A multi-modal approach to monitoring the health of power equipment

Abstract

The present invention discloses a multimodal power equipment health status monitoring method, which relates to the field of power equipment detection technology. The power equipment monitoring method is based on the fusion of multimodal data and knowledge graphs. By collecting sensor data such as oil chromatography, temperature, current, and vibration, cross-checking and expert rule cleaning and label correction are implemented to output high-confidence data. Autoencoder dimensionality reduction fusion is adopted and a graph model is constructed based on the equipment topology. Graph neural networks are used to achieve early anomaly detection and generate anomaly alarms. Mechanism matching and consistency assessment are carried out based on the fault mechanism knowledge graph, and an explainable diagnosis is output. When the diagnostic result deviates significantly from the actual operation and maintenance conclusion, online incremental and transfer learning are triggered to update the model and expand the knowledge graph, forming a closed-loop self-learning mechanism. This method significantly improves the accuracy of fault detection, reduces false alarms and missed alarms, and improves the efficiency of operation and maintenance decision-making, while strengthening intelligent operation and maintenance and real-time alarms.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and in particular to a multimodal-based power equipment health status monitoring method. Background Art

[0002] Existing online monitoring systems for power transformers are gradually evolving from single-sensor acquisition to multimodal fusion, utilizing multi-dimensional data such as dissolved gas chromatography in oil, winding temperature, load current, vibration, and noise to dynamically perceive the health status of the transformer and determine faults. This multimodal online monitoring system is typically deployed in high-voltage substations, power plants, and large-scale industrial power consumption scenarios. To adapt to increasingly complex operating environments, the system is often equipped with distributed sensor nodes and uploads data via IoT gateways or dedicated communication networks. At the same time, operations and maintenance personnel will also incorporate text information such as inspection reports, maintenance records, and historical event libraries into the data platform, thereby forming a comprehensive information resource library covering structural characteristics, mechanism knowledge, and dynamic monitoring.

[0003] Leveraging this multi-source data, it's theoretically possible to comprehensively analyze the transformer's internal insulation status, signs of partial discharge, and winding fault risks, providing early warnings. Due to differences in operating conditions and significant fluctuations in load conditions in industrial scenarios, the patterns of change in key indicators for different transformers in different regions or during different power usage periods often vary significantly, posing challenges to data standardization, consistency, and model generalization. However, in practice, if the advantages of multimodal data can be fully utilized and deeply mined through a combination of operational and maintenance experience and mechanistic knowledge, higher accuracy and decision-making efficiency can still be achieved in many areas, such as remote monitoring and troubleshooting.

[0004] However, current multimodal transformer monitoring technologies, when deployed in practice, often face diagnostic uncertainty due to uneven data quality, a lack of fault samples, and inaccurate labeling. Specifically, during normal operation, most sensor readings are typically within healthy ranges, while data indicating severe fault conditions is extremely scarce and often concentrated in a limited number of typical faults. For some rare faults (such as inter-turn short circuits and high-temperature burnout of tap changers), there are no established precedents for comparison. Furthermore, due to environmental interference, sensor aging, or inaccurate manual inspections, the data collected by the monitoring system may be mislabeled or missed. Some potential faults may even manifest as only subtle symptoms, making them difficult to identify using existing rules or experience. Failure to effectively address these data imbalances and labeling uncertainties can lead to misjudgments and omissions during model training or operation. At best, this can lead to frequent false alarms, increasing the burden on operations and maintenance (O&M) and eroding the trust of O&M personnel in the monitoring system. At worst, critical fault precursors may be missed, resulting in sudden transformer shutdowns or widespread power outages, severely impacting power grid security and industrial stability.

[0005] This problem is particularly prominent against the backdrop of rapid growth in the scale of power systems and increasing demand for intelligent operation and maintenance. It urgently requires a comprehensive technical solution that combines small sample processing capabilities, real-time update mechanisms, and mechanism integration to solve it. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In response to the shortcomings of the existing technology, the present invention provides a multimodal power equipment health status monitoring method. By collecting sensor data such as oil chromatography, temperature, current, and vibration, cross-checking, expert rule cleaning, and label correction are implemented to output high-reliability data. Autoencoder dimensionality reduction fusion is adopted and a graph model is constructed based on the equipment topology. Graph neural networks are used to achieve early anomaly detection and generate anomaly alarms. Mechanism matching and consistency assessment are carried out based on the fault mechanism knowledge graph to output an explainable diagnosis. When the diagnostic result deviates significantly from the actual operation and maintenance conclusion, online incremental and transfer learning are triggered to update the model and expand the knowledge graph to form a closed-loop self-learning mechanism. This can reduce false alarms and missed alarms, improve the efficiency of operation and maintenance decision-making, and strengthen operation and maintenance intelligence and real-time alarms. It solves the technical problems recorded in the background technology.

[0008] (2) Technical solution

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multimodal power equipment health status monitoring method, including: when detecting noise or label imbalance in multimodal data, correcting fault labels and eliminating outliers, thereby generating cleaned and corrected high-reliability multimodal data;

[0010] After acquiring high-confidence multimodal data, the system combines the autoencoder dimensionality reduction matrix with the graph neural network adjacency matrix to perform denoising and fusion on the high-confidence multimodal data through a multi-layer mapping function. It also uses anomaly scoring to detect early signs of faults and outputs abnormality alerts and feature descriptions.

