Machine learning based power system inspection method

Abstract

The application relates to the technical field of power inspection management, and particularly discloses a power system inspection method based on machine learning, which comprises the following steps: an edge end reads self-checking data of a sensing device in a historical inspection period and determines availability, according to which, only qualified data is used to extract original flow and construct a window observation set of the same device and the same time period, the window observation set is sent to a lightweight model together with the availability state to generate an initial defect vector, then, only suspicious or uncertain samples are used to generate associated data packets and upload according to an uplink strategy, the cloud end performs regression fusion on the uplink samples to obtain a defect probability and performs gating, risk assessment is performed according to an event label to output a risk score, finally, the risk score is used to limit an adaptive inspection period and automatically configure a next round of inspection and a collection mode, high-risk encryption and low-risk frequency reduction are realized, invalid mileage and repeated dispatching are reduced, and a service level agreement is achieved and overall operation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of power system inspection and management technology, specifically a power system inspection method based on machine learning. Background Technology

[0002] Existing power system inspection methods typically begin with asset ledger modeling, establish fixed cycles and inspection routes based on operation and maintenance procedures, and employ manual foot patrols, vehicle patrols, drone / robot patrols, and fixed monitoring points to collect images, videos, and time-series data. After denoising, focusing, or occlusion quality checks and time alignment are completed at the edge, feature or evidence frames are transmitted back to the central platform. Anomalies are identified and alarms are classified according to thresholds, rules, or template libraries. Work orders are generated by combining ledger data and geographical topology. After on-site verification and handling, retesting and archiving are performed, forming a closed loop of report and ledger updates.

[0003] For example, Chinese invention patent application CN118840098A discloses a method for inspecting microservices in a power system. This method includes acquiring microservice data information of the target power system and extracting semantic features from abnormal logs, as well as filtering information based on comprehensive user tag associations to obtain abnormal log information for the microservices; constructing a prediction model and predicting inspection indicators for the power system microservices; analyzing the microservice results of the target power system; generating an inspection report for the target power system microservices, and completing the inspection of the target power system microservices. A system for implementing this method for inspecting power system microservices is also disclosed, which achieves inspection through processing, prediction, and analysis of microservice data information from the target power system.

[0004] For example, Chinese invention patent CN120471409B discloses a power inspection system and method, including: collecting equipment status data, constructing a third-order state transition tensor field to model the temporal evolution characteristics of the equipment's spatial state, and identifying short-term high-risk equipment based on a weighted anomaly score and dynamic threshold strategy; recording historical state peaks through a state-time matrix fusion exponential decay model, calculating the forgetting risk factor in conjunction with the current inactivity level, and proactively re-inspecting equipment that has previously experienced severe anomalies but has recently become inactive; constructing a dynamic evolution graph structure, optimizing the inspection path using a genetic perturbation mechanism, and parallel scheduling multiple teams through task cluster partitioning and queue matching; establishing an equipment anomaly linkage propagation model to predict cascading risks, and automatically matching the optimal handling solution in conjunction with a strategy template library.

[0005] Based on the above technical solutions, it was found that most existing power system inspection methods are based on fixed-cycle inspections, experience thresholds, or fixed threshold judgments and templated keyword screening. However, due to the complex and ever-changing actual application scenarios of power systems, such as weather and seasonal fluctuations, load and topology switching, sensor noise and link instability, rule-based and template-based methods are difficult to adapt to non-stationary and cross-scenario conditions. This leads to negative impacts such as increased false alarms / missed alarms, disordered alarm priorities, suboptimal inspection paths and resource scheduling, increased burden of manual review, and decreased timeliness and accuracy of handling. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a power system inspection method based on machine learning, which can effectively solve the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a power system inspection method based on machine learning, comprising: an edge processing terminal acquiring self-inspection data sets of power equipment sensing devices within a historical inspection cycle to determine the availability status of the power equipment sensing devices; based on the availability status determination result, extracting the raw flow data of the power equipment to construct a window observation set; inputting the window observation set and availability status into a machine learning model to output an initial defect vector of the window observation set, used to determine whether the window observation set executes an uplink strategy; if the uplink strategy is executed, the edge processing terminal generates associated data packets of the window observation set; the edge processing terminal uploads the associated data packets of the window observation set to a cloud processing terminal; the cloud processing terminal performs regression fusion to obtain the defect probability of the window observation set; based on the defect probability, performing gating decisions to generate event labels; performing risk assessment on the power equipment based on the gating decision event labels to obtain a risk score for the power equipment; limiting the adaptive inspection cycle based on the risk score of the power equipment and configuring the next round of inspection process.

[0008] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0009] (1) This invention provides a power system inspection method based on machine learning. The edge device first reads the self-test data of the sensor device within the historical inspection cycle and determines whether it is available or unavailable, which serves as the first quality gate to intercept degraded data and reduce subsequent false alarms. Based on this, only the qualified data is extracted into the original stream and a window observation set of the same device and the same time period is constructed to ensure that the input evidence is complete and the bandwidth overhead is controllable. The window observation set, together with the availability status, is sent into a lightweight model to generate an initial defect vector, realizing the rapid and uniform defect clue extraction at the edge. Then, according to the uplink strategy, only suspicious or uncertain samples are generated into associated data packets. The data is then sent up, prioritizing the use of the link and cloud computing power on the samples being processed. The cloud performs regression fusion on the uplink samples to obtain the defect probability and performs gating to stabilize the false alarm rate, directly converting high-confidence risks into event labels, and transferring boundary samples to manual review. Risk assessment is performed based on the event labels to generate risk scores, unifying the environment, load, and trends into comparable dimensions to facilitate priority ranking. Finally, the risk scores are used to limit the adaptive inspection cycle and automatically configure the next round of inspection and data collection modes, achieving high-risk encryption and low-risk frequency reduction, reducing invalid mileage and duplicate dispatches, and improving SLA achievement and overall operation and maintenance efficiency.

[0010] (2) The present invention first evaluates the self-test quality indicators of the sensing device on the edge processing side and determines whether it is usable or unusable. It can isolate low-quality observations at the data source and participate the quality vector in subsequent fusion and gating, so that the defect judgment is naturally immune to anomalies such as clock deviation, sampling jitter, signal-to-noise ratio drop, and frame loss. The cloud-based risk score output based on the gating event label quantifies factors such as environmental pressure, load stress and trend slope in a unified manner. Through adaptive inspection cycle, high-risk equipment is re-inspected with encryption and low-risk equipment is down-inspected, thereby reducing false alarms / missed alarms and invalid inspection mileage, improving alarm hit rate and handling timeliness, and ensuring compliance with minimum inspection frequency and SLA (service level agreement).

