted. For instance, when CPU or memory utilization is high, to avoid system overload due to processing too many low-value anomalies, the value judgment threshold can be appropriately increased, thus processing only those more important anomalies. Conversely, when resource utilization is low, the threshold can be appropriately decreased to more sensitively capture subtle potential anomalies. This dynamic adjustment mechanism ensures that the system remains efficient and robust even in resource-constrained environments.

[0062] When the value assessment score is greater than or equal to the value determination threshold, the micro-slice storage mechanism is triggered. This mechanism involves temporarily caching multi-channel sensor data as local data slices for a preset duration before and after the trigger moment. For example, when the calculated value assessment score reaches or exceeds the dynamically adjusted value determination threshold, the system immediately triggers a data retrieval operation. A preset duration, such as 30 seconds before the trigger moment to 10 seconds after, can be used to extract all multi-channel sensor data (including temperature, partial discharge, and humidity data) from the circular buffer during this period. These extracted data form a "local data slice" and are temporarily stored in the cache area of ​​the edge computing unit. This mechanism avoids the enormous storage pressure of storing all the original data while retaining sufficient information before and after the anomaly occurs for subsequent in-depth analysis.

[0063] Subsequently, by performing multi-dimensional feature extraction on local data slices, a refined feature set representing the anomaly pattern is obtained. For example, various signal processing and data analysis techniques can be applied to extract features from the extracted local data slices. This can include trend analysis and fluctuation amplitude calculation for temperature data; pulse counting, discharge quantity analysis, and spectrum analysis (such as FFT transformation to obtain the dominant frequency component, energy concentration, and harmonic components) for partial discharge data; and mean and variance calculation for humidity data. In addition, cross-channel feature extraction can be performed, such as analyzing the cross-correlation, phase difference, or frequency coupling between partial discharge signals and temperature fluctuations. These extracted features, such as dominant frequency components, energy concentration, harmonic components, and inter-channel frequency domain phase or amplitude correlation features, together constitute a "refined feature set." This set aims to retain the key information of the anomaly pattern to the greatest extent possible with the minimum amount of data, thereby effectively addressing the challenge of limited storage space in edge units.

[0064] Finally, an anomaly description containing a refined feature set is reported, and the cache space occupied by the local data slice is released. For example, once the refined feature set is successfully extracted, the edge computing unit generates an anomaly description packet. This packet may include the time and location of the anomaly (high-voltage cable connector number), the type of anomaly pattern identified, the value assessment score, and, most importantly, the refined feature set. This packet is then reported to a remote monitoring center or cloud platform via a network (e.g., cellular network, LoRaWAN, or Ethernet) for further diagnosis and analysis. Because a refined feature set is reported instead of raw data, network bandwidth consumption is significantly reduced. After the anomaly description information is successfully reported, to optimize the resource utilization of the edge computing unit, the cache space occupied by the previously temporarily cached local data slice is immediately released, freeing up space for subsequent potentially abnormal data.

[0065] The sensor data anomaly pattern processing method proposed in this application aims to address the challenges faced by traditional edge computing systems in identifying novel, subtle, and evolving high-voltage cable joint faults. Traditional methods often rely on preset thresholds or strong correlation patterns, making it difficult to capture subtle, non-linear correlation shifts between multi-channel sensor data. The core innovation of this application lies in introducing a mechanism of "correlation evaluation rules" and "dynamically adjusting value judgment thresholds," enabling edge computing units to identify potential anomalies more intelligently and flexibly.

[0066] Specifically, this application first ensures the accuracy of data analysis by performing time synchronization processing on multi-channel sensor data. Subsequently, based on "correlation evaluation rules" such as the co-occurrence rules of partial discharge signals and temperature fluctuation signals, the non-random jitter rules of single-channel sensor data, and the phase or frequency correlation rules between sensor data from different channels, the data is analyzed in real time to identify "correlation offset patterns" between sensor data from different channels. This method overcomes the limitations of single-sensor threshold alarms and can capture weak anomalies that only become apparent after the fusion of multi-sensor information. For example, when a local discharge signal is extremely weak and insufficient to trigger a traditional threshold alarm, but it exhibits a previously unknown co-occurrence pattern with minute temperature fluctuations, the method of this application can identify this potential precursor to a fault.

[0067] Furthermore, this application introduces a mechanism to dynamically adjust the "value judgment threshold" based on the "real-time resource occupancy status" of the edge computing unit. This enables the system to intelligently adjust its sensitivity to anomalies according to its own computing and storage capabilities. When resources are scarce, the system prioritizes processing anomalies of higher value to avoid missing critical information due to overload; when resources are abundant, it can more sensitively capture subtle potential anomalies. This adaptive threshold adjustment mechanism significantly improves the robustness and efficiency of the edge computing unit in resource-constrained environments.

[0068] When the "value assessment score" of an identified abnormal pattern reaches or exceeds the dynamically adjusted "value judgment threshold," this application triggers a "micro-slice storage mechanism." This mechanism only extracts multi-channel sensor data for a preset duration before and after the triggering time as "local data slices" for temporary caching, rather than storing all the original data. This greatly alleviates the problem of limited storage space in edge units while retaining sufficient information before and after the anomaly occurs. Subsequently, a "refined feature set" is obtained by performing multi-dimensional feature extraction on the local data slices. These refined feature sets contain key information for identifying new types of faults with a very small amount of data, effectively solving the problem of the original data being too large to upload. Finally, an anomaly description information containing the refined feature set is reported, and the cache space occupied by the local data slices is released.

[0069] Compared to the closest existing technologies, this application has the advantage that existing technologies often "trapped" raw data at the edge, failing to effectively identify and utilize subtle, evolving cross-sensor correlation changes that contain key information about novel faults. This application, however, through innovative correlation evaluation rules and dynamic threshold adjustment mechanisms, enables edge computing units to proactively identify these "hidden" fault precursors. Simultaneously, the micro-slice storage and refined feature extraction mechanisms solve the problem of limited storage and uploading of raw data by edge units, ensuring that key anomaly information can be efficiently identified, extracted, and reported to the monitoring center. Therefore, this application significantly improves the early warning capability for high-voltage cable joint faults, avoiding missed optimal intervention opportunities due to information lag, and providing a more reliable guarantee for the safe and stable operation of the power system.

[0070] In some embodiments, the specific steps in step S2 include: S21. Obtain structural vibration data of the protective box; S22. Calculate the root mean square value of the structural vibration data within a preset sliding time window, and identify the low-frequency energy peak in the structural vibration data by performing low-frequency feature identification on the structural vibration data; S23. When the root mean square value exceeds the preset vibration threshold and a low-frequency energy peak is detected, an external vibration indication signal is generated; S24. Based on the correlation evaluation rules, the correlation offset patterns between different channels of sensor data are identified by real-time analysis of multi-channel sensor data, and a preliminary value evaluation score is calculated based on the identification results. S25. Determine whether the associated offset mode and the external vibration indication signal overlap in the time domain. If they overlap in the time domain and their frequency characteristics are consistent, reduce the preliminary value assessment score; otherwise, increase the preliminary value assessment score. S26. The adjusted preliminary value assessment score shall be used as the final output value assessment score.

