Fault detection method and device for electrical equipment, equipment and storage medium

Abstract

The invention discloses a fault detection method and device for electrical equipment, equipment and a storage medium, which are applied to the technical field of power systems, and comprise the following steps: in the real-time operation process of the electrical equipment, obtaining real-time operation data, carrying out fault analysis on the real-time operation data based on a fault knowledge graph, and obtaining a fault detection result. The construction process of the fault knowledge graph comprises the following steps: acquiring sample data of electrical equipment; obtaining sample enhanced data; based on the sample enhancement data, obtaining an optimized fault detection model, and inputting historical operation data of the electrical equipment into the optimized fault detection model to obtain operation fault data; extracting multi-dimensional fault data in the initial fault database to obtain operation fault integration data, and updating the initial fault database to obtain an updated fault database; and processing the updated fault database to obtain a fault knowledge graph. According to the method provided by the invention, comprehensive and timely detection of the operation state of the electrical equipment is realized.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a fault detection method, apparatus, device, and storage medium for electrical equipment. Background Technology

[0002] With the expansion of the power grid and the increase in electricity load, the operating status of electrical equipment is directly related to the safe operation of the power grid and the reliability of power supply.

[0003] In existing technologies, when a fault is detected in electrical equipment, the power supply to the faulty equipment needs to be disconnected, and manual inspection is required. However, manual inspection has blind spots and cannot accurately identify the unstructured features of the faulty equipment, easily leading to one-sided fault diagnosis. Furthermore, disconnecting the power supply before inspecting the faulty equipment can result in delayed fault diagnosis and uncontrolled electricity costs. Therefore, achieving comprehensive and timely detection of the operating status of electrical equipment is crucial. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for fault detection of electrical equipment, in order to solve the problem that it is difficult to capture the multi-dimensional correlation of electrical equipment in the prior art, so as to achieve comprehensive and timely detection of the operating status of electrical equipment.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a fault detection method for electrical equipment, comprising: During the real-time operation of the target electrical equipment, real-time operating data of the target electrical equipment is acquired. Fault analysis is then performed on the real-time operating data based on a constructed fault knowledge graph to obtain the fault detection results of the target electrical equipment. The construction process of the fault knowledge graph includes: Obtain sample data for all the target electrical devices, wherein the sample data includes first fault sample data and second normal sample data; Sample augmentation data is obtained, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic mask layer; Based on the sample augmentation data, the initial fault detection model is trained to obtain an optimized fault detection model. The historical operating data of each of the target electrical devices is input into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data. Extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database. Based on multimodal mapping, realize the entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data to obtain integrated operational fault data. Update the initial fault database based on the integrated operational fault data to obtain the updated fault database. The updated fault database is subjected to knowledge structuring processing to obtain the fault knowledge graph.

[0006] As one preferred embodiment, the obtained sample augmentation data includes: A gradient penalty condition is introduced into the adversarial network layer to constrain the first fault sample data, thereby obtaining the third sample data; A regression model is introduced into the semantic mask layer to perform semantic imputation on the second normal sample data to obtain the fourth sample data. The third sample data and the fourth sample data are integrated to obtain the sample augmentation data.

[0007] As one preferred embodiment, the initial fault detection model is trained based on the sample augmentation data to obtain an optimized fault detection model. Historical operating data of each of the target electrical devices is then input into the optimized fault detection model to obtain operational fault data, including: The initial fault detection model is constructed based on a gradient boosting residual network. The sample augmentation data is input into the initial fault detection model for training to obtain the optimized fault detection model; Obtain the historical operating data of each of the target electrical devices; The historical operation data is filtered and denoised to obtain the historical operation data to be detected; Based on the optimized fault detection model, fault detection is performed on the historical operating data to be detected to obtain the operating fault data.