[0011] (3) In this invention, the self-inspection quality index is used for both availability determination and as a reliability adaptation weight in regression fusion, and is linked to gating tightening / loosening, uplink strategy, and evidence frame sampling ratio; the uncertainty obtained from the window observation set and the historical defect probability distribution simultaneously drive the adaptive adjustment of the active review ratio and the historical backtracking window length; the risk score not only limits the inspection correction cycle, but also infers the path / modal configuration. This one-time calculation and multiple reuse reduces redundant modeling and communication overhead, maintains strategy consistency and interpretability, and provides a unified index closed loop for subsequent rolling calibration.

[0012] (4) Compared with the traditional scheme of fixed period, experience / fixed threshold and templated keyword screening, the present invention realizes end-to-cloud collaboration of end-to-end self-testing, quality perception fusion, uncertainty gating, event classification and risk-driven scheduling: only related data packets are sent up in weak network conditions to adapt the bandwidth; for unseen scenarios, uncertainty guides manual review to avoid high confidence errors; for common environmental factors, the adjacent consistency and regression fusion suppress group false alarms; for sensor unavailability, isolation and redundancy replacement are provided and automatic order dispatch is performed. In the end, while maintaining the detection rate, the false alarm rate and operating costs are significantly reduced, the overall throughput and stability from discovery to handling are improved, and the anti-drift, scalability and auditability are stronger. Attached Figure Description

[0013] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0015] Figure 2 This is a schematic diagram of the power system inspection process.

[0016] Figure 3 The inspection management interface of the power inspection and control platform. Figure 1 .

[0017] Figure 4 The inspection management interface of the power inspection and control platform. Figure 2 . Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] This method is designed for power system inspection, focusing on multimodal inspection of transmission and distribution equipment (such as transmission line insulators, switchgear, main transformers / GIS, and distribution switchgear). After completing batch automatic inspections, the power inspection system assigns anomalies requiring verification to the smart wearable devices (such as AR / VR smart glasses) of the operators. On-site personnel wear the glasses and follow the prompts to reach their designated positions, achieving efficient collaboration from automatic detection to manual closed-loop verification.

[0024] The specific process of power system inspection method is based on Figure 2 As shown, Figure 2 This diagram illustrates the power system inspection process, starting from the edge: First, sensor self-checks determine availability. Then, a window observation set is constructed from qualified data, and a lightweight model generates initial defect vectors. Based on the uplink strategy, only suspicious samples are packaged and uploaded, while normal samples are allowed locally to save bandwidth and computing power. Upon reaching the cloud, quality-aware fusion is performed on the uplink samples to obtain defect probabilities, and gating is used to label events as normal, requiring review, or alarm. The review-required task is sent to manual review (wearable devices), and on-site supplementary sampling and judgment results are fed back to calibrate gating and strategies. The process then enters the planning / execution phase: a risk score is obtained based on risk assessment, the adaptive inspection cycle is automatically adjusted, and the next round of inspections (route, modality, sampling parameters) is configured. The plan is then sent back to the edge. Overall, the process integrates discovery, judgment, review, and scheduling: high-risk samples are prioritized, low-value data is processed locally, and the linkage between review and planning reduces false alarms and invalid mileage, improving alarm hit rate, handling timeliness, and SLA achievement rate.

[0025] Reference Figure 1As shown, the present invention provides a power system inspection method based on machine learning, including: an edge processing terminal acquiring a self-inspection data set of power equipment sensing devices within a historical inspection cycle, used to determine the availability status of the power equipment sensing devices; based on the availability status determination result, extracting the raw streaming data of the power equipment to construct a window observation set, thereby obtaining the window observation set of the power equipment.

[0026] Power equipment refers to various devices and auxiliary facilities used in power systems for power generation, transformation (step-up / step-down / rectifier-inverter), transmission, distribution, protection and control, energy storage and power consumption. It includes both primary equipment that directly carries electrical energy and secondary equipment such as measurement, protection, control and communication equipment.

[0027] In this embodiment, taking the common device of insulator strings in transmission lines as an example, during the inspection, the end side first performs a sensor self-test to obtain a quality vector (clock deviation, sampling jitter, SNR, frame loss, etc.), and then collects and aligns multi-modal data for the same string of insulators within a time window to form a "spatiotemporally aligned multi-modal observation unit". The cloud side assigns credibility weights to each mode based on the quality vector and merges them into a defect probability. Combined with uncertainty gating, uncertain samples are sent to manual review. Then, the consistency of adjacent towers / phase circuits and environmental pressure are superimposed to calculate the risk score, distinguishing between common causes of temperature rise / wind-induced fluctuations and individual faults. Finally, the inspection cycle is automatically adjusted according to the risk to complete the inspection closed loop.

[0028] The aforementioned windowed observation set (referred to as a spatiotemporally aligned multimodal observation unit) represents all observations from the same device within a short period of time: visible light / infrared / partial discharge / SCADA / meteorology, etc., aligned along the same time axis and packaged together with contexts such as load, weather, and topology into the smallest data unit required for a single decision. Since single-frame / single-point inspection data is easily misjudged due to interference from glare, jitter, and short-term load fluctuations, the window provides trends and persistence (slope, out-of-bounds duration), multimodal consistency, and comparison with adjacent devices, while carrying quality indicators and uncertainties. This facilitates quality-aware fusion and gating by the model, and provides sufficient evidence to support alarms, risk scoring, and adaptive inspection cycles, significantly reducing false alarms / missed alarms and ensuring traceability.

[0029] Specifically, the process for determining the availability status of power equipment sensors is as follows:

[0030] The self-test data set of the power equipment sensor includes the maximum clock deviation, sampling interval jitter, acquisition signal-to-noise ratio, and number of repeated frame acquisitions within the historical inspection cycle. This self-test data set can be extracted from the sensor's self-test records. The sampling interval jitter is obtained by processing the standard deviation of each sampling interval.

[0031] The self-test data set of the power equipment sensing device is compared with the preset self-test data definition threshold set in the power system inspection database to obtain the self-test data comparison result, which includes the first comparison result and the second comparison result.

