Power equipment health state monitoring method based on multiple modes

Abstract

The invention discloses a multi-modal-based power equipment health state monitoring method, and relates to the technical field of power equipment detection.The power equipment monitoring method is based on multi-modal data and knowledge graph fusion, cross check, expert rule cleaning and label correction are implemented by collecting sensing data such as chromatography, temperature, current and vibration in oil, and the detection result is obtained. Outputting high-credibility data; a graph model is constructed based on equipment topology by adopting self-encoder dimension reduction fusion, early anomaly detection is realized by utilizing a graph neural network, and an anomaly alarm is generated; mechanism matching and consistency evaluation are carried out based on the fault mechanism knowledge graph, and interpretable diagnosis is output; and when the diagnosis result is significantly deviated from the actual operation and maintenance conclusion, triggering an online increment and transfer learning updating model and expanding the knowledge graph to form a closed-loop self-learning mechanism. According to the method, the fault detection accuracy is remarkably improved, false alarms and missing alarms are reduced, the operation and maintenance decision-making efficiency is improved, and meanwhile operation and maintenance intelligence and real-time alarm are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and specifically to a method for monitoring the health status of power equipment based on multi-modalities. Background Technique

[0002] The existing on-line monitoring system for power transformers is gradually developing from single-sensor acquisition to multi-modal fusion, and uses multi-dimensional data including dissolved gas in oil chromatography, winding temperature, load current, vibration and noise, etc. to dynamically perceive and fault-determine the health status of the transformer. This multi-modal on-line monitoring system is usually configured in high-voltage substations, power plants and large industrial electricity consumption scenarios. In order to adapt to the increasingly complex operating environment, the system is often equipped with distributed sensor nodes and uploads data through an Internet of Things gateway or a dedicated communication network. At the same time, maintenance personnel will also incorporate text information such as inspection reports, maintenance records and historical event libraries into the data platform, thus forming a full-scale information resource library covering structural characteristics, mechanism knowledge and dynamic monitoring.

[0003] With the help of these multi-source data, in theory, it is possible to realize the comprehensive analysis and early warning of the internal insulation state, partial discharge signs and winding fault risks of the transformer. Due to the significant differences in operating conditions and load conditions in industrial scenarios, the change laws of key indicators of different transformers often vary greatly in different regions or different power consumption periods, posing challenges to data standardization, consistency and model generalization. However, in practice, if the advantages of multi-modal data can be fully utilized and deeply mined in combination with maintenance experience and mechanism knowledge, higher accuracy and decision-making efficiency can still be obtained in many aspects such as remote monitoring and fault troubleshooting.

[0004] However, the current multi-modal transformer monitoring technology generally faces the problem of diagnostic uncertainty caused by unbalanced data quality, lack of fault samples and inaccurate labels in actual deployment. Specifically, during the normal operation of the equipment, most sensor readings are often within the healthy range, while data related to serious fault situations is extremely scarce and often concentrated in a limited number of typical faults; there are even fewer precedents for some rare faults (such as inter-turn short circuit of windings, high-temperature ablation of tap changers, etc.). At the same time, due to environmental interference, sensor aging or inaccurate manual inspection, the data collected by the monitoring system may be mislabeled or missed, and even some potential faults only show weak signs and are difficult to be identified by existing rules or experience. If these data imbalances and annotation uncertainties cannot be effectively solved, misjudgments and missed judgments will occur during the model training or operation stage: at best, it will cause the system to frequently give false alarms, increase the maintenance burden and weaken the trust of maintenance personnel in the monitoring system, and at worst, it may miss key fault precursors, resulting in sudden shutdown of the transformer or large-scale power outages, thus having a serious impact on the safety of the power grid and the stability of industrial production.

[0005] This problem is particularly prominent against the backdrop of the rapid growth of the power system scale and the increasing demand for intelligent operation and maintenance. There is an urgent need for a comprehensive technical solution that combines small-sample processing capabilities, real-time update mechanisms, and mechanism integration to address it. Summary of the Invention

[0006] (I) Technical Problems to be Solved In view of the deficiencies of the prior art, the present invention provides a method for monitoring the health status of power equipment based on multi-modalities. By collecting sensing data such as chromatogram in oil, temperature, current, vibration, etc., cross-checking is implemented, and expert rules are used for cleaning and label correction to output highly reliable data; an autoencoder is adopted for dimensionality reduction and fusion, and a graph model is constructed based on the equipment topology. A graph neural network is used to achieve early anomaly detection and generate anomaly alarms; relying on the fault mechanism knowledge graph for mechanism matching and consistency evaluation, an interpretable diagnosis is output; when the diagnosis result significantly deviates from the actual operation and maintenance conclusion, an online incremental and transfer learning is triggered to update the model and expand the knowledge graph, forming a closed-loop self-learning mechanism. False alarms and missed alarms can be reduced, the efficiency of operation and maintenance decision-making can be improved, and at the same time, the intelligence of operation and maintenance and real-time alarm can be strengthened; the technical problems recorded in the background art are solved.