[0011] When receiving early warnings and feature descriptions and discovering conflicts with fault types, the system uses knowledge graphs and mechanism matching functions to locate the mechanism of the feature descriptions, outputs interpretable diagnostic results and consistency assessment results, and marks new fault nodes.

[0012] If there is a significant difference between the explainable diagnostic results and the actual operation and maintenance conclusions, the model will be modified based on the difference and incremental learning rate, and the relationship weights will be integrated to dynamically update the knowledge graph and multimodal data to form a self-learning closed loop.

[0013] Furthermore, when dynamic label correction is performed based on cross-comparison of sensor readings and historical maintenance reports, the benchmark distribution and credibility coefficient constructed by historical samples are used to perform secondary screening of suspicious data labels to correct possible omissions or mislabeling and form high-credibility multimodal data.

[0014] Furthermore, when using small sample learning or synthetic oversampling to perform limited amplification on scarce fault data, over-simulated fault samples are filtered in combination with expert rule constraints, where:

[0015] When the label consistency is lower than the consistency threshold, the record is determined to have a label anomaly; after confirmation, adaptive adjustments are made and the corresponding data together with the label are stored in a high-confidence multimodal dataset.

[0016] Furthermore, when using a deep autoencoder to perform dimensionality reduction and feature fusion on multi-dimensional sensor data, the hidden layer parameters of the encoder-decoder network are used to denoise the features of each modality to obtain the low-dimensional feature vector of the sample in the hidden layer, which is then aggregated to generate a low-dimensional feature matrix of the autoencoder.

[0017] Furthermore, a graph neural network is used to perform multi-layer propagation on the graph structure formed by transformer components or monitoring points. Based on the adjacency matrix and anomaly scoring mechanism, early suspected fault nodes are identified, and early anomaly warnings and feature descriptions are generated.

[0018] Among them, if the comprehensive scoring function is greater than the balance threshold, the sample or time window corresponding to the judgment node is abnormal and it is regarded as an abnormal node.

[0019] Furthermore, for early abnormal alarms and feature descriptions, the similarity between the fault mechanism node and the feature description is used to construct a mechanism matching score. The fault node is located in the knowledge graph based on the mechanism matching score, and is marked when the match is uncertain.

[0020] Furthermore, after obtaining the fault mechanism location results, the consistency assessment is used to measure the credibility of the diagnostic conclusion and output an explainable diagnostic result and consistency assessment. When the credibility is judged to be insufficient, manual intervention or the collection of supplementary monitoring data is automatically triggered; the optimal mechanism location node and its corresponding typical inducement in the knowledge graph, recommended maintenance links, corresponding core abnormal feature descriptions and fault consistency scores are output.

[0021] Furthermore, based on the difference detection between the diagnostic output and the actual operation and maintenance conclusion, if the difference between the two exceeds the difference threshold, the misjudgment samples will be included in the incremental learning and combined with the online learning rate for small batch correction. The corrected model parameters will be output to continuously optimize the diagnostic performance, select the new optimal mechanism node and output an explainable diagnostic result.

[0022] Furthermore, after completing the incremental correction of the local model, the empirical parameters of other transformer devices are inherited between the relational weights of the graph neural network or the autoencoder structure through hierarchical migration to accelerate the model convergence in new environments or rare fault modes.

[0023] Furthermore, by incrementally updating and automatically clustering the knowledge graph, new fault nodes and their relationship edges are automatically inserted into the existing mechanical network through the mapping function, and the newly added labels and multimodal data are re-cleaned and re-detected after backtracking to complete the closed-loop self-learning update.

[0024] (3) Beneficial effects

[0025] The present invention provides a multi-modal power equipment health status monitoring method, which has the following beneficial effects:

[0026] The first step uses data cleaning and dynamic label correction technology to output high-confidence multimodal data, effectively improving the deviations caused by data imbalance and incorrect labeling, and ensuring the basic quality of subsequent deep learning model training.

[0027] The second step is to combine autoencoders with graph neural networks to obtain high-confidence multimodal matrices Extract the low-dimensional feature matrix from the encoder , and based on Characterize the component association structure, and then use the anomaly scoring function Early warning of faults can be provided, so that potential faults can be caught in time before they cause serious consequences.

[0028] The third step is to combine the above statistical and deep representation results with the knowledge graph Perform fusion and define the mechanism matching score and consistency assessment This system measures the consistency between model judgments and professional mechanistic knowledge, thereby addressing the difficulties of purely data-driven interpretation and enhancing diagnostic transparency. When unknown faults or matching conflicts arise, uncertain nodes can be marked through a feedback mechanism, and combined with manual review, this ensures that the final interpretable diagnostic results and consistency assessments are highly reliable and clearly traceable.

[0029] The fourth step is to build a closed-loop system of online update and transfer learning: if the final judgment result is found and the actual operation and maintenance conclusions The difference is significant, and it is sent to the incremental correction of sub-step 401 to adjust the model parameters. Make real-time adjustments; and further call sub-step 402 based on relationship weight The hierarchical migration of other transformers or historical working conditions can be transferred to the current equipment to shorten the adaptation period of rare faults. Finally, sub-step 403 uses the mapping function Incrementally update new fault types or mechanism nodes to the knowledge graph , and go back to the first and second steps to achieve a complete self-learning closed loop.