[0032] The first comparison result of the self-test data indicates that the following conditions are met simultaneously: the maximum clock deviation of the power equipment sensor is less than or equal to the clock limit deviation; the sampling interval jitter of the power equipment sensor is less than or equal to the sampling interval limit jitter; the acquisition signal-to-noise ratio of the power equipment sensor is greater than or equal to the limit signal-to-noise ratio; and the number of repeated acquisition frames of the power equipment sensor is less than or equal to the number of repeated frame limits.

[0033] The second comparison result of the self-test data indicates that there is a maximum clock deviation of the power equipment sensor that is greater than the clock limit deviation, or the sampling interval jitter of the power equipment sensor is greater than the sampling interval limit jitter, or the acquisition signal-to-noise ratio of the power equipment sensor is less than the limit signal-to-noise ratio, or the number of repeated acquisition frames of the power equipment sensor is greater than the number of repeated frame limits.

[0034] The availability status of the power equipment sensing device is determined to be available only if the self-test data comparison result shows the first self-test data comparison result; otherwise, it is determined to be unavailable.

[0035] If the availability status of the power equipment sensor is determined to be unavailable, the edge processing terminal will report the unavailable status of the sensor to trigger active verification.

[0036] In this embodiment, the self-test data set of the power equipment sensing device is compared with the preset threshold set and its availability / unavailability is determined in real time. This can intercept deteriorated / abnormal observations at the data source, prevent low-quality data from entering alarms and model fusion, and significantly reduce false alarms / missed alarms. At the same time, it drives the quality perception weight and gating adaptation to ensure the stability and interpretability of inspection judgment, reduce invalid verification and bandwidth occupation, form a traceable quality closed loop, and support adaptive inspection cycle and accurate operation and maintenance dispatch.

[0037] Furthermore, the raw flow data of the power equipment is extracted and a window observation set is constructed. The specific analysis process for the window observation set of the power equipment is as follows:

[0038] When the availability status of the power equipment sensor is determined to be available, the self-test quality index of the power equipment sensor is obtained by jointly evaluating and fusing the self-test data set and the self-test data definition threshold set.

[0039] Based on the self-inspection quality indicators of power equipment sensing devices, a mapping relationship table between the preset self-inspection quality indicators and window limiting factors in the power system inspection database is extracted. The self-inspection quality indicators of power equipment sensing devices are then substituted into the mapping relationship table to obtain the window limiting factors.

[0040] It should be explained that the preset relationships stored in the aforementioned power system inspection database include, but are not limited to, preset, matching, and mapping relationships. Specifically, this embodiment uses a mapping relationship table between self-inspection quality indicators and window-limiting factors as an example to illustrate how these preset relationships are obtained. First, the self-inspection quality indicators and window-limiting factors are aligned on a large number of historical samples. An objective function J is set (combining false alarm rate, missed detection rate, bandwidth / computing power overhead, and SLA overdue penalty). Under monotonic priors and boundary constraints, binning statistics and quantile regression / equimetric fitting are performed. An equidistant table lookup combined with anisotropic monotonic splines / equidistant isotropic regression or a machine learning model with monotonic constraints is used to obtain the monotonic mapping of each indicator to a factor and the joint mapping of multiple indicators. Then, the continuous mapping is discretized into a versioned lookup matrix (reference table) and verified and stress-tested using an independent backtesting set. Rolling calibration is performed according to the indicator drift cycle. This acquisition method is also applicable to other "X and Y" preset mappings in the database (such as quality vectors and fusion weights), requiring only the input and output to be replaced and the same table building and calibration process reused.

[0041] The window constraint factor is associated with and merged with the default window. Specifically, the window constraint factor is multiplied by the default window to obtain the adapted window.

[0042] When the sensing device determines that the device is available, the default window is converged into an adaptive window. This dynamically divides the historical inspection cycle into more suitable time slices based on the quality and fluctuation characteristics of the device and the scene. This ensures that each time slice covers sufficient evidence (trends, out-of-bounds duration, modal consistency) while avoiding noise contamination due to excessively long slices or insufficient evidence due to excessively short slices. By dividing and extracting the raw streaming data under each adaptive window in this way, a window observation set is formed, which directly improves the robustness and comparability of subsequent judgments: indicators can be aligned between similar windows, resulting in more stable defect probabilities; uplink and review focus only on windows with high information density, reducing bandwidth and manpower consumption; risk scoring and adaptive cycles can also be updated based on evidence with consistent granularity, thereby reducing false alarms / false negatives and accelerating the closed loop from discovery to handling.

[0043] Based on the power equipment sensing devices under available conditions, the raw flow data of the power equipment within the historical inspection cycle is extracted. The historical inspection cycle is divided by an adaptive window, and the raw flow data of the power equipment under each adaptive window is extracted and recorded as the window observation set of the power equipment.

[0044] Specifically, the self-inspection quality indicators of the power equipment sensing devices are obtained, and the specific analysis process is as follows:

[0045] The self-test data definition threshold group includes clock deviation, sampling interval jitter, signal-to-noise ratio, and number of repeated frames.

[0046] First, calculate the deviation processing results for each parameter of the self-test data set: including the deviation processing results for the maximum clock deviation relative to the clock limit deviation, the deviation processing results for the sampling interval jitter relative to the sampling interval limit jitter, the deviation processing results for the acquisition signal-to-noise ratio relative to the limit signal-to-noise ratio, and the deviation processing results for the number of repeated acquisition frames relative to the number of repeated frame limits. Then, weight and summarize the above deviation processing results with the corresponding weight parameters to obtain the self-test quality index of the power equipment sensing device. The deviation of the maximum clock deviation relative to the clock limit deviation is recorded as clock health, the deviation of the sampling interval jitter relative to the sampling interval limit jitter is recorded as sampling stability, the deviation of the acquisition signal-to-noise ratio relative to the limit signal-to-noise ratio is recorded as signal-to-noise quality, and the deviation of the number of repeated acquisition frames relative to the number of repeated frame limits is recorded as data uniqueness parameter. The specific self-test quality index analysis process is as follows:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052] In the formula, Q represents the self-inspection quality index of the power equipment sensing device. 1 represents the clock health status of the power equipment's sensing devices. 2 represents the sampling stability of the power equipment sensing device. 3 represents the signal-to-noise quality of power equipment sensing devices. 4 represents the unique data parameter of the power equipment sensing device. T represents the maximum clock deviation of the power equipment's sensing device. T' represents clock boundary deviation, Ji represents sampling interval jitter of the power equipment sensor, Ji' represents sampling interval boundary jitter, SNR represents acquisition signal-to-noise ratio of the power equipment sensor, SNR' represents boundary signal-to-noise ratio, Re represents the number of repeated acquisition frames of the power equipment sensor, Re' represents the number of repeated frame boundaries, q1 represents the weight parameter corresponding to the predefined clock health in the power system inspection database, q2 represents the weight parameter corresponding to the predefined sampling stability in the power system inspection database, q3 represents the weight parameter corresponding to the predefined signal-to-noise quality in the power system inspection database, and q4 represents the weight parameter corresponding to the predefined data uniqueness parameter in the power system inspection database.