[0008] As a preferred embodiment, the entity-relationship mapping of all operational fault data with all multi-dimensional fault data based on multimodal mapping is used to obtain integrated operational fault data, including: Extract the multi-dimensional fault data corresponding to each target electrical device from the initial fault database, wherein the multi-dimensional fault data consists of at least the electrical device model, fault code, maintenance record, inspection report and test indicators; The operational fault data is subjected to semantic vector transformation to obtain operational fault data to be mapped; The multi-dimensional fault data is transformed into semantic vectors to obtain the multi-dimensional fault data to be mapped. Determine the cosine similarity between the operational fault data to be mapped and the multi-dimensional fault data to be mapped; Based on the cosine similarity and sequence labeling model, entity-relationship mapping is performed on the operational fault data and the multi-dimensional fault data to obtain integrated operational fault data.

[0009] As one preferred embodiment, updating the initial fault database based on the integrated operational fault data to obtain an updated fault database includes: The hash values ​​of the integrated data of each operational failure are obtained based on the hash algorithm; Based on all the hash values, the corresponding duplicate integrated data of the operation failures are removed to obtain integrated data of operation failures without redundancy. The initial fault database is updated based on the non-redundant operational fault integration data to obtain the updated fault database.

[0010] As one preferred embodiment, the step of performing knowledge structuring processing on the updated fault database to obtain the fault knowledge graph includes: Extract the entities from the updated fault database and convert the entities into node information; Extract the relationships from the updated fault database and convert the relationships into edge information; Based on the graph database, an index is established for the node information and the edge information, and a uniqueness constraint is applied to the node information and the edge information to obtain the fault knowledge graph.

[0011] As one preferred embodiment, after obtaining the fault detection result of the target electrical equipment, the method further includes: Based on the causal paths of the fault knowledge graph, the cause of the fault, the location of the fault, and the repair plan for the target electrical equipment are determined. The cause of the fault, the location of the fault, and the repair plan are uploaded to the user terminal, and the cause-and-effect path is visualized on the user terminal.

[0012] Another embodiment of the present invention provides a fault detection device for electrical equipment, comprising: A fault detection module is used to acquire real-time operating data of the target electrical equipment during its real-time operation, perform fault analysis on the real-time operating data based on a constructed fault knowledge graph, and obtain fault detection results for the target electrical equipment. The construction process of the fault knowledge graph includes: The data acquisition module is used to acquire sample data of all the target electrical devices, wherein the sample data includes first fault sample data and second normal sample data; A data augmentation module is used to obtain sample augmentation data, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic masking layer; The model optimization module is used to train the initial fault detection model based on the sample augmentation data to obtain an optimized fault detection model, and input the historical operating data of each of the target electrical devices into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data; The data optimization module is used to extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database, realize entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data based on multi-modal mapping, obtain integrated operational fault data, and update the initial fault database based on the integrated operational fault data to obtain the updated fault database. The graph construction module is used to perform knowledge structuring processing on the updated fault database to obtain the fault knowledge graph.

[0013] Another embodiment of the present invention provides a fault detection device for electrical equipment, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a fault detection method for electrical equipment as described above.

[0014] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements a fault detection method for electrical equipment as described above.

[0015] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: First, the technical solution of this application performs data augmentation processing on the sample data through adversarial network layer and semantic mask layer. Then, it optimizes the fault detection model based on the sample augmentation data to reduce the impact of sample data bias on the fault detection results. Second, it performs entity-relationship mapping on the operational fault data and multi-dimensional fault data to update the fault database. Finally, it constructs a fault knowledge graph based on the updated fault database, thereby realizing fault detection in the real-time operation of electrical equipment and comprehensive and timely detection of the operating status of electrical equipment. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a fault detection method for electrical equipment in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a transformer health intelligent management method according to one embodiment of the present invention; Figure 3 This is a schematic diagram of the fault diagnosis process of a fault detection device for electrical equipment in one embodiment of the present invention; Figure 4 This is a schematic diagram of the RoBERTa-BiLSTM-CRF model in a fault detection method for electrical equipment according to one embodiment of the present invention; Figure 5 This is a schematic diagram of the "multi-point grounding of iron core" node and its associated query in a fault detection method for electrical equipment in one embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the fault cause query and solution measures of a fault detection method for electrical equipment in one embodiment of the present invention; Figure 7 This is a framework diagram of a fault diagnosis reasoning system in a fault detection method for electrical equipment according to one embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a fault diagnosis device for electrical equipment in one embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a fault detection device for electrical equipment in one embodiment of the present invention.

