An electric power equipment health state artificial intelligence integration method and system

Abstract

A method and system for integrating artificial intelligence into the health status of power equipment belongs to the field of intelligent diagnostic technology. This method aligns multi-channel online monitoring data through time windows to generate window sequences, missing mask matrices, and quality matrices. These are then input into a self-supervised representation model based on blind spot mask reconstruction to obtain window representation vectors. Target buckets are determined in a normal reference library according to equipment type and operating condition labels, and the top k normal reference representations are retrieved from either the target bucket or the fallback bucket to form prototype representations. Anomaly standard scores are calculated based on the residuals between the window representation vectors and the prototype representations, combined with a set of historical normal residuals. Dual-indicator triage is performed based on the anomaly standard scores and uncertainties, and a review or retest is initiated when trigger conditions are met. An admission-based closed-loop update is performed based on the review or retest conclusions. This invention can improve the accuracy, reliability, and long-term online operational stability of health status assessment under conditions of missing data, noise, and changing operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and intelligent diagnosis technology, and particularly to an artificial intelligence integration method and product for power equipment health status. More specifically, it relates to an intelligent integration technology for assessing, warning, and continuously optimizing the health status of power equipment under conditions of irregular sampling, missing data, changing operating conditions, and scarce fault labels, utilizing blind spot mask reconstruction of self-supervised representations, multi-prototype consistency reference comparison, dual-indicator diversion of anomaly degree and credibility, and an admission-based closed-loop update mechanism, for multi-channel online monitoring data. Background Technology

[0002] Power equipment typically operates in complex environments for extended periods, resulting in online monitoring data characterized by multiple channels, inconsistent sampling frequencies, variable time series lengths, significant noise interference, and widespread data gaps. Particularly for transformers, circuit breakers, cables, switchgear, and secondary equipment, the distribution of monitoring data varies considerably across different equipment types, load levels, and environmental conditions. This makes it difficult for traditional health status assessment methods based on fixed thresholds, fixed rules, or single supervised classification models to simultaneously achieve accuracy, stability, and transferability.

[0003] In existing technologies, one type of method mainly relies on manual experience to set alarm thresholds and makes separate judgments on monitoring indicators such as temperature, current, voltage, vibration, partial discharge, and gas characteristic quantities. This type of method is simple to implement, but it is difficult to fully utilize the correlation between multi-channel data, and it is prone to false alarms or missed alarms when operating conditions change, equipment ages, or environmental disturbances are significant. Another type of method uses supervised learning models to directly classify or regress health status, but this type of method usually relies on sufficient and accurate fault labeling samples; however, in real power scenarios, fault samples are scarce, labeling costs are high, and fault mode evolution is complex, resulting in insufficient model training foundation and limited generalization ability.

[0004] Furthermore, online monitoring data often exhibits irregular sampling and significant missing data. Directly interpolating, truncating, or simply padding the raw data with zeros can easily confuse true low values ​​with missing values, thus affecting subsequent feature extraction and health assessment results. Even when some existing solutions incorporate deep learning models for time-series modeling, they often lack explicit expression of missing data locations and data quality, making it difficult to maintain stable performance in terms of noise resistance, missing data resistance, and drift resistance.

[0005] Meanwhile, while many existing anomaly detection methods can output anomaly scores, they fail to adequately consider the differences in data distribution between different equipment types and operating conditions, easily misjudging changes in normal operating conditions as anomalies. Furthermore, they lack effective quantification of the reliability of model outputs and struggle to establish a reasonable hierarchy among "direct alarm," "continued observation," and "manual verification." In addition, simply writing subsequent manual conclusions back into the model or sample library could contaminate the normal baseline with latent fault samples, affecting the reliability of long-term online operation.

[0006] Therefore, there is an urgent need for a new AI-integrated solution for the health status of power equipment. This solution should be able to construct a stable equipment status representation under conditions of scarce fault labels, irregular sampling, missing data, and significant changes in operating conditions. It should also form a normal reference baseline associated with operating conditions, achieve health status diversion driven by the degree of anomaly and the credibility of conclusions, and improve the system's continuous adaptive capability and long-term operational reliability through a constrained closed-loop update mechanism. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems existing in the prior art, such as strong dependence on fault labeling, poor adaptability to irregular sampling and missing data, sensitivity to changes in operating conditions, insufficient credibility of health conclusions, and easy contamination of normal baselines during long-term online updates. The invention provides an artificial intelligence integration method and system for the health status of power equipment to achieve stable assessment, anomaly warning, reliable diversion, and closed-loop optimization of the health status of power equipment.

[0008] To achieve the aforementioned objective, this invention provides, in one aspect, an artificial intelligence integration method for the health status of power equipment, comprising: Time window alignment is performed on multi-channel online monitoring data to generate window sequences, missing mask matrices, and quality matrices. These window sequences, missing mask matrices, and quality matrices are then input into a self-supervised representation model based on blind spot mask reconstruction to obtain window representation vectors. ; Determine the target bucket in the normal reference library according to equipment type and operating condition label, and search the previous bucket or rollback bucket. A normal reference characterization Based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation According to the window representation vector With the prototype representation The residuals and combined with the bucket Historical normal residual set Calculate the abnormal standard score Based on the window representation vector and the aforementioned abnormal standard score Output uncertainty and with the aforementioned abnormal standard score As an indicator of the degree of anomaly, the aforementioned uncertainty The system performs dual-indicator traffic splitting as a credibility indicator; it initiates a review or retest when trigger conditions are met, and performs an admission-based closed-loop update based on the review or retest conclusions. Only window representation vectors that meet the normal conclusion and satisfy quality constraints, missing rate constraints, and no-fault constraints during the observation period are written into the normal reference library; the system outputs a health conclusion and an anomaly standard score. Uncertainty And evidence package.