[0053] In the formula, clip is the cutoff / clamping function: it restricts a certain value to a given range.

[0054] For example, clip(0,1,x) means that if x is less than 0, it takes the value 0; if x is greater than 1, it takes the value 1; and if x is between 0 and 1, it keeps x unchanged.

[0055] This refers to a predefined, extremely small positive number used as a safety net for fractions with a denominator of 0, preventing calculation errors; it does not change the physical meaning.

[0056] The items that are "the smaller the better" (deviation, jitter, duplicate frames) naturally get a perfect score of 1 when the actual measurement is 0.

[0057] In this embodiment, a multivariate analysis is conducted on clock skew, sampling interval jitter, acquisition signal-to-noise ratio (SNR), and the number of repeated acquisition frames. Specifically, the correlation between these parameters is considered. Clock skew disrupts multi-sensor time synchronization, leading to increased alignment / interpolation and thus amplifying sampling interval jitter. Increased jitter fragments the same physical change, effectively raising the noise floor during resampling and de-jittering, resulting in a decrease in acquisition SNR. On the other hand, when the link is congested or the buffer is blocked, the acquisition end often experiences repeated acquisition frames or repeated reporting. This reduces the effective information density and causes timestamp sequences to stagnate or surge, further worsening jitter and time synchronization assessment, forming a feedback loop of skew, jitter, and repeated frames. Low SNR scenarios (low light, haze, strong noise) are prone to triggering automatic retake / multi-frame stacking strategies, which also increases the probability of repeated frames.

[0058] The window observation set and availability status are input into the machine learning model, which outputs the initial defect vector of the window observation set. It then determines whether the window observation set should execute the uplink strategy. If the uplink strategy is executed, the edge processing end generates the associated data packets of the window observation set.

[0059] The aforementioned machine learning model (which can be based on a lightweight convolutional neural network) uses historical window observation sets as samples, and combines historical sensor self-inspection quality vectors with work orders / review records to generate window-level multi-label ground truth (defect family vectors); it enhances and normalizes images / infrared / partial discharge / SCADA, etc., and retains missing / low-quality masks; it adopts a multimodal encoder combined with a quality-aware fusion architecture (such as: visual CNN / segmentation head, temporal Transformer / TCN, partial discharge PRPD convolutional encoder, with quality vectors tuned through gating / FiLM). Each modality channel is configured, and a shared representation is generated and an initial defect vector (multi-label probability) is output in the fusion layer using attention / weighted stacking. The loss function combines Focal / weighted BCE (anti-class imbalance) with necessary localization / segmentation loss (Dice / IoU) and consistency / contrast loss (to improve cross-modal and cross-device generalization). Quality-weighted sampling (high-quality samples are learned first, and low-quality samples are gradually released) and device-grouped cross-validation (grouped by device / line to prevent information leakage) are introduced. After training, probability calibration (temperature scaling) and uncertainty estimation (MC Dropout / small ensemble) are performed, and quantization / distillation is carried out to meet edge inference. In this way, real-time window observation set and self-inspection quality indicators can be input during inference to stably output the initial defect vector, including the initial defect probability and defect probability uncertainty, which is seamlessly connected with the inspection process.

[0060] Furthermore, the determination of whether the window observation set should execute the uplink strategy is as follows:

[0061] The initial defect vector of the window observation set is specifically represented as a two-dimensional vector, including the initial defect probability of the window observation set and the defect probability uncertainty of the window observation set. The defect probability uncertainty is a quantitative measure of how reliable the probability is when the machine learning model gives a defect probability for a certain sample data. It measures the confidence level of this judgment; the smaller the uncertainty, the greater the confidence.

[0062] The initial defect probability of the window observation set is compared with the predefined defect baseline probability to obtain the defect probability comparison result. The uncertainty of the defect probability of the window observation set is compared with the predefined uncertainty baseline to obtain the uncertainty comparison result.

[0063] The defect probability comparison results include the first defect probability comparison result and the second defect probability comparison result, and the uncertainty comparison results include the first uncertainty comparison result and the second uncertainty comparison result.

[0064] The first comparison result of the defect probability indicates that the initial defect probability of the window observation set is less than the baseline defect probability, and the second comparison result indicates that the initial defect probability of the window observation set is greater than or equal to the baseline defect probability. The first comparison result of the uncertainty indicates that the uncertainty of the defect probability of the window observation set is less than the baseline uncertainty, and the second comparison result indicates that the uncertainty of the defect probability of the window observation set is greater than or equal to the baseline uncertainty.

[0065] The decision window observation set will not execute the uplink strategy if and only if both the defect probability comparison result and the uncertainty comparison result are the first comparison result. Otherwise, the decision window observation set will execute the uplink strategy. Specifically, the uplink strategy refers to the process by which the edge processing end uploads the associated data packets of the window observation set to the cloud processing end for cloud processing.

[0066] When the window observation set does not execute the uplink strategy, the edge processing end performs local release for the window observation set, that is, it only records the execution status of the power equipment under the window observation set. After the edge processing end has traversed all window observation sets, it transmits the complete execution status of the power equipment under the historical inspection cycle to the cloud processing end for status storage.

[0067] In this embodiment, bandwidth and cloud computing power are prioritized for suspicious / high-risk or uncertain observation samples, significantly reducing invalid data uplinks and central queuing, and lowering link congestion and alarm latency. In weak network or cost-constrained scenarios, bandwidth adaptability and on-demand computing power are achieved. At the same time, by allowing local data to pass through without uploading, low-value and low-quality data is avoided from polluting cloud judgments and reports, improving overall stability and interpretability. The implementation of the uplink strategy ensures that important alarms are not interrupted or overloaded, thereby optimizing end-to-cloud resource utilization and shortening the actual alarm handling time without sacrificing the detection rate.

[0068] The edge processing end uploads the associated data packets of the window observation set to the cloud processing end. The cloud processing end performs regression fusion to obtain the defect probability of the window observation set, and performs gating decision to generate event labels.