[0030] Through the close coupling between steps (such as high-confidence data-supported feature extraction, abnormal alarm matching knowledge graph, online update to feed back the diagnosis results to the front link), the synergistic gain of cross-modal fusion and mechanism explanation is achieved; by strengthening the accurate recognition of misjudgment samples and using layered migration to the relationship weights of the graph neural network It flexibly inherits external experience at all levels, achieving better results in multimodal fault monitoring, real-time learning, and explainable mechanism decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure is a flow chart of the multi-modal power equipment health status monitoring method of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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.

[0033] See also Figure 1 The present invention provides a multi-modal power equipment health status monitoring method, comprising:

[0034] Step 1: When it is detected that the multimodal data contains noise or is imbalanced and cannot meet the fault monitoring accuracy requirements, sensor reading cross-verification, dynamic comparison of historical maintenance reports, and minority class amplification are used to correct fault labels and remove outliers, producing cleaned and corrected high-confidence multimodal data;

[0035] The step 1 includes the following:

[0036] Step 101: Multimodal data cleaning

[0037] In order to eliminate the interference of unreasonable values ​​or extreme outliers on the diagnostic model, a cleaning method based on the earth's transport distance is introduced. A corresponding reference distribution is constructed for each sensor mode to measure the degree of deviation between the current observation data and the reference distribution. The specific process is as follows:

[0038] For the For each sensor modality, we first build a baseline distribution based on its historical reliable data. , the distribution can be confirmed by experts or extracted from long-term steady-state operation data; for the single data currently observed , defining the Earth transport distance to measure the overall deviation between the current sensor reading distribution and the reference healthy state distribution: ; measures the overall deviation of the observation from the reference distribution, The larger the value, the more likely this observation is noise or anomaly. Based on the above distance, the reliability coefficient of the observation is defined as:

[0039] ;

[0040] in is the sensitivity coefficient (which can be set based on operation and maintenance experience), It is between 0 and 1, and the higher the value, the more reliable the data;

[0041] When the reliability coefficient Below the preset reliability threshold When the reliability coefficient is When the sensor is within the preset threshold critical interval, it can make a comprehensive assessment based on the reliability of other sensors at the same timestamp and then decide whether to make corrections, and finally output the cleaned data; for example, when a sensor is at the timestamp Reliability Falling into the critical interval When the system first calculates the weighted average reliability of all other sensors at the same time ,like , indicating that the overall data credibility is high, in this case the sensor reading will be eliminated or interpolated; if , the overall credibility is considered low, the original reading is retained and the global abnormal flag is triggered; if , then further statistical reliability is higher than The number of sensors is determined by majority voting to decide whether to modify or retain, thereby making reliable judgments based on multimodal information in the case of a single ambiguous reading.

[0042] Defining 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 modes, while improving data quality and avoiding excessive filtering, through the sensitivity coefficient and threshold , which can provide more room for sensitive detection of different fault characteristics and filtering out false alarms.

[0043] Step 102: Dynamic label correction

[0044] After obtaining the cleaned data output from step 101, step 102 proposes a dynamic correction strategy that integrates expert judgment and prediction distribution comparison for possible missed or incorrect labeling of the same fault type. The process is as follows:

[0045] For any record in the cleaned data set, let its original label be If there is no clear label or the label is missing, the operation and maintenance expert can give a temporary label; select a reference model or expert library with a high correlation with the fault category, and judge the fault type according to the current sensor characteristics to obtain a reference judgment result ,, define the label consistency function:

[0046] ;

[0047] in According to actual needs, it can be designed into advanced criteria such as polynomial distance and symbol error count to measure the difference between two tags; is the adjustment coefficient, the value is greater than 0;

[0048] Obtained label consistency The higher the value, the more consistent the original label is with the reference label. Below the consistency threshold If the record is judged to have label anomalies, manual verification or automatic correction is required; for records judged to be abnormal, if the original label is confirmed to have errors in combination with the operation and maintenance report or on-site maintenance information, the original label will be Replace with judgment result Or adjust to a more appropriate fault type; after the correction is completed, the data is stored together with its new label in a high-confidence multimodal dataset; thereby completing the correction of dynamic labels and ultimately outputting high-confidence multimodal data;

[0049] By comparing labels with reference models or expert libraries, potential mislabeling or omissions can be effectively discovered, especially for rare fault types that are not common in historical samples; comparing and integrating the label correction process with reference models or expert criteria can make label improvement more real-time and flexible.