[0017] Figure label: Among them, 11. Fault detection module; 12. Data acquisition module; 13. Data augmentation module; 14. Model optimization module; 15. Data optimization module; 16. Map construction module; 21. Processor; 22. Computer program. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] In existing technologies, once an abnormality is detected in electrical equipment, the standard procedure is to immediately disconnect the power supply to the faulty equipment, completely taking it out of operation, before maintenance personnel conduct further inspections. However, manual inspection can only obtain scalar data from discrete measurement points, failing to continuously sample the unstructured characteristics accumulated throughout the equipment's lifecycle. This can easily lead to a one-sided approach to fault diagnosis. Furthermore, disconnecting the power supply to the faulty equipment before inspection can cause some latent defects to temporarily become dormant, resulting in the bizarre phenomenon of no abnormalities offline, only to reappear upon reconnection. This further amplifies the diagnostic lag, and repeated power outages and restorations can lead to uncontrolled electricity costs. Therefore, achieving comprehensive and timely monitoring of the operating status of electrical equipment is crucial.

[0022] Furthermore, in this embodiment of the invention, in order to achieve comprehensive and timely fault detection of electrical equipment, real-time operating data of the electrical equipment during real-time operation is acquired, and fault analysis is performed on the real-time operating data based on the constructed fault knowledge graph to obtain the fault detection results of the electrical equipment. The electrical equipment includes at least converter transformers, distribution transformers, coupling capacitors, multilevel converters, and capacitive voltage transformers.

[0023] For details, please see Figure 1 , Figure 1 The diagram shown illustrates a fault detection method for electrical equipment according to one embodiment of the present invention, the method comprising steps S1 to S6: S1. During the real-time operation of the target electrical equipment, acquire the real-time operating data of the target electrical equipment, perform fault analysis on the real-time operating data based on the constructed fault knowledge graph, and obtain the fault detection results of the target electrical equipment. The construction process of the fault knowledge graph includes: S2. Obtain sample data for all the target electrical devices, wherein the sample data includes first fault sample data and second normal sample data; S3. Obtain sample augmentation data, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic mask layer; S4. Based on the sample augmentation data, train the initial fault detection model to obtain an optimized fault detection model, and input the historical operating data of each of the target electrical devices into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data; S5. Extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database. Based on multi-modal mapping, realize the entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data to obtain integrated operational fault data. Update the initial fault database based on the integrated operational fault data to obtain the updated fault database. S6. Perform knowledge structuring processing on the updated fault database to obtain the fault knowledge graph.

[0024] Further, in step S1, during the real-time operation of the target electrical equipment, real-time operating data of the target electrical equipment is acquired. Fault analysis is performed on the real-time operating data based on the constructed fault knowledge graph to obtain the fault detection results of the target electrical equipment. Preferably, a transformer health intelligent management method can be derived based on the fault detection method. Preferably, this solution is specifically applied to transformers in electrical equipment, and a related transformer monitoring intelligent management method is designed based on the above method. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram illustrates a flowchart of a transformer health intelligent management method in one embodiment of the present invention. This method diagnoses transformer fault types using a fault diagnosis network based on CWGAN-GP (Conditional Wasserstein GAN with Gradient Penalty) and a gradient boosting residual network. It extracts knowledge using RoBERTa (Robustly Optimized BERT Pretraining Approach) combined with bidirectional LSTM (Long Short-Term Memory) and a conditional random field model to achieve entity-relation mapping of the transformer. The fault diagnosis results and entity-relation mapping results are input into a data mining-based transformer health intelligent management system to construct a fault knowledge graph. Based on the constructed fault knowledge graph, knowledge reasoning is performed to obtain fault reasoning results.

