Transformer fault diagnosis method and system based on multi-source data and root cause analysis

Abstract

The invention discloses a transformer fault diagnosis method and system based on multi-source data and root cause analysis, and relates to the field of fault diagnosis, and the method comprises the specific steps: obtaining multi-source monitoring data in the operation process of a transformer, the multi-source monitoring data comprising position-time data, physical parameter data and record-behavior data; performing feature extraction on the multi-source monitoring data, and fusing the extracted features to obtain fused features; and inputting the fusion features into a trained fault discrimination model, outputting a fault classification result, and determining a root cause path from a preset candidate fault chain based on probabilistic reasoning. According to the method, probability weighted correction is carried out on the preliminary judgment result by adopting the confidence mapping function, so that the diagnosis output not only comprises the fault category, but also provides the confidence index to greatly improve the accuracy of the output result, and the conditions of fuzzy and conflict diagnosis are reduced.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis, and in particular to a method and system for diagnosing transformer faults based on multi-source data and root cause analysis. Background Technology

[0002] Currently, in the manufacturing field, especially in the fault diagnosis of power transformers, traditional root cause analysis (RCA) methods rely heavily on human experience and are inefficient when processing exponentially increasing manufacturing data. Existing automated root cause analysis (ARCA) methods mostly rely on only a single type of data, such as using only location-time data to locate faulty equipment or relying only on physical parameter data to detect anomalies, resulting in incomplete root cause analysis.

[0003] In the scenario of power transformer fault diagnosis, location-time data is used for spatiotemporal fault location, physical parameter data is used for abnormal status detection, and record-behavioral data provides operation and maintenance management information. However, the fragmentation of data makes it difficult for the analysis to go from phenomenon diagnosis to the root cause of management.

[0004] Therefore, how to develop a method to integrate multi-source data to achieve in-depth analysis from phenomena to root causes is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a transformer fault diagnosis method and system based on multi-source data and root cause analysis, which overcomes the above-mentioned defects.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a transformer fault diagnosis method based on multi-source data and root cause analysis, the specific steps of which are as follows: Acquire multi-source monitoring data during transformer operation, including location-time data, physical parameter data, and record-behavior data; Feature extraction is performed on the multi-source monitoring data, and the extracted features are fused to obtain fused features; The fused features are input into the trained fault discrimination model, the fault classification results are output, and the root cause path is determined from the preset candidate fault chain based on probabilistic reasoning.

[0007] Optionally, the feature extraction step of the physical parameter data is as follows: The physical parameter data is normalized to generate normalized data; The normalized data is converted into an image using Gram angular field. The deep feature extraction module is used to extract deep features from the converted image; The deep features are filtered based on a two-layer feature optimization strategy, and a physical parameter feature set is constructed based on the filtered deep features.

[0008] Optionally, the feature extraction step of the record-behavioral data is as follows: Based on the power named entity recognition model, triples are constructed using key elements extracted from the record-behavior data; the key elements include the device object, action type, and event result. Based on the cross-document knowledge modeling mechanism, learn the semantic or logical relationships between different documents and model them to construct a document relationship graph; The triples are mapped to entities and relations in a knowledge graph to form an entity-relation-entity chain. The entities and relations are then converted into vector representations using a graph embedding algorithm. These vector representations are then fused with cross-document semantic vectors obtained based on the document association graph to generate record-behavioral features.

[0009] Optionally, the fault discrimination model includes an input layer, multiple hidden layers, and an output layer, wherein the hidden layers introduce an improved self-attention mechanism, and the output layer outputs the probability distribution of fault categories.

[0010] Optionally, the principle of the improved self-attention mechanism is as follows: The fused features are divided into multiple subspaces, and the attention scores of each subspace are calculated in parallel. Differential attention scores are calculated based on the attention scores of each subspace, and the features of each subspace are weighted and fused based on the differential attention scores.

[0011] Optionally, the step of obtaining the candidate failure chain is as follows: Identifying abnormal events in historical multi-source monitoring data based on a sliding time window and threshold adaptive mechanism; Calculate the connection scores between each anomalous event based on chronological order, spatial correlation, and causal dependency strength. The abnormal events whose connection scores exceed a preset threshold are connected sequentially to generate the candidate fault chain.

[0012] Optionally, a feature enhancement mechanism is also introduced during the root cause path acquisition process, specifically: The comprehensive similarity is calculated based on feature distance, temporal morphological similarity and event semantic consistency, and similar cases are extracted from the historical fault database based on the comprehensive similarity. Feature extraction is performed on the similar cases, and the extracted similar features are fused into the current feature vector to obtain enhanced discriminative features. The enhanced discriminative features are then used to determine the root cause path from the preset candidate fault chain.

[0013] Optionally, a chain pattern constraint is also introduced when generating candidate failure chains, specifically: Establish a causal knowledge base based on the operating principles and historical failure evolution of the equipment; The generated initial candidate fault chains are matched with the causal knowledge base, and the initial candidate chains are filtered, corrected or eliminated based on similarity to obtain the candidate chains.