[0069] Specifically, the cloud processing unit performs regression fusion to obtain the defect probability of the window observation set. The specific analysis process is as follows:

[0070] The associated data packets of the window observation set include the initial defect vector of the window observation set and the self-inspection quality indicators of the power equipment sensing device.

[0071] Based on the initial defect probability of the window observation set, it is matched with the historical backtracking time length corresponding to each predefined initial defect probability interval to determine the specific interval of the initial defect probability of the window observation set, and thus obtain the historical backtracking time length corresponding to the above specific interval.

[0072] The cloud processing end extracts the credibility weights corresponding to several historical window observation sets based on the historical backtracking time length coverage, performs mean processing, and obtains and records them as the credibility preset weights. It should be explained that if the historical backtracking time length does not completely cover a certain historical window observation set during the historical window observation set extraction process, then that historical window observation set is discarded.

[0073] Based on the correlation set between self-inspection quality indicators and weight correction factors, the self-inspection quality indicators of power equipment sensing devices are substituted into the correlation set to obtain weight correction factors. The weight correction factors are then correlated and merged with the preset confidence weights. Specifically, the weight correction factors are multiplied by the preset confidence weights to obtain the confidence adaptation weights.

[0074] Based on the monotonically positive relationship between self-inspection quality indicators and weight correction factors, when the quality of the power equipment sensing device is low, the reliability of the mode is low; when the quality of the power equipment sensing device is high, the reliability of the mode is high. This yields a reliability-adaptive weight, giving high-quality observations more weight in the fusion process and automatically reducing the weight of low-quality observations. By mapping data quality to decision weights in real time in an interpretable, calibrable, and low-overhead manner, false alarms / misjudgments caused by defocusing, frame loss, clock skew, and low signal-to-noise ratio are significantly suppressed, improving the robustness and consistency of multimodal fusion. It also exhibits good transferability and auditability across different devices / scenarios.

[0075] The cloud processing unit performs regression fusion on the initial defect vector of the window observation set and the confidence adaptation weights to obtain the defect probability of the window observation set. Specifically, the cloud processing unit receives the window observation set, first performs consistency calibration on the initial defect vector to make it comparable in scale; then multiplies it with the corresponding confidence adaptation weights to obtain quality-weighted features, which are then fed into a lightweight regressor (such as weighted logistic regression / GBDT stacking) and fused with a bias term learned by a lightweight convolutional neural network. The regressor outputs a comprehensive score, which is mapped to the overall defect probability of the window via a sigmoid function.

[0076] The learned bias term is essentially an automatically calibrated baseline knob. This fusion unit contains a constant term that can be understood as the system's average risk baseline for the current scenario. If a certain type of device has historically been more prone to problems, this baseline will be learned to be slightly higher; conversely, slightly lower.

[0077] Furthermore, gating decision generation of event labels is performed, and the specific analysis process is as follows:

[0078] Based on the correlation set between the defect probability uncertainty and the defect probability adjustment element, the defect probability uncertainty of the window observation set is substituted into the correlation set to obtain the defect probability adjustment element. The defect probability threshold is then corrected by multiplying the defect probability adjustment element by the defect probability threshold to obtain the defect probability fitting threshold.

[0079] It's important to explain that in the above process, the adjustment element is set to increase with increasing uncertainty, with a minimum value of 1. This means that when the model is unsure about the outcome, the adjustment element will be greater than 1, proportionally raising the default threshold, resulting in a suitable threshold that is not less than the default threshold. When the model is confident, the adjustment element will be close to 1, and the suitable threshold will be the same as the default threshold. This way, samples falling between the default and suitable thresholds are uniformly included in the manual review zone, avoiding the direct triggering of automatic alarms in uncertain scenarios and thus preventing an increase in false alarms. Simultaneously, high-confidence obvious anomalies still trigger rapid alarms, while clearly normal samples are allowed to pass, achieving a better balance between sensitivity, stability, and manpower / bandwidth investment.

[0080] Based on the defect probability of the window observation set, the predefined default threshold and the defect probability adaptation threshold are verified respectively to obtain the defect probability verification results. The defect probability verification results include the first verification result, the second verification result, and the third verification result.

[0081] The first verification result of the defect probability is expressed as the defect probability of the window observation set being less than or equal to the default defect probability threshold. The second verification result of the defect probability is expressed as the defect probability of the window observation set being greater than the default defect probability threshold and less than or equal to the defect probability adaptation threshold. The third verification result of the defect probability is expressed as the defect probability of the window observation set being greater than the defect probability adaptation threshold.

[0082] If the defect probability verification result shows the first defect probability verification result, then the generated gated decision event label is the normal observation label.

[0083] If the defect probability verification result shows the second defect probability verification result, then the gating decision event label generated is the active review label.

[0084] If the defect probability verification result is displayed as the third defect probability verification result, then the gating decision event label generated is the preliminary alarm label.

[0085] In this embodiment, a hierarchical mechanism of event labels (normal / review / alarm) is adopted. Based on quality perception and uncertainty gating, the window observation set is automatically divided into three streams: low-risk samples are directly allowed (reducing invalid uplinks and interference), boundary / uncertain samples enter the manual review channel (stabilizing false alarm rate and avoiding mishandling), and high-confidence, high-risk samples are immediately alarmed and trigger risk assessment. This hierarchical approach allows bandwidth and human resources to be allocated according to risk priority, significantly reducing false alarms / missed alarms and duplicate dispatches, and improving on-site response efficiency. At the same time, the accompanying traceability and threshold band interpretability form an auditable and traceable closed-loop decision-making process.

[0086] Risk assessment of power equipment is performed based on gating decision event labels to obtain risk scores for the power equipment. The adaptive inspection cycle is then limited based on the risk scores of the power equipment, and the next round of inspection process is configured.

[0087] Specifically, risk assessment of power equipment is conducted based on gating decision event labels. The specific analysis process is as follows:

[0088] When the cloud processing terminal identifies the generated gating decision event label as a normal observation label, it will report back to the inspection interaction terminal that the current window observation set is a normal observation and record the gating decision log for this time.

[0089] When the cloud processing terminal identifies the gating decision event label as an active review label, it generates a manual review signal from the current window observation set and feeds it back to the inspection interaction terminal. At the same time, it extracts the total probability of defects in the window observation set under the historical inspection cycle, maps it to obtain the manual review ratio of the raw stream data, segments the raw stream data of the power equipment based on the manual review ratio, obtains the raw stream data volume of the power equipment, and packages it into a review dataset of the power equipment and feeds it back to the inspection interaction terminal.