[0009] To achieve the aforementioned objective, the present invention also provides an artificial intelligence-integrated system for the health status of power equipment, comprising: The window alignment module aligns multi-channel online monitoring data through time windows, generating a window sequence, a missing mask matrix, and a quality matrix. The blind spot representation module inputs the window sequence, the missing mask matrix, and the quality matrix into a self-supervised representation model based on blind spot mask reconstruction to obtain window representation vectors. The operating condition reference module is used to determine the target bucket in the normal reference library according to the equipment type and operating condition label, and to retrieve the previous bucket from the target bucket or the rollback bucket. A normal reference characterization Based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation The residual exception module is used to determine the window representation vector. With the prototype representation The residuals and combined with the bucket Historical normal residual set Calculate the abnormal standard score The evidence inference module is used to infer evidence based on the window representation vector. and the aforementioned abnormal standard score Output uncertainty and execute based on the aforementioned abnormal standard score. As an indicator of the degree of anomaly, with the aforementioned uncertainty The system employs a dual-indicator splitting mechanism for credibility metrics; a review and closed-loop module initiates review or retesting upon meeting trigger conditions, and performs admission-based closed-loop updates based on the review or retesting conclusions. Only window representation vectors that meet the normal conclusion and satisfy quality constraints, missing rate constraints, and no-fault constraints during the observation period are written into the normal reference library; the result output module outputs the health conclusion and the abnormal standard score. Uncertainty And evidence package.

[0010] Compared with the prior art, the present invention has at least the following beneficial effects: Firstly, this invention aligns data using a unified time window and generates a missing mask matrix and a quality matrix, enabling irregular sampling, missing data, and data with quality fluctuations to be processed in a consistent manner. This effectively reduces feature distortion caused by unstable data acquisition and improves the stability of health status assessment.

[0011] Secondly, this invention adopts a self-supervised representation learning method based on blind spot mask reconstruction, which can still obtain robust window representation vectors even when there are insufficient fault labeling samples, thereby reducing the dependence on manual labeling and improving the adaptability to noise, missing and drift scenarios.

[0012] Thirdly, this invention establishes a normal reference library based on equipment type and operating condition labels, and is based on the previous... A normal reference characterization forms a prototype characterization, which can build a normal baseline that matches the operating conditions, reduce false alarms caused by equipment differences and operating condition fluctuations, and enhance the consistency of anomaly detection results.

[0013] Fourth, this invention robustly standardizes the window representation residuals by combining them with a set of historical normal residuals to obtain anomaly standard scores. This gives anomaly detection better identifiability and transferability, which is beneficial for deployment across devices, operating conditions, and time periods. Fifth, by combining abnormal standard scores with uncertainty for dual-indicator triage, this invention can not only output health conclusions, but also trigger alarms, monitor alarms, review or retest based on the degree of abnormality and the difference in credibility, thereby improving the rationality and reliability of operation and maintenance decisions.

[0014] Sixth, by setting an admission-based closed-loop update mechanism, this invention only writes samples that meet the normal conclusion and satisfy the quality constraints, missing rate constraints, and no-fault constraints during the observation period into the normal reference library. This can reduce the risk of latent fault samples contaminating the normal baseline and improve the system's long-term online adaptability and security.

[0015] Seventh, by outputting evidence packages, this invention provides a basis for manual review, maintenance and disposal, and post-audit, which can improve the interpretability, traceability and engineering application value of health status assessment results. Attached Figure Description

[0016] Figure 1 This is a flowchart of the artificial intelligence integration method for the health status of power equipment provided by the present invention.

[0017] Figure 2 This is a diagram comparing the overall detection performance of different methods.

[0018] Figure 3 This is a schematic diagram comparing the robustness of different methods under missing and noisy conditions.

[0019] Figure 4 This is a diagram comparing the false alarm rates of different methods under changing operating conditions.

[0020] Figure 5 This is a comparative diagram showing the results of dual-index diversion using different methods.

[0021] Figure 6 This is a diagram comparing the early warning capabilities and closed-loop stability of different methods. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions or conventional modifications to some of the technical features, and such substitutions or modifications should all fall within the scope of protection of the present invention.

[0023] Figure 1 This is a flowchart of the artificial intelligence integration method for the health status of power equipment provided by the present invention, such as... Figure 1 As shown, the artificial intelligence integration method for the health status of power equipment provided by this invention is executed by a processor and includes steps S1 to S7. Correspondingly, it can also be implemented collaboratively by a data acquisition module, a window alignment module, a blind spot characterization module, an operating condition reference module, a residual anomaly module, an evidence inference module, a review and closed-loop module, and a result output module. The method is applicable to scenarios involving health status assessment, anomaly early warning, review and diversion, and closed-loop optimization of power equipment such as transformers, circuit breakers, GIS, switchgear, cables, and their accessories.

[0024] S1. Online monitoring data collection, unified coding and window alignment S1 specifically includes: acquiring multi-channel online monitoring data of the target power equipment and uniformly encoding the equipment identifier and channel identifier; wherein, the multi-channel online monitoring data includes at least an operating condition channel, an environmental channel, and a critical status channel, wherein the operating condition channel can be load current or active power, the environmental channel can be ambient temperature or ambient humidity, and the critical status channel can be one or more of partial discharge, vibration, SF6 micro-water, and key gases in oil chromatography; for each device, collecting the observation sequence of its corresponding channels at different time stamps, and according to a preset window length. Window step size and resampling interval Perform uniform time window alignment to form a window sequence.

