A power equipment-oriented multi-source perception anomaly identification diagnosis and reliability closed-loop evaluation method

Abstract

The application discloses a kind of multi-source perception abnormality identification diagnosis and reliability closed-loop evaluation methods for power equipment, comprising: the management and millisecond level time alignment of offline inspection data and online monitoring data are carried out, unified time index is constructed and alignment quality score is output;Based on equipment-component index, structured data repository and traceable sample slice are constructed;Online time series signal and offline inspection evidence are jointly represented using spatiotemporal fusion encoder;Through the output of early warning level, failure probability, suspected component and evidence fragment of modularized abnormality identification diagnosis model;Uncertainty estimation and probability calibration are carried out to the diagnosis result, and generate credibility score and trigger re-investigation or artificial review when low credibility;The diagnosis result and credibility score are mapped to reliability level, and the mapping weight and model threshold are dynamically updated according to the maintenance feedback.The application can improve the multi-source data fusion quality, abnormal diagnosis stability and the closed-loop adaptive ability of reliability evaluation.

Description

Technical Field

[0001] This invention relates to a method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment, belonging to the field of power equipment condition monitoring technology. Background Technology

[0002] Substations, power plants, power transmission and distribution channels, and critical electrical equipment are susceptible to problems such as abnormal temperature rise, partial discharge, abnormal vibration, appearance defects, or abnormal switch operation during long-term operation due to factors such as load fluctuations, environmental changes, component aging, insulation degradation, and mechanical wear. To detect these anomalies promptly, existing operation and maintenance systems typically employ both online monitoring systems and offline inspection methods. Online monitoring systems continuously collect time-series signals such as current, voltage, temperature, partial discharge, and vibration; offline inspection methods utilize inspection robots, drones, and handheld testing devices to acquire evidence such as visible light, infrared, acoustic, partial discharge, and positional signals.

[0003] However, existing technologies typically handle online monitoring and early warning, offline defect identification, and reliability assessment separately. Online threshold-based early warning or single-source model-based early warning relies on a single or limited number of measurement points, making it sensitive to changes in operating conditions, sensor noise, and individual equipment differences. Offline image or acoustic recognition can detect localized defects, but it is difficult to analyze them simultaneously with online time-series anomalies within the same event window. Reliability assessments often rely on human experience, static scoring tables, or historical fault statistics, making it difficult to dynamically update based on real-time early warning results, reliability, and maintenance feedback.

[0004] Therefore, existing technologies have at least the following shortcomings: offline inspection data and online monitoring data have inconsistent formats, asynchronous times, and large differences in noise, making it difficult to construct fusion samples; the accuracy of anomaly identification is unstable under multiple operating conditions and different equipment types; the anomaly diagnosis conclusions lack credible quantification and re-evidence mechanisms, making it difficult to directly support maintenance decisions; the early warning results are difficult to map into executable reliability levels and maintenance priorities, and there is also a lack of a closed-loop mechanism for continuous updates based on actual maintenance results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for anomaly identification and reliability assessment of power equipment, which addresses the problems in the prior art such as the difficulty in unifying and integrating offline and online data, insufficient generalization ability of anomaly identification, lack of reliable quantification of diagnostic conclusions, and lack of closed-loop updates in reliability assessment.

[0006] To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution: A method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment, the method comprising: S1: Acquire offline inspection data and online monitoring data of power equipment, standardize the format of the offline inspection data, remove outliers, complete metadata and perform pose calibration, filter and reduce noise, repair missing data, calibrate timestamps and label operating conditions of the online monitoring data, and generate a unified time index and alignment quality score based on equipment identifier, component identifier, data mode and acquisition time. S2: Based on the unified time index and aligned quality score, construct a structured data repository oriented towards equipment and components, generate traceable sample slices according to the working condition window, event window or risk window, and use a spatiotemporal fusion encoder to jointly represent online time series signals and offline inspection evidence to obtain fused feature vectors; S3: Input the fused feature vector into the modular anomaly identification and diagnosis model, and output the warning level, fault probability, suspected components, fault cause chain and related evidence fragments through a three-level pipeline of feature extraction, early warning reasoning and diagnostic analysis. S4: Perform uncertainty estimation, probability calibration and multi-source evidence consistency verification on the warning level, failure probability, suspected components and related evidence fragments, generate a credibility score, and trigger a re-evidence collection or manual review process when the credibility score is lower than a preset threshold or when there is a conflict in the multi-source evidence. S5: Map the warning level, failure probability, suspected components, related evidence fragments and credibility score to reliability risk score and reliability level, and dynamically update the reliability level mapping weight, diagnostic model threshold and re-evidence threshold based on maintenance feedback, false alarm statistics and missed alarm statistics.