[0071] Specifically, acquiring structural vibration data of the protective box refers to directly collecting mechanical vibration information experienced by the protective box during operation by deploying vibration sensors on the box. This vibration data can serve as a direct basis for determining the existence and intensity of external interference. For example, a microelectromechanical system (MEMS) accelerometer can be used, which features low power consumption and small size, making it suitable for deployment in edge computing environments; or a piezoelectric vibration sensor can be used, converting mechanical vibration into an electrical signal through the piezoelectric effect. The root mean square (RMS) value of the structural vibration data within a preset sliding time window is calculated, and low-frequency energy peaks in the structural vibration data are identified through low-frequency feature recognition, aiming to quantify the vibration intensity and identify its main frequency components. The RMS value can effectively reflect the overall energy of the vibration; for example, it can be obtained by squaring, summing, averaging, and then taking the square root of the vibration data collected within the sliding time window. Low-frequency feature recognition can be performed by performing Fourier transform (FFT) or wavelet transform on the vibration data to analyze its spectral distribution, thereby identifying whether there are significant energy peaks in a specific low-frequency range (e.g., 1Hz to 10Hz). When the root mean square value exceeds a preset vibration threshold and the low-frequency energy peak is detected, an external vibration indication signal is generated, indicating that the system has confirmed the existence of external vibration interference that may affect sensor data. This indication signal can be a Boolean flag or an event message containing vibration intensity and dominant frequency information. The core logic of this scheme is to determine whether the associated offset pattern and the external vibration indication signal overlap in the time domain. If they overlap in the time domain and their frequency characteristics are consistent, the preliminary value assessment score is reduced; otherwise, the preliminary value assessment score is increased. Time domain overlap can be determined by comparing whether the timestamp of the associated offset pattern and the timestamp of the external vibration indication signal are within a preset time window (e.g., one second before or after). Frequency characteristic consistency can be determined by comparing whether the spectral characteristics of the sensor data (e.g., temperature fluctuations) involved in the associated offset pattern are similar to the spectral characteristics of the external vibration indication signal. If they are highly matched, the associated offset pattern is considered likely to be caused by external vibration, and its value as a precursor to internal faults should be reduced; conversely, if they do not match, its value as a precursor to internal faults should be increased. Specifically, spectral features typically refer to key information extracted in the frequency domain after performing a Fourier Transform (FFT) or other time-frequency analysis methods (such as wavelet transform) on a time-domain signal. These features can include: 1. Dominant frequency component: The frequency point or frequency range in which the energy is most concentrated in the signal.

[0072] 2. Energy concentration: The distribution of signal energy within a specific frequency range.

[0073] 3. Harmonic components: Frequency components that are integer multiples of the fundamental frequency in a signal.

[0074] 4. Spectral envelope or shape: The overall trend and shape of the entire frequency distribution.

[0075] Based on these spectral characteristics, the similarity of spectral characteristics can be determined in the following ways: 1. Frequency comparison: Compare whether the frequency components of two signals are similar.

[0076] 2. Energy band comparison: Compare the energy distribution or energy ratio of two signals in a specific frequency band (such as low-frequency vibration band, power frequency harmonic band, etc.).

[0077] 3. Spectral Correlation Analysis: Calculate the correlation coefficient between the power spectral density (PSD) or energy spectral density (ESD) of two signals.

[0078] 4. Feature vector distance: Construct feature vectors from spectral features (such as main frequency, energy concentration, etc.), and then calculate the distance (such as Euclidean distance) or similarity (such as cosine similarity) between these vectors.

[0079] A high-level match is, for example: Main frequency close: If the frequency difference between the main frequency component of the sensor data in the associated offset mode and the main frequency component of the external vibration indication signal is less than a preset minimum frequency threshold (e.g., less than 0.5 Hz or 1 Hz), they are considered to be highly matched in terms of main frequency.

[0080] Overlapping energy distribution: If two signals have very similar energy concentration or proportion in one or more key frequency bands (e.g., external vibrations are usually in the lower frequency range), or their power spectral density curves exhibit a highly consistent shape in these frequency bands, for example, the correlation coefficient between the power spectral density curves is higher than a high threshold (e.g., 0.8 or 0.9), then the energy distribution is considered to be highly matched.

[0081] Feature vector similarity: If the distance between the feature vectors extracted from the spectral features of two signals is very small, or the cosine similarity is very high (e.g., greater than 0.95), then they are considered to be highly matched.

[0082] When these highly matched conditions are met, the system considers the associated offset pattern to be likely caused by external vibration, and therefore its value as a precursor to internal faults should be reduced. This is because signals caused by external interference may have similar spectra to internal fault signals, and this matching judgment can effectively eliminate false alarms caused by external interference.

[0083] Conversely, a mismatch is, for example: Large frequency difference: If there is a significant frequency difference between the main frequency component of the sensor data in the associated offset mode and the main frequency component of the external vibration indication signal, which far exceeds the preset minimum frequency threshold.

[0084] Energy distribution non-overlapping: If two signals have significantly different energy distributions or energy proportions in key frequency bands, or their power spectral density curves exhibit significantly different shapes in these frequency bands, for example, if the correlation coefficient between the power spectral density curves is below a low threshold (e.g., 0.5 or 0.6), then the energy distributions are considered mismatched.

[0085] Large differences in feature vectors: If the distance between feature vectors is large, or the cosine similarity is low, they are considered to be mismatched.

[0086] When a mismatch occurs, the system considers the associated offset pattern unlikely to be caused by external vibration, or at least not as the primary cause, thus enhancing its value as a precursor to internal faults. This helps highlight anomalies truly caused by internal equipment defects, preventing them from being masked by external interference.

[0087] In practical applications, these specific thresholds and judgment logics typically need to be pre-configured and continuously optimized based on historical data, field experience, and expert knowledge. The system will intelligently adjust the initial value assessment score based on these quantitative results, thereby more accurately reflecting the true source of the abnormal patterns.

[0088] The adjusted preliminary value assessment score is used as the final output value assessment score to ensure that subsequent anomaly handling processes (such as microslice storage) can make decisions based on more accurate and reliable assessment results.

[0089] This solution introduces an external vibration detection step to determine the source of interference in identified associated offsets and dynamically adjusts the value assessment score. This effectively distinguishes between false associated offsets caused by external vibrations and those caused by real defects, improving the accuracy of value assessment and adapting to the limited resources of edge computing units. Acquiring structural vibration data of the protective box directly provides the actual vibration signal generated by external vibration acting on the protective box, offering a direct and reliable basis for subsequent interference determination and avoiding inaccuracies caused by indirect inference of interference based solely on multi-channel sensor data. Calculating the root mean square (RMS) value of the structural vibration data within a preset sliding time window reflects the overall intensity of the current vibration, determining whether the vibration intensity is sufficient to interfere with multi-channel sensor data acquisition. Furthermore, low-frequency feature identification is performed on the structural vibration data, extracting low-frequency energy peaks. This is because common external environmental interference vibrations are mostly low-frequency vibrations; this processing can distinguish external interference vibrations from the high-frequency noise of the sensor acquisition itself, further improving the accuracy of external vibration interference judgment. When the root mean square value exceeds the preset vibration threshold and a low-frequency energy peak is detected, an external vibration indication signal is generated, effectively marking external vibration interference and providing a clear basis for subsequent adjustment of the value assessment score. The system determines whether the associated offset pattern and the external vibration indication signal overlap in the time domain and whether their frequency characteristics are consistent. Combining the matching of the time and frequency domains ensures the synchronization of the interference and associated offset in time and further verifies their homogeneity through frequency characteristics, significantly improving the accuracy of interference detection. When the associated offset is determined to be a false anomaly caused by external vibration, the score is reduced to eliminate the influence of interference; when the associated offset is determined to be unrelated to external vibration, the score is increased to amplify the value of the real anomaly, allowing the value assessment score to accurately reflect the probability that the associated offset belongs to a real defect anomaly. The adjusted preliminary value assessment score is used as the final output value assessment score, providing accurate input for subsequent anomaly trigger determination. This ensures that only anomalies that are truly likely caused by a real defect will trigger the subsequent storage and reporting process, saving the limited computing and storage resources of the edge computing unit, reducing the probability of false alarms, and improving the accuracy of early latent defect identification.