[0007] (II) Technical Solutions To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for monitoring the health status of power equipment based on multi-modalities, including, when the noise content or annotation imbalance of multi-modal data is detected, correcting the fault labels and removing outliers to produce highly reliable multi-modal data that has been cleaned and corrected; After obtaining the highly reliable multi-modal data, combining the autoencoder dimensionality reduction matrix and the graph neural network adjacency matrix, denoising and fusing the highly reliable multi-modal data through a multi-layer mapping function, and detecting early fault signs by means of anomaly scores, outputting early anomaly alarms and feature descriptions; When receiving the early alarms and feature descriptions and finding conflicts in matching with the fault types, using the knowledge graph and the mechanism matching function to locate the mechanism of the feature descriptions, outputting interpretable diagnosis results and consistency evaluation results, and marking new fault nodes; If the interpretable diagnosis result significantly differs from the actual operation and maintenance conclusion, the model is corrected according to the degree of difference and the incremental learning rate, and the relationship weights are fused to dynamically update the knowledge graph and multi-modal data, forming a self-learning closed loop.

[0008] Furthermore, when performing dynamic label correction based on cross-comparison of sensor readings and historical maintenance reports, the suspicious data labels are secondarily screened using the reference distribution and credibility coefficient constructed from historical samples to correct possible missing labels or mislabels and form highly reliable multi-modal data.

[0009] Furthermore, when using few-shot learning or synthetic oversampling to perform bounded amplification on scarce fault data, filter the overly simulated fault samples by combining expert rule constraints, where: When the label consistency is lower than the consistency threshold, it is determined that there is a label anomaly in the record; after confirmation, perform adaptive adjustment and store the corresponding data together with the label in the high-confidence multi-modal dataset.

[0010] Furthermore, when using a deep autoencoder to perform dimensionality reduction and feature fusion on multi-dimensional sensor data, use the hidden layer parameters of the encoder-decoder network to perform denoising processing on the features of each modality to obtain the low-dimensional feature vector of the sample in the hidden layer, and generate an autoencoder low-dimensional feature matrix after summarization.

[0011] Furthermore, perform multi-layer propagation on the graph structure formed by transformer components or monitoring points through a graph neural network, determine early suspected fault nodes based on the adjacency matrix and the anomaly score mechanism, and output early anomaly alarms and feature descriptions; Among them, if the comprehensive scoring function is greater than the balance threshold, it is determined that the sample or time window corresponding to the node is abnormal, and it is used as an abnormal node.

[0012] Furthermore, for the early anomaly alarms and feature descriptions, use the similarity between the fault mechanism nodes and the feature descriptions to construct a mechanism matching score, locate the fault nodes in the knowledge graph based on the mechanism matching score, and mark them when the matching is uncertain.

[0013] Furthermore, after obtaining the fault mechanism localization result, use consistency evaluation to perform credibility measurement on the diagnosis conclusion and output an interpretable diagnosis result and consistency evaluation. When it is determined that the credibility is insufficient, automatically trigger manual intervention or supplement the collection of monitoring data; output the optimal mechanism localization node and its corresponding typical incentives, recommended maintenance links, corresponding core anomaly feature descriptions, and fault consistency scores in the knowledge graph.

[0014] Furthermore, based on the difference detection between the diagnostic output and the true operation and maintenance conclusion, if the difference between the two exceeds the difference threshold, incorporate the misjudged samples into incremental learning and perform mini-batch correction in combination with the online learning rate, output the corrected model parameters to continuously optimize the diagnostic performance, select a new optimal mechanism node, and output an interpretable diagnosis result.

[0015] Furthermore, after completing the incremental correction of the local model, inherit the empirical parameters of other transformer devices between the relationship weights of the graph neural network or the autoencoder structure through hierarchical migration to accelerate the model convergence in the new environment or rare fault modes.

[0016] Furthermore, through incremental updating and automatic clustering of the knowledge graph, new fault nodes and their relationship edges are automatically inserted into the existing mechanism network through a mapping function, and after backtracking, the newly added labels and multimodal data are re-cleaned and re-detected to complete the closed-loop self-learning update.

[0017] (III) Beneficial Effects The present invention provides a method for monitoring the health status of power equipment based on multimodality, having the following beneficial effects: In the first step, data cleaning and dynamic label correction technologies are adopted to output high-confidence multimodal data, effectively improving the biases caused by data imbalance and mislabeling, and ensuring the basic quality of subsequent deep learning model training.

[0018] In the second step, an autoencoder is combined with a graph neural network to extract the low-dimensional feature matrix of the autoencoder from the high-confidence multimodal matrix , and based on the component association structure is characterized, and then an anomaly scoring function is used to give early warnings for incipient faults. Thus, potential faults can be timely captured before they cause serious consequences.

[0019] In the third step, the above statistical and deep representation results are fused with the knowledge graph , a mechanism matching score and a consistency evaluation are defined to measure the degree of agreement between the model judgment and professional mechanism knowledge, thereby making up for the defects that are difficult to explain by pure data-driven methods and enhancing the diagnostic transparency. When unknown faults or matching conflicts occur, uncertain nodes can be marked through a feedback mechanism and combined with manual review to ensure that the final interpretable diagnostic results and consistency evaluation have high confidence and clear traceability.

[0020] In the fourth step, a closed-loop system of online update and transfer learning is constructed: if it is found that the final judgment result is significantly different from the actual operation and maintenance conclusion , it is sent to the incremental correction in sub-step 401 to adjust the model parameters in real time; and the hierarchical transfer based on the relationship weight in sub-step 402 can be further called to transfer the experience accumulated from other transformers or historical operating conditions to this equipment, shortening the adaptation period for rare faults. Finally, in sub-step 403, new fault types or mechanism nodes are incrementally updated to the knowledge graph through a mapping function , and backtracking to the first and second steps to achieve a complete self-learning closed loop.