[0025] Furthermore, the number of time steps within the window satisfy: .

[0027] Furthermore, a regular time grid is constructed for each time window, and a forward hold-and-resample rule is adopted: for any grid time point With any channel Selecting those that meet the tolerance threshold from the original observations The most recent observation time point of the constraint As the corresponding observed value; when there is no satisfying At the observation time point, the window sequence is marked as missing at that time step and at that channel position to generate a missing mask matrix. and the quality matrix The corresponding position is set to 0; when a quality marker exists, the quality marker is mapped to a quality matrix. The reliability value at the corresponding position in the matrix; when no quality marker is present, the quality matrix is ​​set to 1 by default.

[0028] Furthermore, the resampled window sequence is normalized by channel to obtain the normalized window sequence. Its normalization satisfies:

[0029] in, For time steps ,aisle Resampled value at that point, For channel The training set mean, For channel The standard deviation of the training set. To avoid stable terms with a denominator of zero, This is the normalized value.

[0030] When the mask matrix is ​​missing When the value is 1 at the corresponding position, the normalized value on the input side is set to the preset missing fill value, preferably 0, while the missing mask matrix is ​​retained. and mass matrix As an explicit input to the subsequent model, it distinguishes between "true values ​​of 0" and "missing values ​​imputed to 0". When the working condition channel is missing, the working condition label is fixed as medium working condition; when the environment channel is missing, the environment label is fixed as medium environment, to ensure that the subsequent bucketing, retrieval and inference processes can be executed.

[0031] Through step S1, this invention can uniformly convert raw online monitoring data with inconsistent sampling frequencies, misaligned timestamps, missing data, and high noise into data representations with consistent time scales and explicit missing data expressions. This provides a unified input basis for subsequent characterization learning, operating condition retrieval, and anomaly detection, and reduces false alarms and missed alarms caused by inconsistent preprocessing standards.

[0032] S2. Self-supervised representation learning based on blind spot mask reconstruction Specifically, this includes: the normalized window sequence obtained in step S1. Missing mask matrix and mass matrix Construct a blind spot indicator matrix and blind spot set and the normalized window sequence Missing mask matrix quality matrix and blind spot indicator matrix The data is then pieced together to form the model input. Then, input the model. The self-supervised representation model for blind spot mask reconstruction is used to obtain the window representation vector corresponding to the current time window. .

[0033] The model input satisfy:

[0034] in, Indicates splicing.

[0035] Furthermore, the set of blind spots It includes both dot masks and segment masks. Dot masks are used to randomly obscure discrete locations, while segment masks are used to simulate missing data over continuous time periods. For any selected blind spot, only the input-side values ​​are replaced, preserving the missing mask matrix. With the mass matrix Unchanged, and utilizes the blind spot indicator matrix. Mark blind spot locations to prevent the model from directly leaking target information.

[0036] Furthermore, the self-supervised representation model can employ a temporal convolutional network, a temporal attention network, or a combination of both encoding structures; preferably, a temporal convolutional network is used as the encoder to extract features from the input along the time dimension, and temporal pooling is performed on the encoder output to obtain a window representation vector. Simultaneously, the reconstruction head is configured to output the reconstruction values ​​for each blind spot location. During the training phase, only the blind spot set is processed. Location calculation reconstruction loss ,satisfy:

[0037] in, The number of elements in the blind spot set. For time step ,aisle The reconstruction value at that location, This corresponds to the true normalized value.

[0038] Furthermore, the training method of the self-supervised representation model can be a combination of offline pre-training and online fine-tuning: first, a training set is constructed using historical online monitoring data, and offline self-supervised pre-training is performed under conditions of no fault labeling or a small number of labels to obtain initial encoder parameters; then, after the system goes online, periodic fine-tuning is performed based on newly added normal window samples or samples that have been verified and closed-loop confirmed, in order to adapt to changes in data distribution caused by equipment aging, environmental changes and sensor drift.

[0039] Through step S2, this invention enables the learning of stable and robust window representation vectors from temporal windows with missing and noisy data, even under conditions where faulty samples and high-quality annotations are scarce. This reduces reliance on manual fault labels and improves adaptability to irregular sampling, missing, and drift scenarios.

[0040] S3, Construction of Normal Reference Library for Working Condition Bucketing and Multi-Prototype Consistency Retrieval Specifically, this includes: the window representation vector obtained in step S2. Establish a normal reference library based on equipment type and operating condition labels, and determine the target bucket corresponding to the current time window in the normal reference library. From the target bucket Or backtrack to the bucket before searching A normal reference characterization And based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation .

[0041] The operating condition labels include at least load labels and environment labels. Load labels can be divided into low, medium, and high load buckets based on the quantile thresholds of the operating condition channels in the training set, and environment labels can be divided into low, medium, and high environment buckets based on the quantile thresholds of the environment channels in the training set; target buckets... This can be determined by the equipment type, load bin, and environment bin. Furthermore, a normal reference bin is maintained for each bin, and a maximum capacity is set for each bin to control the storage scale.

[0042] When the target bucket When the number of normal reference samples in the target bucket is less than the preset empty bucket threshold, the search is performed by switching to the global bucket of the same device type or the global bucket of the entire site as the fallback bucket according to the preset fallback order. When the number of normal reference samples in both the target bucket and the fallback bucket is less than the preset empty bucket threshold, the output of a deterministic health conclusion is prohibited, and a review or retest suggestion is output.