[0050] Step 2: After obtaining high-confidence multimodal data, combine the autoencoder dimensionality reduction matrix Adjacency Matrix with Graph Neural Networks , through multi-layer mapping functions and Denoise and fuse high-confidence multimodal data, and use anomaly scoring Detect early signs of faults, output abnormal alarms and feature descriptions;

[0051] The second step includes the following:

[0052] Step 201: Feature fusion of deep autoencoder

[0053] In step 201, based on the high-confidence multimodal data output in the first step, a deep autoencoder is used to perform dimensionality reduction and feature fusion to generate a low-dimensional feature vector that can be used for fault precursor identification, where:

[0054] make represents the high-confidence multimodal matrix output from the first step, where is the total sample size, is the multimodal feature dimension after fusion (each modality is cleaned, aligned and then spliced ​​together); is the first value obtained by the autoencoder in the hidden layer The low-dimensional representation of the sample, is the hidden layer dimension ;make and are the encoder and decoder network parameter sets of the autoencoder, respectively; the autoencoder is a deep neural network model used for multimodal feature dimensionality reduction and denoising;

[0055] Introduce the following advanced measurement functions Measure the global difference between the original input and the reconstructed output, let is the decoder output, then the overall cost function of the autoencoder is defined as :

[0056] ;

[0057] in and Respectively represent The original input and reconstruction results of the samples; in order to improve the ability to capture multimodal features, You can choose to use structured correlation or other advanced indicators that can highlight cross-modal differences, such as Mahalanobis distance and canonical correlation coefficient, and weight them.

[0058] Forward propagation: original input Via encoder Mapping to low-dimensional feature vector , followed by the decoder Output reconstruction results ;

[0059] Backward update: Minimize using optimization algorithms such as gradient descent , gradually update the network parameter set of the encoder and the decoder's network parameter set ;

[0060] Convergence judgment: When the overall cost function The training ends when the convergence reaches the preset threshold or the number of iterations reaches the upper limit, and the final mapping function is used to extract the low-dimensional feature vector ;

[0061] After training, each sample can obtain its low-dimensional feature vector in the hidden layer. , these representations integrate the core features of different modal information and are recorded as the low-dimensional feature matrix of the autoencoder ; The matrix Output to step 202 as the basic attribute vector of each node in the subsequent graph neural network construction.

[0062] By measuring the reconstruction differences of multimodal data in a structured way, the information related to each modality can be fully retained; the deep autoencoder achieves denoising and feature fusion while reducing the dimension, which not only reduces redundancy but also effectively retains the weak features that may be contained in rare fault modes. The low-dimensional feature vector output by step 201 is This provides a more robust feature foundation for step 202, significantly improving the accuracy of subsequent graph construction and anomaly detection.

[0063] Step 202: Early anomaly detection based on graph neural network

[0064] In step 202, in order to further capture the potential fault propagation relationship between different sensors or transformer components, the low-dimensional feature matrix of the autoencoder output in step 201 is used. Build a graph neural network to ultimately achieve real-time detection and identification of early signs of failure. The specific process is as follows:

[0065] make Represents a node set, each node Corresponding to a sample in step 201 (or representing a time window in time series); to establish associations between nodes, define the adjacency matrix ,in It can be determined based on the proximity of sensor measurement points, historical correlation or actual physical structure mapping. , then it means the node and There is a connection;

[0066] Node feature matrix , that is, directly use the low-dimensional feature vector output in step 201 As the initial node embedding representation of the graph neural network; let For graph neural network The trainable parameters of the layer, Indicates that the node representation matrix of this layer, then the node In the The representation of a layer can be defined as:

[0067] ;

[0068] in for The set of neighboring nodes, A nonlinear activation function (such as ELU or LeakyReLU) to enhance feature expression capabilities;

[0069] After multiple layers of propagation, the final node representation matrix can be obtained ,remember For nodes The final eigenvector of node (The final embedding vector is recorded as ) Define the comprehensive scoring function ;

[0070] in: express The manifold distance from the embedded manifold formed by the normal state or the main group can be calculated by measuring the nearest point on the manifold (or its approximation), the shortest geodesic length, etc.

[0071] when The larger the value, the greater the deviation of the node from the normal distribution in the high-dimensional feature space;

[0072] is the amplification factor of manifold deviation, which can be set according to operation and maintenance requirements; Indicates that the node and its neighboring node set Weighted feature differences Norm accumulation, where Dimension The weight of is used to highlight or weaken certain key features; or higher; is the balance coefficient of the neighborhood weighted norm;

[0073] Pre-set balance threshold ,like , then the decision node If an abnormality occurs in the corresponding sample (or time window), it will be treated as an abnormal node and trigger an early alarm.

[0074] In addition, the relative threshold can be divided by combining ranking or quantile mechanisms, and all The abnormal node index and the corresponding final feature vector This data is packaged into early anomaly warnings and feature descriptions, which are then fed into the third step: knowledge graph fusion and explainable decision making for subsequent reasoning. If no nodes meet the anomaly criteria, the system is temporarily considered healthy, but the updated node representations are still stored for subsequent cumulative analysis.

[0075] In multimodal dimensionality reduction features The graph neural network is constructed based on the , which not only fully captures the interaction between sensors, but also detects potential fault propagation paths from the perspective of the overall structure. The introduction of the dual metrics of information sparsity norm and relative distribution deviation significantly improves sensitivity to early, atypical failure modes. When using graph neural networks for early anomaly detection, they not only consider the outliers in individual node features but also incorporate the correlation structure between nodes, forming a scalable graph-level anomaly recognition capability.