[0021] Through the tight coupling between steps (such as high-confidence data-supported feature extraction, anomaly alarm matching knowledge graph, and online update to feedback the diagnostic results back to the previous steps), the collaborative gain of cross-modal fusion and mechanism interpretation is achieved; by strengthening the accurate identification of misjudged samples and flexibly inheriting external experience at the relationship weight level of the graph neural network through hierarchical transfer, better effects are achieved in multi-modal fault monitoring, real-time learning, and interpretable mechanism decision-making. At the level, it plays a better role in multi-modal fault monitoring, real-time learning, and interpretable mechanism decision-making. Brief Description of the Drawings

[0022] Figure 1 It is a schematic flow diagram of the method for monitoring the health status of power equipment based on multi-modal of the present invention. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] Please refer to Figure 1 , the present invention provides a method for monitoring the health status of power equipment based on multi-modal, including, Step 1: When it is detected that the multi-modal data has noise content or annotation imbalance and cannot meet the fault monitoring accuracy requirements, use sensor reading cross-checking, historical maintenance report dynamic comparison, and minority class amplification to correct the fault labels and eliminate outliers, and produce high-confidence multi-modal data that has been cleaned and corrected; The content of the first step is as follows: Step 101: Multi-modal data cleaning In order to eliminate the interference of unreasonable values or extreme outliers on the diagnostic model, a cleaning method based on the Earth Mover's Distance is introduced. For each sensor modality, a corresponding reference distribution is constructed to measure the deviation degree between the current observed data and the reference distribution. The specific process is as follows: For the th sensor modality, first construct a reference distribution based on its historical reliable data, which can be confirmed by experts or extracted from long-term steady-state operation data; for the single piece of data currently observed, define the Earth Mover's Distance to measure the overall deviation degree between the current sensor reading distribution and the reference healthy state distribution: ; measure the overall deviation of this observation from the reference distribution, the larger the value of, the more likely this observation is noise or an anomaly. Based on the above distance, define the reliability coefficient of this observation: ; where is the sensitivity coefficient (which can be set according to operation and maintenance experience), ranges between 0 and 1, and the higher the value, the more credible the data; When the reliability coefficient is lower than the preset reliability threshold , it is regarded as extremely abnormal data and directly excluded or marked as invalid; if the reliability coefficient is within the preset threshold critical interval, it can be decided whether to correct it after comprehensive evaluation according to the reliability of other sensors at the same timestamp, and finally the cleaned data is output; for example, when the reliability of a certain sensor at timestamp falls into the critical interval , the system will first calculate the weighted average of the reliability of all other sensors at the same moment. If , it indicates that the overall data credibility is high, and at this time, the reading of this sensor will be excluded or interpolated and corrected; if , it is considered that the overall credibility is low, and the original reading is retained and a global anomaly mark is triggered; if , then further count the number of sensors with a reliability higher than , and use majority voting to decide whether to correct or retain, so as to make a reliable judgment relying on multi-modal information in the case of a single fuzzy reading.

[0025] Define the reliability coefficient and set the sensitivity coefficient , so that operation and maintenance personnel can flexibly adjust the cleaning rules according to the importance of different modalities, avoid excessive filtering while improving data quality, and provide more room for sensitive detection of different fault characteristics and filtering of false alarms through the sensitivity coefficient and the threshold .

[0026] Step 102, Dynamic label correction After obtaining the cleaned data output in step 101, step 102 proposes a dynamic correction strategy that combines expert criteria and comparison of prediction distributions for possible missing or mislabeled situations of the same fault type. The process is as follows: For any record in the cleaned dataset, let its original label be . If there is no clear label or the label is missing, a temporary annotation can be given by the operation and maintenance expert; select a reference model or expert library with a high correlation with this fault category, and perform fault type discrimination according to the current sensor characteristics to obtain the reference judgment result , and define the label consistency function: ; Among them it can be designed as advanced criteria such as polynomial distance and symbol error counting according to actual needs to measure the difference between two tags; is an adjustment coefficient, and its value is greater than 0; The obtained tag consistency The higher it is, the more consistent the original tag is with the reference tag; when the tag consistency is lower than the consistency threshold it is determined that there is an abnormal tag in this record, and manual verification or automatic correction is required; for the records determined to be abnormal, if it is confirmed that the original tag is incorrect by combining the operation and maintenance report or on-site maintenance information, then the original tag is replaced with the determination result or adjusted to a more appropriate fault type; after the correction is completed, this data together with its new tag is stored in the high-confidence multi-modal dataset; thus, the dynamic tag correction is completed, and finally the high-confidence multi-modal data is output; By comparing the tags with the reference model or the expert database, potential mislabeling or missing labeling can be effectively discovered, especially for rare fault types that are not common in historical samples; comparing and fusing the tag correction process with the reference model or expert criteria makes the tag improvement more real-time and flexible.