[0043] Furthermore, the operating condition reference module is based on a distance function. Select the bucket with the smallest distance from the reference representation set of the target bucket or back bucket. A normal reference characterization To form a candidate set, the distance function shall employ at least one of the following:

[0044] or

[0045] in, The cosine similarity function is used. The distance is Euclidean.

[0046] Furthermore, regarding the aforementioned... A normal reference characterization Aggregation is performed to form a prototype representation. Preferably, mean aggregation, weighted mean aggregation, or aggregation based on similarity weights can be used; for example, when using similarity-weighted mean, it can be calculated based on each reference representation and window representation vector. Weights are assigned based on the similarity scores to improve the prototype representation after aggregation. Representative of the normal state under the current working conditions.

[0047] Through step S3, this invention can construct a normal baseline that matches the operating conditions under different equipment types, load levels, and environmental conditions. By using multi-prototype consistency retrieval, it reduces the impact of single-sample bias on anomaly identification, thereby reducing misjudgments caused by changes in normal operating conditions and improving the consistency of health assessment results.

[0048] S4. Residual characterization calculation and robust standardization of outlier standard scores Specifically, this includes: the window representation vector obtained in step S3. and prototype representation Calculate the anomaly score for the current time window. ; and then combine with the target bucket Or the set of historical normal residuals corresponding to the backoff bucket Regarding the abnormal scores Robust standardization is performed to obtain outlier standard scores. .

[0049] The abnormal score satisfy:

[0050] in, This is the cosine similarity function.

[0051] The abnormal standard score satisfy:

[0052] in, For the historical normal residual set the median of For the historical normal residual set The absolute deviation of the median, It is a stable term.

[0053] Furthermore, the historical normal residual set From the corresponding bucket The internal review and closed-loop confirmation process identifies the normal window sample residual composition, which can be dynamically expanded or cleaned up as the closed-loop is updated to ensure the accuracy of outlier standard scores. The statistical benchmarks are clearly derived, reproducible, and continuously adaptable.

[0054] Through step S4, this invention can uniformly map the degree of deviation in representation under different time windows, different devices, and different operating conditions to comparable anomaly standard scores. This makes the abnormal threshold more calibrable and transferable, which is conducive to maintaining relatively consistent alarm criteria across devices, operating conditions and time periods.

[0055] S5. Inference of Health Evidence and Two-Indicator Triage Decision Specifically, this includes: the window representation vector obtained in step S2. and the abnormal standard score obtained in step S4 Construct health inference inputs, and infer the network output health level probability and uncertainty based on evidence. Subsequently, the abnormal standard scores were... As an indicator of the degree of anomaly, uncertainty As a credibility indicator, a dual-indicator split is implemented to output health conclusions, alerts, alert conclusions, or review / retest recommendations.

[0056] Furthermore, the evidence inference network uses a window to represent vectors. with abnormal standard scores The combined input vector serves as the input, and the output is the non-negative evidence value for each health level. In one embodiment, the first Class Health Level Probability and uncertainty They can be satisfied respectively:

[0057]

[0058] in, For health level, The larger the value, the more uncertain the model is.

[0059] Furthermore, when window-level health labels exist, the evidence inference network can be trained under supervision using cross-entropy loss, focus loss, or classification loss with calibration constraints; when labels are missing, the uncertainty threshold and health probability output can be quantized based on the window set that has been verified to be normal through closed-loop testing, so that the source of the threshold is clear and reproducible.

[0060] Furthermore, the dual-indicator diversion includes at least the following rules: when and When, output the alarm conclusion; when and When, trigger a review or retest; when and At that time, output an alert of concern or a conclusion to continue observation; among them, This is the threshold for triggering abnormal events. This represents the uncertainty threshold. Optionally, to improve timing stability, the triggering condition may further include a continuous triggering constraint:

[0061] in, To calculate the window length, The threshold for the number of consecutive triggers. For indicator functions, For the first The abnormal standard score corresponding to each time step.

[0062] Step S5 avoids rigidly judging health status based on a single threshold. Instead, it comprehensively considers the degree of anomaly and the reliability of the model, forming a more reasonable triage decision among "direct alarm", "continued observation" and "manual review / retest", thereby improving the pertinence and reliability of operation and maintenance.

[0063] S6, Review / Retest Trigger, Access-Based Closed-Loop Update and Evidence Package Generation Specifically, this includes: when the dual-indicator splitting in step S5 meets the triggering conditions for review or retesting, generating a review or retesting task and outputting the corresponding evidence package; after obtaining the review or retesting conclusion, applying the closed-loop conclusion to the normal reference library and the historical normal residual set. An admission-based closed-loop update is performed on the abnormal threshold and inference model.

[0064] Furthermore, the corresponding window representation vector is only allowed to be used if the sample corresponding to the current time window is confirmed to be normal after review or retesting, and simultaneously meets the quality constraints and missing rate constraints, and no fault closure event occurs within the preset observation period. Write to target bucket Alternatively, it can roll back to the normal reference library of the bucket; at the same time, the corresponding residuals can be written to the bucket. Historical normal residual set The anomaly detection baseline is updated; if the review or retest results in an anomaly, the corresponding sample is written into the anomaly sample pool for subsequent threshold recalibration and evidence inference network retraining.