[0014] Optionally, an adaptive confidence mechanism is used during root cause analysis, specifically: When the highest predicted probability of any candidate chain exceeds the first threshold, any candidate chain is output as a single diagnostic result. When the highest predicted probability of any candidate chain is between the first threshold and the second threshold, multiple candidate chains are retained, and the multiple candidate chains are weighted and corrected to output the diagnostic result.

[0015] Secondly, this application provides a transformer fault diagnosis system based on multi-source data and root cause analysis, including: The data acquisition module is used to acquire multi-source monitoring data during transformer operation. The multi-source monitoring data includes location-time data, physical parameter data, and record-behavior data. The feature extraction module is used to extract features from the multi-source monitoring data and fuse the extracted features to obtain fused features; The fault diagnosis module is used to input the fused features into the trained fault discrimination model, output the fault classification result, and determine the root cause path from the preset candidate fault chain based on probabilistic reasoning.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: 1. By fusing multi-source data, the accuracy of fault classification was improved; 2. A confidence mapping function is used to perform probability-weighted correction on the preliminary judgment results, so that the diagnostic output not only includes the fault category, but also provides a confidence index, which greatly improves the accuracy of the output results and reduces the situation of fuzzy and conflicting diagnoses. 3. It can introduce traditionally underutilized operation and maintenance management data (such as historical maintenance and defect records) into the modeling process, thereby achieving information gain, improving the robustness of the diagnostic model under insufficient sample conditions, and enhancing the stability index of the model under small sample conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a method flow provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] This embodiment discloses a transformer fault diagnosis method based on multi-source data and root cause analysis, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Acquire multi-source monitoring data during transformer operation. Multi-source monitoring data includes location-time data, physical parameter data, and record-behavior data. Step 2: Extract features from multi-source monitoring data and fuse the extracted features to obtain fused features; Step 3: Input the fused features into the trained fault discrimination model, output the fault classification results, and determine the root cause path from the preset candidate fault chains based on probabilistic reasoning.

[0022] In one embodiment, the set of monitoring parameters (including multi-source monitoring data) acquired by the online monitoring device during the operation of the transformer is defined as: ; In the formula, For the first time series data collected... Several monitoring parameters, such as oil temperature, load current, and winding temperature rise. Monitoring parameter set. This provides a data foundation for subsequent feature extraction and modeling.

[0023] Furthermore, the three different types of data have different methods of collection and representation.

[0024] The location-time data is represented as follows: ; In the formula, For location-time data sets; For the first The device location is recorded in the record; It is the corresponding timestamp; This represents the total number of location-time data. For the set of device locations; Let be the set of real numbers.

[0025] The physical parameter data is represented as follows: ; In the formula, A collection of physical parameter data; For the first A vector of physical parameters (such as gas concentration, multi-channel partial discharge signal); The number of physical parameter samples; The dimension of the physical parameters.

[0026] Records – behavioral data, represented as follows: ； In the formula, For recording – a collection of behavioral data; For the first Vector representation of each maintenance / operation record; To record feature dimensions Indicates the number of record entries.

[0027] In this embodiment, data acquisition encompasses three aspects: location-time data, physical parameter data, and record-behavioral data, ensuring comprehensive and multi-dimensional location-time information sources for fault diagnosis. Location-time data primarily records the transformer's installation location (e.g., substation number, bay location) and the timestamp of the fault occurrence, enabling rapid location of faulty equipment and supporting spatiotemporal distribution analysis of fault modes. For example, it can identify patterns of frequent faults in transformers within a specific substation bay during the rainy season. Physical parameter data includes operational monitoring indicators such as oil chromatography (DGA), partial discharge quantity, winding temperature, and insulation resistance. This type of data directly reflects the equipment's health status and can be used for anomaly detection and initial fault type assessment. For example, a sudden increase in acetylene (C2H2) content can diagnose discharge-related faults, while a sustained abnormal rise in oil temperature suggests potential overheating. Record-behavioral data covers equipment maintenance records, operation logs, and historical defect ledgers, carrying management information throughout the equipment's entire lifecycle. By comparing and analyzing physical parameters, this type of data can reveal the management-related root causes behind operational anomalies. For example, it can link the deterioration of insulating oil to the failure to replace oil within its expiration period, or corroborate bushing discharge with the failure to conduct special inspections during the rainy season. Based on the comprehensive collection and correlation mining of these three types of data, a step-by-step progression from "operational phenomenon diagnosis" to "management root cause analysis" can be achieved, providing complete information support for the accurate diagnosis and root cause tracing of transformer faults.

[0028] In one embodiment, to enhance the characterization capability of different operating conditions, the original set of monitoring parameters is... The input is fed into the feature mapping function to obtain the feature vector: ; In the formula, For the first Each feature component is generally obtained by extracting the aforementioned data using a feature function. These features can be statistical characteristics (mean, variance), frequency domain characteristics (harmonic components), or time-frequency characteristics (wavelet energy, etc.). Feature vector As input for fault diagnosis.