[0027] Step 2. After obtaining the high-confidence multi-modal data, combine the autoencoder dimensionality reduction matrix and the graph neural network adjacency matrix , and pass through the multi-layer mapping functions and to perform denoising fusion on the high-confidence multi-modal data, and detect early fault signs by means of anomaly scoring , and output early anomaly alarms and feature descriptions; The said Step 2 includes the following contents: Step 201. Feature fusion of the deep autoencoder In Step 201, based on the high-confidence multi-modal data output in the first step, use the deep autoencoder for dimensionality reduction and feature fusion to generate a low-dimensional feature vector that can be used for fault precursor identification, where: Let represent the high-confidence multi-modal matrix output in the first step, where is the total number of samples, is the dimension of the fused multi-modal features (obtained by splicing each modality after cleaning and alignment); denote as the low-dimensional representation of the th sample obtained by the autoencoder in the hidden layer, is the hidden layer dimension ; let and They are respectively the sets of network parameters of the encoder and decoder of the autoencoder; the autoencoder is a deep neural network model for multi-modal feature dimensionality reduction and denoising; Introduce the following advanced metric functions Measure the global difference between the original input and the reconstructed output. Let be the output of the decoder, then define the overall cost function of the autoencoder : ; where and respectively represent the original input and the reconstruction result of the th sample; to improve the ability to capture multi-modal features, advanced metrics such as structural correlation or other metrics that can highlight cross-modal differences can be selected, for example, including the Mahalanobis distance term and the canonical correlation coefficient term, and both are weighted; Forward propagation: The original input is mapped to the low-dimensional feature vector via the encoder , and then the decoder outputs the reconstruction result ; Backward update: Use optimization algorithms such as gradient descent to minimize , and gradually update the set of network parameters of the encoder and the set of network parameters of the decoder; Convergence judgment: When the overall cost function converges to a preset threshold or the number of iterations reaches the upper limit, end the training, and finally obtain the mapping function for subsequent extraction of low-dimensional feature vectors ; After the training is completed, each sample can obtain its low-dimensional feature vector in the hidden layer. These represent the core features that integrate different modal information, denoted as the autoencoder low-dimensional feature matrix ; Output the matrix to step 202 as the basic attribute vector of each node in the subsequent graph neural network construction.

[0028] By measuring the reconstruction difference of multi-modal data in a structured manner, the information interrelated between modalities can be fully retained; the deep autoencoder realizes denoising and feature fusion while reducing dimensionality, reducing redundancy and effectively retaining the weak features that may be contained in rare fault modes. The low-dimensional feature vector output by step 201 provides a more robust feature basis for step 202, greatly improving the accuracy of subsequent graph construction and anomaly detection.

[0029] Step 202: Early anomaly detection based on graph neural network In step 202, to further capture the potential fault propagation relationships between different sensors or transformer components, based on the low-dimensional feature matrix of the autoencoder output in step 201 A graph neural network is constructed to finally achieve real-time detection and identification of early fault signs. The specific process is as follows: Let represent the node set, and each node corresponds to a sample in step 201 (or represents a time window in time series); to establish associations between nodes, define the adjacency matrix , where can be determined according to the proximity of sensor measurement points, historical correlation, or actual physical structure mapping. If , it means that node is connected to ; The node feature matrix , that is, directly adopt the low-dimensional feature vector output in step 201 as the initial node embedding representation of the graph neural network; let be the trainable parameters of the -th layer of the graph neural network, represents the node representation matrix of this layer, then the representation of node in the -th layer can be defined as: ; where is the set of neighboring nodes of , is a non-linear activation function (such as ELU or LeakyReLU) to enhance the feature expression ability; After multiple layers of propagation, the final node representation matrix can be obtained. Denote as the final feature vector of node . For node (whose final embedding vector is denoted as ), define the comprehensive scoring function ; where: represents the manifold distance between and the embedding manifold formed by the normal state or the main group, which can be calculated by measuring the nearest point to this manifold (or its approximation), the length of the shortest geodesic, etc.; When the larger the value, the higher the degree of deviation of this node from the normal distribution in the high-dimensional feature space; is the amplification coefficient of the manifold deviation, which can be set according to operation and maintenance requirements; denotes the node and its set of neighborhood nodes weighted sum of the feature differences, where the norm is accumulated, and is the dimension weight, used to highlight or weaken certain key features; or higher; is the balance coefficient of the neighborhood weighted norm; The balance threshold is set in advance , if , then it is determined that the node corresponding sample (or time window) is abnormal, and it is used as an abnormal node to trigger an early warning.

[0030] In addition, relative threshold division can also be performed by combining mechanisms such as ranking or quantiles. All abnormal node indices and corresponding final feature vectors exceeding the threshold are packaged to form early abnormal warnings and feature descriptions, and output to the third step of knowledge graph fusion and interpretable decision-making for subsequent reasoning. If no node meets the abnormal conditions, the system is temporarily regarded as being in a healthy state, but the updated node representations will still be stored for subsequent cumulative analysis.

[0031] Based on the multi-modal dimensionality-reduced features , a graph neural network is constructed, which not only fully captures the interactions between sensors, but also can detect potential fault propagation paths from the perspective of the overall structure. By introducing a dual measure of information sparsity norm and relative distribution deviation through a comprehensive scoring function , the sensitivity to early and atypical fault patterns is significantly improved. When using the graph neural network for early anomaly detection, not only the feature outliers of individual nodes are considered, but also the association structure between nodes is combined to form an extensible graph-level anomaly recognition ability.