[0090] As a specific implementation, an extremely low-power, low-cost micro-mechanical vibration sensor can be integrated into the edge computing unit of the high-voltage cable connector protection box, such as an STM32H7 series microcontroller, in addition to the existing temperature sensor, partial discharge detection unit, and humidity sensor. This vibration sensor can be a single-axis MEMS accelerometer, such as the ADXL345, which has extremely low power consumption and can accurately capture weak, low-frequency vibrations. The accelerometer connects to the STM32H7 microcontroller via an SPI or I2C interface, continuously acquiring weak vibration data from the protection box structure at a low sampling rate (e.g., 100 times per second). This vibration data, like the original sensor data, is precisely timestamped to ensure the time synchronization of all data streams, which is crucial for subsequent evaluation of their correlation. The software program in the edge computing unit processes the data stream from the micro-mechanical vibration sensor in real time. A lightweight low-frequency feature extraction process is performed on this vibration data. For example, by calculating the root mean square (RMS) value of the vibration signal within a sliding time window (e.g., 2 seconds) and performing a simple Fast Fourier Transform (FFT) on the data within that window, the presence of a dominant low-frequency component (e.g., a peak value in the range of 1 Hz to 10 Hz) can be identified. If the RMS value remains above a preset weak vibration threshold (e.g., 0.01 g), and an energy peak exceeding a preset energy threshold is detected in the low-frequency band, the edge computing unit generates an "external vibration indication signal" and records its intensity and dominant frequency. When the original "value density assessment module" identifies a potential weak pattern based on its heuristic rule set (e.g., "partial discharge and temperature micro-fluctuation co-occurrence rule," "sensor data non-random jitter rule," or "cross-sensor phase or frequency correlation rule"), it no longer immediately determines its "value density," but instead simultaneously refers to the aforementioned generated "external vibration indication signal." The specific judgment logic is as follows: If the detected weak pattern (e.g., non-random temperature fluctuations or low-frequency modulation of partial discharge background noise) highly overlaps with the "external vibration indication signal" in time (e.g., appearing within the same second) and exhibits consistency in frequency characteristics (e.g., the dominant frequency of temperature fluctuations is similar to the dominant frequency of the vibration indication signal), then the weak pattern is initially judged as "external interference tendency." In this case, the "confidence" of this pattern as a precursor to an internal fault will be significantly reduced. Conversely, if the detected weak pattern does not overlap with the "external vibration indication signal" in time, or has significant differences in frequency characteristics (e.g., a weak harmonic at 500kHz appears in the partial discharge signal, while the vibration indication signal only shows low-frequency vibration), then the weak pattern is initially judged as "internal fault tendency." In this case, the "confidence" of this pattern as a precursor to an internal fault will be significantly increased.The "Value Density Assessment Module" calculates a final "Fault Prediction Confidence" based on the preliminary judgment results and the original "Value Density" score. This confidence level is a value between 0 and 1, representing the probability that the current data segment truly originates from an internal fault. For example, if the "Partial Discharge and Temperature Micro-Fluctuation Co-occurrence Rule" is triggered, and a strong "External Vibration Indication Signal" is also present, the initial "Fault Prediction Confidence" for this mode may drop from 0.8 to 0.3. If the dynamic threshold is 0.5 at this time, a microslicing will not be triggered. Conversely, if there is no external vibration indication and the confidence level remains at 0.8, a microslicing will be triggered.

[0091] It should be noted that "weak patterns" refer to extremely weak and constantly changing correlation offset patterns between sensor data from different channels. These patterns are early warning signs of new, slowly developing cable joint defects. They are characterized not by drastic fluctuations in data from a single sensor, but by subtle shifts in the relationships between data from multiple channels. These changes are often difficult to detect using traditional fault analysis methods based on thresholds or predefined strong correlation patterns.

[0092] The specific method for identifying these weak patterns is based on preset correlation evaluation rules, using a "value density evaluation module" within the edge computing unit to perform real-time analysis of multi-channel sensor data. These correlation evaluation rules are specifically designed to capture weak, nonlinear, and constantly changing cross-sensor information correlation patterns, mainly including the following three types: 1. Co-occurrence rules of partial discharge signals and temperature fluctuation signals: This rule identifies weak patterns by continuously monitoring the amplitude of the partial discharge signal and the fluctuation range of ambient temperature or connector surface temperature. Specifically, the identification conditions are: if the amplitude of the partial discharge signal remains at a very low level, for example, above the background noise but far below the traditional alarm threshold; and simultaneously, within a very short time window (e.g., a continuous 10-second data segment), the fluctuation range of the ambient temperature or connector surface temperature (which can be measured by calculating the standard deviation of the temperature within this time window) also exhibits small but continuous, non-random changes. When both conditions are met simultaneously, the system considers this co-occurrence rule triggered, identifying a weak pattern. This pattern can capture the subtle connections between different physical quantities that arise in the very early stages of novel faults, when individual data changes are not obvious.

[0093] 2. Non-random jitter rules for single-channel sensor data: This rule focuses on minute but frequent, non-random jitter or jumps in one or a group of sensor data (e.g., vibration or humidity sensor data) within a very short period (e.g., consecutive 5-second data segments). The identification method involves calculating the number of sign changes in the first-order difference (i.e., the difference between adjacent data points) of the data sequence. Specifically, if the number of sign changes in the first-order difference exceeds a preset jitter threshold within a fixed-length sliding window, and the absolute amplitude of the data itself does not change significantly, then non-random jitter is considered to exist. This jitter is not simply measurement noise, but may indicate subtle changes in the physical state inside the cable joint, such as the propagation of a tiny crack or slight material displacement.

[0094] 3. Phase or frequency correlation rules between sensor data from each channel: This rule aims to detect subtle but persistent phase shifts or synchronized changes in specific frequency components between different sensor signals. The identification method involves performing a lightweight Fast Fourier Transform (FFT) on the partial discharge signal to analyze its frequency components, searching for specific frequencies with low amplitude but persistent occurrences. Simultaneously, if phase information of the AC voltage is available, the instantaneous correlation between the specific frequency components of the partial discharge signal and the AC voltage phase is calculated. If this correlation (e.g., by calculating the Pearson correlation coefficient) remains consistently above a very low threshold within a short time window, even if their absolute values ​​are low, it indicates a potential anomalous correlation pattern. This method can identify hidden connections between different physical phenomena in the early stages of novel faults.

[0095] By comprehensively applying the above three correlation evaluation rules, the edge computing unit can analyze multi-channel sensor data in real time and identify weak correlation offset patterns that are difficult to detect by traditional methods, thus providing a foundation for subsequent value assessment score calculation and anomaly handling.

[0096] Through the above technical solution, this application can effectively distinguish between false anomaly modes caused by external environmental vibrations and associated offset modes caused by real defects inside high-voltage cable joints, thereby significantly reducing the false alarm rate of edge computing units. This not only avoids wasting limited computing and storage resources due to false alarms, but also ensures that only data with genuine potential fault value is further processed and reported, greatly improving the accuracy and reliability of early latent defect identification.

[0097] In some embodiments, step S24, which involves calculating a preliminary value assessment score based on the identification results, includes the following specific steps: S241. For each associated offset pattern in the identification results, determine the trigger strength of the association evaluation rule; the trigger strength is determined based on the duration or trigger frequency of the associated offset pattern. S242. Perform a weighted summation or weighted product operation on all determined trigger strengths to obtain a preliminary value assessment score.

[0098] Specifically, for each associated offset pattern identified during real-time analysis of multi-channel sensor data, the system independently quantifies it to determine its trigger strength. An associated offset pattern refers to a specific manifestation of a deviation from the normal correlation between multi-channel sensor data, such as the co-occurrence of partial discharge signals and temperature fluctuation signals, non-random jitter in single-channel sensor data, or phase or frequency correlation between different channels. This method of determining trigger strength separately ensures that each detected anomalous pattern can be quantified independently and meticulously, avoiding confusion between different anomalous patterns, thus guaranteeing that all anomalous information brought about by associated offsets can be accurately captured and quantified. The quantification of trigger strength is based on the duration or trigger frequency of the associated offset pattern. Duration refers to the length of time from the onset to the end of an associated offset pattern; a longer duration usually indicates a more severe anomaly. Trigger frequency refers to the number of times an associated offset pattern occurs within a certain time window; a higher trigger frequency may also indicate a more significant anomaly. This quantization method based on duration or trigger frequency is highly consistent with the actual anomalous characteristics of novel, slowly developing defects. Early manifestations of these defects are often weak but persistent, or intermittently occurring correlation shifts. Therefore, by measuring their duration or trigger frequency, the development and severity of these early anomalies can be accurately reflected, thus more effectively capturing fault precursors that are difficult to detect using traditional methods. After determining the trigger strength of all correlation shift patterns, the system performs a weighted summation or weighted product operation on these strength values ​​to obtain a preliminary value assessment score. This weighted operation mechanism allows different weights to be assigned to different correlation assessment rules or their triggered correlation shift patterns to reflect their relative importance in predicting faults. For example, some correlation patterns may be proven by experience or models to be more reliable fault indicators and therefore can be given higher weights. In this way, the final preliminary value assessment score can comprehensively reflect the overall anomaly of all detected correlation shifts, providing a comprehensive and objective quantitative indicator of anomaly severity. This score calculation method not only provides reliable input for subsequent dynamic value determination thresholds but also significantly improves the ability of edge computing units to identify early weak correlation anomalies under limited resource conditions through its refined and comprehensive characteristics.