[0065] Furthermore, this can be based on the updated historical normal residual set. The threshold for abnormal standard scores can be recalibrated; for example, a quantile update method or a smooth update method can be used to dynamically adjust the threshold as the distribution of normal samples evolves.

[0066] Furthermore, to improve the auditability of the conclusions, key channels and key time periods can be determined based on occlusion sensitivity when generating the evidence package: For each channel, the corresponding channel input is set to 0 and the anomaly standard score is recalculated to obtain the channel contribution; for each time period, the window is divided into several segments according to time, and the anomaly standard score is recalculated after occlusion of each segment to determine the key time period with the greatest impact on the anomaly results. The channel contribution can satisfy:

[0067] in, For the first The contribution of each channel To cover the first Abnormal standard scores after each channel The abnormal standard score before occlusion.

[0068] The evidence package includes at least the device identifier, time window start and end dates, bucket identifier, missing rate, quality statistics, window representation summary, prototype representation summary, and anomaly standard score. Uncertainty And the trigger reason marker.

[0069] Step S6 enables timely manual review or on-site retesting when the model is uncertain or significantly abnormal. The access mechanism of "normal conclusion + quality constraint + missing rate constraint + no fault during the observation period" reduces the risk of latent fault samples contaminating the normal reference library. At the same time, the output of the evidence package enhances the interpretability, traceability and auditability of the health conclusion.

[0070] S7. Results Output and Online Inference Deployment Specifically, this includes: outputting health conclusions and abnormal standard scores for each device at each time window. Uncertainty The system includes evidence packages and suggested actions; the suggested actions must include at least one of the following: normal operation, continued observation, monitoring alarms, review suggestion, retest suggestion, and maintenance suggestion. During online inference, the system executes steps S1 to S6 sequentially: first, it collects online monitoring data and completes time window alignment, missing mask matrix generation, and quality matrix generation; second, it performs blind spot self-supervised representation inference to obtain the window representation vector. Then, perform multi-prototype consistency retrieval in the target bucket or fallback bucket to obtain the prototype representation. Then calculate the abnormal standard scores. and uncertainty The system performs dual-indicator triage; finally, it generates review / retest tasks and evidence packages when needed, and performs admission-based closed-loop updates after obtaining closed-loop conclusions.

[0071] In one embodiment, the system can be deployed in a station-side edge computing device, a master station server, or a cloud-edge collaborative platform. The station-side device is responsible for basic data acquisition, window alignment, and local inference, while the master station or cloud platform is responsible for model training, reference library management, cross-site calibration, and global threshold maintenance, thereby meeting the real-time performance and computing power requirements under different deployment conditions. This implementation does not change the technical essence of the present invention: "multi-prototype consistency + dual-index diversion + admission-based closed-loop update."

[0072] Step S7 enables the formation of a complete online operation chain from data acquisition, representation learning, operating condition reference, anomaly detection, reliable traffic diversion to closed-loop update, giving the system continuous self-adaptation capabilities during long-term operation and enabling more stable, reliable and auditable health status assessment under conditions of missing data, noise, drift and unknown anomalies.

[0073] Corresponding to the above-described method implementation, the present invention also provides an artificial intelligence integrated system for the health status of power equipment. The system includes: a window alignment module for aligning multi-channel online monitoring data into time windows to generate a window sequence, a missing mask matrix, and a quality matrix; and a blind spot representation module for inputting the window sequence, the missing mask matrix, and the quality matrix into a self-supervised representation model based on blind spot mask reconstruction to obtain a window representation vector. The operating condition reference module is used to determine the target bucket in the normal reference library according to the equipment type and operating condition label, and to retrieve the previous bucket from the target bucket or the rollback bucket. A normal reference characterization Based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation The residual exception module is used to determine the window representation vector. With the prototype representation The residuals and combined with the bucket Historical normal residual set Calculate the abnormal standard score The evidence inference module is used to infer evidence based on the window representation vector. and the aforementioned abnormal standard score Output uncertainty and execute based on the aforementioned abnormal standard score. As an indicator of the degree of anomaly, with the aforementioned uncertainty The system employs a dual-indicator splitting mechanism for credibility metrics; a review and closed-loop module initiates review or retesting upon meeting trigger conditions, and performs admission-based closed-loop updates based on the review or retesting conclusions. Only window representation vectors that meet the normal conclusion and satisfy quality constraints, missing rate constraints, and no-fault constraints during the observation period are written into the normal reference library; the result output module outputs the health conclusion and the abnormal standard score. Uncertainty And evidence package.

[0074] Through the specific embodiments described above, this invention improves the consistency of processing irregular sampling, missing, and noisy data by unifying time window alignment and explicit expression of missing data; enhances robust representation capabilities under conditions of scarce fault labeling through self-supervised representation learning via blind spot mask reconstruction; constructs a normal baseline related to operating conditions and reduces single-sample bias by using a working condition-based bucketed normal reference library and multi-prototype consistency retrieval; improves the reliability of anomaly detection and decision-making by robust standardization of representation residuals and dual-index diversion; and enhances the system's adaptability, interpretability, and auditability during long-term online operation through admission-based closed-loop updates and evidence package output.

[0075] Comparison Experiment Embodiments of the present invention This embodiment is used to verify the effectiveness of the artificial intelligence integration method for power equipment health status described in this invention in assessing health status under complex operating conditions, data gaps, and long-term online operation conditions.