[0029] In one embodiment, location-time data mainly includes transformer installation location, substation number, interval information, and fault occurrence time. This embodiment extracts temporal and spatial features using a multi-scale spatiotemporal analysis method. In the time dimension, the mean time between failures (MTBF) is first calculated to reflect the stability of the equipment over different operating cycles. The expression for calculating the mean time between failures is: ; In the formula, For the first The timestamp of the fault; For the first The timestamp of the fault; This represents the total number of failures. This feature reflects the operational stability of the equipment over time.

[0030] In the spatial dimension, a spatial distribution model is constructed based on geographic coordinates or electrical topology location, using geographic coordinates as the basis. Construct a spatial adjacency matrix Extract spatiotemporal joint features and calculate spatial clustering: ; It can measure the concentration trend of faults in geospatial space. Mean time between failures and spatial clustering describe the evolution of faults in the temporal and locational dimensions, providing an input basis for subsequent root cause correlation.

[0031] In one embodiment, the feature extraction step of the physical parameter data is as follows: The physical parameter data is normalized to generate normalized data; Normalized data is converted into an image using Gram angular field; The deep feature extraction module is used to extract deep features from the converted image; A two-layer feature optimization strategy is used to filter deep features, and a physical parameter feature set is constructed based on the filtered deep features.

[0032] Furthermore, physical parameter data includes oil chromatography (DGA), partial discharge quantity, winding temperature, insulation resistance, and vibration signals. This embodiment employs a fault diagnosis method combining image-based feature extraction and dual-layer feature selection, aiming to achieve more efficient and robust feature construction and screening. Specifically: First, the various physical monitoring signals are normalized to convert the original sequences into normalized data. Mapping to interval To eliminate dimensional differences, the Gramian Angular Field (GAF) method was then used to transform the time series data into an image format. The specific steps include: (1) Polar coordinate mapping: ; In the formula, Indicates the polar angle of the data point; Indicates the first i One original feature; This represents the normalized time index.

[0033] (2) Construction of Gram matrix: ; This matrix reflects the phase correlation between different gas parameters and can preserve the time-dependent structure and amplitude variation trend of the signal.

[0034] Will The matrix is ​​linearly scaled to the [0, 255] range to generate a grayscale image, and the sample is expanded by means of brightness perturbation, additive noise and Gaussian blur to enhance the robustness of the model to abnormal signal perturbation.

[0035] Subsequently, the system employs a deep feature extraction module based on the VGG16 transfer learning network. The convolutional layer weights obtained from ImageNet training are retained, and only the classification layer structure is replaced, enabling the network to automatically learn deep spatial features in transformer DGA images. The extracted intermediate layer feature vector is denoted as: ; This vector comprehensively reflects the implicit coupling relationships between various gas parameters.

[0036] To mitigate the redundancy and overfitting risks associated with high-dimensional features, a two-layer feature optimization strategy is further introduced. In the two-layer feature optimization stage, the system sequentially performs two levels of screening to improve the effectiveness of the feature set. The first layer uses the Random Forest (RF) algorithm to calculate the importance score of each feature, selecting a subset of features with high importance by evaluating the rate of change of out-of-bottle (OOB) error, and removing redundant features. The second layer, based on Pearson correlation analysis, re-screens the features initially rejected, selecting features with high correlation to the retained features and potential contributions as supplementary features, forming the final optimal feature set (i.e., the physical parameter feature set). This process can be represented as: ; In the formula, The feature subset selected by the random forest; Pearson correlation coefficient; This is the correlation threshold.

[0037] Finally, the feature vectors extracted and filtered through the above process constitute the final high-dimensional representation of the DGA features: ; In the formula, This is a deep feature extraction operator based on transfer learning; For Gram corner field conversion operation; This represents the feature subset obtained after two-layer optimization. This feature representation can effectively integrate the temporal patterns and correlation structures of gas changes, providing highly recognizable input information for subsequent fault diagnosis and chain construction.

[0038] Besides oil chromatography analysis, this step also focuses on extracting time-domain, frequency-domain, and time-frequency features. In the time domain, statistics such as mean and variance are calculated to capture the overall trend of the signal. The expression for calculating the time-domain signal information is as follows: ; In the formula, For a moment The monitoring value, The mean, Variance is used to characterize the stationarity and volatility of a signal.

[0039] At the frequency domain level, features such as the dominant frequency component, harmonic amplitude, and spectral energy distribution are obtained through Fast Fourier Transform (FFT) to detect periodic anomalies. The spectrum is calculated first, followed by the dominant frequency energy. The formula is: ; ; In the formula, For frequency components, This represents the corresponding energy intensity.

[0040] In one embodiment, the feature extraction step for record-behavioral data is as follows: The power named entity recognition model constructs triples using key elements extracted from record-behavior data; the key elements include device object, action type, and event result. Based on the cross-document knowledge modeling mechanism, learn the semantic or logical relationships between different documents and model them to construct a document relationship graph; The triples are mapped to entities and relations in the knowledge graph to form an entity-relation-entity chain. The entities and relations are then converted into vector representations using a graph embedding algorithm. These vector representations are then fused with cross-document semantic vectors obtained from the document association graph to generate record-behavioral features.