[0025] Further, in step S2, sample data of all target electrical equipment is obtained, wherein the sample data includes first fault sample data and second normal sample data. Preferably, the first fault sample data is the status data of the target electrical equipment when it is operating abnormally, and the second normal sample data is the status data of the target electrical equipment when it is operating normally.

[0026] Furthermore, in step S3, since the first fault sample data and the second normal sample data are unbalanced in terms of categories, the fault detection results of the fault identification model directly constructed based on the first fault sample data and the second normal sample data are often inaccurate. Therefore, a gradient penalty condition is introduced in the adversarial network layer to constrain the first fault sample data and generate high-quality first fault sample data. A regression model is introduced in the semantic mask layer to perform semantic imputation on the second normal sample data and obtain high-quality second normal sample data. Then, the high-quality first fault sample data and the high-quality second normal sample data are integrated to obtain sample augmentation data.

[0027] Specifically, the core structure of a traditional adversarial network consists of two parts: a generator and a discriminator. The generator receives a random noise vector as input and outputs a set of virtual data that is as realistic as possible. The discriminator is responsible for determining whether the input data is virtual data generated by the generator or real data. Through generator-discriminator game training, the generator can synthesize realistic electrical equipment data with statistical distribution characteristics even when class samples are scarce. The generator and discriminator iterate continuously through this game process, constantly optimizing their performance. The objective function of the adversarial network is defined as follows: The loss function of the discriminator is defined as follows: in, This represents the generated data. With real data Degree of difference When the generator is fixed, the degree of difference is maximized. The mathematical expectation of the distribution function. Represents logarithmic operations.

[0028] To further improve training stability and convergence speed, a gradient penalty mechanism is introduced into the traditional adversarial network, resulting in the CWGAN-GP adversarial network. Here, the Wasserstein distance is used instead of the Jensen-Shannon divergence of the traditional adversarial network as the optimization objective, providing a smooth and differentiable loss function. This effectively enhances the generator's ability to learn the distribution of electrical equipment data under constraints. The formula for calculating the Wasserstein distance is as follows: in, Represents the joint distribution of all possible combinations. , gather, Represents real sample data. This indicates the generation of sample data. This represents the distance between the real sample and the generated sample.

[0029] In WGAN-GP (Wasserstein GAN with Gradient Penalty), the gradient penalty term is defined as follows: in, Indicates the penalty coefficient. Indicates interpolated samples Data distribution Represents the distribution along the real data and generating data Random interpolation sampling of the lines connecting them.

[0030] The loss functions for the generator and discriminator are as follows: CWGAN-GP minimizes the difference between the generated distribution and the real distribution by minimizing the generator loss function L(G), so that the generated electrical equipment feature vector distribution Pf gradually approaches the real monitoring data distribution Pr, thereby improving the physical consistency of the data.

[0031] Further, in step S4, an initial fault detection model is constructed based on a gradient boosting residual network. Sample augmentation data is input into the initial fault detection model for training, resulting in an optimized fault detection model. Historical operating data for each target electrical device is acquired, and the historical operating data is filtered and denoised to obtain the historical operating data to be detected. Based on the optimized fault detection model, fault detection is performed on the historical operating data to obtain the operating fault data. Specifically, the gradient boosting residual network combines residual approximation of the function space with the residual structure of a deep network, enabling the model to achieve higher accuracy in feature representation and fault identification under multimodal complex data. The gradient boosting residual network can effectively fuse sensor data from electrical equipment, such as dissolved gas concentration from DGA (Dissolved Gas Analysis), winding / oil temperature and ambient temperature, load / harmonic electrical parameters, time-frequency diagrams of partial discharge and ultrasonic waveforms, and ultraviolet / thermal imaging images. Residual networks (ResNet) are used to automatically extract deep spatiotemporal features from waveforms and images. Gradient boosting trees (GBDT) are used to process tabular operating conditions and historical records of electrical equipment. The two are then combined through residual fitting or feature concatenation. This approach leverages the ability of deep networks to represent complex signals while retaining the advantages of GBDT in handling heterogeneous data, missing values, and interpretability. Specifically, in the gradient boosting framework, given a dataset... The loss function is The iterative process is calculated as follows: in, For input features; The true label of the i-th sample; Predicted values ​​for initial constants; It refers to the number of samples.