[0090] It's important to explain that when the inspection system detects a suspected anomaly on a device, the edge device packages the raw streaming data of that device during the critical period into a review dataset, sends it back to the inspection interaction terminal, and pushes it to the power operation personnel on-site or in the background. In this way, the inspection is responsible for broad coverage and high efficiency in discovering suspicious points, while the manual review is responsible for making a final decision based on sufficient evidence (confirmation / exclusion / suggestion for supplementary data collection). The two form a human-machine collaboration of discovery, evidence collection, judgment, handling, and feedback: it avoids creating false alarms by automatically triggering a small number of difficult samples, and it focuses human time on the high-value segments that most need to be judged. The review conclusion then feeds back into the threshold and strategy mapping, updates work orders and adaptive inspection cycles, making the system more accurate, timely, and compliant with regulations and traceable.

[0091] In this embodiment, the cloud processing terminal sends the power equipment tasks marked as requiring review to the smart glasses of the power workers, pre-loaded with evidence frames and model prompts. After the workers arrive at the site, the glasses use positioning and alignment (such as site marking or environmental mapping) to overlay the equipment name, suspected defect location, operation instructions, and safety boundaries in the field of vision. The workers use voice / gestures / buttons to gradually complete the visible light, infrared, or UV supplementary sampling and magnification of the same equipment, and collect the necessary angles and distances as prompted. The equipment displays data quality prompts (clarity, jitter, overexposure) locally and allows one-click reshooting or supplementary shooting. Voice / text annotations and pass / fail judgments can be added directly on site. After completion, the glasses automatically package and send back the original stream segments, keyframes, sensor quality indicators, and review conclusions. If offline, they are first encrypted and cached locally. When online, they are synchronized to the work order system and trigger subsequent processing or order closure.

[0092] Once the cloud determines that the current window observation set requires manual review, it immediately adapts the review ratio by combining the total probability distribution of defects within the historical inspection cycle, and converts it into a clear indication of the amount of raw streaming data at the front end. This allows for precise allocation of manpower and bandwidth within the constraints of a predetermined false alarm rate and labeling budget. This enables dynamic matching of the review task scale and risk intensity, avoiding congestion and cost waste caused by full upload / blind sampling; at the same time, it ensures sufficient evidence, improves the review hit rate and decision stability, shortens the alarm-to-handling link, and forms a traceable closed-loop feedback.

[0093] When the cloud processing terminal identifies the generated gating decision event label as a preliminary alarm label, it generates an abnormal risk signal from the current window observation set and feeds it back to the inspection interaction terminal. It then performs a risk assessment on the power equipment under the current window observation set, obtains the risk score of the power equipment, limits the adaptive inspection cycle, configures the next round of inspection process, and outputs a risk aggregation node based on the risk score of the power equipment.

[0094] After calculating the risk score of the power equipment, the cloud processing unit compares it with preset thresholds in real time. If the score exceeds the preset threshold, a risk aggregation node is generated in the geographical or electrical area where the equipment is located. This node aggregates similar or adjacent high-risk points in the area into an actionable hotspot marker, along with representative evidence and handling priorities in real time. The risk aggregation node serves as the direct basis for manual re-inspection, used to assign tasks to wearable devices, plan nearby inspection routes, merge duplicate alarms, and avoid scattered trips and omissions. Simultaneously, it dynamically adds, deletes, and updates the intensity as the risk changes, helping the re-inspection team quickly identify priority re-inspection areas and collect additional parameters on-site in complex scenarios, thereby improving the re-inspection hit rate and handling timeliness.

[0095] Furthermore, a risk score is obtained for the electrical equipment. The specific assessment process is as follows:

[0096] The characteristic parameter set of the power equipment is obtained, including the vibration trend characteristic slope of the power equipment, the real-time load rate of the power equipment, and the real-time wind direction alignment of adjacent power equipment over the time span of the window observation set. The characteristic parameter set can be extracted from the monitoring records of the power equipment sensing device.

[0097] The analysis process of the vibration trend characteristic slope of power equipment is as follows: the vibration signal intensity of power equipment at the beginning time of the window observation set is compared with the vibration signal intensity of power equipment at the end time of the window observation set to obtain the vibration trend characteristic slope of power equipment.

[0098] The analysis process for the real-time wind direction alignment of adjacent power equipment is as follows: the real-time wind line angle of the power equipment is compared with the real-time wind line angle of the adjacent power equipment to obtain the real-time wind direction alignment of the adjacent power equipment.

[0099] The term "adjacent equipment" here refers to equipment that is spatially and topologically close, such as adjacent towers on the same line, components with the same circuit and phase, or equipment located within the same wind farm grid.

[0100] The vibration trend characteristic slope of the power equipment, the real-time load rate of the power equipment, and the real-time wind direction alignment of adjacent power equipment are normalized respectively. The normalized results are combined with the defect probability of the window observation set to obtain the risk score of the power equipment. The specific analysis process is as follows:

[0101]

[0102] In the formula, R is the risk score of the power equipment, P* is the defect probability of the window observation set, and T is the slope of the vibration trend characteristic of the power equipment. (t) represents the load rate of the power equipment at time t, where t is a time variable. [t0, t1], where t0 is the start time of the window observation set and t1 is the end time of the window observation set. (t) represents the wind alignment of adjacent power equipment at time t, w1 is the weight parameter corresponding to the predefined defect probability in the power system inspection database, w2 is the weight parameter corresponding to the predefined vibration trend characteristic slope in the power system inspection database, w3 is the weight parameter corresponding to the predefined load rate in the power system inspection database, and w4 is the weight parameter corresponding to the predefined wind alignment in the power system inspection database.

[0103] It should be explained that when the wind direction alignment is large, it indicates that the wind is coming from the same direction over a large area, which is more like a common environmental phenomenon and therefore not an independent failure of a certain piece of equipment. The relative risk of individual equipment should be reduced.

[0104] In this embodiment, multivariate analysis is performed using defect probability, vibration trend slope, real-time load rate, and real-time wind alignment. Specifically, the correlation between these parameters is considered. The real-time load rate provides the background of electrothermal / mechanical stress; the higher the load, the more dangerous the same vibration increase, thus amplifying the driving effect of vibration on defect probability. The wind alignment of adjacent equipment is an indicator of common environmental factors. When the alignment is high, it is more like a group fluctuation caused by synchronous wind-induced excitation, and the system should suppress the defect probability (favoring environmental factors rather than single-point failures). When the alignment is low and the vibration increases, it is more like a local defect, and the defect probability is adjusted accordingly. At the same time, the superposition of high load and high wind alignment will amplify the risk of wind-induced vibration (such as galloping / impact), further increasing the defect probability and increasing the risk score.