[0007] Optionally, the offline inspection data includes at least one of visible light images, infrared images, acoustic data, partial discharge detection data, and inspection equipment pose data; the online monitoring data includes at least one of current, voltage, temperature, vibration, partial discharge, gas concentration, and switch status data.

[0008] Optionally, the generation of the unified time index and alignment quality score specifically includes: Estimate the time offset between the offline inspection data collection time and the online monitoring data stream, resample, interpolate or window aggregate data at different sampling frequencies, and calculate the alignment quality score based on the time offset residual, missing rate, noise level and metadata integrity.

[0009] Optionally, the equipment-component structured data repository includes at least an equipment ledger index, a component topology index, a data modality index, a working condition label index, a time window index, and a quality identifier index; the traceable sample slice includes at least online time-series segments, offline evidence segments, equipment identifiers, component identifiers, working condition labels, aligned quality scores, and feedback labels.

[0010] Optionally, the spatiotemporal fusion encoder includes a temporal coding unit for encoding the dynamic evolution characteristics of online temporal signals, and a spatial topology coding unit for encoding the spatial relationships and component topology relationships of offline inspection evidence; the outputs of the temporal coding unit and the spatial topology coding unit are combined into a fusion feature vector after attention fusion, gating fusion or splicing mapping.

[0011] Optionally, the modular anomaly identification and diagnostic model includes: The feature extraction module is used to extract key frequency bands, temperature anomalies, partial discharge anomalies, acoustic anomalies, or image defect features. The early warning reasoning module is used to output the early warning level and fault probability based on the spatiotemporal characteristics of the operating conditions; The diagnostic analysis module is used to locate suspected components and form a fault cause chain based on the fault correlation factor attention mechanism.

[0012] Optionally, the associated evidence fragment includes the online time series interval that triggers the early warning, the offline inspection image or acoustic fragment aligned with the online time series interval, the corresponding component topology location, and the fault association factor weights that affect the diagnostic conclusion; the associated evidence fragment is used to support manual review, generation of re-evidence collection tasks, and maintenance feedback backfilling.

[0013] Optionally, the generation of the credibility score specifically includes: A comprehensive credibility score is constructed based on the entropy of the output probability distribution of the modular anomaly identification and diagnosis model, the model integration variance, the calibration error, the alignment quality score, and the consistency of multi-source evidence. When the credibility score is lower than a first threshold, a re-evidence collection task is generated to supplement infrared images, acoustic data, partial discharge data, or retest key measurement points. When the credibility score is lower than a second threshold or the evidence conflict exceeds a preset range, a manual review task is generated.

[0014] Optionally, the reliability risk score is obtained by weighting the failure probability, anomaly severity, failure frequency, runtime, maintenance cost, and credibility score; the reliability level is mapped to at least two levels based on the reliability risk score and is used to determine maintenance priority, inspection frequency, or operation and maintenance strategy.

[0015] Optionally, the dynamic closed-loop update specifically includes: The actual inspection results are compared with the diagnostic results output in step S3. The actual fault, false alarm and missed alarm samples are statistically analyzed. Based on the statistical results, the reliability level mapping weight, early warning threshold, credibility threshold and sample slice quality gating parameters are adjusted, and the updated parameters are synchronized to the structured data repository and modular anomaly identification and diagnosis model.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: First, by implementing offline-online multi-source data governance, unified time indexing, and alignment quality scoring, we can improve the quality of data fusion and reduce the impact of time drift, missing data, and noise differences on anomaly identification.

[0017] Second, by using a structured data repository for equipment and components, sample slices, and a spatiotemporal fusion encoder, online time-series anomalies and offline inspection evidence can be correlated and analyzed within the same equipment, the same component, and the same event window, thereby improving the stability and traceability of anomaly identification.

[0018] Third, through a modular model architecture for feature extraction, early warning reasoning, and diagnostic analysis, as well as a fault correlation factor attention mechanism, the model can output early warning levels, fault probabilities, suspected components, and evidence fragments, thereby enhancing the interpretability of diagnostic conclusions.

