Comprehensive power supply test method and system integrating fault prediction

Abstract

The invention discloses an integrated power supply test method and system integrating fault prediction, and relates to the technical field of power supply test. A comprehensive power supply test system integrating fault prediction comprises a test parameter acquisition module, a reference state updating module, a data fusion module, a fault prediction module and a retest recording module. According to the method, the first prediction channel and the second prediction channel are arranged, and the model path is flexibly selected in combination with the initial anomaly judgment result, so that the prediction accuracy is improved while the prediction efficiency is ensured, and operation requirements in different complexity scenes are met; by setting a circulation analysis mechanism and a dynamic circulation limit value and based on weight increasing operation of a maximum influence data item, intelligent verification and self-adaptive adjustment of model result consistency are realized, and the stability and robustness of a fault judgment process are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply testing, and in particular to a comprehensive power supply testing method and system with integrated fault prediction. Background Art

[0002] Power supply equipment, as critical infrastructure in modern industrial and information systems, has a direct impact on system stability and security. With the rapid development of intelligent manufacturing, data centers, and communications equipment, power supply systems face increasing operational loads and stability requirements. However, traditional power supply testing and fault diagnosis methods rely heavily on periodic inspections and rule-based anomaly detection, making it difficult to promptly identify potential fault trends or capture multi-factor anomalies in complex scenarios.

[0003] Existing power supply testing methods primarily focus on measuring electrical parameters, such as real-time monitoring of voltage, current, and frequency. Some systems incorporate auxiliary methods like infrared thermal imaging for image recognition after a fault occurs. However, these methods commonly suffer from the following technical deficiencies: single data dimension, lack of deep fusion analysis, inadequate predictive capabilities, weak model generalization, poor adaptability to individual device differences, fragmented knowledge accumulation, and a lack of closed-loop learning mechanisms.

[0004] Therefore, there is an urgent need for a power supply testing system and method that can integrate multimodal data, has adaptive and fault prediction capabilities, and supports knowledge accumulation and dynamic updating, so as to improve the ability to identify and respond to complex faults and achieve the transition from passive response to active prediction. Summary of the Invention

[0005] A comprehensive power supply testing method with integrated fault prediction, comprising: Acquire multimodal data of power supply equipment during power supply testing, including electrical parameter information, operating environment information, thermal imaging information, and operation log information; Based on historical multimodal data, a reference state transition diagram corresponding to the power supply equipment is constructed, including the fault evolution path from "normal state" to "abnormal trend" and then to "fault state", and a fault knowledge graph corresponding to the "fault type-parameter performance-handling suggestion" relationship for each standard fault case; The stable fluctuation period of each data point in the electrical parameter information is calculated, and the weighted average of the stable fluctuation period of each data point is taken as the fusion period. All multimodal data within the fusion period is obtained and fused to obtain the period fusion feature. The multidimensional health status profile of the power supply equipment is obtained through feature engineering processing. The multidimensional health status profile is initially judged for abnormalities, and the results include normal and abnormal. The periodic fusion features are input into the fault prediction model for fault prediction. The fault prediction model selects a prediction channel based on the results of the initial abnormality judgment. The model includes two independent prediction channels. The first prediction channel performs fast analysis based on a lightweight model, and the second prediction channel performs in-depth analysis based on a convolutional neural network. The results of the prediction channels are circulated and analyzed within the fault prediction model to obtain the initial test results.

[0006] As a preferred technical solution of the present invention, a comprehensive power supply testing method with integrated fault prediction further includes: Match the fault knowledge graph based on the initial test results, obtain the re-verification weight based on the corresponding parameter performance, test the power supply equipment again, use the re-verification weight to weight the periodic fusion features to obtain the re-verified periodic fusion features, and use the fault prediction model to predict the fault. The output of the correction result or the final result is determined based on whether the number of flow analysis exceeds the flow limit. If the output result is a correction result, the correction result is used to match the fault knowledge graph, and the re-verification weight is obtained, and the test is performed again until the final result is obtained or the number of re-tests exceeds the set threshold; The re-verification weight is integrated into the personalized device test sample after weighting the periodic fusion feature each time; the personalized device test sample is used for weighting in the data fusion process; By performing cluster analysis on personalized device test samples and constructing personalized device features in the device twin digital model based on the cluster analysis results, the personalized device features are used to correct the fault evolution path.