[0032] The formula for calculating the negative gradient is as follows: Among them, by training the m-th base learner To fit the residuals in, It is a deep residual network used to extract nonlinear mappings from input features and approximate residuals. The optimal step size is... The solution formula is as follows: The updated model is as follows: Furthermore, the augmented sample data is input into the initial fault detection model for training, resulting in an optimized fault detection model. Based on this optimized model, fault detection is performed on historical operational data to obtain operational fault data. The training process includes generating a training subset, training the learner, and integrating the prediction results. For details, please refer to [link to relevant documentation]. Figure 3 , Figure 3 The diagram illustrates a fault diagnosis process of an electrical equipment fault detection device according to one embodiment of the present invention. Specifically, a training subset is generated using Bootstrap sampling technology. n samples are randomly selected with replacement from an original dataset containing m samples. During each sampling, when a record is selected, all modal data corresponding to that record are simultaneously added to the subset, thereby constructing an independent training subset for each weak learner (equivalent to...). Figure 3 The training subsets are divided into training subset 1, training subset 2, and training subset n. These subsets are both distinct and partially overlapping. Each subset contains key data such as DGA table data, temperature data, electrical parameters, partial discharge waveforms or time-frequency diagrams, environmental and metadata data, maintenance / work order records and tags, etc., for electrical equipment. A learner is trained for each subset generated in the first step. The learner used in this invention is a deep residual network (equivalent to...). Figure 3 The residual networks (residual network 1, residual network 2, and residual network n) are trained using training subsets extracted by Bootstrap sampling. Since each training subset is different, the features learned by each residual network exhibit both differences and similarities. A voting strategy is used to fuse the prediction results of each learner for its respective training subset to obtain the operational fault data. Preferably, a weighted soft voting method is used, averaging the probability distributions of different fault categories output by multiple learners, and then taking the category with the highest probability as the final prediction. The process is as follows: First, a sample x to be diagnosed is input. Each deep residual network outputs a set of probability distributions for different fault types, resulting in multiple probability distributions for different fault types across multiple deep residual networks. Then, the weight coefficients are determined. The weighting coefficients are normalized based on the probability distribution values ​​of each group. The normalization formula is as follows: Let be the fusion weight of the m-th model; M is the total number of models.

[0033] Then, according to the formula for the fusion probability distribution: in, This represents the prediction probability of the m-th learner for category c.

[0034] Finally, the operational failure data represents the failure category with the highest probability: Furthermore, in step S5, RoBERTa-BiLSTM-CRF is used to extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database. For details, please refer to [link to relevant documentation]. Figure 4 , Figure 4 The diagram illustrates a RoBERTa-BiLSTM-CRF model in an electrical equipment fault detection method according to one embodiment of the present invention. This model mainly consists of three layers: the first layer is a RoBERTa model for data preprocessing, converting each character in the sentence into a low-dimensional vector pattern; the second layer is a BiLSTM (Bidirectional Long Short-Term Memory) model for extracting features from electrical equipment fault entities, with the RoBERTa layer outputting its input vector, automatically extracting semantic and temporal features from the context; the third layer is a CRF (Conditional Random Field) model for sequence labeling of named entities in the sentence, considering the order relationship between labels and resolving dependencies between output labels to obtain the globally optimal labeling sequence. Specifically, RoBERTa serves as a pre-trained language model, employing the BiLSTM-CRF model to extract electrical equipment entities. Input textual data related to electrical equipment operation, including unstructured information such as equipment inspection records, fault work orders, maintenance logs, and repair reports, enables accurate extraction of various key entities and their relationships. Preferably, RoBERTa's function is to transform the input text into context-dependent word vectors after word segmentation. These word vectors are then used as input to a BiLSTM task structure. The BiLSTM models from both forward and backward directions, capturing the contextual order information in the text sequence and outputting a feature vector that combines the context and the order of elements. This feature vector is then input into a CRF (Contextual Representation Rendering Function) to jointly decode the entire label sequence, considering the dependencies between labels (e.g., the label "equipment" cannot be directly followed by the label "maintenance plan"). The result is the output of the optimal entity label sequence, achieving globally optimal annotation. The output of the BiLSTM-CRF model is a vector representation combined with the context.