[0041] Furthermore, log-action data, including unstructured information such as maintenance logs, operation and maintenance records, and historical defect ledgers, is first processed using text parsing and semantic encoding methods to extract structured features. Addressing the unstructured nature of log-action data, this embodiment employs a knowledge graph (KG)-driven fusion method. The system transforms the implicit operation and maintenance events, device objects, and operational relationships in the logs into structured knowledge entities and relationship chains through text parsing, semantic extraction, and graph structure mapping, thereby achieving a unified representation of operation and maintenance semantic information, physical monitoring data, and location-time data. Specifically, this involves four implementation steps: event and entity extraction, graph construction and fusion mapping, graph embedding and feature fusion, and dynamic updating and incremental learning.

[0042] (1) In the event and entity extraction, this embodiment extracts key elements from operation and maintenance records and maintenance logs, such as "equipment objects" (e.g., main transformer, oil pump), "action types" (e.g., replacement, shutdown, oil filling), and "event results" (e.g., recovery to normal, detection of anomalies). After the text content is preprocessed (e.g., removal of stop words, word segmentation, and key event identification), a BERT-based named entity recognition model is used to output a triple of (equipment object, action type, result). .

[0043] Because the general-purpose BERT model has limited applicability in the power operation and maintenance field, the system is fine-tuned and trained based on domain-specific corpora (including maintenance records, inspection logs, fault reports, etc.) to construct a dedicated model with power semantic understanding capabilities (i.e., a power named entity recognition model). Its feature extraction process can be represented as follows: ; In the formula, For a domain-fine-tuned language model, The input text is after sentence segmentation, semantic parsing, and encoding.

[0044] (2) In the cross-document semantic association modeling stage, the cross-document knowledge modeling mechanism of LinkBERT is used as a reference. The operation and maintenance log corpus is regarded as a "document graph" (i.e. a document association graph), in which nodes correspond to various operation and maintenance records, maintenance reports and defect ledgers, and edges represent semantic or logical associations between documents (such as the same device, multiple maintenance or operation events within adjacent time windows).

[0045] During input construction, the model not only uses text fragments from single records, but also randomly samples fragments from related documents to form the input format: ; in, This represents a user's natural language query or the main text fragment to be understood. This corresponds to supplementary information, such as candidate SQL fragments, database table schema descriptions, or contextual constraints. The input sequence begins with a special marker [CLS], which, after encoding, aggregates the semantic representation of the entire sequence and is used for subsequent classification or matching tasks. Two text segments are separated by [SEP] to explicitly prompt the model to distinguish different semantic domains during attention computation. The final sequence ends with a second [SEP], used to construct a complete input structure that can be processed by the Transformer encoder. A Document Relation Prediction (DRP) objective is introduced in the pre-training task to determine the relationship type between fragments (same document, random document, related document). This mechanism enables the model to learn contextual connections across logs, capturing potential causal and temporal patterns. For example, when "abnormal main transformer oil temperature" and "cooling pump shutdown" belong to different logs but have semantic links, the model can automatically learn the potential operational dependencies between them.

[0046] (3) In the knowledge graph mapping and embedding stage, the extracted triples are mapped to graph structure nodes and relations, forming an "entity-relationship-entity" chain. The TransE embedding algorithm maps the entities and relations in the graph to a vector space, realizing the vectorized representation of action semantics. ; In the formula, Indicates the first Each identified entity corresponds to a device name, component identifier, or operating object in the power grid scenario; This represents the action associated with the entity, typically corresponding to behavioral descriptions such as faults, operations, and state changes; This is used to characterize the result caused by the action. Embedding functions for knowledge graphs; For the first A semantic vector representation of an action event.

[0047] In this process, the system uses the cross-document semantic vectors generated by LinkBERT as additional semantic features and combines them with the knowledge graph embedding results to preserve cross-document reasoning and contextual semantic dependencies, thereby obtaining a more comprehensive representation of operational behavior.

[0048] (4) During the dynamic incremental update phase, the knowledge graph supports adaptive expansion and continuous learning. With the input of new operation and maintenance records, the graph structure is dynamically adjusted through the local reconstruction and link prediction mechanism based on the adjacency matrix, updating the entity node weights and relationship strengths, and maintaining the model's sensitivity and timeliness to new event patterns.

[0049] The feature extraction results of record-behavioral data can be represented as: ; In the formula, A semantic enhancement function for integrating cross-document relationship modeling; This is a finely tuned domain-specific language model; Embedding operators for knowledge graphs; These are the extracted entity set and relation set, respectively. This design can simultaneously capture the semantic features of the text, the logical relationships of the operation chain, and the potential correlations across logs, thereby improving the accuracy and robustness of operational behavior modeling.