[0007] As a preferred technical solution of the present invention, the fault prediction model uses the first prediction channel for analysis when the initial abnormality judgment result is normal; and uses the second prediction channel for analysis when the initial abnormality judgment result is abnormal; If the analysis result of the prediction channel is consistent with the initial abnormality judgment result, the output is the initial test result, otherwise the flow analysis is performed. The flow analysis is specifically to switch the prediction channel for analysis. If the analysis result of the prediction channel after the flow analysis is inconsistent with the analysis result of the prediction channel before the flow analysis, the flow analysis is continued until the results are consistent or the number of flow analyses exceeds the flow limit, and the initial test result is output.

[0008] As a preferred technical solution of the present invention, before the flow analysis is confirmed, the fault prediction model weights the input periodic fusion features according to the maximum influencing data item corresponding to the result of the prediction channel, that is, the set weight value is added to the corresponding part of the periodic fusion feature of the maximum influencing data item.

[0009] As a preferred technical solution of the present invention, the flow limit is a dynamic value, which is dynamically adjusted by combining the initial setting value with the number of retests as independent variables and setting a mathematical function.

[0010] As a preferred technical solution of the present invention, the data fusion includes: During the fusion cycle, feature extraction is performed on electrical parameter information, operating environment information, thermal imaging information, and operation log information to form a structured feature set. The feature importance evaluation method is used to assign initial fusion weights to each feature, and the structured feature set is fused to generate a high-dimensional fusion vector in a unified format. The t-SNE method is used for dimensionality reduction to obtain the final periodic fusion feature. The periodic fusion features are matched with the device features in the personalized device test sample for similarity, and the fusion weight is dynamically adjusted according to the similarity threshold to achieve adaptive feature reconstruction for different device states.

[0011] As a preferred technical solution of the present invention, the initial abnormality judgment includes: Construct a time series feature trajectory diagram based on multimodal data, and use a sliding window mechanism to perform trend fitting on electrical parameter information and operating environment information to extract statistical features; Performing discriminant analysis on the above statistical features using an abnormal trend detection model. The abnormal trend detection model is an unsupervised anomaly detection model built based on the isolation forest algorithm, which is used to identify data segments that have not yet triggered hard failure thresholds but have potential abnormal trends; The initial abnormality judgment result is judged and output according to the results of the hard fault threshold and the abnormal trend detection model. The judgment method is that when only two results are normal, the initial abnormality judgment result is output as normal, otherwise it is abnormal.

[0012] A comprehensive power supply test system with integrated fault prediction, comprising: Test parameter acquisition module: obtains multi-modal data of power supply equipment during power supply testing; Reference state update module: used to obtain and update the reference state transition diagram corresponding to the power supply equipment; Data fusion module: performs data fusion on all multimodal data within the fusion cycle; Fault prediction module: used to calculate the periodic fusion data according to the fault prediction model to obtain the initial test results, correction results and final test results of the fault prediction; Retest record module: used to obtain and record the retest weight, personalized equipment test samples and personalized equipment characteristics during the retest process of the power supply equipment.

[0013] The present invention has the following advantages: The present invention constructs a state transition diagram and a fault knowledge graph, and combines periodic fusion features and multi-dimensional health status portrait analysis to achieve dynamic modeling of power supply equipment operating trends and accurate characterization of fault evolution paths, significantly improving the ability to identify abnormal trends.

[0014] The present invention sets a first prediction channel and a second prediction channel, and flexibly selects a model path in combination with the initial abnormality judgment result, thereby improving the prediction accuracy while ensuring prediction efficiency and adapting to the operation requirements under scenarios of different complexity; by setting a flow analysis mechanism and a dynamic flow limit, as well as a weighting operation based on the most influential data item, intelligent verification and adaptive adjustment of the consistency of model results are achieved, thereby enhancing the stability and robustness of the fault judgment process.

[0015] The present invention introduces personalized device test samples and device feature clustering analysis mechanism to construct a device digital twin model, realizes adaptive fault judgment based on individual differences of devices, and improves the generalization ability and applicability of the model; through the similarity matching and feature reconstruction mechanism between periodic fusion features and personalized device samples, it supports dynamic optimization of fusion weights for specific operating states, further improving the pertinence and accuracy of feature expression.