[0035] Furthermore, entity-relationship mapping is implemented between all operational fault data and all multi-dimensional fault data to obtain integrated operational fault data. Specifically, the RoBERTa-BiLSTM-CRF model is used to extract unstructured text. Basic entity information with category labels has not yet been merged and may contain redundancy or ambiguity. Therefore, after obtaining preliminary structured entities, similar or duplicate entities from different text records, historical databases, and maintenance manuals are matched and merged. By calculating the semantic similarity of entities or descriptions, it is determined whether entries from different sources point to the same electrical equipment, fault type, or location, thereby completing entity alignment and disambiguation to form a consistent and complete fault database. Sentence-BERT uses a contrastive loss function as its loss function calculation formula, as follows: Where Y represents the sample class; Y=0 indicates that the two input samples x1 and x2 belong to the same class, and vice versa. m is the marginal value. DW is the Euclidean distance between the two input texts, calculated using the following formula: in, Embedding vectors generated for Sentence-BERT.

[0036] If Y=1, the distance between two input samples is greater than m, and parameter optimization is not required. However, if the distance between two samples is less than m, but they are actually different entities (semantic conflict or different contexts), incorrect merging may occur, leading to semantic confusion in the knowledge base. For example, "winding overheat" and "oil temperature too high" might be incorrectly merged into the same fault phenomenon. Parameters should be adjusted promptly to increase the distance between them. To accurately calculate these distances in the semantic space, each piece of text related to electrical equipment needs to be converted into a fixed-length sentence vector. This step is accomplished through the pooling mechanism. Sentence-BERT uses a pooling mechanism to convert BERT's variable-length word vectors into fixed-length sentence vectors. Average pooling involves averaging two word vectors to form corresponding sentence vectors u and v; u and v, along with their absolute difference |uv|, are concatenated to form a combined feature vector; the fused features are input into a fully connected network, and the class probability distribution is output through the Softmax function. The mathematical expression for the class probability is as follows: in, It is a weight parameter matrix.

[0037] In semantic similarity calculation, the similarity between two sentences is determined by measuring the distance between their corresponding vectors u and v. When two sentences are semantically similar, the distance between their corresponding vectors u and v is small; conversely, the distance is large. The calculation formula is as follows: Let the similarity threshold be... like If u and v are considered to be the same entity, they will be merged. This is the result of the entity integration decision.

[0038] The similarity threshold fluctuates between 0 and 1. The closer the similarity is to 1, the more similar the two sentences are semantically; the closer the similarity is to 0, the greater the semantic difference between the two sentences.

[0039] The model maps each text entity to a high-dimensional semantic vector space, making texts with similar semantics or pointing to the same entity close in distance in the vector space, while texts with different semantics are far apart. This enables semantic matching and disambiguation of cross-source entities, which can reduce duplicate entries and ambiguous information, making structured knowledge more reliable and providing a standardized and computable knowledge base for subsequent fault cause analysis, maintenance plan recommendation and intelligent operation and maintenance.