[0032] Step 3: When receiving early warnings and feature descriptions and finding conflicts with the matching of fault types, use the knowledge graph and the mechanism matching function to perform mechanism localization on the feature description , output an interpretable diagnosis result and a consistency evaluation result, and mark suspected new fault nodes; The content of Step 3 is as follows: Step 301: Knowledge graph matching and fault mechanism localization In Step 301, based on the early abnormal warnings and feature descriptions output in Step 202, combined with the power equipment fault knowledge graph constructed in this solution, matching and localization are performed at the mechanism level, where: Denote as the fault knowledge graph, where represents the set of nodes in the knowledge graph (such as mechanism concepts like winding overheating, core grounding fault, and increasing concentration in oil, etc.), ; represents the relationship between nodes (such as cause - effect or feature - fault type); each node is attached with multiple attributes, such as the correlation weight with gas component changes, the corresponding relationship with historical maintenance conclusions, etc.; Let represent the feature description output for the suspected fault node , where is the dimension after uniformly encoding information such as sensor readings, graph embedding features, and anomaly scores, etc.; Let represent the category label of the fault node (or moment) initially identified as a fault suspect (if it is only an anomaly without a specific classification, it is recorded as unknown); Introduce the following mechanism matching score , used to quantify the degree of conformity between the node and the feature description : ; where: is the reference fault feature vector corresponding to the knowledge graph node (summarizing common gas concentration distributions, temperature thresholds, topological associations, etc. under this fault mechanism); is a high - order similarity metric function, such as a function that can combine isomorphic embedding or differential geometry tools to measure the distance between the feature description and the reference fault feature vector on the feature manifold; is a penalty term that penalizes the deviation between and the preliminary category according to the fault type - mechanism node relationship in the knowledge graph; is a non - linear mapping (such as Softplus, ELU), which transforms the difference into the final score, and the higher the score, the higher the matching degree; For each abnormal node , search and calculate the maximum value of in the knowledge graph node set to obtain the optimal mechanism location node : ; According to the optimal mechanism location node The most likely corresponding fault mechanism node can be locked, and the associated background knowledge (such as common causes and recommended maintenance items) can be obtained; this result is then passed to step 302 for diagnostic explanation and visualization.

[0033] Through the construction of the mechanism matching score The statistical features extracted by deep learning are effectively connected with the professional mechanism information in the knowledge graph, avoiding black-box diagnosis; under the global relationship constraints of the knowledge graph (reflected by the penalty term ), when some anomalies are detected but the categories are not clear, fine inferences can also be made with the help of the most similar mechanism nodes, and rich background information can be provided for subsequent explanations.

[0034] Step 302, Explainable Reasoning and Decision Output After the fault mechanism is located in step 301, step 302 further connects the optimal mechanism location node with the specific sensing anomaly information, outputs an explainable diagnostic conclusion for the operation and maintenance personnel, and verifies it with the actual operation experience, where: For each abnormal node that has been located to the optimal mechanism location node Collect the set of feature-fault or component-mechanism relationship edges connected to the optimal mechanism location node in the knowledge graph , such as the obvious increase in oil risk of winding overheating or CO concentration superimposed with temperature exceeding the limit risk of multi-point overheating; map the above structure to the feature description output in the second step of step 202, such as the current concentration increases by X ppm, the CO concentration increases by Y ppm, and display the potential mechanism path in natural language or a graphical interface. For example: because the feature description has a high similarity with the optimal mechanism location node , and both H2 and CO in the oil exceed the threshold, it is inferred as the risk of winding overheating or multi-point overheating.

[0035] Define the fault consistency function to measure the degree of agreement between this diagnostic conclusion and historical precedents or expert rules. If there is a large difference from the existing operation and maintenance records, the operation and maintenance experts can be prompted for a second review, which can be formalized as: ; where: is the anomaly score defined in the second step (or its improved version), represents the typical anomaly score threshold related to the optimal mechanism location node in the knowledge graph; If the fault consistency function has a value lower than expected, it indicates that the degree of abnormality of the final feature vector does not match the reference level of the located mechanism in the knowledge graph, and a secondary review is required; Finally, an interpretable diagnosis result and consistency evaluation are formed and output, including: The optimal mechanism location node and its corresponding typical inducements and recommended maintenance links in the knowledge graph; the description of the core abnormal features that match it; The fault consistency score , as well as a mark indicating whether expert review or further on-site detection is required; This output is passed to the fourth step of online model update and transfer learning to update the knowledge graph and model parameters in a timely manner after receiving the actual maintenance results or new fault samples in the future.

[0036] Based on the fault mechanism location in step 301, step 302 provides complete traceability from principle to feature, and quantifies the diagnostic credibility through the fault consistency score ; through the relationship edges of the knowledge graph and the specific mechanism nodes that have been matched, the operation and maintenance personnel can directly see the cause inference logic behind the data, thereby increasing the trust in the system output and providing a clear direction for subsequent improvement or troubleshooting. Combining the anomaly score produced by deep learning with the typical anomaly score threshold based on mechanism priors in the knowledge graph in the fault consistency score function realizes the comprehensive judgment of statistical anomaly degree and mechanism adaptability. Using a graph database or an inference engine, the fault explanation can be visualized, and the associated information of the mechanism nodes can be dynamically called, enabling the decision-making layer to generate more in-depth and extensible interpretability.