[0016] The present invention introduces the isolation forest algorithm in the initial abnormality judgment to perform discriminant analysis on trend statistical characteristics, which can achieve early warning of abnormal trends before the traditional fault threshold is reached, and has stronger foresight and initiative. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only schematic diagrams of the present invention. Those skilled in the art can also derive other drawings based on the provided drawings without inventive effort. Figure 1 This is a schematic diagram of the structure of a comprehensive power supply test system with integrated fault prediction adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1, a comprehensive power supply testing method with integrated fault prediction, comprising the following steps: Step S1: Acquire multimodal data of the power supply device during the power supply test, including electrical parameter information, operating environment information, thermal imaging information, and operation log information; The electrical parameter information includes at least voltage, current, active power, reactive power, power factor and harmonics; The operating environment information includes external environmental indicators such as temperature, humidity, altitude, wind speed, and dust concentration, and is obtained by real-time sampling using a multifunctional environmental sensor installed in the test environment; The thermal imaging information captures the thermal distribution image of the target device at a certain time interval (e.g., 5 seconds to 60 seconds) through a thermal imaging camera, and converts the thermal image into temperature matrix data using an infrared temperature calibration model; The operation log information includes manual operation records, test program execution logs and system exception logs, and supports data synchronization capture of the operation and maintenance system through standard interface protocols (such as MODBUS, IEC61850, etc.); All types of data are collected synchronously and aligned with timestamps to ensure consistency of data of different modalities in the time dimension; To enhance data integrity, the system also includes a data quality assessment module to perform integrity checks on the collected data, remove outliers, and fill in missing values. The missing value filling methods include K-nearest neighbor interpolation, time series autoregressive interpolation, and multimodal information collaborative filling methods.

[0020] Step S2: Construct a reference state transition diagram corresponding to the power supply equipment based on historical multimodal data, including the fault evolution path from "normal state" to "abnormal trend" and then to "fault state", and a fault knowledge graph corresponding to the "fault type - parameter performance - processing suggestion" relationship for each standard fault case; The reference state transition diagram is constructed based on cluster analysis and state annotation. The historical data is divided into blocks using a segmented time window. The characteristic trend change amount, including the mean change rate, fluctuation index, abnormal frequency, etc., is calculated for each data block to calibrate the equipment operation status. The transition paths between states are modeled using a state transition probability matrix. The state classification includes at least five categories: normal, sub-healthy, abnormal trend, critical fault, and failure. The transition paths can be fitted and optimized using Markov chain modeling. The fault knowledge graph is a triplet graph structure, containing entity nodes (fault type, parameter expression, and processing suggestion) and relationship edges ("trigger," "correspondence," and "suggestion"), and supports storage in a graph database (Neo4j). The fault knowledge graph has an automatic update mechanism, including: After the fault prediction model outputs the final result, if the result is marked as a misjudgment or an uncovered new fault type after manual review or subsequent actual operation and maintenance verification, the current cycle's fusion features, prediction labels, actual fault type, and treatment measures are written as new samples into the queue to be updated; Perform semantic extraction and parameter standardization on the data in the update queue, and automatically generate a triplet of "fault type - parameter performance - processing suggestion"; Adopting a graph embedding-based entity matching and fusion algorithm, the similarity between the newly added triples and the entity nodes in the original graph is determined, and based on the set threshold, it is decided whether to add new nodes, merge existing nodes, or update edge weights. After the knowledge graph structure is updated, the graph version management module records the update information and triggers the re-scoring and path weight adjustment of the prediction model's associated paths to ensure the consistency of the graph and model reasoning logic; The update process is set to be triggered regularly (every 24 hours) or executed after manual confirmation (such as in combination with an expert review mechanism) to ensure the accuracy and controllability of the atlas update.