[0040] Further, in step S6, the updated fault database undergoes knowledge structuring processing to obtain a fault knowledge graph. Specifically, after obtaining high-quality structured entities and relationships that have undergone fusion and disambiguation, these entities are treated as nodes, and the relationships between entities are treated as edges. These are then organized and stored in the graph database Neo4j to form a queryable, reasonable, and visualizeable fault knowledge graph. The knowledge graph not only preserves entities and their attributes but also expresses complex causal, relational, and auxiliary judgment relationships, enabling intelligent retrieval and decision support. Triples in the knowledge graph are the basic units constituting the knowledge graph, consisting of a subject, predicate, and object, describing the relationships between entities in a structured form. The fused entities and their relationships are organized in the form of triples: Where h is the head entity, t is the tail entity, and r is the relation type. It is a collection of merged entities. A collection of relation types.

[0041] Import the triples into the Neo4j graph database, where each entity is mapped to a node and each relation is mapped to an edge. Attribute fields can be attached to both nodes and edges.

[0042] Preferably, based on the constructed fault knowledge graph, reasoning is performed using a rule engine and the py2neo interface to achieve automatic linkage analysis of alarm information and diagnostic results. When a specific alarm is received, the graph can quickly locate the historical fault modes and maintenance plans of electrical equipment, and provide interpretable causal chains or decision-making basis. A set of domain rules is defined on the knowledge graph. If an electrical device exhibits a specific fault symptom and the cause of that symptom is known to be unique, then the cause of the fault is deduced. If the cause of the fault matches a known repair solution, then the corresponding handling suggestion is returned. This is formally represented as: in, This is a new relationship derived through reasoning.

[0043] Preferably, the key functions in Py2neo are as follows: `nodes()` primarily calls the `NodeMatch` object, executes a query equivalent to `MATCH(n) RETURN n`, and returns a full node generator, supporting `count()` to count the total number of nodes or to iterate through the node set; `relationships()` works similarly, obtaining the entire database's relationship chain, which can be sorted chronologically using `order_by()`; `run()` directly executes native Cypher statements, returning a `Result` object containing nodes / relationships / paths; `data()` automatically converts the query results into a Python dictionary structure for easy JSON serialization; `find_one()` or `find()` can be used to find specific nodes, while `match` or `match_one` functions are used to query specific relationships. Utilizing the Python interface provided by py2neo, the graph database can be queried and updated, with query results including not only diagnostic conclusions but also inference paths. This allows maintenance personnel to trace the basis of their reasoning. Therefore, using the query function of the py2neo library, the required electrical equipment fault knowledge graph data can be efficiently retrieved. When querying related knowledge about electrical equipment, the "multi-point grounding of iron core" node and all its associated nodes can be obtained. For details, please refer to [link to relevant documentation]. Figure 5 , Figure 5 The diagram illustrates the "multi-point grounding of iron core" node and its associated query in a fault detection method for electrical equipment according to one embodiment of the present invention. In this diagram, blue nodes, with blue edges pointing to the center, represent the cause of the fault; red edges, pointing to red nodes, represent the fault phenomenon; green edges, pointing to green nodes, represent the inspection method; yellow edges, pointing to yellow nodes, represent auxiliary judgment; and orange edges, pointing to orange nodes, represent the faulty component.

[0044] When electrical equipment malfunctions, fault knowledge graphs can also be used to deduce and query the cause of the fault, the location of the fault, auxiliary diagnostic methods, and repair procedures. For details, please refer to [link to relevant documentation]. Figure 6 , Figure 6 This diagram illustrates the fault cause query and solution measures of a fault detection method for electrical equipment according to one embodiment of the present invention. Figure 6 The meanings represented by the different colored edges and nodes in the middle are... Figure 5 Consistent and preferred, when electrical equipment shows a rapid increase in CO and CO2 content, the query results indicate that the fault location is the winding. The cause may be "poor contact in the eddy current winding connection (or welding) of the leakage magnetic circuit" or "short circuit in structural components or magnetic shielding." When the cause cannot be determined, the auxiliary judgment method provided by the graph database can be used for further judgment. When the cause is "poor contact in the eddy current winding connection (or welding) of the leakage magnetic circuit," "a 1.1 times overcurrent will cause a significant increase in dissolved gases in the oil." When the cause is "short circuit in structural components or magnetic shielding," "the DC resistance is unstable and has a large deviation; even under low voltage excitation, total hydrocarbons will continue to be generated." After determining the cause and location of the fault, the repair method is determined according to whether the fault is an overheating fault or a discharge fault. The repair method for "short circuit in structural components or magnetic shielding" found in the graph database is "inspect the magnetic shielding fixing bolts and insulating pads, test the resistance to ground, repair the short circuit and strengthen the insulation."