[0105] Specifically, based on the risk score of power equipment, the adaptive inspection cycle is limited, and the next round of inspection process is configured. The specific analysis process is as follows:

[0106] Based on the correlation set between risk scores and inspection cycle shortening factors, the risk scores of power equipment are substituted into an adaptive truncation function to obtain the inspection cycle shortening factor.

[0107] The inspection cycle shortening factor is associated and merged with the historical inspection cycle. Specifically, the inspection cycle shortening factor is multiplied by the historical inspection cycle to obtain the inspection correction cycle. The inspection correction cycle is then compared with the predefined minimum allowable inspection cycle to obtain the cycle permission comparison result. The cycle permission comparison result includes the first cycle permission comparison result and the second cycle permission comparison result.

[0108] If the cycle permit comparison result shows the first cycle permit comparison result, then the next round of inspection process is configured based on the inspection correction cycle. If the cycle permit comparison result shows the second cycle permit comparison result, then the next round of inspection process is configured based on the minimum allowable inspection cycle.

[0109] The first comparison result of the cycle permission indicates that the inspection correction cycle is greater than or equal to the minimum allowable inspection cycle, and the second comparison result of the cycle permission indicates that the inspection correction cycle is less than the minimum allowable inspection cycle.

[0110] In this embodiment, risk-adaptive execution is performed when the correction period is greater than or equal to the minimum allowable period; otherwise, the minimum allowable period is used instead. This achieves optimal resource allocation—encrypting high-risk areas and reducing low-risk areas—while ensuring a minimum inspection frequency that guarantees safety and compliance. This avoids excessive inspections, route congestion, and waste of manpower, equipment, and materials caused by instantaneous fluctuations, while ensuring timely verification and handling of high-risk equipment. Simultaneously, it enhances the predictability of the lower limit guardrail improvement plan and the service level agreement achievement rate, stabilizes scheduling and shift load, reduces invalid work orders driven by false alarms, and overall improves detection timeliness and operational cost efficiency.

[0111] In this embodiment of the invention, the inspection process is specifically based on a power inspection and management platform, and the specific inspection management display is as follows: Figure 3 , Figure 4 As shown, Figure 3 The inspection management interface of the power inspection and control platform. Figure 1 , Figure 4 The inspection management interface of the power inspection and control platform. Figure 2 The inspection management overview dashboard summarizes key performance indicators (KPIs) such as the number of valid devices, active inspections, uplink events per hour, alarm pass rate, manual review rate, and upcoming SLA work orders. It also displays today's pending, ongoing, and completed tasks. The equipment and risk page allows filtering by region, voltage, type, and risk at the top. A table in the middle displays the current risk score, trend slope, real-time load rate, and last handling time for each device and component. Risk distribution and hotspots are listed by region, line, and equipment type, providing a top list of high-risk and top alarm sites. One-click task assignment, adding to review, and common cause masking are supported. The inspection management interface allows maintenance personnel to quickly locate high-risk devices and hotspot areas, monitor inspection health and capacity, and then issue handling and review tasks with one click, improving alarm hit rate and SLA achievement.

[0112] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A power system inspection method based on machine learning, characterized in that, include: The edge processing end acquires the self-inspection data set of the power equipment sensing device within the historical inspection cycle, which is used to determine the availability status of the power equipment sensing device. Based on the availability status determination result, the raw stream data of the power equipment is extracted to construct a window observation set, thus obtaining the window observation set of the power equipment. The window observation set and availability status are input into the machine learning model, and the initial defect vector of the window observation set is output. This vector is used to determine whether the window observation set should execute the uplink strategy. If the uplink strategy is executed, the edge processing end generates the associated data packets of the window observation set. The edge processing end uploads the associated data packets of the window observation set to the cloud processing end. The cloud processing end performs regression fusion to obtain the defect probability of the window observation set, and generates event labels based on the defect probability for gating decisions. Risk assessment of power equipment is performed based on gating decision event labels to obtain risk scores for the power equipment. The adaptive inspection cycle is then limited based on the risk scores of the power equipment, and the next round of inspection process is configured.

2. The power system inspection method based on machine learning according to claim 1, characterized in that: The specific process for determining the availability status of the power equipment sensing device is as follows: The self-test data set of the power equipment sensor includes the maximum clock deviation of the power equipment sensor, the sampling interval jitter of the power equipment sensor, the acquisition signal-to-noise ratio of the power equipment sensor, and the number of repeated acquisition frames of the power equipment sensor during the historical inspection cycle. The self-test data set of the power equipment sensing device is compared with the preset self-test data definition threshold set to obtain the self-test data comparison result, which includes a first self-test data comparison result and a second self-test data comparison result. The availability status of the power equipment sensing device is determined to be available only if the self-test data comparison result shows the first self-test data comparison result; otherwise, it is determined to be unavailable.

3. The power system inspection method based on machine learning according to claim 2, characterized in that: The process of extracting raw stream data from power equipment to construct a windowed observation set is as follows: When the availability status of the power equipment sensing device is determined to be available, the self-test quality index of the power equipment sensing device is obtained by jointly evaluating and fusing the self-test data group and the self-test data definition threshold group. Based on the self-inspection quality indicators of the power equipment sensing device, a mapping relationship table between the preset self-inspection quality indicators and the window limiting factor is extracted. The self-inspection quality indicators of the power equipment sensing device are then substituted into the mapping relationship table to obtain the window limiting factor. The window constraint factor is associated with and merged with the default window to obtain the adaptive window; Based on the power equipment sensing devices under available conditions, the raw flow data of the power equipment within the historical inspection cycle is extracted. The historical inspection cycle is divided by an adaptive window, and the raw flow data of the power equipment under each adaptive window is extracted and recorded as the window observation set of the power equipment.

4. The power system inspection method based on machine learning according to claim 3, characterized in that: The specific analysis process for obtaining the self-inspection quality index of the power equipment sensing device is as follows: The self-test data definition threshold group includes clock definition deviation, sampling interval definition jitter, definition signal-to-noise ratio, and definition number of repeated frames; First, calculate the deviation processing results of each parameter in the self-test data set: including the deviation processing results of the maximum clock deviation relative to the clock limit deviation, the deviation processing results of the sampling interval jitter relative to the sampling interval limit jitter, the deviation processing results of the acquisition signal-to-noise ratio relative to the limit signal-to-noise ratio, and the deviation processing results of the number of repeated acquisition frames relative to the number of repeated frame limits. Then, the above deviation processing results are weighted and summarized with the corresponding weight parameters to obtain the self-test quality index of the power equipment sensing device.