[0021] Step S3: Calculate the stable fluctuation period of each data item in the electrical parameter information, take the weighted average of the stable fluctuation period of each data item as the fusion period, obtain all multimodal data within the fusion period and perform data fusion to obtain the period fusion feature, and obtain a multidimensional health status profile of the power supply equipment through feature engineering processing; perform initial abnormality judgment on the multidimensional health status profile, and the results include normal and abnormal. The calculation of the stable fluctuation period is based on a signal analysis method. Sliding window Fourier transform (STFT) and autocorrelation function analysis are performed on each electrical parameter information to extract its dominant frequency or periodic pattern. The stable period of each signal is identified by evaluating the energy concentration and peak spacing of the periodic fluctuation. The calculation of the weighted average takes into account the weight of each electrical parameter's sensitivity to the fault, which is determined by the rate of change of different parameters in historical data before the fault occurs and the importance ranking in the prediction model; The data fusion includes: During the fusion cycle, feature extraction is performed on electrical parameter information, operating environment information, thermal imaging information, and operation log information to form a structured feature set. The characteristics of electrical parameter information include statistical features (mean, variance, maximum value, kurtosis, etc.) and frequency domain features (harmonic content, spectral energy, etc.); operating environment information features include mean, volatility, and mutation event statistics; thermal imaging information features obtain the temperature distribution of equipment hotspot areas through image segmentation and extract regional temperature differences, maximum / minimum temperatures, temperature rise gradients, etc.; operation log information uses natural language processing technology (such as TF-IDF and BERT encoding) to extract keyword vectors; The feature importance evaluation method (XGBoost backtracking weight analysis) is used to assign initial fusion weights to each feature, and the structured feature set is fused to generate a high-dimensional fusion vector in a unified format. The t-SNE method is used for dimensionality reduction to obtain the final periodic fusion feature. The periodic fusion features are matched with the device features in the personalized device test sample by similarity, and cosine similarity or Euclidean distance is used as a similarity measurement indicator. The fusion weight is dynamically adjusted according to the similarity threshold to achieve adaptive feature reconstruction for different device states.

[0022] The initial abnormality judgment includes: Construct a time series feature trajectory diagram based on multimodal data, and use a sliding window mechanism to perform trend fitting on electrical parameter information and operating environment information to extract statistical features (including at least trend slope, fluctuation amplitude, and coefficient of variation); Trend fitting methods include linear least squares, polynomial fitting, or smoothing prediction based on Kalman filtering; Performing discriminant analysis on the above statistical features using an abnormal trend detection model. The abnormal trend detection model is an unsupervised anomaly detection model built based on the isolation forest algorithm, which is used to identify data segments that have not yet triggered hard failure thresholds but have potential abnormal trends; The initial abnormality judgment result is judged and output based on the results of the hard fault threshold and the abnormal trend detection model. The judgment method is that when only two results are normal, the initial abnormality judgment result is output as normal, otherwise it is abnormal; The hard thresholds are set based on the power supply equipment's factory specifications, historical safety limits, and expert experience, and can be customized by device type.

[0023] Step S4: Input the periodic fusion features into the fault prediction model for fault prediction. The fault prediction model selects a prediction channel based on the result of the initial abnormality judgment. The model includes two independent prediction channels. The first prediction channel performs fast analysis based on a lightweight model, and the second prediction channel performs in-depth analysis based on a convolutional neural network. The results of the prediction channels are circulated and analyzed within the fault prediction model to obtain the initial test results.

[0024] The first prediction channel is a lightweight model combination based on ensemble learning algorithms, including random forest, gradient boosted tree (GBDT) and support vector machine (SVM), which quickly classifies the input period fusion features; The second prediction channel is a deep analysis model based on a convolutional neural network (CNN). The input layer receives periodic fusion features, extracts local temporal features through a one-dimensional or two-dimensional convolutional layer, and then connects to a fully connected layer to output the fault category probability. Multiple channels can be set up in the CNN model to process data from different modalities, and the importance of each modality feature can be dynamically adjusted through the attention mechanism module; The fault prediction model uses the first prediction channel for analysis when the initial abnormality judgment result is normal; and uses the second prediction channel for analysis when the initial abnormality judgment result is abnormal; If the analysis result of the prediction channel is consistent with the initial abnormality judgment result, the output is the initial test result. Otherwise, the flow analysis is performed. The flow analysis specifically switches the prediction channel for analysis. If the analysis result of the prediction channel after the flow analysis is inconsistent with the analysis result of the prediction channel before the flow analysis, the flow analysis is continued until the results are consistent or the number of flow analysis exceeds the flow limit, and the initial test result is output; Before the flow analysis is confirmed, the fault prediction model weights the input periodic fusion features according to the maximum impact data item corresponding to the result of the prediction channel, that is, the set weight value is added to the corresponding part of the periodic fusion feature of the maximum impact data item; The most influential data item is determined by ranking the feature importance in the model, such as extracting it through Grad-CAM heat map analysis in CNN; The turnover limit is a dynamic value, which is dynamically adjusted by setting a mathematical function using an initial set value combined with the number of retests as independent variables; the function can be in the form of an exponential decay function or a linear weighted function, which is used to balance test efficiency and prediction accuracy; The fault prediction model has a dynamic migration mechanism, including: When the statistical characteristics of the cluster to which the personalized device test sample belongs deviate significantly, or the historical prediction accuracy falls below the preset threshold continuously within a set period, the model migration evaluation mechanism is triggered; The model transfer evaluation mechanism calculates the distribution distance between the source domain model and the target domain data based on the current device category, historical sample distribution and the latest fusion features, including at least the maximum mean difference (MMD) and kernel mapping similarity; When the distribution distance exceeds the set migration trigger threshold, the model migration module is activated and the existing model parameters are updated using a transfer learning strategy based on domain adaptation technology. The strategy includes at least freezing and fine-tuning some weight layers in the convolutional neural network, or using a domain discriminator for adversarial training optimization. The updated model after migration replaces the original model to participate in subsequent fault prediction analysis. At the same time, the migration process and results are recorded in the model evolution log, and update prompts are regularly output to the administrator.