[0019] Fourth, through uncertainty estimation, probability calibration and credibility threshold triggering mechanisms, re-evidence or manual review tasks are automatically generated for low-credibility results, reducing the impact of false alarms and missed alarms on maintenance decisions.

[0020] Fifth, through reliability risk scoring, reliability level mapping, and closed-loop updates of maintenance feedback, the results of abnormal early warnings can be transformed into maintenance priorities, and the diagnostic thresholds, credibility thresholds, and indicator weights can be continuously optimized. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the offline-online multi-source sensing data governance and millisecond-level alignment process in an embodiment of the present invention. Figure 2 This is a flowchart of the data spatiotemporal fusion encoder and structured storage in an embodiment of the present invention; Figure 3 This is a flowchart of the anomaly identification and diagnosis model in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the uncertainty / credibility assessment and re-forensics triggering process in an embodiment of the present invention. Figure 5 This is a flowchart of reliability level quantification and closed-loop weight update in an embodiment of the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0023] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0024] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances. Example

[0025] This embodiment uses substation circuit breakers, main transformers, or GIS equipment as targets. On the online side, it collects time-series signals such as current, voltage, temperature, vibration, partial discharge, gas concentration, and switching status. On the offline side, inspection robots, drones, or handheld detection devices collect visible light images, infrared images, acoustic data, partial discharge detection data, and pose information. The system first performs multi-source data processing and time-series alignment, then constructs structured samples of equipment and components, followed by anomaly identification and diagnosis, credibility assessment, and reliability level mapping. Maintenance feedback is then fed back into the sample library and model parameters. Figures 1-5 As shown, a method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment is disclosed. The method includes: S1: Acquire offline inspection data and online monitoring data of power equipment, standardize the format of the offline inspection data, remove outliers, complete metadata and perform pose calibration, filter and reduce noise, repair missing data, calibrate timestamps and label operating conditions of the online monitoring data, and generate a unified time index and alignment quality score based on equipment identifier, component identifier, data mode and acquisition time. S2: Based on the unified time index and aligned quality score, construct a structured data repository oriented towards equipment and components, generate traceable sample slices according to the working condition window, event window or risk window, and use a spatiotemporal fusion encoder to jointly represent online time series signals and offline inspection evidence to obtain fused feature vectors; S3: Input the fused feature vector into the modular anomaly identification and diagnosis model, and output the warning level, fault probability, suspected components, fault cause chain and related evidence fragments through a three-level pipeline of feature extraction, early warning reasoning and diagnostic analysis. S4: Perform uncertainty estimation, probability calibration and multi-source evidence consistency verification on the warning level, failure probability, suspected components and related evidence fragments, generate a credibility score, and trigger a re-evidence collection or manual review process when the credibility score is lower than a preset threshold or when there is a conflict in the multi-source evidence. S5: Map the warning level, failure probability, suspected components, related evidence fragments and credibility score to reliability risk score and reliability level, and dynamically update the reliability level mapping weight, diagnostic model threshold and re-evidence threshold based on maintenance feedback, false alarm statistics and missed alarm statistics.

[0026] In step S1, each data source is converted into a unified record format, which includes at least device identifier, component identifier, acquisition time, data mode, feature value, quality identifier, and source identifier. For offline inspection data, outliers can be removed using statistical thresholds, isolated forests, local outliers, or manual rules, and metadata such as inspection device number, shooting angle, measurement point location, and ambient temperature and humidity can be supplemented. For online monitoring data, noise can be reduced using sliding filtering, wavelet denoising, Kalman filtering, or median filtering, and short-term missing data can be repaired using interpolation, forward imputation, or model prediction.

[0027] Further, step S1 estimates the time offset based on the offline evidence collection time, online sampling time, device event time, and inspection pose information, and resamples, interpolates, or aggregates the data at different sampling frequencies to generate a unified time index. The alignment quality score can be calculated comprehensively based on the time offset residual, missing rate, noise level, and metadata integrity. Samples with low alignment quality scores can be downweighted, removed, or trigger supplementary sampling.