[0099] This application's solution, by clearly defining the specific process of score calculation and combining the actual characteristics of the associated offset patterns to determine the intensity and then integrating them to obtain the score, can obtain a preliminary assessment score that objectively reflects the overall degree of anomaly, providing a reliable basis for subsequent anomaly judgment and adapting to the need for early weak associated anomalies identified at the edge. Determining the trigger intensity for each associated offset pattern in the identification results can distinguish the anomalies of different associated offset patterns, avoiding confusion between different anomalies, ensuring that the characteristics of each anomaly pattern are accurately quantified, and not missing any abnormal information brought by associated offsets. The trigger intensity is determined based on the duration or trigger frequency of the associated offset pattern, which aligns with the actual anomaly characteristics of novel, slowly developing defects. These defects exhibit weak, continuous, or intermittent associated offsets; using duration or trigger frequency to measure the trigger intensity can accurately reflect the actual development degree of these defects, better matching the characteristics of early anomalies and accurately quantifying the severity of the anomaly. The initial score is obtained by weighted summation or weighted product operation of all determined trigger strengths. The weighting method can assign different weights to the importance of different association evaluation rules, so that the final score can comprehensively reflect the overall anomaly of all association offsets. At the same time, two operation methods (weighted summation operation and weighted product operation) are provided, which can be selected according to the actual computing resources and evaluation needs, adapting to the limited computing resources of edge computing units, without bringing too much additional computing burden.

[0100] As a specific implementation, suppose that when the edge computing unit monitors the high-voltage cable joint, it identifies two correlation offset modes by analyzing multi-channel sensor data in real time: the first is "co-occurrence of partial discharge signals and temperature fluctuation signals," characterized by the high temporal synchronization between weak pulses of partial discharge signals and small periodic fluctuations in temperature; the second is "non-random jitter of humidity sensor data," characterized by frequent but small jumps in humidity data within a short period of time. For the first correlation offset mode, "co-occurrence of partial discharge signals and temperature fluctuation signals," the system determines its trigger strength. For example, if this co-occurrence mode lasts for 15 seconds, its duration is 15 seconds. If the system sets the strength to increase by 1 unit for every second, then the trigger strength of this mode is 15. For the second correlation offset mode, "non-random jitter of humidity sensor data," the system determines its trigger strength. For example, within a sliding window of the past 60 seconds, this jitter mode was triggered 5 times. If the system sets the strength to increase by 2 units for each trigger, then the trigger strength of this mode is 10. After determining the trigger strength of these two modes, the system performs a weighted calculation to obtain a preliminary value assessment score. Assuming that, based on experience or preset configuration, the co-occurrence of "partial discharge signals and temperature fluctuation signals" is considered more indicative of a fault, its weighting coefficient is set to 0.7; while the weighting coefficient for "non-random jitter in humidity sensor data" is set to 0.3. If a weighted summation calculation is used, the preliminary value assessment score = (15 * 0.7) + (10 * 0.3) = 10.5 + 3 = 13.5. Through this method, the system can quantify the anomaly degree of different types of weakly correlated offset patterns and synthesize them into a preliminary value assessment score, providing a reliable basis for subsequent anomaly determination.

[0101] Through the above technical solution, this application can determine the trigger strength of each identified associated offset pattern and quantify this strength based on the duration or trigger frequency of the associated offset pattern. This meticulous quantification method enables the system to accurately capture the early, weak, and continuous or intermittent abnormal features of novel, slowly developing defects, avoiding the confusion of different abnormal patterns and ensuring that all potential fault information can be effectively identified and quantified. Furthermore, by performing weighted summation or weighted product operations on these determined trigger strengths, this application can comprehensively reflect the overall abnormal situation of all associated offsets, obtaining an objective and representative preliminary value assessment score. This score calculation method not only provides a more reliable and refined input for the subsequent dynamic value judgment threshold, but also significantly improves the ability to identify early weak associated anomalies under the condition of limited edge computing unit resources, thereby effectively solving the problem that traditional methods are difficult to accurately quantify early weak faults and improving the timeliness and accuracy of high-voltage cable joint fault early warning.

[0102] In some embodiments, step S3, which involves dynamically adjusting the preset value judgment threshold based on the real-time resource occupancy status, includes the following specific steps: S3A1. Determine whether the CPU utilization rate or memory utilization rate exceeds the preset first threshold; S3A2. If the CPU utilization or memory utilization exceeds the first threshold, the value judgment threshold will be increased by a preset first step length. S3B1. Determine whether the CPU utilization rate or memory utilization rate is lower than the preset second threshold; S3B2. If the CPU utilization rate or memory utilization rate is lower than the second threshold, the value judgment threshold is reduced by a preset second step.

[0103] CPU utilization is a metric measuring the workload of the edge computing unit's central processing unit. It can be obtained through APIs provided by the operating system or embedded real-time operating system (RTOS), for example, by reading the percentage of CPU idle task execution time. Memory utilization reflects the degree to which the edge computing unit's memory resources are used. It can be obtained by querying the memory allocation of the memory management unit (MMU) or RTOS, for example, by calculating the ratio of allocated memory to total memory. The preset first threshold is a critical value used to determine whether the edge computing unit's resources are under strain. Its setting can be based on the edge computing unit's hardware performance, expected load, and system stability requirements. For example, it can be set at a CPU utilization of 80% or a memory utilization of 85%. The preset second threshold is a critical value used to determine whether the edge computing unit's resources are in a surplus state. Its setting can be based on the system's normal operating state under low load. For example, it can be set at a CPU utilization of less than 40% or a memory utilization of less than 30%. The value judgment threshold is a key parameter used to filter abnormal patterns. Its initial value can be set based on historical data or expert experience. The first step length and the second step length are used to adjust the increment and decrement of the value judgment threshold. Their size can be determined according to the system's response speed to resource changes and the fineness of threshold adjustment. For example, it can be set as a percentage of the threshold base value or a fixed value.

[0104] This application's solution optimizes the efficiency of anomaly capture within limited edge resources by dynamically adjusting the screening criteria for anomaly patterns based on real-time monitoring of edge computing unit resource usage. Specifically, the edge computing unit continuously monitors its CPU and memory utilization rates. When the system detects that either CPU or memory utilization exceeds a preset first threshold, it indicates that the edge computing unit's resources are becoming strained, and the system proactively raises the value judgment threshold. This raising ensures that only anomaly patterns with higher value assessment scores pass the screening, reducing the load on subsequent micro-slice storage and feature extraction, effectively preventing the edge computing unit from being affected by resource overload. Conversely, when the system detects that either CPU or memory utilization is below a preset second threshold, it indicates that the edge computing unit's resources are relatively abundant, and the system proactively lowers the value judgment threshold. This lowering allows more anomaly patterns with lower value assessment scores but still potentially containing fault information to pass the screening, thereby improving the system's sensitivity to early, subtle anomalies and fully utilizing abundant edge resources to capture more valuable fault clues. This dynamic adjustment mechanism, combined with the basic sensor data anomaly pattern processing method, enables the edge computing unit to intelligently balance the sensitivity of anomaly detection and the consumption of system resources based on its real-time resource status when executing the S3 step. This ensures that in resource-constrained environments, it can effectively cope with resource shortages while making full use of surplus resources, thereby improving the adaptability and robustness of monitoring the operating status of high-voltage cable joints.