[0076] The experiment involved 186 transformers, GIS equipment, and switchgear in a substation scenario, from which multi-channel online monitoring data were continuously collected for six months. The multi-channel online monitoring data included one or more of the following: load current, ambient temperature, ambient humidity, partial discharge amplitude, root mean square vibration value, SF6 moisture content, and key gas characteristic quantities in oil chromatography. The collected raw data was divided into preset time windows, resulting in 52,800 time window samples, of which 49,920 were normal samples and 2,880 were abnormal samples. Irregular sampling, local missing data, and quality fluctuations were retained in the raw samples, with an average missing rate of 12.4%, and consecutive missing segments accounting for 31.7% of all missing data.

[0077] The method of this invention is used to process the above-mentioned time window samples, specifically including: firstly, performing unified time window alignment on the multi-channel online monitoring data to generate a window sequence, a missing mask matrix, and a quality matrix; then, inputting the window sequence, missing mask matrix, and quality matrix into a self-supervised representation model based on blind spot mask reconstruction to obtain a window representation vector; and finally, establishing a normal reference library according to equipment type and operating condition labels, and retrieving previous data from the target bucket or backtracking bucket. A normal reference representation is used to form a prototype representation; then, the abnormal standard score is calculated based on the residual between the window representation vector and the prototype representation and the historical normal residual set of the corresponding bucket; then, a dual-indicator diversion is performed based on the abnormal standard score and uncertainty; when the triggering condition is met, a review or retest is initiated, and an admission closed-loop update is performed based on the review or retest conclusion; finally, a health conclusion, abnormal standard score, uncertainty and evidence package are output.

[0078] The experiment used accuracy, precision, recall, F1 score, false alarm rate, false negative rate, average early warning time, verification hit rate, and the fluctuation range of the false alarm rate within 30 days after the closed-loop update as evaluation indicators. The average early warning time was defined as the average time difference between the system's first output of an anomaly alarm and the actual confirmation of the fault; the verification hit rate was defined as the percentage of samples that triggered verification or retesting, which were subsequently confirmed to be abnormal or requiring maintenance; and the fluctuation range of the false alarm rate within 30 days after the closed-loop update was defined as the difference between the maximum and minimum daily false alarm rates within 30 consecutive days after the closed-loop update.

[0079] Comparative Example 1 Comparative Example 1 employs a manual experience-based threshold method. Specifically, fixed alarm thresholds are set for each monitoring channel. When any critical monitoring channel exceeds its corresponding threshold, the device corresponding to the current time window is directly determined to be in an abnormal state. This method does not model the time-series correlation between multiple channels, nor does it introduce operating condition binning, normal reference library, self-supervised representation, uncertainty diversion, or closed-loop update mechanisms.

[0080] Comparative Example 2 Comparative Example 2 employs a traditional supervised classification model. Specifically, it extracts statistical features from the data of each channel within the time window, including mean, extreme values, variance, skewness, kurtosis, and rate of change, and uses a random forest classifier to output normal / abnormal classification results. This method does not explicitly introduce a missing mask matrix and a quality matrix, does not use blind spot masks to reconstruct self-supervised representations, and does not employ a working condition-based bucketed normal reference library, multi-prototype consistency retrieval, dual-index splitting, or admission-based closed-loop update mechanisms.

[0081] Comparative Example 3 Comparative Example 3 employs a single-prototype anomaly detection method. Specifically, the data after time window alignment is encoded, the nearest-neighbor normal reference representation is retrieved from the global normal sample library, and an anomaly score is output based on the residual between the current window representation and the nearest-neighbor normal reference representation. Although this method introduces reference representation comparison, it does not distinguish between the target bucket and the backtracking bucket, and does not use the previous... A normal reference representation forms a prototype representation, without outputting uncertainty, nor is it subject to verification or retesting triggers or admission-based closed-loop updates.

[0082] Table 1 Comparison of overall detection performance of different methods

[0083] Based on Table 1, the following diagram was drawn. Figure 2 The diagram showing the overall detection performance comparison is as follows. Figure 2 In Figure 2 In the figure, the letter X represents the horizontal axis, and A, B, C, and D in the horizontal axis represent Comparative Example 1, Comparative Example 2, Comparative Example 3, and the embodiment of the present invention, respectively. The letter Y represents the vertical axis, and the letter L represents the legend. In the legend, the letters E, F, G, and H represent accuracy, precision, recall, and F1 score, respectively. The F1 score is a combined evaluation metric for precision and recall, commonly used in classification tasks, and particularly suitable for imbalanced datasets. Its formula is:

[0084] in, For accuracy, This refers to the recall rate.

[0085] From Table 1 Figure 2 As shown, the embodiments of the present invention outperform comparative examples 1 to 3 in terms of accuracy, precision, recall, and F1 score, and significantly reduce both false positive and false negative rates. Compared to comparative example 1, the false positive rate of the embodiments of the present invention decreased from 9.8% to 2.9%, and the false negative rate decreased from 27.6% to 9.2%. Compared to comparative example 3, the false positive rate of the embodiments of the present invention further decreased by 2.3 percentage points, indicating that the present invention, through the dual-indicator diversion of multi-prototype consistency, anomaly standard score, and uncertainty, can more effectively distinguish between operating condition fluctuations and true anomalies.

[0086] Table 2 Comparison of robustness under missing and noisy conditions To further verify the robustness of the invention under missing and noisy scenarios, 10%, 20%, and 30% random missing values ​​were injected into the test set, respectively, and Gaussian noise with a mean of 0 and a standard deviation of 3% of the original range was superimposed. The experimental results are shown in Table 2.