[0050] Its advantage lies in the fact that using the above steps to construct record-behavioral data for the knowledge graph not only more explicitly represents the relationships between different data nodes, but also facilitates the introduction of new data. As new maintenance records and operation logs are input, the system dynamically expands the knowledge graph through incremental graph update algorithms (such as local reconstruction based on the adjacency matrix), enabling the model to continuously learn the operation and maintenance behavior patterns of the equipment.

[0051] In one embodiment, the trained discriminant function is used to analyze the feature vector. The faults are classified to obtain the fault diagnosis results: ; In the formula, For fault detection models, which are typically large models based on deep learning, the output results are... Where 0 represents the normal state. This indicates different types of faults (such as overheating, insulation breakdown, partial discharge, etc.).

[0052] In one embodiment, the fault discrimination model includes an input layer, multiple hidden layers, and an output layer, wherein the hidden layers introduce an improved self-attention mechanism, and the output layer outputs the probability distribution of fault categories.

[0053] In one embodiment, the principle of the improved self-attention mechanism is as follows: The fused features are divided into multiple subspaces, and the attention scores of each subspace are calculated in parallel. Differential attention scores are calculated based on the attention scores of each subspace, and the features of each subspace are then weighted and fused based on the differential attention scores. Furthermore, this embodiment employs a classification model based on a deep neural network (DNN) as the fault detection model, capturing the complex correlations between different features through multi-layer nonlinear mapping. The model consists of an input layer, several hidden layers, and an output layer, with the output of each layer defined as: ; In the formula, and The first Layer weight matrix and bias terms; The activation function is (e.g., ReLU). The output layer uses the Softmax function to convert the network output into a probability distribution for various faults.

[0054] This embodiment integrates data from multiple modalities, including text, location, and time. To enhance the correlation modeling capability between multimodal features, an improved self-attention mechanism is introduced. In the modified attention fusion layer, this embodiment first employs a grouped query / key mechanism to construct attention representations from different perspectives. Specifically, the fused feature vector is divided into several subspaces according to channels. For example, the physical parameter subspace is mapped to the first group of query and key (…). ), mapping the operation and maintenance semantic subspace to the second set of queries and keys ( This perspective division allows attention computation to focus on both the temporal signals of device operating status and independently assess the semantic relationships of operational behaviors, thereby characterizing the interactions and differences between different modalities in subsequent differential operations. In the differential attention computation process, the system first calculates the conventional normalized attention scores for both perspectives, and then obtains the final weighted attention distribution through differentiation. The attention scores for the first and second groups are as follows: ; ; Next, the differential attention score is calculated and a weighted output is generated: ; Finally use Weighted V vector output: ; To avoid negative weights from differential calculations affecting inference stability, in engineering implementation, [the following can be done]: Nonnegation is applied (e.g., ReLU or clamp to the lower bound of zero), and the outputs of each head are normalized and fused in a multi-head parallel structure.

[0055] The application of this differential attention mechanism in fault diagnosis can improve the model's response to key cross-modal features, thereby reducing misjudgments caused by single-modal noise. In specific applications, differential attention can be placed in the fusion layer after feature concatenation, responsible for weighted filtering of interactions from different perspectives, such as physical channels and maintenance channels. This ensures that when an event is significant in both physical measurements and maintenance records (e.g., a sudden increase in acetylene content and records of maintenance delays), the attention weight corresponding to that event is amplified in the differential output. Conversely, for isolated signals that appear only in a single path and lack cross-modal support, their differential weight tends to decrease, reducing their impact on the final judgment. This mechanism can also extract multi-scale dependencies from different subspaces through parallel multi-head differential attention, thereby improving diagnostic robustness and interpretability under complex and variable power conditions.

[0056] In addition, to balance local and global dependencies, a multi-head attention structure can be used in the hidden layer to extract multi-scale dependency information between features from different subspaces in parallel, thereby improving the overall robustness and diagnostic accuracy of the model.

[0057] Finally, during model training, the cross-entropy loss function is used to measure the difference between the predicted results and the true labels: ; In the formula, This is a real label; For the predicted results; Input data into the model.

[0058] Finally, the parameters are continuously updated using the backpropagation algorithm and the gradient descent optimizer Adam.

[0059] In one embodiment, the step of obtaining candidate failure chains is as follows: Identifying abnormal events in historical multi-source monitoring data based on a sliding time window and threshold adaptive mechanism; Calculate the connection scores between each anomalous event based on chronological order, spatial correlation, and causal dependency strength. Abnormal events with connection scores exceeding a preset threshold are connected sequentially to generate candidate fault chains.

[0060] Furthermore, after obtaining the preliminary fault identification results, it is necessary to further locate the potential root causes of the fault. To this end, the system performs root cause analysis based on probabilistic reasoning in the fused feature space to determine the most likely fault chain pattern. In the context of root cause analysis, the root cause analysis formula based on fused features is: ; In the formula, For a candidate chain among all candidate failure chain patterns, and respectively with chain The corresponding parameter vector and bias, For the chain The system ultimately selects the predicted probability from all candidate chains. The highest value is taken as the final root cause path, realizing a closed-loop analysis from multi-source data to causal inference.