[0037] Step Four. When there are significant differences between the generated interpretable diagnosis result and the actual operation and maintenance conclusion , the model is corrected according to the degree of difference and the incremental learning rate , and the relationship weights are fused , and finally the knowledge graph and multi-modal data are dynamically updated to form a self-learning closed loop; The content of the above step Four includes the following: Step 401. Incremental correction and difference detection In step 401, the interpretable diagnosis result and consistency evaluation output in the third step are compared with the actual operation and maintenance conclusion (such as subsequent maintenance reports, inspection results) to identify possible biases in the model and perform incremental correction, where: Denoted as The final determination result of the third - step diagnosis model for the fault / health state represents the true operation and maintenance conclusion (such as the fault diagnosis type); if or the diagnosis deviation is large, it is considered that there are obvious errors in the current model parameters in this fault scenario; all error samples are recorded and collected into a misjudgment sample set for subsequent incremental learning.

[0038] To measure the difference between the actual output of the model and the true conclusion in the high - dimensional space, a high - order difference metric can be defined For example: represents the prediction given by the system for the th sample (such as the fault probability distribution, the set of key points of the feature vector over time, etc.), represents the vector representation mapped by its corresponding actual operation and maintenance result (which can be obtained by the embedding / alignment method used in the first, second, or third step); Denote as the non - linear mapping function from the input space to the manifold embedding space, which has been defined and trained in the second or third step, where: the mapping function is the encoder of the auto - encoder in sub - step 201 of the second step — that is, the part that compresses the original dimensional multi - modal input to the dimensional hidden space. It is defined and trained in the following situations: Training stage: In the multi - modal feature extraction link of sub - step 201, by minimizing the auto - encoder reconstruction cost: ; At the same time, iteratively update the encoder parameters (the decoder is trained simultaneously).

[0039] After training convergence, it becomes the mapping which can project any new sample onto the same hidden space for use in the fourth - step high - order difference metric ; therefore, is not a call to a fixed offline function library, but a non - linear embedding mapping learned along with the joint training of the auto - encoder and can be directly reused after the second step is completed; Define the high - order difference metric function as follows: ; Where: A straight-line interpolation path defined on the original input space: ; When then (i.e., the vector corresponding to the true result); when then (i.e., the vector corresponding to the predicted result), any on the path can be mapped into the manifold space through ; denotes the Jacobian matrix of the mapping at the point , whose size is , and can be understood as the set of partial derivatives of the input components at this point; different correspond to different , so varies dynamically with , and can more accurately characterize the sensitivity to input increments along the path; is a certain advanced norm or measure (the Frobenius norm, 2-norm, or other designed higher-order norms can be selected) to measure . The operation actually measures how much difference is generated after the input makes a small offset in the direction at the point and is mapped into the manifold space, so that the difference evaluation has more differential geometry characteristics.

[0040] When the degree of difference exceeds the difference threshold , it is determined that the sample is seriously misjudged in the diagnostic model and needs to enter the incremental learning process; if the degree of difference is lower than the difference threshold but there is still a certain deviation, the corresponding parameters of the model will be reduced or corrected according to the weights.

[0041] Extract several small batches of samples from the misjudged sample set , and perform small-scale gradient updates using the loss function originally defined in this scheme (such as the combined loss of autoencoder reconstruction loss, graph neural network anomaly detection loss, and mechanism matching loss in the third step); Set the online learning rate , balance the influence of historical model knowledge and the introduction of new samples, and ensure that the update process neither overfits nor can correct obvious errors in a timely manner; after several small-batch iterations, obtain the corrected model parameter set . Combining this model with the knowledge graph in the third step can complete the new fault determination. Specifically, after the model update is completed, through incremental learning (based on the misjudged sample set ) The latest auto - encoder parameters obtained and the graph neural network parameters are deployed into the same inference pipeline as in the second and third steps; For the real - time multi - modal input to be diagnosed, first use the new auto - encoder to map and obtain the hidden - layer vector , and then use the new GNN parameters to complete message passing in the graph structure constructed by the knowledge graph to obtain the final node embedding ; Compare with the pre - constructed mechanism nodes through a matching function or consistency evaluation to select the new optimal mechanism node and output the interpretable diagnostic result; In this way, the new model obtained in sub - step 401 not only corrects the error at the data level but also forms an end - to - end re - diagnosis process when fusing with the knowledge graph in the third step, ensuring more accurate and interpretable fault determination of the equipment status using the latest learning results.

[0042] By comprehensively comparing the diagnostic output and the true result in the manifold space through the difference degree , errors in complex dimensions can be captured in a timely manner, rather than relying only on whether a single label matches; the incremental learning method avoids large - scale full - volume re - training every time new data is obtained, improving the efficiency and scalability of model updates; the online learning rate can be dynamically adjusted according to the fault type and data characteristics to form a differential priority correction strategy for different rare faults.

[0043] Step 402, Transfer learning strategy and cross - device knowledge sharing In step 402, to accelerate the adaptation of the model in a new environment or different transformer devices, a transfer learning strategy is introduced to effectively transfer the diagnostic experience learned in other devices or historical working conditions to the current scenario; This step is executed after the incremental correction in step 401 to ensure good connection between the corrected model parameter set and external auxiliary knowledge: Let be the model parameters pre - trained in the source device or source environment (which may include auto - encoder weights, graph neural network structures, knowledge graph association strengths, etc.), be the model parameters to be optimized for the target device currently, with the initial value of ; By detecting sensor distribution, operation duration, failure type frequency, etc., judge the similarity level between the source and the target as follows: First, construct probability distributions for the respective key sensors (such as oil temperature, gas concentration, vibration, etc.) of the source device and the target device on historical operation data, and calculate the similarity of each modal distribution through the normalized Earth Mover's Distance or KL divergence. At the same time, count the load duration distribution of the two devices and evaluate the alignment degree of working conditions with cosine similarity. Then, perform similarity measurement on the failure type frequency vectors, and finally synthesize the overall similarity score in a weighted manner. When this score exceeds a preset threshold (e.g., 0.85), the system determines that the operating environments and failure characteristics of the source domain and the target domain are highly consistent, and the parameters of the graph neural network and the weights of the autoencoder can be safely migrated with high weights.