[0076] Step 3: When receiving early warnings and feature descriptions and finding conflicts with fault types, use the knowledge graph and mechanism matching function Description of features Locate the fault mechanism, output interpretable diagnostic results and consistency assessment results, and mark suspected new fault nodes;

[0077] The step three includes the following:

[0078] Step 301: Knowledge graph matching and fault mechanism location

[0079] In step 301, based on the early abnormality warning and feature description output in step 202, combined with the power equipment fault knowledge graph constructed by this solution, matching and positioning at the mechanism level are performed, wherein:

[0080] remember is the fault knowledge graph, where Represents a node set in the knowledge graph (e.g. winding overheating, core grounding fault, oil concentration increase, etc.), Represents the relationship between nodes (such as cause-effect or feature-fault type); each node It comes with various attributes, such as the weight associated with changes in gas composition and the corresponding relationship with historical maintenance conclusions;

[0081] make Indicates the suspected fault node Output feature description, where is the dimension after uniformly encoding the information such as sensor readings, graph embedding features and anomaly scores; The category label indicating that the fault node (or moment) is initially identified as a suspected fault (if it is only an anomaly without a specific classification, it is recorded as unknown);

[0082] Introduce the following mechanism matching score , used to quantize nodes and feature description The degree of compliance between:

[0083] ;

[0084] in: Knowledge graph node The corresponding reference fault feature vector (which summarizes the common gas concentration distribution, temperature threshold, topological association, etc. under this fault mechanism);

[0085] It is a high-order similarity metric function, which can be combined with isomorphic embedding or differential geometry tools to measure feature descriptions. and the reference fault feature vector distance on the characteristic manifold; According to the fault type-mechanism node relationship in the knowledge graph With preliminary categories A penalty term that penalizes the deviation between For nonlinear mapping (such as Softplus and ELU), the difference is transformed into a final score, where a higher score indicates a better match;

[0086] For each abnormal node , in the knowledge graph node set Search and calculate The maximum value of the optimal mechanism to locate the node :

[0087] ;

[0088] Position nodes according to the optimal mechanism The most likely corresponding fault mechanism node can be locked and the background knowledge associated with it (such as common causes and recommended maintenance items) can be obtained; the result is then passed to step 302 for diagnostic interpretation and visualization.

[0089] Matching scores by mechanism The construction of deep learning effectively connects the statistical features extracted by deep learning with the professional mechanism information in the knowledge graph to avoid black box diagnosis; under the global relationship constraints of the knowledge graph (constrained by the penalty term When certain anomalies are detected but the categories are unclear, it can also make detailed inferences with the help of the most similar mechanism nodes and provide rich background information for subsequent explanations.

[0090] Step 302: Explainable Reasoning and Decision Output

[0091] After the fault mechanism is located in step 301, step 302 further locates the optimal mechanism node By linking it with specific sensor anomaly information, it outputs interpretable diagnostic conclusions for operation and maintenance personnel, and verifies them with actual operation experience.

[0092] For each node that has been located to the optimal mechanism Abnormal nodes Collect nodes in the knowledge graph that are related to the optimal mechanism The set of connected feature-fault or component-mechanism relationship edges , such as oil Significantly increased Winding overheating risk or CO concentration and temperature exceeding the limit Multiple points of overheating are possible; the above structure is combined with the feature description output from step 202 Mapping, such as the current When the concentration increases by Xppm, the CO concentration increases by Yppm, and the potential mechanism pathway is displayed in natural language or graphical interface. For example: due to the characteristic description Positioning nodes with the optimal mechanism The similarity is high, and the H2 and CO in the oil exceed the threshold at the same time, which is inferred to be the risk of winding overheating or multi-point overheating.

[0093] Defining a failure consistency function The degree of consistency between the diagnostic conclusion and historical precedents or expert rules is measured. If there is a significant difference with the existing operation and maintenance records, the operation and maintenance experts can be prompted to conduct a second review. This can be formalized as follows:

[0094] ;

[0095] in: Score the anomaly defined in step 2 (or its modified version), Represents the optimal mechanism for locating nodes in the knowledge graph Related typical anomaly scoring thresholds;

[0096] If the fault consistency function The value is lower than expected, indicating that the final eigenvector The degree of abnormality does not match the reference level of the located mechanism in the knowledge graph and requires a second review;

[0097] Finally, interpretable diagnostic results and consistency assessment are generated and output, including:

[0098] Optimal mechanism positioning node Its corresponding typical causes and recommended maintenance steps in the knowledge graph; and the corresponding core abnormality feature description;

[0099] Fault consistency score , and a flag indicating whether expert review or further on-site testing is required;

[0100] The output is passed to the fourth step of online model updating and transfer learning, so that the knowledge graph and model parameters can be updated in time after receiving actual maintenance results or new fault samples.

[0101] Step 302 provides complete principle-to-feature traceability based on the fault mechanism location in step 301, and uses the fault consistency score Quantify the credibility of the diagnosis; through the relationship edges of the knowledge graph and the specific mechanism nodes that have been matched, operation and maintenance personnel can directly see the reasoning logic behind the data, thereby improving the trust in the system output and providing a clear direction for subsequent improvements or troubleshooting. Compared with the typical anomaly scoring threshold based on mechanism prior in knowledge graph Fault consistency score By combining the two functions, a comprehensive judgment of statistical anomaly and mechanism adaptability is achieved. The fault explanation can be visualized by using a graph database or inference engine, and the mechanism node association information can be dynamically called, which can enable the decision-making layer to produce deeper and scalable explainability.