[0061] To obtain the fault chain pattern, relevant abnormal events are combined into several candidate chains in chronological order, with each chain containing several event nodes. Specifically, abnormal events are first extracted based on multi-source monitoring data and operation and maintenance records, including but not limited to sudden changes in gas concentration, abnormal oil temperature, partial discharge, excessive equipment vibration, maintenance delays, and abnormal operation commands.

[0062] The event detection phase employs a combination of time window sliding and threshold adaptive algorithms to achieve dual identification of short-term fluctuations and long-term trend changes. The system monitors indicators including oil chromatography gas concentration, winding temperature, partial discharge, and operation delay in maintenance logs by sliding a time window with a fixed step size (e.g., a 1-hour window, sliding once every 10 minutes). Within each sliding window, the system calculates the mean μ and standard deviation σ of the current indicator and defines the threshold as follows: ; In the formula, A coefficient (usually 2 or 3) is set for experience. Furthermore, to address long-term trend deviations, an Exponentially Weighted Moving Average (EWMA) model is used to smooth and predict historical indicator values. An anomaly is considered when the current value deviates from the predicted value by more than the dynamically adjusted percentage (±10%). Within a window, if the indicator exceeds a threshold number or percentage (more than 3 times within any 6-minute period), the result is considered an anomaly. An event is triggered if an anomaly occurs. To suppress false alarms, the system requires that an anomaly occur simultaneously in multiple channels or that there are violations within two consecutive windows before it is considered a critical anomaly event. The sliding window length can be adaptively adjusted at different stages of equipment operation; for example, the window can be shortened to 30 minutes during periods of increased load to improve early warning capabilities. This mechanism can capture both short-term mutations and identify long-term drifts, thus providing a reliable starting point for the construction of subsequent candidate chains.

[0063] Subsequently, based on the chronological order, spatial correlation, and causal dependency strength of the abnormal events, related events are grouped into several candidate chains. After detecting multiple abnormal events, candidate fault chains are formed according to the following specific steps: First, the monitored abnormal events are sorted by timestamp from earliest to latest according to their occurrence time, forming a preliminary sequence. Second, the spatial or equipment correlation between events is measured based on spatial correlation—for example, events occurring within the same substation, in the same bay, or along the same path have a higher correlation if their spatial distance or electrical topological distance is small. Finally, the causal dependency strength is evaluated, i.e., the conditional probability / causal probability between event pairs obtained statistically or empirically from the historical fault database and expert knowledge base, such as event... e i What happened next e i+1 When the probability of either event exceeds a threshold θ, the two events are considered to have a causal dependency. The system calculates scores for the three dimensions mentioned above: time sequence score. Spatial correlation score Causal probability score And through weighted public ; Determine the event Can the two events be connected as adjacent nodes in a chain? When the score is higher than a set threshold, the system connects the two events. This connection method is then used to generate several candidate chains across the entire event set. Each chain consists of several sequentially connected event nodes and is assigned a confidence level or score for the entire chain, which is then used for subsequent root cause analysis.

[0064] Each candidate chain consists of multiple event nodes, and the connections between nodes are based on historical statistical co-occurrence relationships and physical logic rules defined by the expert knowledge base (such as "oil temperature rise → insulation aging → enhanced partial discharge"). The chain construction time window can be adaptively adjusted according to the equipment type and operating frequency. In this embodiment, it is generally set to 24 to 72 hours to cover the entire process from the occurrence of the fault to the manifestation of the abnormality.

[0065] In this way, the system can reproduce the evolution path of a fault in a temporal dimension and provide structured candidate pattern inputs for subsequent root cause discrimination models, thereby achieving interpretable and traceable causal reasoning. For the specific calculation process of the chain, the confidence level of the entire chain is calculated by multiplying the probabilities of occurrence between adjacent events to reflect the reliability of the entire chain as a potential root cause path. ; In the formula, For candidate chains, The product operation represents the probability of adjacent anomalous events in the chain, reflecting the contribution of the event sequence to the overall chain confidence.

[0066] In one embodiment, a feature enhancement mechanism is also introduced during the root cause path acquisition process, specifically: The comprehensive similarity is calculated based on feature distance, temporal morphological similarity and event semantic consistency, and similar cases are extracted from the historical fault database based on the comprehensive similarity. Feature extraction is performed on similar cases, and the extracted similar features are fused into the current feature vector to obtain enhanced discriminative features. The enhanced discriminative features are then used to determine the root cause path from the preset candidate fault chains.