[0044] If the similarity is high, a large proportion of the pre-trained model parameters can be introduced. If the similarity is low, selective parameter fusion is required; at the autoencoder level, initialize the weights of the encoder part (such as , and the decoder can be partially retained or randomly reset; at the graph neural network level, perform weighted fusion on the relationship weights as follows: ; where represents the migration trade-off coefficient, and the larger it is, the more the experience of the source device is trusted; is all the trainable parameters (including the weight matrix and bias vector) of the th layer of the graph neural network pre-trained on the source device, representing the feature extraction and transfer rules learned under the source domain failure mode; is the parameter of the th layer of the graph neural network of the target device before migration, reflecting the existing model state on the data of the target device; the same symbol on the left side of the equal sign refers to the new parameter after migration; In the knowledge graph mapping part, based on the mechanism node mapping relationship mentioned in the third step, directly incorporate the highly similar failure mechanism nodes into the target graph, or perform incremental update of the relationship weights, as follows: In the knowledge graph mapping link, the newly discovered or migrated failure mechanisms in the source graph will be integrated into the target graph according to the following process: First, calculate the vector embedding similarity (such as cosine similarity) between each source domain mechanism node located in the third step and all mechanism nodes in the target graph to obtain the corresponding similarity scores ; then for each source node, if the highest similarity (such as ), it is considered to be the same source as the existing node in the target graph, and the weight of the associated edge between this pair of nodes is incrementally updated, for example, with the new weight: ; To smoothly integrate new and old knowledge; if the similarity is lower than the preset threshold, it is considered that the source mechanism does not exist in the target graph, and a new node is added accordingly and the typical relationship edge (such as cause → result) connected to it and its weight are initialized. In this way, the target graph not only retains the original expert knowledge structure, but also can dynamically absorb new fault mechanisms or adjust edge strength to achieve adaptive expansion and optimization of knowledge.

[0045] The fused model parameters Perform a small batch online correction, and the set of misjudged samples identified in step 401 Compare and if the performance is significantly improved after migration, keep the new parameters; if there is a conflict or performance degradation, reduce the migration trade-off coefficient Record the failure modes of successful migration and their corresponding mechanism nodes to accumulate more prior data for subsequent migration of new equipment.

[0046] Through hierarchical migration and parameter reorganization, the degree of inheritance of external experience by different network parts can be finely controlled, accelerating the convergence of the model under different working conditions; the migrated model parameters output by step 402 can be further combined with the actual data of this device, significantly shortening the learning cycle on rare fault types; combined with the online update of step 401, a two-pronged optimization path of local increment + external migration is formed, taking into account both accuracy and generalization ability.

[0047] Step 403: Dynamic expansion of knowledge graph and closed feedback loop In step 403, the knowledge graph corresponding to the third step It also needs to be updated and expanded with new data and migration information to achieve a true closed-loop progression. This step takes over the migration adjustment results completed in step 402 and synchronizes the results back to the key data structures required in the first, second and third steps.

[0048] When a new fault type or new combination feature is found in step 401 or step 402 (for example, some abnormal gas ratios have not appeared in previous equipment), the knowledge graph Insert corresponding new fault mechanism nodes or update existing relationship edges; Defining the mapping function , the newly added or modified node information is compared with the existing mechanism nodes for similarity, and automatically assigned to the most suitable level or cluster to maintain the orderliness of the knowledge graph structure.

[0049] For newly added sensor data or new fault tags, it is necessary to go back to the first step: multi-modal data cleaning and dynamic tag correction for re-screening and annotation correction. Insert the new fault features into the training data of the graph neural network or autoencoder in the second step: multi-modal feature extraction and early anomaly detection, so that new faults of the same type can be detected in a timely manner during subsequent operation.

[0050] When entering the third step for interpretable decision-making again, there are more and updated fault mechanisms and associated edges in the knowledge graph, which can improve the precision and credibility of the next round of diagnosis; if operation and maintenance conclusions are continuously collected subsequently, the loop iteration can continue in steps 401, 402, and 403, forming a self-learning and self-evolving closed-loop process.

[0051] Mapping function Used to compare newly discovered fault mechanism nodes with existing nodes in the target knowledge graph : First, calculate the cosine similarity of the feature vectors between the newly discovered fault mechanism node and each node in the graph. If the highest similarity is not lower than the preset threshold (e.g., 0.8), it is considered that the newly discovered fault mechanism node is homologous to that node, and only the weight of its relationship edge is updated smoothly incrementally; otherwise, a new node will be added to the graph and its connection with related mechanism nodes will be initialized. In this way, the mapping function can not only maintain the structural consistency of the knowledge graph but also dynamically absorb new fault mechanisms, realizing the orderly expansion of the knowledge base.