[0105] The following is a concrete example. As a specific implementation, the edge computing unit can be a microcontroller based on the ARM Cortex-M7 core, such as the STM32H7 series. This microcontroller obtains CPU utilization in real time through its built-in Performance Monitoring Unit (PMU) or APIs provided by the operating system, and obtains memory utilization by querying the Memory Management Unit (MPU) or the memory allocator of the RTOS. Assume a preset first threshold is set to CPU utilization exceeding 80% or memory utilization exceeding 85%, and a preset second threshold is set to CPU utilization below 40% or memory utilization below 30%. The initial value judgment threshold is 0.5. When the CPU utilization of the edge computing unit consistently reaches 82%, the system determines that it exceeds the first threshold of 80%. At this point, the value judgment threshold is increased by a preset first step, for example, by 0.1, making the new value judgment threshold 0.6. This means that only abnormal patterns with a value assessment score greater than or equal to 0.6 will trigger the microslice storage mechanism. In another scenario, if the CPU utilization of an edge computing unit drops to 35% while the memory utilization remains at 25%, the system determines that this is below the second threshold of 40% (or 30%). In this case, the value judgment threshold will be lowered by a preset second step, for example, by 0.05, resulting in a new value judgment threshold of 0.55 (if previously 0.6) or 0.45 (if previously 0.5). In this way, the edge computing unit can flexibly adjust its screening criteria for abnormal patterns based on its real-time resource status.

[0106] Through the above technical solution, this application can effectively solve the problems caused by resource limitations and fixed thresholds when edge computing units process abnormal sensor data patterns. When the CPU or memory utilization rate of the edge computing unit exceeds a preset first threshold, increasing the value judgment threshold can effectively reduce the triggering frequency of abnormal patterns, reduce the computation and storage overhead of micro-slice storage and subsequent feature extraction, thereby avoiding system instability or crashes due to resource overload of the edge computing unit, while not hindering the normal reporting of serious faults, ensuring the continuous operation of core monitoring functions. Conversely, when the CPU or memory utilization rate of the edge computing unit is lower than a preset second threshold, decreasing the value judgment threshold can improve the system's sensitivity to potential abnormal patterns, enabling even weak abnormal information with low value assessment scores to be captured and processed, making full use of the surplus resources of the edge computing unit, and avoiding missing early, hidden fault signs. This dynamic adjustment mechanism enables edge computing units to intelligently balance resource utilization and anomaly detection sensitivity, significantly improving the adaptability and early warning capabilities of high-voltage cable joint monitoring systems in complex and ever-changing operating environments. In particular, it can provide more timely and reliable early warnings for early-stage novel faults that develop slowly and have indistinct characteristics.

[0107] In some embodiments, the specific steps in step S4 include: S41. Extract multi-channel sensor data before and after the current trigger time for a preset duration to form a local data slice to be cached; S42. Obtain the time range of the local data slice to be cached; S43. Determine whether the time range of the local data slice to be cached overlaps with the time range of the local data slices already existing in the current cache; S44. If there is overlap, adjust the time range of the existing local data slice to include the time range of the local data slice to be cached, and update the data content of the existing local data slice. S45. If there is no overlap, the local data slice to be cached will be temporarily cached as a new local data slice. When the cache space is insufficient, the local data slices will be prioritized according to their value assessment score or cache duration, and the local data slice with the lowest priority will be released.

[0108] The purpose of extracting multi-channel sensor data for a preset duration before and after the current trigger moment is to accurately capture the data context before and after the occurrence of an abnormal event, providing a complete time window for subsequent in-depth analysis. This can be achieved through a circular buffer, copying data from before and after the current moment when the trigger event occurs; or by recording the trigger moment and then backtracking from the original data stream to extract data for the corresponding time period. The preset duration can be configured according to the actual application scenario and the evolution speed of the abnormal pattern; for example, it can be set to 5 seconds, 10 seconds, or longer to ensure coverage of the complete development process of the abnormal pattern. Forming a local data slice to be cached refers to encapsulating the extracted multi-channel sensor data into an independent data unit for subsequent storage, management, and processing. This data slice can be a data structure containing metadata such as timestamps, sensor type, and data values, or it can be a contiguous block of memory. The aim is to operate on data related to a specific abnormal event as a whole, improving the efficiency and accuracy of data management.

[0109] Obtaining the time range of the local data slice to be cached is to clarify the time interval covered by the data slice, typically including the start and end timestamps of the slice. This can be obtained by reading the metadata in the data slice or by calculating based on the slice length and sampling rate. Obtaining the time range is the basis for subsequent determination of whether there is time overlap between data slices, ensuring the accuracy of time overlap determination.

[0110] Determining whether the time range of the newly generated local data slice to be cached overlaps with the time range of existing local data slices in the current cache aims to identify whether there is any temporal overlap between the newly generated local data slice and existing data slices in the current temporary cache. Overlap can be determined by comparing the start and end times of the two time ranges. For example, if the start time of the slice to be cached is earlier than the end time of an existing slice, and the end time of the slice to be cached is later than the start time of an existing slice, then overlap is considered to exist. This determination is a crucial step in avoiding data redundancy and optimizing cache space management.

[0111] If overlap exists, the time range of existing local data slices is adjusted to include the time range of the local data slice to be cached. This means that when time overlap is detected, the time boundaries of existing data slices are extended to cover the time range of the new slice. For example, the start time of an existing slice is updated to the earlier of the two, and the end time is updated to the later of the two. This adjustment ensures that the merged data slice can completely contain data on all relevant anomalies, avoiding information loss. The data content of existing local data slices is also updated. This means that after the time range adjustment, the data content of the data slice to be cached is merged into the existing data slice. This can be achieved by appending new data to the existing slice or by overwriting overlapping data in the existing slice with new data. Updating data content aims to ensure that existing slices contain the latest and most complete data, while avoiding the storage of duplicate data, thereby optimizing cache space utilization.

[0112] If there is no overlap, the local data slice to be cached is temporarily cached as a new local data slice. This means that when the newly generated local data slice has no time overlap with any slice in the existing cache, it is added to the temporary cache as an independent entity. This ensures that all non-overlapping exception event data can be stored and processed independently. Temporary caching is usually implemented using mechanisms such as circular buffers or first-in-first-out (FIFO) queues. Furthermore, when cache space is insufficient, priority is ranked according to the value assessment score or cache duration of the local data slices. This aims to provide an intelligent data eviction strategy when the temporary cache space of the edge computing unit reaches its limit. Priority ranking can be based on the value assessment score obtained by the data slice in step S2; a higher score indicates more important data and should be retained first. Alternatively, it can be based on cache duration; the data slice with the longest cache time is considered relatively less valuable and should be evicted first. This mechanism ensures that limited cache resources can prioritize retaining the most valuable exception data. Releasing the lowest-priority local data slice means removing the lowest-priority data slice from the cache according to the priority ranking result to free up space for storing new data slices. The release operation typically involves marking the memory occupied by the data slice as available for subsequent writes. This strategy effectively manages the limited cache space of edge units, avoiding the problem of important data failing to be cached due to insufficient space.

[0113] This application's solution effectively addresses the problems of limited cache space in edge computing units, redundant data occupancy during temporary micro-slice storage, and inability to properly clean up insufficient cache by introducing an intelligent micro-slice storage mechanism during abnormal mode handling. This solution goes beyond simply creating a new data slice for each triggered event; instead, it shifts the focus to judging and managing the temporal overlap between data slices. When the "value density assessment module" of the edge computing unit identifies a potential abnormal mode and triggers the micro-slice storage mechanism, the system first extracts multi-channel sensor data for a preset duration before and after the current trigger time, forming a local data slice to be cached. Subsequently, the system obtains the time range of this local data slice to be cached and compares it with the time range of existing local data slices in the current cache. If overlap exists, it indicates that the newly detected abnormal event may be related to or a continuation of an already cached event. In this case, the system intelligently adjusts the time range of the existing local data slices to expand it to include the time range of the local data slice to be cached, and updates the data content of the existing local data slices. This merging operation effectively avoids redundant storage of data in overlapping time periods, integrating multiple closely related anomalies into a more comprehensive and context-rich data slice. This significantly reduces data redundancy in the cache and improves cache space utilization. If there is no overlap, the local data slice to be cached is treated as an independent anomaly and temporarily cached as a new local data slice. Furthermore, to address the limited cache space of the edge unit, when cache space is insufficient, the system prioritizes local data slices based on their value assessment score or cache duration, releasing the lowest-priority local data slices. This dynamic priority management mechanism ensures that, in resource-constrained environments, critical data with higher anomaly analysis value or more recent occurrences are prioritized, avoiding the risk of losing important early warning information due to indiscriminate cleanup. This integrated processing flow enables the edge computing unit to autonomously and effectively manage the temporary storage of local data slices within limited computing power and storage space. This ensures that all relevant anomaly patterns, especially early signs of new and slowly developing cable joint defects, are fully captured and effectively preserved, providing a solid data foundation for subsequent in-depth analysis and early warning.