[0102] Step 4: When an interpretable diagnostic result is generated and actual operation and maintenance conclusions Produce significant differences, based on the degree of difference and incremental learning rate Modify the model and integrate relationship weights , and finally dynamically update the knowledge graph Form a self-learning closed loop with multimodal data;

[0103] The step 4 includes the following contents:

[0104] Step 401: Incremental correction and difference detection

[0105] In step 401, the interpretable diagnostic results and consistency assessment output from step 3 are compared with the actual operation and maintenance conclusions (e.g., subsequent maintenance reports, inspection results) to identify possible deviations in the model and perform incremental corrections, where:

[0106] Recorded as The third step is the final judgment result of the diagnostic model on the fault / health status. Indicates the actual operation and maintenance conclusion (such as the fault diagnosis type); if Or if the diagnostic deviation is large, it is considered that the current model parameters have obvious errors in this fault scenario; all error sample records are collected into a misjudgment sample set , for subsequent incremental learning.

[0107] To measure the difference between the actual output of the model and the true conclusion in high-dimensional space, a high-order difference metric can be defined ,For example: Indicates that the system predictions given by samples (e.g., failure probability distribution, key point sets of feature vectors over time, etc.), The vector representation of the actual operation and maintenance result (which can be obtained by the embedding / alignment method used in the first, second or third step)

[0108] remember is a nonlinear mapping function from the input space to the manifold embedding space, which has been defined and trained in the second or third step, where: This is the encoder of the autoencoder in the second sub-step 201 ——That is, the part that dimensional multimodal input is compressed to dimensional latent space network. It is defined and trained in the following situations:

[0109] Training phase: In the multimodal feature extraction phase of sub-step 201, the reconstruction cost of the autoencoder is minimized:

[0110] ;

[0111] Iteratively update the encoder parameters simultaneously (Decoder training simultaneously).

[0112] After training converges, It becomes a mapping , which can transform any new sample Projected into the same latent space , for the fourth step high-order difference measurement Use; therefore, It is not a fixed offline function library call, but a nonlinear embedding mapping learned during the joint training of the autoencoder that can be directly reused after the second step is completed;

[0113] Define a higher-order difference metric function as follows:

[0114] ;

[0115] in: Define a straight line interpolation path on the original input space:

[0116] ;

[0117] when hour, (i.e. the real result corresponds to the vector); when hour, (i.e., the prediction result corresponds to the vector), any Can be passed Mapping into manifold space;

[0118] Representation Mapping At the point The Jacobian matrix at , whose size is , which can be understood as The set of partial derivatives of the input components at this point; different Corresponding to different ,therefore Follow However, dynamic changes can more accurately characterize the sensitivity to input increments along the path;

[0119] Use some advanced norm or measure (Frobenius norm, 2-norm or other designed high-order norms can be used to measure The size of the operation is actually measured when the input is Chuyan The direction is slightly shifted, and the difference produced after mapping to the manifold space is determined, so that the difference evaluation has more differential geometry characteristics.

[0120] When the difference Exceeding the difference threshold If the difference is Below the difference threshold However, if there is still a certain deviation, the corresponding parameters of the model will be reduced or corrected according to the weight.

[0121] From the misclassified sample set Extract several small batches of samples from the algorithm and perform small-scale gradient updates using the loss functions originally defined in this scheme (such as the autoencoder reconstruction loss, the graph neural network anomaly detection loss, and the combination of the mechanism matching loss in the third step);

[0122] Setting the online learning rate , balance the influence of historical model knowledge and the introduction of new samples, ensure that the update process is neither overfitting nor correcting obvious errors in time; after several small batch iterations, the corrected model parameter set is obtained The model is combined with the knowledge graph in the third step to complete the new fault judgment. Specifically, after the model is updated, through incremental learning (based on the misjudgment sample set ) get the latest autoencoder parameters and graph neural network parameters Deploy to the same inference pipeline in steps 2 and 3;

[0123] For the real-time multimodal input to be diagnosed, the new autoencoder is first used to map the hidden vector , and then use the new GNN parameters in the knowledge graph Complete message passing in the constructed graph structure to obtain the final node embedding ;

[0124] Will With pre-built mechanism nodes By matching function or consistency assessment Compare and select the new optimal mechanism node And output interpretable diagnostic results;

[0125] 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 integrated with the knowledge graph in the third step, ensuring that the latest learning results are used to make more accurate and explanatory fault judgments on the equipment status.

[0126] Through the difference A comprehensive comparison of diagnostic outputs and real results in manifold space can timely capture errors in complex dimensions, rather than relying solely on whether a single label matches. The incremental learning method avoids the need for large-scale full retraining every time new data is acquired, improving the efficiency and scalability of model updates. The online learning rate It can be dynamically adjusted according to the fault type and data characteristics to form differentiated priority correction strategies for different rare faults.