[0025] Step S5: Match the fault knowledge graph based on the initial test results, obtain the re-verification weight based on the corresponding parameter performance, test the power supply equipment again, use the re-verification weight to weight the periodic fusion features to obtain the re-verified periodic fusion features, and perform fault prediction using the fault prediction model. The output of the correction result or the final result is determined based on whether the number of flow analysis exceeds the flow limit. If the output result is a correction result, the correction result is used to match the fault knowledge graph, and a re-verification weight is obtained, and the test is performed again until the final result is obtained or the number of re-tests exceeds the set threshold; the re-verification weight is weighted each time the periodic fusion feature is weighted and integrated into the personalized device test sample; the personalized device test sample is used for weighting in the data fusion process; The re-verification weight is calculated based on the degree of match between the initial test results and the historical fault types and parameter performance in the knowledge graph; the match degree is quantified using the similarity calculation formula (cosine similarity) to obtain the weighted coefficient of each fault type; The generation of re-verification weights also considers the frequency and severity of historical fault cases. If a certain fault type occurs frequently and has a greater impact, the weight of that type will be higher; The revalidation weights are weighted using either a linear weighting method or an adaptive weighting method. The linear weighting method is a linear combination of the weighting coefficients of each feature dimension, while the adaptive weighting method uses a machine learning algorithm (weighted least squares regression) to self-learn from historical data to obtain the optimal weighting coefficient. The weighted periodic fusion features are re-input into the fault prediction model for fault prediction. The model output results will be corrected and judged based on the number of flow analysis. If the number of flow analysis exceeds the flow limit and there is a significant difference between the revised result and the initial test result, the revised result is output; otherwise, the final result is output and all flow processes and analysis details are recorded.

[0026] Step S6: Perform cluster analysis on the personalized device test samples, and construct personalized device features in the device twin digital model based on the cluster analysis results. The personalized device features are used to correct the fault evolution path.

[0027] The cluster analysis uses an unsupervised learning algorithm to first standardize the multidimensional features of the device test samples and then reduce the dimensionality of the high-dimensional data using the t-SNE method to reveal the potential structure in the data; Then, the device samples are clustered based on the hierarchical clustering algorithm, and similar device samples are classified into the same category. The clustering results are evaluated using indicators such as the silhouette coefficient and the Davies-Bouldin index to ensure the effectiveness of the clustering effect. Each cluster represents a type of equipment with similar operating characteristics and fault characteristics. The personalized equipment characteristics are the cluster center or the most representative equipment characteristics of this type of equipment. In the device twin digital model, the personalized device characteristics of each device are modeled through digital twin technology, combined with historical operating data, failure modes and device characteristics to form a virtual mirror image of the device; The personalized device characteristics constructed in the digital twin model include not only electrical parameters at the hardware level, but also dynamically changing factors such as environment, temperature, and load. By monitoring and analyzing these characteristics in real time, the digital model can be updated in real time when the device changes, identifying potential device failure trends in advance. By updating personalized device features in real time, the model can dynamically adjust its inference of the fault evolution path, especially when the device status changes significantly (such as drastic fluctuations in the working environment or changes in operating load), and automatically correct the judgment basis of the fault prediction model.

[0028] The corrected fault evolution path will be compared with actual fault history data, continuously optimized and added to the fault knowledge graph to ensure that the model and graph are updated synchronously, further improving the accuracy of the prediction.