[0114] The following is a concrete example. When the edge computing unit (e.g., an STM32H7 series microcontroller) in the high-voltage cable joint protection box identifies a potential abnormal pattern inside the high-voltage cable joint through its "value density assessment module" and calculates the corresponding value assessment score, reaching the trigger threshold, it will activate the microslice storage mechanism. Specifically, the microcontroller will first extract high-resolution, multi-channel raw sensor data within a very short time window before and after the current trigger moment (e.g., from 5 seconds before triggering to 5 seconds after triggering, a total of 10 seconds), forming a "local data slice to be cached." This slice contains data from temperature sensors, partial discharge sensors, and humidity sensors. Subsequently, the system will obtain the precise time range of this "local data slice to be cached," for example, the start time is T0, and the end time is T0+10 seconds. Next, the edge computing unit will check the existing local data slices in its internal high-speed random access memory (RAM). Assume that there is already a local data slice A in the cache, with a time range of T0-2 seconds to T0+8 seconds. The system determines whether the newly generated "local data slice to be cached" (T0 to T0+10 seconds) overlaps with slice A (T0-2 seconds to T0+8 seconds). In this example, they overlap. Therefore, the system adjusts the time range of the existing local data slice A, expanding it to T0-2 seconds to T0+10 seconds to include all time periods of the new slice. Simultaneously, the system merges the data from the "local data slice to be cached" into slice A. For example, it appends the data from T0+8 seconds to T0+10 seconds of the new slice to the end of slice A and updates the data from T0 to T0+8 seconds in slice A to ensure it contains the latest collected data. If there is no slice in the cache that overlaps with the "local data slice to be cached" in time—for example, the time range of the new slice is T1 to T1+10 seconds, and the time range of all slices in the cache is earlier than T1 or later than T1+10 seconds—then the "local data slice to be cached" will be added directly to the temporary cache as a completely new local data slice. When temporary cache space is limited (e.g., total RAM capacity is 2MB), the system initiates a cleanup mechanism when there is insufficient space to cache new local data slices. For example, the system sorts each local data slice according to its "value assessment score" calculated in step S2. Assume slice B has a value assessment score of 0.6, slice C has a value assessment score of 0.8, and slice D has a value assessment score of 0.4. The system will identify slice D as having the lowest value assessment score and release it from the cache to make room for new, high-value data slices. Alternatively, the system can sort according to cache duration, prioritizing the release of slices with the longest cache duration. For example, if slice E has been cached for 30 minutes, while other slices have only been cached for 10 minutes, slice E will be released.This mechanism ensures that edge units can prioritize retaining the critical data that best indicates new types of faults.

[0115] Through the above technical solutions, edge computing units can autonomously and effectively manage the temporary storage of local data slices in resource-constrained environments, avoiding data redundancy or the overwriting of critical information. This ensures that all relevant abnormal patterns, especially early signs of new and slowly developing cable joint defects, can be fully captured and effectively preserved, providing a solid data foundation for subsequent in-depth analysis and early warning.

[0116] In some embodiments, the specific steps in step S5 include: S51. Frequency domain characteristics are obtained by performing frequency domain energy distribution analysis on the sensor data of each channel in the local data slice; the frequency domain characteristics include the dominant frequency component, energy concentration and harmonic components; S52. By performing frequency domain phase or amplitude correlation analysis on the sensor data of different channels in a local data slice, the frequency domain correlation characteristics between channels are obtained; S53. Combine the frequency domain features and the inter-channel frequency domain correlation features to obtain a refined feature set.

[0117] Frequency domain energy distribution analysis is performed on the sensor data of each channel in a local data slice to obtain frequency domain features, including dominant frequency components, energy concentration, and harmonic components. Frequency domain energy distribution analysis aims to transform time-domain sensor data into the frequency domain to reveal the intensity distribution of different frequency components in the data. This transformation can identify periodic or quasi-periodic patterns that are difficult to detect in the time domain; these patterns are often related to subtle changes in the device's operating state. Implementation methods may include, but are not limited to, using Fast Fourier Transform (FFT) to convert time-domain signals into frequency-domain signals, or using wavelet transform for time-frequency analysis. Frequency domain features are information extracted from the original sensor data through frequency domain energy distribution analysis that characterizes the data's properties in the frequency dimension. These features are highly condensed and can effectively capture the essential attributes of the signal. The dominant frequency component refers to the frequency point or frequency range where energy is most concentrated in the frequency domain energy distribution. It represents the most important periodic or vibrational mode in the signal, and identifying the dominant frequency component helps to quickly locate the main dynamic characteristics of the signal. Energy concentration describes how tightly signal energy is distributed in the frequency domain. For example, a signal's energy may be concentrated at a few frequency points or dispersed over a wide frequency range. Energy concentration can quantify this distribution characteristic, reflecting the signal's "purity" or "complexity." Harmonic components refer to components in a signal whose frequencies are integer multiples of the fundamental frequency, excluding the fundamental frequency. In power systems, harmonic components are often associated with abnormal phenomena such as nonlinear loads, insulation aging, or partial discharge. Analyzing harmonic components helps identify specific types of fault modes.

[0118] Frequency-domain phase or amplitude correlation analysis is performed on sensor data from different channels within a local data slice to obtain inter-channel frequency-domain correlation characteristics. Frequency-domain phase or amplitude correlation analysis aims to assess the interrelationships between different sensor channels in the frequency domain, including their phase difference or amplitude ratio at specific frequencies. This correlation can reveal whether there are synchronous changes or causal relationships between different physical quantities, which is crucial for identifying weak cross-channel correlation shifts in novel faults. Implementation methods may include, but are not limited to: calculating the cross-spectral density function of signals from different channels in the frequency domain, or analyzing the time delay between signals by calculating the cross-correlation function, thereby deriving the phase relationship. Inter-channel frequency-domain correlation characteristics are information characterizing the interdependence between different sensor channels in the frequency domain, obtained through frequency-domain phase or amplitude correlation analysis. These characteristics can capture anomalous patterns of multi-channel coordinated changes that cannot be detected by single-channel analysis.

[0119] The refined feature set is obtained by combining the frequency domain features and the inter-channel frequency domain correlation features. Combination refers to integrating various frequency domain features extracted from single-channel and multi-channel analysis to form a unified and more comprehensive feature vector or set. This combination can be a simple concatenation or a more complex fusion method, aiming to maximize the retention of key information about abnormal patterns. The refined feature set is a highly condensed and representative feature set obtained after multi-dimensional feature extraction. It contains the key information needed to identify abnormal patterns in high-voltage cable joints, while significantly reducing the amount of data to fit the limited storage space of edge computing units.

[0120] This application addresses the characteristic of early, weak faults in high-voltage cable joints manifesting as cross-channel data correlation offsets. It designs a hierarchical feature extraction method to efficiently capture and retain key anomaly information within the limited storage space of edge computing units. When the micro-slice storage mechanism is triggered and local data slices are temporarily cached, the method first performs frequency domain energy distribution analysis on the sensor data of each channel within the local data slice. This analysis aims to convert the time-domain signal into the frequency domain, thereby revealing potential periodic or quasi-periodic patterns in the data of each sensor channel, which may be masked by noise in the time domain. By extracting frequency domain features such as dominant frequency components, energy concentration, and harmonic components, early indicators of anomalies can be extracted from the data of a single channel, such as weak specific frequency components in partial discharge signals or abnormal energy distribution in temperature fluctuations. These single-channel frequency domain features effectively compress the data volume while retaining the anomaly information of the signal itself. Based on this, to capture the core characteristic of novel faults—the correlation offset between sensor data from different channels—this application further performs frequency domain phase or amplitude correlation analysis on the sensor data of different channels within the local data slice. This analysis focuses on evaluating the frequency domain relationships between different sensor signals, such as the phase difference or amplitude ratio between partial discharge signals and temperature fluctuation signals at specific frequencies. Through this cross-channel correlation analysis, weak anomaly patterns involving the coordinated changes of multiple physical quantities can be identified, which are undetectable by single-channel analysis. For example, there may be hidden phase synchronization or amplitude coupling between partial discharge and temperature fluctuations. Finally, the frequency domain features obtained from single-channel analysis are combined with the inter-channel frequency domain correlation features obtained from cross-channel analysis to form a refined feature set. This combination ensures that key information about anomaly patterns is comprehensively integrated, including both the anomalous behavior of individual physical quantities and the interactions between different physical quantities. This refined feature set is stored in a highly compressed form, significantly reducing storage space requirements and effectively addressing the challenge of limited storage space in edge computing units. Through this hierarchical and targeted feature extraction, edge computing units can retain the key information needed to identify new early-stage faults without storing large amounts of raw data, providing effective and concise data for subsequent in-depth analysis at remote monitoring centers.