[0067] Furthermore, in practical applications, this invention improves inference accuracy through the following detailed optimizations: First, through a feature enhancement mechanism, the feature vector not only includes the original parameter features of the current device operation but also integrates the context vectors of historical similar cases, thereby enhancing the ability to discriminate complex chains. Specifically, the system retrieves several cases most similar to the current features from the historical fault database, extracts their fused feature vectors, and integrates these context vectors into the current feature vector through weighted averaging, thereby enhancing the ability to discriminate complex fault chains. The discrimination criteria for similar cases comprehensively consider three aspects: feature distance, temporal morphological similarity, and event semantic consistency. Specifically, the feature distance similarity is calculated using weighted Euclidean distance, which is obtained by weighting the distance between the sample features and historical features. Similarity is used as a distance feature. Secondly, the temporal morphology is mainly based on the Dynamic Time Warping (DTW) algorithm to calculate the dynamic alignment difference of signals under different operating conditions, obtaining... Finally, event semantic consistency is calculated using cosine similarity based on BERT encoding of the operation and maintenance logs. This is used to measure the degree of consistency in the described content. The final similarity score is calculated as follows: ; in Weighting coefficients determined empirically. The system selects... The highest-valued historical cases are used as similar samples, and their corresponding feature vectors are then weighted by attention and fused into the context representation of the current sample to enhance the model's ability to perceive potential causal patterns. In this project, a balanced weighting ratio is used. Both are 0.33.

[0068] In one embodiment, a chain pattern constraint is also introduced when generating candidate fault chains, specifically: Establish a causal knowledge base based on the operating principles and historical failure evolution of the equipment; The generated initial candidate fault chains are matched with the causal knowledge base, and the initial candidate chains are filtered, corrected or eliminated based on similarity to obtain candidate chains.

[0069] Secondly, a chain pattern constraint is introduced, so that the candidate chain is jointly defined by expert knowledge and the results generated by the large model, ensuring that the inference chain is both complete and meets physical constraints. The specific steps are divided into four stages: expert knowledge extraction, model chain generation, consistency constraint fusion, and constraint optimization.

[0070] In the expert knowledge extraction phase, the system first establishes a causal knowledge base based on the operating mechanisms and fault evolution experience of power equipment. This knowledge base is represented in the form of triples, such as... "Increased oil temperature" leads to "insulation aging". It covers knowledge such as energy transfer between equipment components, degradation paths, and maintenance logic.

[0071] In the model chain generation phase, a domain-adjusted BERT model is used to analyze multi-source monitoring data and text records to generate several potential failure chain patterns. The BERT model is responsible for discovering hidden causal patterns from complex data, compensating for gaps in expert knowledge coverage.

[0072] During the consistency constraint fusion phase, the chains generated by the model are matched and filtered against the expert knowledge base. First, the candidate chains generated by the model are structurally parsed, decomposing each chain into several event-causal pairs. e i →e j Then, extract the event entities (such as "oil temperature rises" or "sudden increase in discharge") and their relationship types (such as "caused" or "associated with"). Next, search the expert knowledge base for knowledge triples that correspond to or are similar to the event pair. Entity 1, Relationship, Entity 2 And calculate semantic similarity (Based on embedding vectors or semantic distance metrics).

[0073] If the event relationships in the model generation chain contradict the knowledge base (e.g., a violation of physical logic such as "current decreases → temperature increases"), the system automatically removes or corrects them. If the confidence level of a conflicting relationship is low (the model output probability is below the threshold θ1=0.53), the relationship is directly removed; if the confidence level of a conflicting relationship is high (above the threshold θ2=0.83), the system attempts to replace the relationship or reverse its direction based on similar patterns in expert knowledge; for marginal relationships with confidence levels between the two, they are marked as "pending review" for manual confirmation by experts.

[0074] If the model generates new relations that are not defined in the knowledge base but are statistically significant, they are marked as "candidate new chains" for expert review and then incorporated into the knowledge base update. During the constraint optimization phase: a logical consistency scoring function is introduced onto the final candidate chain set. ; In the formula, The degree to which the chain conforms to physical constraints. To determine the degree of matching with semantic logic, These are the weighting coefficients. In this embodiment, a balanced weighting ratio is used. All scores are 0.5. The system only retains chain patterns with scores above the threshold as input for subsequent reasoning stages.

[0075] Through this constraint mechanism, the reasoning process not only relies on data-driven results, but also ensures that the generation chain conforms to the operating rules and engineering logic of the power system, achieving a deep integration of data intelligence and expert experience.

[0076] In one embodiment, an adaptive confidence mechanism is used during root cause analysis, specifically: When the highest predicted probability of any candidate chain exceeds the first threshold, any candidate chain is output as a single diagnostic result. When the highest predicted probability of any candidate chain is between the first threshold and the second threshold, multiple candidate chains are retained, and a weighted correction is applied to these multiple candidate chains to output the diagnostic result.

[0077] Furthermore, through a confidence-adaptive mechanism, the prediction results are... A dynamic threshold is introduced. When the highest probability chain reaches a preset threshold (e.g., ≥ 0.7), it can be directly used as a single diagnostic result. If the highest probability is insufficient to support a single conclusion, multiple high-confidence chains (e.g., probability ≥ 0.5) are retained for subsequent manual review or model iteration optimization. This mechanism ensures that potential root causes are not missed in high-uncertainty scenarios, and the threshold can be dynamically adjusted to adapt to different device types or data quality.