[0029] Example 2, a comprehensive power supply test system with integrated fault prediction, see Figure 1 As shown, it includes the following modules: Test parameter acquisition module: obtains multi-modal data of power supply equipment during power supply testing; Reference state update module: used to obtain and update the reference state transition diagram corresponding to the power supply equipment; Data fusion module: performs data fusion on all multimodal data within the fusion cycle; Fault prediction module: used to calculate the periodic fusion data according to the fault prediction model to obtain the initial test results, correction results and final test results of the fault prediction; Retest record module: used to obtain and record the retest weight, personalized equipment test samples and personalized equipment characteristics during the retest process of the power supply equipment.

[0030] Example 3, power supply equipment fault diagnosis and prediction based on a comprehensive power supply test bench, includes: The integrated power supply test bench adopts a method that combines multimodal data acquisition with intelligent prediction models. By analyzing the input and output characteristics, electrical parameters, environmental information, thermal imaging data and operation logs of the power supply equipment, it realizes intelligent prediction of the health status and fault diagnosis of the power supply equipment.

[0031] 1. Multimodal data acquisition: During the power supply test, the integrated power supply test bench collects electrical parameters, environmental parameters, thermal imaging information and operation log information in real time by configuring a variety of sensors and data acquisition modules.

[0032] 2. Data fusion and periodic feature calculation: By fusing the periodic features of multimodal data, we calculate the stable fluctuation periods of electrical parameters, environmental changes, and equipment operation to generate fused periodic features. Feature engineering methods are used to extract and weightedly fuse the data to generate high-dimensional fused features, ensuring that the multidimensional information of the data is comprehensively considered.

[0033] 3. Preliminary abnormality judgment and intelligent prediction: After preliminary data processing, an unsupervised anomaly detection model based on the isolation forest algorithm is used to determine the health status of the power supply equipment. If the equipment status shows an abnormal trend, the next step is in-depth analysis.

[0034] If the initial judgment is normal, a quick assessment is performed using a lightweight prediction model; If it is initially judged to be abnormal, a convolutional neural network is used to further analyze the abnormal type and cause of the equipment.

[0035] 4. Fault knowledge graph and dynamic update: Integrating with the fault knowledge graph, after the fault prediction model outputs results, the prediction results, actual fault type, and treatment measures are automatically added to the queue to be updated. Using the graph update algorithm, fault triplets are automatically extracted and analyzed through graph embedding to refine and update the device's fault evolution path and diagnostic rules.

[0036] 5. Equipment health status profiling and continuous optimization: Based on personalized device test samples, cluster analysis technology is used to classify devices into different categories and build personalized device signatures. Digital twin technology is used to update the health status of devices in real time, and fault prediction paths are continuously optimized based on changes in device status.

[0037] 6. Fault prediction and intelligent adjustment: During the operation of the integrated power supply test bench, the equipment fault prediction model intelligently predicts potential equipment failures based on real-time collected feature data and an updated fault knowledge graph. If the prediction results are consistent with the initial anomaly judgment, a fault warning is issued. If the results are inconsistent, the flow analysis mechanism switches the prediction channel to further ensure prediction accuracy.

[0038] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A comprehensive power supply testing method with integrated fault prediction, characterized in that: include: Acquire multimodal data of power supply equipment during power supply testing, including electrical parameter information, operating environment information, thermal imaging information, and operation log information; Based on historical multimodal data, a reference state transition diagram for power supply equipment is constructed, including the fault evolution path from "normal state" to "abnormal trend" and then to "fault state", and a fault knowledge graph corresponding to the "fault type-parameter performance-handling suggestions" relationship for each standard fault case; The stable fluctuation period of each data in the electrical parameter information is calculated, and the weighted average of the stable fluctuation period of each data is taken as the fusion period. All multimodal data within the fusion period is obtained and fused to obtain the period fusion feature. Feature engineering is then used to obtain a multi-dimensional health status portrait of the power supply equipment. Perform initial abnormality judgment on the multi-dimensional health status portrait, with the results including normal and abnormal; The periodic fusion features are input into the fault prediction model for fault prediction. The fault prediction model selects a prediction channel based on the results of the initial abnormality judgment. The model includes two independent prediction channels. The first prediction channel performs fast analysis based on a lightweight model, and the second prediction channel performs in-depth analysis based on a convolutional neural network. The results of the prediction channels are circulated and analyzed within the fault prediction model to obtain the initial test results.