[0052] Through the incremental update and automatic clustering of the knowledge graph structure, the system can continuously absorb new fault modes and incorporate them into the mechanism layer for interpretation; the dynamically updated knowledge graph output by step 403 and the synchronously updated multi-modal data will ensure that when the first step, the second step, and the third step are executed next time, the latest fault and mechanism information can be directly utilized, greatly improving the overall adaptation speed of the system. Connecting the knowledge graph with the data preprocessing and model training links into a closed-loop update transcends the traditional way of separating static graphs and offline models; through the mapping function automatically inserts new faults into the correct positions, reducing the burden of manual maintenance of large graphs and also improving the expansion ability in ultra-large multi-modal monitoring scenarios.

[0053] Those of ordinary skill in the art can realize that, combining the units and algorithm steps of each example described in the embodiments disclosed in this article, they can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0054] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0055] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0056] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0057] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application and should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for monitoring the health status of power equipment based on multimodality, characterized in that: including When the noise content or annotation imbalance of multimodal data is detected, correct the faulty labels and remove outliers, and produce highly reliable multimodal data that has been cleaned and calibrated. After obtaining highly reliable multimodal data, combine the autoencoder dimensionality reduction matrix and the graph neural network adjacency matrix, and perform denoising fusion on the highly reliable multimodal data through a multi-layer mapping function, and detect early fault signs by means of anomaly scoring, and output early anomaly warnings and feature descriptions. When receiving the early warning and feature description and finding a conflict with the fault type match, use the knowledge graph and the mechanism matching function to perform mechanism positioning on the feature description, output an interpretable diagnosis result and a consistency evaluation result, and mark the new fault node. If there is a significant difference between the interpretable diagnosis result and the actual operation and maintenance conclusion, correct the model according to the difference degree and the incremental learning rate, and fuse the relationship weights, and dynamically update the knowledge graph and multimodal data to form a self-learning closed loop.

2. The multimodal-based power equipment health status monitoring method according to claim 1, wherein: When performing dynamic label correction based on cross-comparison of sensor readings and historical maintenance reports, use the reference distribution and credibility coefficient constructed from historical samples to perform secondary screening on suspicious data labels to correct missing or mislabeled data and form highly reliable multimodal data.

3. The multimodal-based power equipment health status monitoring method according to claim 2, wherein: When using few-shot learning or synthetic oversampling to perform limited amplification on scarce fault data, filter the over-simulated fault samples by combining expert rule constraints, where: When the label consistency is lower than the threshold, it is determined that there is a label anomaly in the record, and after confirmation, an adaptive adjustment is performed, and the corresponding data and label are stored in the highly reliable multimodal data set.

4. The multimodal-based power equipment health status monitoring method according to claim 3, wherein: When using a deep autoencoder to perform dimensionality reduction and feature fusion on multi-dimensional sensor data, use the hidden layer parameters of the encoding and decoding network to perform denoising processing on each modal feature, obtain the low-dimensional feature vector of the sample in the hidden layer, and generate an autoencoder low-dimensional feature matrix after summarization.

5. The multimodal-based power equipment health status monitoring method according to claim 4, wherein: Perform multi-layer propagation on the graph structure formed by transformer components or monitoring points through a graph neural network, determine early suspected fault nodes according to the adjacency matrix and the anomaly score mechanism, and output early anomaly warnings and feature descriptions; Among them, if the comprehensive scoring function is greater than the balance threshold, it is determined that the sample or time window corresponding to the node is abnormal, and it is used as an abnormal node.

6. The multimodal-based power equipment health status monitoring method according to claim 5, wherein: For the early anomaly warning and feature description, use the similarity between the fault mechanism node and the feature description to construct a mechanism matching score, locate the fault node in the knowledge graph according to the mechanism matching score, and mark it when the matching is uncertain.

7. The multimodal-based power equipment health status monitoring method according to claim 6, wherein: After obtaining the fault mechanism localization result, use consistency evaluation to measure the credibility of the diagnosis conclusion and output the interpretable diagnosis result and consistency evaluation. When the determined credibility is insufficient, automatically trigger manual intervention or supplement the collection of monitoring data; Output the optimal mechanism localization node, its corresponding typical inducement in the knowledge graph, the recommended maintenance link, the corresponding core abnormal feature description, and the fault consistency score.

8. The multimodal-based power equipment health status monitoring method according to claim 7, wherein: Based on the difference detection between the diagnostic output and the true operation and maintenance conclusion, if the difference between the two exceeds the difference threshold, incorporate the misjudged samples into incremental learning and perform mini-batch correction in combination with the online learning rate, output the corrected model parameters to continuously optimize the diagnostic performance, select a new optimal mechanism node, and output the interpretable diagnosis result.

9. The multimodal-based power equipment health status monitoring method according to claim 8, wherein: After completing the incremental correction of the local model, inherit the empirical parameters of other transformer devices between the relationship weights of the graph neural network or the autoencoder structure through hierarchical migration to accelerate the model convergence in the new environment or rare fault modes.

10. The multimodal-based power equipment health status monitoring method according to claim 9, wherein: By performing incremental update and automatic clustering on the knowledge graph, automatically insert the new fault node and its relationship edges into the existing mechanism network in combination with the mapping function, and perform re-cleaning and re-detection on the newly added labels and multimodal data after backtracking to complete the closed-loop self-learning update.

Images