[0078] Furthermore, to improve the reliability of diagnostic results, a confidence fusion mechanism is introduced to perform weighted correction on the discrimination results: ; In the formula, For the confidence level of the diagnostic results, This is a confidence mapping function based on probability distribution weights. Indicates the first The probability of each discrimination result. The system's final output. The combined results enable robust diagnosis of transformer faults.

[0079] In summary, this invention achieves accurate diagnosis of transformer faults through a step-by-step modeling process of data acquisition, feature extraction, fault identification, and confidence fusion.

[0080] This embodiment also discloses a transformer fault diagnosis system based on multi-source data and root cause analysis, including: The data acquisition module is used to acquire multi-source monitoring data during transformer operation. The multi-source monitoring data includes location-time data, physical parameter data, and record-behavior data. The feature extraction module is used to extract features from multi-source monitoring data and fuse the extracted features to obtain fused features; The fault diagnosis module is used to input fused features into the trained fault discrimination model, output fault classification results, and determine the root cause path from the preset candidate fault chain based on probabilistic reasoning.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A transformer fault diagnosis method based on multi-source data and root cause analysis, characterized in that, The specific steps are: Obtain multi-source monitoring data in the operation process of the transformer, the multi-source monitoring data including position-time data, physical parameter data and record-behavior data; Feature extraction is performed on the multi-source monitoring data, and the extracted features are fused to obtain fused features; The fused features are input into a trained fault discrimination model to output a fault classification result, and a root cause path is determined from a pre-set candidate fault chain based on probabilistic reasoning.

2. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, The feature extraction step of the physical parameter data is: Normalizing the physical parameter data to generate normalized data; Converting the normalized data into an image using a Gram angle field; Deep feature extraction is performed on the converted image using a deep feature extraction module; The deep features are screened based on a double-layer feature optimization strategy, and a physical parameter feature set is constructed based on the screened deep features. 3.The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, The feature extraction step of the record-behavior data is: Based on the power named entity recognition model, a triple is constructed using the key elements extracted from the record-behavior data; the key elements include device objects, action types and event results; Based on a cross-document knowledge modeling mechanism, the semantic or logical associations between different documents are learned and modeled to construct a document association graph; The triple is mapped to entities and relationships in a knowledge graph to form an entity-relation-entity chain, and the entities and relationships are converted into vector representations through a graph embedding algorithm, which are fused with the cross-document semantic vectors obtained based on the document association graph to generate record-behavior features.

4. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, The fault discrimination model includes an input layer, multiple hidden layers and an output layer, wherein the hidden layers introduce an improved self-attention mechanism, and the output layer outputs the probability distribution of the fault category.

5. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 4, characterized in that, The principle of the improved self-attention mechanism is: The fused features are divided into multiple subspaces, and the attention scores of each subspace are calculated in parallel; Based on the attention scores of each subspace, differential attention score calculation is performed, and the features of each subspace are weighted and fused based on the differential attention scores.

6. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, The acquisition step of the candidate fault chain is: Based on a sliding time window and a threshold adaptive mechanism, abnormal events are identified in historical multi-source monitoring data; According to the time sequence, spatial correlation degree and causal dependence strength, the connection scores between each abnormal event are calculated; The abnormal events with connection scores exceeding a pre-set threshold are connected in order to generate the candidate fault chain.

7. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, During the root cause path acquisition process, a feature enhancement mechanism is also introduced, specifically: Based on feature distance, time sequence pattern similarity and event semantic consistency, a comprehensive similarity is calculated, and similar cases are extracted from a historical fault database based on the comprehensive similarity; Feature extraction is performed on the similar cases, the extracted similar features are fused into the current feature vector to obtain enhanced discrimination features, and the enhanced discrimination features are used to determine the root cause path from the pre-set candidate fault chain.

8. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 6, characterized in that, The chain pattern constraint is also introduced when generating the candidate fault chain, specifically: A causal knowledge base is established based on the operating principle of the device and the historical fault evolution process; The generated initial candidate fault chain is matched with the causal knowledge base, the initial candidate chain is screened, corrected or eliminated based on similarity, and the candidate chain is obtained.

9. The transformer fault diagnosis method based on multi-source data and root cause analysis according to claim 1, characterized in that, A confidence adaptive mechanism is used in root cause analysis, specifically: When the highest prediction probability of any of the candidate chains exceeds a first threshold, any of the candidate chains is taken as a single diagnostic result output; When the highest prediction probability of any of the candidate chains is between the first threshold and a second threshold, multiple candidate chains are retained, and multiple candidate chains are weighted and corrected, and a diagnostic result is output.

10. A transformer fault diagnosis system based on multi-source data and root cause analysis, characterized in that, It comprises: A data acquisition module is configured to acquire multi-source monitoring data during the operation of the transformer, wherein the multi-source monitoring data comprises position-time data, physical parameter data and record-behavior data; A feature extraction module is configured to extract features from the multi-source monitoring data and fuse the extracted features to obtain fused features; A fault diagnosis module is configured to input the fused features into a trained fault discrimination model, output a fault classification result, and determine a root cause path from a pre-set candidate fault chain based on probability reasoning.

Images