2. The comprehensive power supply testing method with integrated fault prediction according to claim 1, characterized in that: Also includes: Match the fault knowledge graph based on the initial test results, obtain the re-verification weight based on the corresponding parameter performance, test the power supply equipment again, use the re-verification weight to weight the periodic fusion features to obtain the re-verified periodic fusion features, and use the fault prediction model to predict the fault. The output of the correction result or the final result is determined based on whether the number of flow analysis exceeds the flow limit. If the output result is a correction result, the correction result is used to match the fault knowledge graph, and the re-verification weight is obtained, and the test is performed again until the final result is obtained or the number of re-tests exceeds the set threshold; The re-verification weight is integrated into the personalized device test sample after weighting the periodic fusion feature each time; the personalized device test sample is used for weighting in the data fusion process; By performing cluster analysis on personalized device test samples and constructing personalized device features in the device twin digital model based on the cluster analysis results, the personalized device features are used to correct the fault evolution path.

3. The comprehensive power supply testing method with integrated fault prediction according to claim 1, characterized in that: The fault prediction model uses the first prediction channel for analysis when the initial abnormality judgment result is normal; When the initial abnormality judgment result is abnormal, the second prediction channel is used for analysis; If the analysis result of the prediction channel is consistent with the initial abnormality judgment result, the output is the initial test result, otherwise the flow analysis is performed. The flow analysis is specifically to switch the prediction channel for analysis. If the analysis result of the prediction channel after the flow analysis is inconsistent with the analysis result of the prediction channel before the flow analysis, the flow analysis is continued until the results are consistent or the number of flow analyses exceeds the flow limit, and the initial test result is output.

4. The comprehensive power supply testing method with integrated fault prediction according to claim 3, characterized in that: Before the flow analysis is confirmed, the fault prediction model weights the input periodic fusion features according to the maximum impact data item corresponding to the result of the prediction channel, that is, increases the set weight value for the corresponding part of the periodic fusion feature of the maximum impact data item.

5. The comprehensive power supply testing method with integrated fault prediction according to claim 3, characterized in that: The flow limit is a dynamic value, which is dynamically adjusted by combining the initial setting value with the number of retests as independent variables and setting a mathematical function.

6. The comprehensive power supply testing method with integrated fault prediction according to claim 1, characterized in that: The data fusion includes: During the fusion cycle, feature extraction is performed on electrical parameter information, operating environment information, thermal imaging information, and operation log information to form a structured feature set. The initial fusion weight is assigned to each feature through the feature importance evaluation method, and the structured feature set is fused to generate a high-dimensional fusion vector in a unified format. The t-SNE method is used for dimensionality reduction to obtain the final periodic fusion feature. The periodic fusion features are matched with the device features in the personalized device test sample for similarity, and the fusion weight is dynamically adjusted according to the similarity threshold to achieve adaptive feature reconstruction for different device states.

7. The comprehensive power supply testing method with integrated fault prediction according to claim 1, characterized in that: The initial abnormality judgment includes: Construct a time series feature trajectory diagram based on multimodal data, and use a sliding window mechanism to perform trend fitting on electrical parameter information and operating environment information to extract statistical features; The above statistical features are discriminantly analyzed using an abnormal trend detection model. The abnormal trend detection model is an unsupervised anomaly detection model built based on the isolation forest algorithm, which is used to identify data segments that have not yet triggered the hard failure threshold but have potential abnormal trends; The initial abnormality judgment result is judged and output according to the results of the hard fault threshold and the abnormal trend detection model. The judgment method is that when only two results are normal, the initial abnormality judgment result is output as normal, otherwise it is abnormal.

8. A comprehensive power supply test system with integrated fault prediction, characterized in that: The system applies any one of the integrated power supply testing methods with integrated fault prediction according to claims 1 to 7, including: Test parameter acquisition module: obtains multi-modal data of power supply equipment during power supply testing; Reference state update module: used to obtain and update the reference state transition diagram corresponding to the power supply equipment; Data fusion module: performs data fusion on all multimodal data within the fusion cycle; Fault prediction module: used to calculate the periodic fusion data according to the fault prediction model to obtain the initial test results, correction results and final test results of the fault prediction; Retest record module: used to obtain and record the retest weight, personalized equipment test samples and personalized equipment characteristics during the retest process of the power supply equipment.

Images