[0028] In step S2, a structured data repository for equipment and components is established based on the equipment ledger and component topology. The structured data repository includes at least an equipment ledger index, a component topology index, a data modality index, a condition label index, a time window index, and a quality identifier index. Sample slices can be generated according to time windows before and after alarm events, load change time windows, maintenance time windows, or risk increase time windows. Each sample slice contains online time-series segments, offline evidence segments, condition labels, equipment / component identifiers, aligned quality scores, and feedback labels.

[0029] The spatiotemporal fusion encoder in step S2 includes a temporal coding unit and a spatial topology coding unit. The temporal coding unit can use a temporal Transformer, a temporal convolutional network, a recurrent neural network, or a state-space model to extract the dynamic evolution features of the online signal; the spatial topology coding unit can use a convolutional neural network, a visual Transformer, an acoustic feature encoder, a graph neural network, or a knowledge graph to encode offline evidence and component relationships. The outputs of the two coding units are fused into a fused feature vector through attention fusion, gating fusion, or concatenation mapping.

[0030] In step S3, the modular anomaly identification and diagnosis model includes a feature extraction module, an early warning inference module, and a diagnostic analysis module. The feature extraction module extracts key frequency bands, temperature rise patterns, partial discharge waveforms, acoustic anomalies, image defects, and component topological features. The early warning inference module outputs early warning levels such as normal, attention, warning, and severe, along with their corresponding fault probabilities. The diagnostic analysis module locates suspected components based on a fault correlation factor attention mechanism and provides the fault cause chain and related evidence fragments. Related evidence fragments may include the online time series interval that triggered the early warning, the corresponding offline image or acoustic fragment, the component topological location, and the weights of the fault correlation factors.

[0031] In step S4, uncertainty estimation and probability calibration are performed on the model output. The credibility score can be obtained by combining the entropy of the model output probability distribution, the model ensemble variance, the calibration error after temperature scaling, the alignment quality score, and the consistency of multi-source evidence. When the credibility score is lower than the first threshold, the system automatically generates a re-evidence collection task, such as adjusting the inspection angle, supplementing infrared images, supplementing acoustic data, retesting partial discharge, or retesting key online measurement points. When the credibility score is lower than the second threshold or the conflict of multi-source evidence exceeds a preset range, the system generates a manual review task. The supplementary collection, retesting, or review results are used as new samples or new labels to backfill the structured data repository.

[0032] In step S5, the system constructs a reliability indicator system. This indicator system may include failure probability, anomaly severity, failure frequency, runtime, maintenance cost, environmental risk factors, alignment quality score, and credibility score. The system calculates a reliability risk score based on these indicators and maps the reliability risk score to a reliability level. The reliability level can be set to high reliability, medium reliability, low reliability, and pending review, or to Level I, Level II, Level III, and Level IV. Different reliability levels correspond to different maintenance priorities, inspection frequencies, and operation and maintenance strategies.

[0033] Step S5 also includes closed-loop weight updates. The system compares the actual maintenance results with the diagnostic results and statistically analyzes the actual faults, false alarms, and missed alarms. If the missed alarm rate for a certain type of fault increases, the weight of the relevant fault factor is increased or the corresponding warning threshold is decreased; if the false alarm rate for a certain operating condition increases, the reliability mapping weight, credibility threshold, or sample quality gating parameter for the corresponding operating condition is adjusted. The updated weights and thresholds are synchronized to the structured data repository, diagnostic model, and reliability assessment module for subsequent sample slicing, early warning diagnosis, and reliability level mapping. Example

[0034] Without altering the core logic of this invention, the time alignment algorithm can be replaced by dynamic time warping, cross-correlation peak alignment, or event anchor alignment; uncertainty estimation can be replaced by model ensemble, Monte Carlo Dropout, confidence interval estimation, or consistency testing; and reliability mapping can be replaced by rule mapping, logistic regression, gradient boosting trees, Bayesian networks, or learnable scoring functions. These replacement schemes still fall within the overall technical concept of this invention: multi-source governance alignment, spatiotemporal fusion diagnosis, credibility re-verification, and reliability closed-loop update.