[0087]

[0088] Based on Table 2, the diagram is as follows: Figure 3 The diagram showing the comparison of robustness of different methods under missing and noisy conditions is as follows: Figure 3 As shown, the letter X represents the horizontal axis, and E, F, and G on the horizontal axis represent the 10% missing and noise scenarios, the 20% missing and noise scenarios, and the 30% missing and noise scenarios, respectively. The letter Y represents the vertical axis, and the letter L represents the legend. The letters B, C, and D in the legend represent Comparative Example 2, Comparative Example 3, and the embodiment of the present invention, respectively. From Table 2 and Figure 3 It can be seen that, under the condition of increased missing rate and noise interference, the performance degradation of the embodiment of the present invention is significantly smaller than that of comparative example 2 and comparative example 3, indicating that the present invention can improve the adaptability to missing and noisy data by reconstructing self-supervised representation through missing mask matrix, quality matrix and blind spot mask.

[0089] Table 3. False alarm suppression effect under changing operating conditions. To verify the false alarm suppression capability of the present invention under varying operating conditions, the test samples were divided into three scenarios according to load and environmental conditions: low load / normal temperature, medium load / high temperature, and high load / high temperature. The experimental results are shown in Table 3.

[0090]

[0091] Based on Table 3, the diagram is as follows: Figure 4 The diagram shows a comparison of false alarm rates for different methods under varying operating conditions. Figure 4 In the figure, the letter X represents the horizontal axis, and H, I, and J in the horizontal axis represent low load / normal temperature scenario, medium load / high temperature scenario, and high load / high temperature scenario, respectively. The letter Y represents the vertical axis, and the letter L represents the legend. The letters A, C, and D in the legend represent comparative example 1, comparative example 3, and the embodiment of the present invention, respectively.

[0092] From Table 3 and Figure 4 It can be seen that in scenarios with large changes in operating conditions such as high load and high temperature, the embodiments of the present invention still maintain a low false alarm rate, indicating that the present invention can effectively suppress misjudgments caused by fluctuations in normal operating conditions by establishing a normal reference library according to equipment type and operating condition label and performing multi-prototype consistency retrieval in the target bucket or rollback bucket.

[0093] Table 4 Comparison of Dual-Indicator Triage and Review Hit Rate To verify the effectiveness of the dual-index traffic diversion mechanism of the present invention, the performance of the embodiments of the present invention and Comparative Example 3 in alarm diversion was compared, and the results are shown in Table 4.

[0094]

[0095] The review hit rate is the percentage of samples that were ultimately confirmed to have anomalies or require maintenance and handling among the samples that triggered the review.

[0096] Based on Table 4, draw the following diagram: Figure 5 The diagram shows a comparison of the results of the two-index diversion method using different approaches. Figure 5 The letter X represents the horizontal axis, with C and D representing Comparative Example 3 and the embodiment of the present invention, respectively. The letter Y represents the vertical axis, and the letter L represents the legend. In the legend, the letters E, F, and G represent the number of direct alarm samples, the number of samples triggered for review, and the number of alarm samples requiring attention, respectively. From Table 4 and Figure 5 As can be seen, the embodiments of the present invention perform dual-indicator diversion by using the abnormal standard score as an indicator of the degree of abnormality and uncertainty as an indicator of credibility. This prevents some highly abnormal but low-credibility samples from being directly judged as alarms, but instead enters the review process. The review hit rate reaches 72.7%, indicating that the present invention can improve the efficiency of review resource utilization and reduce invalid handling.

[0097] Table 5 Comparison of Early Warning Capability and Closed-Loop Stability To verify the effectiveness of the present invention in terms of early warning capability and long-term closed-loop operation stability, the embodiments of the present invention were compared with Comparative Examples 2 and 3, and the results are shown in Table 5.

[0098]

[0099] Based on Table 5, draw the following diagram: Figure 6 The diagram shows a comparison of the early warning capabilities and closed-loop stability of different methods. Figure 6 In the figure, the letter X represents the horizontal axis, and B, C, and D in the horizontal axis represent comparative example 2, comparative example 3, and the embodiment of the present invention, respectively. The letter Y represents the left vertical axis, the letter Z represents the right vertical axis, and the letter L represents the legend. The letters E and F in the legend represent the average early warning time and the fluctuation range of the false alarm rate, respectively.

[0100] From Table 5 and Figure 6It can be seen that the embodiments of the present invention can detect abnormal trends of equipment earlier, with an average early warning time of 23.6 hours. At the same time, the false alarm rate fluctuation within 30 days after the closed-loop update is only 1.1%, which is significantly lower than that of comparative examples 2 and 3. This indicates that the present invention, through the admission-based closed-loop update mechanism that "only allows writing to the normal reference library when the conclusion is normal and meets the quality constraints, missing rate constraints, and no fault constraints during the observation period", can reduce the risk of latent fault samples contaminating the normal baseline and improve the long-term online operation stability.

[0101] Experimental Results Analysis As can be seen from the comparison results between the embodiments of the present invention and Comparative Examples 1 to 3, the present invention has at least the following technical effects: This invention improves the system's adaptability to scenarios involving irregular sampling, missing data, and quality fluctuations by using time window alignment, joint input of the missing mask matrix, and quality matrix; enhances robust representation capabilities under insufficient fault labeling conditions through blind spot mask reconstruction and self-supervised representation learning; reduces false alarms caused by operating condition fluctuations through operating condition binning of the normal reference library and multi-prototype consistency retrieval; improves the verification hit rate and the rationality of operation and maintenance decisions through dual-indicator diversion of anomaly standard scores and uncertainties; and improves the purity and false alarm rate stability of the normal reference library during long-term online operation through an admission-based closed-loop update mechanism.