[0127] Step 402: Transfer learning strategy and cross-device knowledge sharing

[0128] In step 402, to accelerate the adaptation of the model to new environments or different transformer equipment, a transfer learning strategy is introduced to effectively transfer the diagnostic experience learned in other equipment or historical working conditions to the current scenario;

[0129] This step is performed after the incremental correction is completed in step 401 to ensure that the corrected model parameter set Form a good connection with external auxiliary knowledge: Model parameters pre-trained on the source device or source environment (may include autoencoder weights, graph neural network structure, knowledge graph association strength, etc.), The model parameters that need to be optimized for the target device. The initial value is ;

[0130] By detecting sensor distribution, operating time, fault type frequency, etc., the similarity level between the source and target is determined as follows: First, the probability distribution of the key sensors of the source and target devices (such as oil temperature, gas concentration, vibration, etc.) is constructed based on the historical operating data. The similarity of each modal distribution is calculated using the normalized earth transport distance or KL divergence. At the same time, the load duration distribution of the two devices is statistically analyzed and the degree of working condition alignment is evaluated using cosine similarity. The fault type frequency vector is then used for similarity measurement. Finally, the overall similarity score is synthesized in a weighted manner. When the score exceeds a preset threshold (e.g., 0.85), the system determines that the operating environment and fault characteristics of the source and target domains are highly consistent, and can safely perform high-weight migration of graph neural network parameters and autoencoder weights.

[0131] If the similarity is high, a large proportion of pre-trained model parameters can be introduced ; If the similarity is low, selective parameter fusion is required; at the autoencoder level, the encoder part (such as Initialize the weights, the decoder can do partial retention or random reset; at the graph neural network level, the relationship weights Perform weighted fusion:

[0132] ;

[0133] in, It represents the migration trade-off coefficient, the larger it is, the more trust is placed on the source device experience; The graph neural network pre-trained on the source device All trainable parameters of the layer (including weight matrix and bias vector), representing the feature extraction and transfer rules learned under the source domain fault mode; The graph neural network of the target device before migration Layer parameters reflect the existing model status of the target device's own data; the symbols on the left side of the equal sign refer to the new parameters after migration;

[0134] In the knowledge graph mapping part, based on the mechanism node mapping relationship mentioned in the third step, high-similarity fault mechanism nodes can be directly incorporated into the target graph, or the relationship weights can be incrementally updated as follows:

[0135] In the knowledge graph mapping phase, the newly discovered or migrated fault mechanisms in the source graph will be integrated into the target graph according to the following process: First, for each source domain mechanism node located in the third step, Calculate the vector embedding similarity (such as cosine similarity) with all the mechanical nodes in the target graph to obtain the corresponding similarity score ; Then for each source node, if the highest similarity (for example ), 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:

[0136] ;

[0137] This allows for a smooth fusion of new and old knowledge. If the similarity falls below a preset threshold, the source mechanism is considered absent from the target graph. Accordingly, a new node is added, and the typical edge (e.g., cause → effect) and its weight are initialized. This allows the target graph to retain the original expert knowledge structure while dynamically incorporating new fault mechanisms or adjusting edge strengths, achieving adaptive knowledge expansion and optimization.

[0138] The fused model parameters Perform a small batch online correction and compare it with the set of misjudged samples identified in step 401 Compare and contrast. If the performance is significantly improved after migration, retain the new parameters. If conflicts occur or performance degrades, 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.

[0139] Through hierarchical migration and parameter reorganization, the degree to which different network parts inherit external experience can be finely controlled, accelerating the convergence of the model under different working conditions. The migrated model parameters output in step 402 can be further combined with the actual data of this device, significantly shortening the learning cycle for rare fault types. Combined with the online update in step 401, a two-pronged optimization path of local increment + external migration is formed, taking into account both accuracy and generalization capabilities.

[0140] Step 403: Dynamic expansion of knowledge graph and feedback loop

[0141] 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 inherits the migration adjustment results completed in step 402 and synchronizes its results back to the key data structures required in the first, second and third steps.

[0142] 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 new fault type or new combination feature can be found in the knowledge graph. Insert the corresponding new fault mechanism node or update the existing relationship edge;

[0143] Defining a 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.

[0144] For newly added sensor data or fault labels, we need to return to the first step (multimodal data cleaning and dynamic label correction) for further screening and labeling. The new fault features are then incorporated into the graph neural network or autoencoder training data from the second step (multimodal feature extraction and early anomaly detection). This allows for the timely detection of new faults of the same type during subsequent runtime.

[0145] When entering the third step again to make explainable decisions, the knowledge graph already has more updated fault mechanisms and related edges, which can improve the precision and credibility of the next round of diagnosis; if operation and maintenance conclusions continue to be collected, the iterative cycle of steps 401, 402, and 403 can continue, forming a closed-loop process of self-learning and self-evolution.

[0146] Mapping Function Used to integrate newly discovered fault mechanism nodes with the target knowledge graph Comparison of existing nodes in the graph: First, calculate the cosine similarity of the feature vectors of the newly discovered fault mechanism node and each node in the graph. If the highest similarity is not lower than the preset threshold (for example, 0.8), the newly discovered fault mechanism node is considered to be the same as the node, and only the relationship edge weight is updated smoothly. Otherwise, a new node will be added to the graph and its connection with the relevant mechanism node will be initialized. In this way, the mapping function It can not only maintain the structural consistency of the knowledge graph, but also dynamically absorb new failure mechanisms to achieve orderly expansion of the knowledge base.