5. The power system inspection method based on machine learning according to claim 1, characterized in that: The specific determination process for whether the observation set of the determination window executes the uplink strategy is as follows: The initial defect vector of the window observation set is specifically represented as a two-dimensional vector, including the initial defect probability of the window observation set and the defect probability uncertainty of the window observation set. The initial defect probability of the window observation set is compared with the predefined defect baseline probability to obtain the defect probability comparison result. The uncertainty of the defect probability of the window observation set is compared with the predefined uncertainty baseline to obtain the uncertainty comparison result. The defect probability comparison results include a first defect probability comparison result and a second defect probability comparison result, and the uncertainty comparison results include a first uncertainty comparison result and a second uncertainty comparison result. The decision window observation set does not execute the uplink strategy if and only if the defect probability comparison result shows the first defect probability comparison result and the uncertainty comparison result shows the first uncertainty comparison result; otherwise, the decision window observation set executes the uplink strategy.

6. The power system inspection method based on machine learning according to claim 1, characterized in that: The cloud processing terminal performs regression fusion to obtain the defect probability of the window observation set. The specific analysis process is as follows: The associated data packets of the window observation set include the initial defect vector of the window observation set and the self-test quality index of the power equipment sensing device. Based on the initial defect probability of the window observation set, it is matched with the historical backtracking time length corresponding to each predefined initial defect probability interval to determine the specific interval of the initial defect probability of the window observation set, and obtain the historical backtracking time length corresponding to the specific interval. The cloud processing end extracts the credibility weights corresponding to several historical window observation sets based on the historical backtracking time length coverage, performs mean processing, and obtains and records them as the credibility preset weights. Based on the correlation set between self-inspection quality indicators and weight correction factors, the self-inspection quality indicators of power equipment sensing devices are substituted into a linear function to obtain weight correction factors. The weight correction factors are then correlated and merged with the preset confidence weights to obtain confidence adaptation weights. The cloud processing unit performs regression fusion on the initial defect vector of the window observation set and the confidence adaptation weight to obtain the defect probability of the window observation set.

7. The power system inspection method based on machine learning according to claim 6, characterized in that: The specific analysis process for generating event tags through gating decisions is as follows: Based on the linear function between the defect probability uncertainty and the defect probability adjustment element, the defect probability uncertainty of the window observation set is substituted into the linear function to obtain the defect probability adjustment element. The defect probability threshold is then corrected using the defect probability adjustment element to obtain the defect probability fitting threshold. Based on the defect probability of the window observation set, the predefined default threshold and the defect probability adaptation threshold are verified respectively to obtain the defect probability verification result. The defect probability verification result includes the first defect probability verification result, the second defect probability verification result, and the third defect probability verification result. If the defect probability verification result shows the first defect probability verification result, then the gated decision event label is generated as the normal observation label; If the defect probability verification result shows the second defect probability verification result, then the gating decision event label generated is the active review label; If the defect probability verification result is displayed as the third defect probability verification result, then the gating decision event label generated is the preliminary alarm label.

8. The power system inspection method based on machine learning according to claim 1, characterized in that: The specific analysis process for risk assessment of power equipment based on gating decision event labels is as follows: When the cloud processing terminal identifies the generated gating decision event label as a normal observation label, the cloud processing terminal will report back to the inspection interaction terminal that the current window observation set is a normal observation and record the gating decision log for this time. When the cloud processing terminal identifies the gating decision event label as an active review label, it generates a manual review signal from the current window observation set and feeds it back to the inspection interaction terminal. At the same time, it extracts the total probability of defects in the window observation set under the historical inspection cycle, maps it to obtain the manual review ratio of the raw stream data, segments the raw stream data of the power equipment based on the manual review ratio, obtains the raw stream data volume of the power equipment, and packages it into a review dataset of the power equipment and feeds it back to the inspection interaction terminal. When the cloud processing terminal identifies the generated gating decision event label as a preliminary alarm label, it generates an abnormal risk signal from the current window observation set and feeds it back to the inspection interaction terminal. It then performs a risk assessment on the power equipment under the current window observation set, obtains the risk score of the power equipment, limits the adaptive inspection cycle, configures the next round of inspection process, and outputs a risk aggregation node based on the risk score of the power equipment.

9. The power system inspection method based on machine learning according to claim 8, characterized in that: The risk score for the obtained power equipment is specifically assessed through the following process: Obtain a set of characteristic parameters for power equipment, including the slope of the vibration trend characteristics of power equipment, the real-time load rate of power equipment, and the real-time wind direction alignment of adjacent power equipment over the time span of the window observation set. The analysis process of the vibration trend characteristic slope of the power equipment is as follows: the vibration signal intensity of the power equipment at the beginning time of the window observation set is compared with the vibration signal intensity of the power equipment at the end time of the window observation set to obtain the vibration trend characteristic slope of the power equipment. The analysis process of the real-time wind direction alignment of adjacent power equipment is as follows: the ratio of the real-time wind line angle of the power equipment to the real-time wind line angle of the adjacent power equipment is processed to obtain the real-time wind direction alignment of the adjacent power equipment. The vibration trend characteristic slope of the power equipment, the real-time load rate of the power equipment, and the real-time wind direction alignment of adjacent power equipment are normalized respectively. The normalization results are combined with the defect probability of the window observation set to obtain the risk score of the power equipment.

10. The power system inspection method based on machine learning according to claim 1, characterized in that: The adaptive inspection cycle is limited based on the risk score of power equipment, and the next round of inspection process is configured. The specific analysis process is as follows: Based on the correlation set between risk score and inspection cycle shortening factor, the risk score of power equipment is substituted into the adaptive truncation function to obtain the inspection cycle shortening factor. The inspection cycle shortening factor is associated and merged with the historical inspection cycle to obtain the inspection correction cycle. The inspection correction cycle is compared with the predefined minimum allowable inspection cycle to obtain the cycle permission comparison result. The cycle permission comparison result includes the cycle permission first comparison result and the cycle permission second comparison result. If the cycle permit comparison result shows the first cycle permit comparison result, then the next round of inspection process is configured based on the inspection correction cycle. If the cycle permit comparison result shows the second cycle permit comparison result, then the next round of inspection process is configured based on the minimum allowable inspection cycle.

Images