[0121] For example, when the microslice storage mechanism of the edge computing unit is triggered and successfully caches a local data slice containing data from temperature sensors, partial discharge sensors, and humidity sensors, multidimensional feature extraction can be performed as follows: First, for each channel sensor data in the local data slice, such as a partial discharge signal, a refined spectral analysis method, such as the Welch method, can be used to calculate its power spectral density. This method allows the acquisition of the frequency domain energy distribution of the partial discharge signal, identifying the dominant frequency component, such as a persistent low-amplitude frequency; calculating the energy concentration to assess the energy distribution in the frequency domain; and detecting harmonic components, such as the presence of specific high-frequency components related to the power frequency. Similar frequency domain energy distribution analysis can be performed on temperature and humidity sensor data to obtain their respective dominant frequency components, energy concentration, and harmonic components. Second, to capture the correlation between different channel sensor data, frequency domain phase or amplitude correlation analysis can be performed on the partial discharge signal and temperature fluctuation signal in the local data slice. For example, the cross-correlation function of two signals can be calculated, and the delay corresponding to the peak of the cross-correlation function can precisely reveal the potential, weak correlation between them, thus obtaining the frequency domain correlation characteristics between channels. Similarly, the phase or amplitude correlation in the frequency domain between partial discharge signals and humidity signals, or between temperature signals and humidity signals, can be analyzed. Specifically, the cross-correlation function is a mathematical tool for measuring the similarity of two signals at different time delays. For any two channels of sensing data in a local data slice, such as a partial discharge signal and a temperature fluctuation signal, a function curve about time delay can be obtained by calculating their cross-correlation function. The delay corresponding to the peak of the cross-correlation function, that is, the time delay corresponding to the maximum value of the function calculation result, indicates how much time one signal needs to be delayed or advanced to achieve optimal matching with another signal. This "peak-correlation delay" itself is an important inter-channel correlation characteristic. For example, if the peak of the partial discharge signal always lags behind the temperature fluctuation signal by a fixed time, this may suggest a causal relationship or synchronicity of some physical process. This delay value, or its trend of change over different time periods, can be directly used as part of the frequency domain correlation characteristics between channels, because it reflects the relative positional relationship of the signals on the time axis, and this time relationship is expressed as a specific phase difference in the frequency domain.

[0122] Secondly, frequency domain phase or amplitude correlation analysis is performed after the signal has been transformed from the time domain to the frequency domain using a Fourier transform (e.g., Fast Fourier Transform, FFT). In the frequency domain, each signal has an amplitude (or energy) and a phase at each frequency point.

[0123] 1. Frequency Domain Phase Correlation Analysis: For two channels of sensor data in a local data slice, such as a partial discharge signal and a temperature fluctuation signal, their phase spectra at different frequencies can be obtained by performing Fourier transforms on them respectively. Frequency domain phase correlation analysis compares the phase difference between these two signals at the same frequency point. If the two signals maintain a relatively stable phase difference within a certain frequency range, it indicates a strong correlation between them at that frequency. For example, a specific harmonic component of the partial discharge signal may have a fixed phase difference with a certain frequency component of the temperature fluctuation signal, which can be used as a frequency domain correlation feature between channels. The value of this phase difference, its distribution pattern at different frequencies, and its stability are all important characteristics.

[0124] 2. Frequency Domain Amplitude Correlation Analysis: Similarly, in the frequency domain, the relationship between the amplitudes (or energies) of two signals at the same frequency point can be compared. For example, the amplitude ratio at a specific frequency can be calculated, or the amplitude correlation coefficient between them over a certain frequency range can be calculated. If the amplitude changes of two signals at a certain frequency or within a frequency range exhibit synchronicity or a proportional relationship, it indicates that there is an amplitude correlation between them at that frequency. For example, when the energy of a local discharge signal increases at a specific frequency, the energy of a temperature fluctuation signal also increases synchronously at the same frequency; this can be considered a characteristic of frequency domain correlation between channels. The strength, frequency range, and variation pattern of this amplitude correlation are all important characteristics.

[0125] The entire execution process of step S52 is as follows: 1. Data Preparation: First, local data slices containing multi-channel sensor data are obtained from the microslice storage mechanism. This data has already undergone time synchronization.

[0126] 2. Frequency Domain Transformation: Perform a Fourier transform (e.g., FFT) on each sensor data point (e.g., temperature, partial discharge, humidity) in the local data slice to convert it from a time-domain signal to a frequency-domain signal. Each frequency-domain signal will contain a series of frequency components, each with a corresponding amplitude and phase.

[0127] 3. Channel Pair Selection: Based on the preset correlation evaluation rules, select the channel pairs that need to be correlated, such as partial discharge and temperature, partial discharge and humidity, temperature and humidity, etc.

[0128] 4. Cross-correlation analysis (time-domain aid): For each selected channel pair, its cross-correlation function can be calculated in the time domain, and the delay corresponding to the peak value of the cross-correlation function can be extracted. This time delay serves as a feature of inter-channel correlation.

[0129] 5. Frequency Domain Phase Correlation Analysis: For each selected channel pair, compare their phase angles at various frequency points in the frequency domain. Calculate and record the phase differences within a specific frequency or frequency band; these phase differences and their variation patterns with frequency constitute the inter-channel frequency domain correlation characteristics.

[0130] 6. Frequency Domain Amplitude Correlation Analysis: For each selected channel pair, compare their amplitudes at various frequency points in the frequency domain. Amplitude ratios, amplitude differences, or amplitude correlation coefficients within a specific frequency range can be calculated. These values ​​and their variation patterns constitute the frequency domain correlation characteristics between channels.

[0131] 7. Feature set construction: Combine all the features (including delay values, phase differences, amplitude ratios or correlation coefficients, etc.) obtained from cross-correlation analysis, frequency domain phase correlation analysis and frequency domain amplitude correlation analysis to form the final inter-channel frequency domain correlation feature set.

[0132] Through the detailed analysis process described above, the edge computing unit can extract multi-dimensional and refined inter-channel frequency domain correlation features from local data slices. These features can accurately capture the weak, nonlinear and constantly evolving hidden connections between different physical quantities caused by novel faults, thus providing key information for subsequent anomaly pattern recognition and the construction of refined feature sets.

[0133] Finally, the frequency domain features of each channel obtained from the above steps (including the main frequency component, energy concentration, and harmonic components) are combined with the inter-channel frequency domain correlation features (such as the phase difference between partial discharge and temperature, and the amplitude ratio between partial discharge and humidity). This combination can be achieved by concatenating all extracted feature values ​​into a feature vector, forming the final refined feature set. For example, this set can include three frequency domain features of the partial discharge signal, three frequency domain features of the temperature signal, three frequency domain features of the humidity signal, as well as the phase correlation features between partial discharge and temperature, and the amplitude correlation features between partial discharge and humidity. These refined feature sets can then be stored in a highly compressed form, such as using fixed-point representation or differential encoding, in the persistent storage area of ​​the edge cell, such as in an external SPI flash memory chip.