[0147] By incrementally updating and automatically clustering the knowledge graph structure, the system can continuously absorb new fault modes and incorporate them into the mechanism layer explanation; the dynamically updated knowledge graph output in step 403 and the synchronously updated multimodal data will ensure that the latest fault and mechanism information can be directly used when the first, second, and third steps are executed next time, greatly improving the overall adaptation speed of the system. The knowledge graph is connected with the data preprocessing and model training links to form a closed-loop update, which goes beyond the traditional static graph + offline model separation method; through the mapping function Automatically inserting new faults into the correct locations reduces the burden of manual maintenance of large maps and improves scalability in ultra-large multimodal monitoring scenarios.

[0148] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0149] Those skilled in the art will 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 aforementioned method embodiments and will not be repeated here.

[0150] In the several embodiments provided in this 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 schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as 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 mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0151] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0152] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A multi-modal power equipment health status monitoring method, characterized by: include, When multimodal data is detected to contain noise or have imbalanced labels, the faulty labels are corrected and outliers are removed to produce cleaned and corrected high-confidence multimodal data. After acquiring high-confidence multimodal data, the system combines the autoencoder dimensionality reduction matrix with the graph neural network adjacency matrix to perform denoising and fusion on the high-confidence multimodal data through a multi-layer mapping function. It also uses anomaly scoring to detect early signs of failure and outputs early anomaly warnings and feature descriptions. When receiving early abnormality alarms and feature descriptions and finding conflicts with the transformer fault type, the knowledge graph and mechanism matching function are used to locate the mechanism of the feature description, output interpretable diagnosis results and consistency assessment results, and mark the new fault node; If there is a discrepancy between the explainable diagnostic results and the actual operation and maintenance conclusions, the graph neural network model is modified based on the difference and incremental learning rate, and the relationship weights are integrated to dynamically update the knowledge graph and multimodal data to form a self-learning closed loop. Through the graph neural network, multi-layer propagation is performed on the graph structure formed by transformer components or monitoring points. Based on the adjacency matrix and anomaly score, early suspected fault nodes are determined, and early anomaly warnings and feature descriptions are generated. Among them, if the comprehensive score function is greater than the balance threshold, the sample or time window corresponding to the node is judged to be abnormal and it is regarded as an abnormal node. When dynamic label correction is performed based on cross-comparison of sensor readings and historical maintenance reports, a benchmark distribution and credibility coefficient constructed from historical samples are used to perform a secondary screening of suspicious data labels to correct missed or incorrect labels and generate highly reliable multimodal data. When using small sample learning or synthetic oversampling to perform limited amplification on scarce fault data, over-simulated fault samples are filtered in combination with expert rule constraints, where: When the label consistency is lower than the threshold, it is determined that there are label anomalies in the records in the cleaned dataset. After confirmation, adaptive adjustments are made and the corresponding data together with the labels are stored in a high-confidence multimodal dataset.

2. The multi-modal power equipment health status monitoring method according to claim 1, characterized in that: When using a deep autoencoder to perform dimensionality reduction and feature fusion on multi-dimensional sensor data, the hidden layer parameters of the encoder-decoder network are used to denoise the features of each modality to obtain the low-dimensional feature vector of the sample in the hidden layer, which is then aggregated to generate a low-dimensional feature matrix of the autoencoder.

3. The multi-modal power equipment health status monitoring method according to claim 2, characterized in that: For early abnormal alarms and feature descriptions, the similarity between the fault mechanism node and the feature description is used to construct a mechanism matching score. The fault node is located in the knowledge graph based on the mechanism matching score, and marked when the match is uncertain.

4. The multi-modal power equipment health status monitoring method according to claim 3, characterized in that: After obtaining the fault mechanism location results, the system uses consistency assessment to measure the credibility of the diagnostic conclusion and outputs interpretable diagnostic results and consistency assessment results. If the credibility is judged to be insufficient, it automatically triggers manual intervention or the collection of additional monitoring data. Output the optimal mechanism positioning node and its corresponding typical causes in the knowledge graph, recommended maintenance links, corresponding core abnormality feature descriptions and fault consistency scores.

5. The multi-modal power equipment health status monitoring method according to claim 4, characterized in that: Based on the difference detection between the diagnostic output and the actual operation and maintenance conclusion, if the difference between the two exceeds the difference threshold, the misjudgment samples will be included in incremental learning and small-batch correction will be performed in combination with the online learning rate. The corrected model parameters will be output to continuously optimize the diagnostic performance, select the new optimal mechanism node and output an explainable diagnostic result.

6. The multi-modal power equipment health status monitoring method according to claim 5, characterized in that: After completing the incremental correction of the local model, the empirical parameters of other transformer devices are inherited between the relational weights of the graph neural network or the autoencoder structure through layered migration to accelerate the model convergence in new environments or rare fault modes.

7. The multi-modal power equipment health status monitoring method according to claim 6, characterized in that: By incrementally updating and automatically clustering the knowledge graph, combined with the mapping function, new fault nodes and their relationship edges are automatically inserted into the existing mechanical network. After backtracking, the newly added labels and multimodal data are re-cleaned and re-detected to complete the closed-loop self-learning update.

Images