[0035] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment, characterized in that, The method includes: S1: Acquire offline inspection data and online monitoring data of power equipment, standardize the format of the offline inspection data, remove outliers, complete metadata and perform pose calibration, filter and reduce noise, repair missing data, calibrate timestamps and label operating conditions of the online monitoring data, and generate a unified time index and alignment quality score based on equipment identifier, component identifier, data mode and acquisition time. S2: Based on the unified time index and aligned quality score, construct a structured data repository oriented towards equipment and components, generate traceable sample slices according to the working condition window, event window or risk window, and use a spatiotemporal fusion encoder to jointly represent online time series signals and offline inspection evidence to obtain fused feature vectors; S3: Input the fused feature vector into the modular anomaly identification and diagnosis model, and output the warning level, fault probability, suspected components, fault cause chain and related evidence fragments through a three-level pipeline of feature extraction, early warning reasoning and diagnostic analysis. S4: Perform uncertainty estimation, probability calibration and multi-source evidence consistency verification on the warning level, failure probability, suspected components and related evidence fragments, generate a credibility score, and trigger a re-evidence collection or manual review process when the credibility score is lower than a preset threshold or when there is a conflict in the multi-source evidence. S5: Map the warning level, failure probability, suspected components, related evidence fragments and credibility score to reliability risk score and reliability level, and dynamically update the reliability level mapping weight, diagnostic model threshold and re-evidence threshold based on maintenance feedback, false alarm statistics and missed alarm statistics.

2. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The offline inspection data includes at least one of visible light images, infrared images, acoustic data, partial discharge detection data, and inspection equipment pose data; the online monitoring data includes at least one of current, voltage, temperature, vibration, partial discharge, gas concentration, and switch status data.

3. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The generation of the unified time index and alignment quality score specifically includes: Estimate the time offset between the offline inspection data collection time and the online monitoring data stream, resample, interpolate or window aggregate data at different sampling frequencies, and calculate the alignment quality score based on the time offset residual, missing rate, noise level and metadata integrity.

4. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The structured data repository for equipment and components includes at least an equipment ledger index, a component topology index, a data modality index, a working condition label index, a time window index, and a quality identifier index; the traceable sample slice includes at least online time-series segments, offline evidence segments, equipment identifiers, component identifiers, working condition labels, aligned quality scores, and feedback labels.

5. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The spatiotemporal fusion encoder includes a temporal coding unit for encoding the dynamic evolution characteristics of online time-series signals, and a spatial topology coding unit for encoding the spatial relationships and component topology relationships of offline inspection evidence; the outputs of the temporal coding unit and the spatial topology coding unit are combined into a fusion feature vector after attention fusion, gated fusion or splicing mapping.

6. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop evaluation of power equipment according to claim 1, characterized in that, The modular anomaly identification and diagnosis model includes: The feature extraction module is used to extract key frequency bands, temperature anomalies, partial discharge anomalies, acoustic anomalies, or image defect features. The early warning reasoning module is used to output the early warning level and fault probability based on the spatiotemporal characteristics of the operating conditions; The diagnostic analysis module is used to locate suspected components and form a fault cause chain based on the fault correlation factor attention mechanism.

7. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The associated evidence fragments include the online time interval that triggers the warning, the offline inspection image or acoustic fragment aligned with the online time interval, the corresponding component topological location, and the fault association factor weights that affect the diagnostic conclusion. The associated evidence fragments are used to support manual review, generation of re-evidence collection tasks, and feedback filling in maintenance.

8. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The generation of the credibility score specifically includes: A comprehensive credibility score is constructed based on the entropy of the output probability distribution of the modular anomaly identification and diagnosis model, the model integration variance, the calibration error, the alignment quality score, and the consistency of multi-source evidence. When the credibility score is lower than a first threshold, a re-evidence collection task is generated to supplement infrared images, acoustic data, partial discharge data, or retest key measurement points. When the credibility score is lower than a second threshold or the evidence conflict exceeds a preset range, a manual review task is generated.

9. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The reliability risk score is obtained by weighting the failure probability, anomaly severity, failure frequency, runtime, maintenance cost, and credibility score; the reliability level is mapped to at least two levels based on the reliability risk score and is used to determine maintenance priority, inspection frequency, or operation and maintenance strategy.

10. The method for multi-source sensing anomaly identification, diagnosis, and reliability closed-loop assessment of power equipment according to claim 1, characterized in that, The dynamic closed-loop update specifically includes: The actual inspection results are compared with the diagnostic results output in step S3. The actual fault, false alarm and missed alarm samples are statistically analyzed. Based on the statistical results, the reliability level mapping weight, early warning threshold, credibility threshold and sample slice quality gating parameters are adjusted, and the updated parameters are synchronized to the structured data repository and modular anomaly identification and diagnosis model.

Images