[0134] Through the above technical solution, this application effectively solves the technical problem of how to efficiently and accurately capture and retain key information on early anomalies of high-voltage cable joints within the limited storage space of edge computing units. This solution employs hierarchical feature extraction, first extracting frequency domain features from the sensor data of each channel in a local data slice, including the dominant frequency component, energy concentration, and harmonic components. This allows even weak, non-drastic anomaly patterns in single sensor data to be effectively identified and quantified. Based on this, further frequency domain phase or amplitude correlation analysis is performed on the sensor data of different channels to obtain inter-channel frequency domain correlation features. This is particularly relevant to the characteristic of novel early faults with cross-channel correlation offset as a core feature, accurately capturing hidden collaborative changes between multiple physical quantities that are difficult to detect with traditional single-channel analysis. Combining these single-channel frequency domain features and inter-channel frequency domain correlation features forms a refined feature set, which not only fully integrates the key information needed to identify novel early faults but also significantly compresses the data volume because only the refined features are stored, rather than the complete original data. This enables edge computing units to retain valuable fault precursor information even with extremely limited storage space, avoiding the risk of losing critical data due to insufficient storage capacity. Therefore, this solution significantly optimizes storage usage while ensuring feature validity, providing comprehensive and concise data for subsequent in-depth fault diagnosis and early warning at the remote monitoring center, thereby improving the early warning capability and overall reliability of the high-voltage cable joint monitoring system.

[0135] Reference Appendix Figure 2 This invention provides a sensor data anomaly mode processing system (this sensor data anomaly mode processing system adopts the sensor data anomaly mode processing method of the above embodiment, the specific process is referred to the corresponding steps above), used to process the sensing data collected by the sensors inside the protection box through the edge computing unit to realize the monitoring of the high-voltage cable joints inside the protection box; the sensor data anomaly mode processing system includes: The processing module 100 is used to acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint and to perform time synchronization processing on the sensor data of each channel. The analysis module 200 is used to identify the correlation offset patterns between different channels of sensor data by performing real-time analysis on multi-channel sensor data based on preset correlation evaluation rules, and to calculate the corresponding value evaluation score based on the identification results. The adjustment module 300 is used to obtain the real-time resource occupancy status of the edge computing unit and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status. The interception module 400 is used to trigger the micro-slice storage mechanism when the value assessment score is greater than or equal to the value judgment threshold. The micro-slice storage mechanism includes intercepting multi-channel sensor data for a preset duration before and after the triggering time as local data slices for temporary caching. The extraction module 500 is used to extract multidimensional features from local data slices to obtain a refined feature set representing abnormal patterns. The reporting module 600 is used to report anomaly description information containing a refined feature set and to release the cache space occupied by local data slices.

[0136] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0137] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing abnormal sensor data patterns, used to process sensor data collected by sensors inside a protection box through an edge computing unit to monitor the high-voltage cable joints inside the protection box; characterized in that, The method for handling abnormal sensor data patterns includes the following steps: S1. Acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint, and perform time synchronization processing on the sensor data of each channel; S2. Based on preset correlation evaluation rules, the correlation offset patterns between different channels of sensor data are identified by real-time analysis of the multi-channel sensor data, and the corresponding value evaluation scores are calculated based on the identification results. S3. Obtain the real-time resource occupancy status of the edge computing unit, and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status; S4. When the value assessment score is greater than or equal to the value judgment threshold, a micro-slice storage mechanism is triggered. The micro-slice storage mechanism includes extracting the multi-channel sensor data before and after the triggering time for a preset duration as a local data slice for temporary caching. S5. By performing multidimensional feature extraction on the local data slices, a refined feature set representing the abnormal pattern is obtained; S6. Report the anomaly description information containing the refined feature set, and release the cache space occupied by the local data slice.

2. The sensor data abnormality mode processing method according to claim 1, characterized in that, The multi-channel sensing data includes sensing data from a temperature sensor, a partial discharge sensor, and a humidity sensor, respectively.

3. The sensor data anomaly mode processing method according to claim 2, characterized in that, The correlation evaluation rules include the co-occurrence rules of partial discharge signals and temperature fluctuation signals, the non-random jitter rules of single-channel sensor data, and the phase or frequency correlation rules between sensor data of each channel.

4. The sensor data abnormality mode processing method according to claim 1, characterized in that, The specific steps in step S2 include: S21. Obtain the structural vibration data of the protective box; S22. Calculate the root mean square value of the structural vibration data within a preset sliding time window, and identify the low-frequency energy peak in the structural vibration data by performing low-frequency feature identification on the structural vibration data; S23. When the root mean square value exceeds the preset vibration threshold and the low-frequency energy peak is detected, an external vibration indication signal is generated; S24. Based on the aforementioned correlation evaluation rules, the correlation offset patterns between different channels of sensor data are identified by real-time analysis of the multi-channel sensor data, and a preliminary value evaluation score is calculated based on the identification results. S25. Determine whether the associated offset mode and the external vibration indication signal overlap in the time domain. If they overlap in the time domain and their frequency characteristics are consistent, then reduce the preliminary value assessment score; otherwise, increase the preliminary value assessment score. S26. The adjusted preliminary value assessment score shall be used as the final output value assessment score.

5. The sensor data anomaly mode processing method according to claim 4, characterized in that, In step S24, the specific steps for calculating the preliminary value assessment score based on the identification results include: S241. For each of the associated offset patterns in the identification results, determine the trigger strength of the association evaluation rule; the trigger strength is determined based on the duration or trigger frequency of the associated offset pattern; S242. Perform a weighted summation or weighted product operation on all determined trigger strengths to obtain the preliminary value assessment score.

6. The sensor data anomaly mode processing method according to claim 1, characterized in that, The real-time resource usage status includes the CPU usage rate and memory usage rate of the edge computing unit.

7. The sensor data anomaly mode processing method according to claim 6, characterized in that, In step S3, the specific steps for dynamically adjusting the preset value judgment threshold based on the real-time resource occupancy status include: S3A1. Determine whether the CPU utilization rate or the memory utilization rate exceeds a preset first threshold; S3A2. If the CPU utilization rate or the memory utilization rate exceeds the first threshold, then the value judgment threshold is increased by a preset first step length; S3B1. Determine whether the CPU utilization rate or the memory utilization rate is lower than a preset second threshold; S3B2. If the CPU utilization rate or the memory utilization rate is lower than the second threshold, then the value judgment threshold is reduced by a preset second step.

8. The sensor data anomaly mode processing method according to claim 1, characterized in that, The specific steps in step S4 include: S41. Extract the multi-channel sensor data before and after the current trigger time for a preset duration to form a local data slice to be cached; S42. Obtain the time range of the local data slice to be cached; S43. Determine whether the time range of the local data slice to be cached overlaps with the time range of the local data slices already existing in the current cache; S44. If there is overlap, adjust the time range of the existing local data slice to include the time range of the local data slice to be cached, and update the data content of the existing local data slice. S45. If there is no overlap, the local data slice to be cached is temporarily cached as a new local data slice. When the cache space is insufficient, the local data slices are prioritized according to their value assessment score or cache duration, and the local data slice with the lowest priority is released.

9. The sensor data anomaly mode processing method according to claim 1, characterized in that, The specific steps in step S5 include: S51. Frequency domain characteristics are obtained by performing frequency domain energy distribution analysis on the sensing data of each channel in the local data slice; S52. By performing frequency domain phase or amplitude correlation analysis on the sensing data of different channels in the local data slice, the frequency domain correlation characteristics between channels are obtained; S53. Combine the frequency domain features and the inter-channel frequency domain correlation features to obtain the refined feature set.

10. A sensor data anomaly mode processing system, used to process sensor data collected by sensors inside a protection box through an edge computing unit to monitor the high-voltage cable joints inside the protection box; characterized in that, The sensor data anomaly mode processing system includes: The processing module is used to acquire multi-channel sensor data characterizing the operating status of the high-voltage cable joint and to perform time synchronization processing on the sensor data of each channel. The analysis module is used to identify the correlation offset patterns between different channels of sensor data by performing real-time analysis on the multi-channel sensor data based on preset correlation evaluation rules, and to calculate the corresponding value evaluation score based on the identification results. The adjustment module is used to obtain the real-time resource occupancy status of the edge computing unit and dynamically adjust the preset value judgment threshold according to the real-time resource occupancy status. The interception module is used to trigger a micro-slice storage mechanism when the value assessment score is greater than or equal to the value judgment threshold. The micro-slice storage mechanism includes intercepting the multi-channel sensor data before and after the triggering time for a preset duration as local data slices for temporary caching. The extraction module is used to obtain a refined feature set representing abnormal patterns by performing multi-dimensional feature extraction on the local data slices; The reporting module is used to report anomaly description information containing the refined feature set and release the cache space occupied by the local data slice.

Images