[0102] Therefore, the above experimental results demonstrate that the present invention can achieve a more stable, reliable, and suitable long-term online power equipment health status assessment under conditions of absence, noise, and changing operating conditions, thereby verifying the beneficial effects described in the specification of the present invention.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements or combinations 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 integrating artificial intelligence into the health status of power equipment, characterized in that, include: Time window alignment is performed on multi-channel online monitoring data to generate window sequences, missing mask matrices, and quality matrices. The window sequence, the missing mask matrix, and the quality matrix are input into a self-supervised representation model based on blind spot mask reconstruction to obtain the window representation vector. ; Determine the target bucket in the normal reference library according to equipment type and operating condition label, and search the previous bucket or rollback bucket. A normal reference characterization Based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation According to the window representation vector With the prototype representation The residuals and combined with the bucket Historical normal residual set Calculate the abnormal standard score Based on the window representation vector and the aforementioned abnormal standard score Output uncertainty and with the aforementioned abnormal standard score As an indicator of the degree of anomaly, the aforementioned uncertainty The system performs dual-indicator traffic splitting as a credibility indicator; it initiates a review or retest when trigger conditions are met, and performs an admission-based closed-loop update based on the review or retest conclusions. Only window representation vectors that meet the normal conclusion and satisfy quality constraints, missing rate constraints, and no-fault constraints during the observation period are written into the normal reference library; the system outputs a health conclusion and an anomaly standard score. Uncertainty And evidence package.

2. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The time window alignment includes setting the window length. Resampling interval And the window step size, and determine the number of time steps within the window. satisfy: 。 3. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The self-supervised representation model based on blind spot mask reconstruction constructs a blind spot indication matrix. Perform training or inference, and then use the normalized window sequence. Missing mask matrix quality matrix and blind spot indicator matrix splicing to form model input : in, Indicates splicing.

4. The method for integrating artificial intelligence into the health status of power equipment according to claim 3, characterized in that, The self-supervised representation model only exists in the blind spot set. Location calculation reconstruction loss : in, Number of blind spots For time steps ,aisle Reconstruction value at the location, This corresponds to the true normalized value.

5. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The target bucket is determined by the device type, load label, and environment label. When the number of normal reference samples in the target bucket is less than the preset empty bucket threshold, the search is performed by switching to the global bucket of the same device type or the global bucket of the entire site as the fallback bucket according to the preset fallback order. When the number of normal reference samples in both the target bucket and the fallback bucket is less than the preset empty bucket threshold, the output of a deterministic health conclusion is prohibited and a review or retest suggestion is output.

6. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The basis A normal reference characterization With the window representation vector Consistency forms prototype representation This includes: selecting the top representatives with the smallest distance from the reference representation set based on a distance function. A normal reference characterization and the former A normal reference characterization Aggregation is performed to obtain prototype characterization The distance function At least one of the following: or in, The cosine similarity function is used. The distance is Euclidean.

7. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The abnormal standard score The calculation includes: first, based on the window representation vector With prototype representation Calculate abnormal scores Combined with bucket Historical normal residual set For the abnormal scores Perform robust standardization: in, The median, This represents the absolute deviation of the median. It is a stable term.

8. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The dual-index diversion must at least satisfy the following rule: when and Output alarm conclusions in real time; when and Trigger review or retest at any time; when and Output alerts or continue monitoring conclusions in a timely manner; among them, This is the threshold for triggering abnormal events. This represents the uncertainty threshold.

9. The method for integrating artificial intelligence into the health status of power equipment according to claim 1, characterized in that, The admission-based closed-loop update includes: allowing the corresponding window representation vector to be updated only when the sample corresponding to the current window is confirmed to be normal after review or retesting, meets quality constraints and missing rate constraints, and no fault closed-loop event occurs within the preset observation period. Write the normal reference library into the target bucket or backoff bucket; when the review or retest conclusion is abnormal, write the corresponding sample into the abnormal sample pool for subsequent threshold recalibration and evidence inference model retraining.

10. An artificial intelligence-integrated system for the health status of power equipment, characterized in that, include: The window alignment module is used to align time windows in multi-channel online monitoring data and generate window sequences, missing mask matrices, and quality matrices. The blind spot representation module is used to input the window sequence, the missing mask matrix, and the quality matrix into a self-supervised representation model based on blind spot mask reconstruction to obtain the window representation vector. ; The operating condition reference module is used to determine the target bucket in the normal reference library according to the equipment type and operating condition label, and to retrieve the previous bucket from the target bucket or the rollback bucket. A normal reference characterization Based on the aforementioned A normal reference characterization With the window representation vector Consistency forms prototype representation The residual exception module is used to determine the window representation vector. With the prototype representation The residuals and combined with the bucket Historical normal residual set Calculate the abnormal standard score ; The evidence inference module is used to infer evidence based on the window representation vector. and the aforementioned abnormal standard score Output uncertainty and execute based on the aforementioned abnormal standard score. As an indicator of the degree of anomaly, with the aforementioned uncertainty The system employs a dual-indicator splitting mechanism for credibility metrics; a review and closed-loop module initiates review or retesting upon meeting trigger conditions, and performs admission-based closed-loop updates based on the review or retesting conclusions. Only window representation vectors that meet the normal conclusion and satisfy quality constraints, missing rate constraints, and no-fault constraints during the observation period are written into the normal reference library; the result output module outputs the health conclusion and the abnormal standard score. Uncertainty And evidence package.

Images