[0045] Preferably, integrating gradient boosting-based residual networks with various functional algorithms in fault knowledge graphs can construct a framework for a fault diagnosis reasoning system for electrical equipment based on gradient boosting residuals and fault knowledge graphs. For details, please refer to [link to relevant documentation]. Figure 7 , Figure 7 The diagram shows a framework of a fault diagnosis reasoning system in a fault detection method for electrical equipment according to one embodiment of the present invention. It includes three modules: data acquisition, data analysis, and data service.

[0046] Specifically, the data acquisition module, as the foundation of the diagnostic system, transforms semi-structured and unstructured data from electrical equipment faults into structured data, enabling entity relationship extraction, knowledge fusion, and knowledge processing in the data analysis module.

[0047] Specifically, the data analysis module, a key challenge in the diagnostic system, processes the semi-structured and unstructured data generated by electrical equipment faults, constructing a fault diagnosis reasoning model based on a fault knowledge graph. First, the data selected by the data acquisition module is manually labeled into structured knowledge suitable for training. This labeled data is then used as input to the RoBERTa-BiLSTM-CRF algorithm model. Through training and learning, the required entities and relationships are obtained, resulting in triples. Knowledge fusion between entities is performed based on the Sentence-Bert model, and the results are stored in the Neo4j graph database. The diagnostic model then uses entity matching to map real-time status data onto entities in the graph, achieving fault diagnosis.

[0048] Specifically, the data service module is designed for end users, providing knowledge graph-driven intelligent query and diagnostic functions. To improve user-friendliness, this paper wraps Neo4j's Cypher statements and implements pattern-based queries using py2neo. The intelligent query function includes entity fuzzy query and relation query, allowing users to query objects under a specific relation of a single entity, or to query all objects associated with a given entity. The intelligent diagnostic function, through input of fault symptoms, accurately identifies the cause and location of the fault, provides auxiliary diagnostic methods, and offers repair solutions.

[0049] Another embodiment of the present invention provides a fault diagnosis device for electrical equipment. For details, please refer to [link to relevant documentation]. Figure 8 , Figure 8 The diagram shown illustrates the structure of a fault diagnosis device for electrical equipment according to one embodiment of the present invention. The device includes: A fault detection module is used to acquire real-time operating data of the target electrical equipment during its real-time operation, perform fault analysis on the real-time operating data based on a constructed fault knowledge graph, and obtain fault detection results for the target electrical equipment. The construction process of the fault knowledge graph includes: The data acquisition module is used to acquire sample data of all the target electrical devices, wherein the sample data includes first fault sample data and second normal sample data; A data augmentation module is used to obtain sample augmentation data, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic masking layer; The model optimization module is used to train the initial fault detection model based on the sample augmentation data to obtain an optimized fault detection model, and input the historical operating data of each of the target electrical devices into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data; The data optimization module is used to extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database, realize entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data based on multi-modal mapping, obtain integrated operational fault data, and update the initial fault database based on the integrated operational fault data to obtain the updated fault database. The graph construction module is used to perform knowledge structuring processing on the updated fault database to obtain the fault knowledge graph.

[0050] Another embodiment of the present invention provides a fault detection device for electrical equipment, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a fault detection method for electrical equipment as described above. For details, please refer to... Figure 9 , Figure 9 The diagram shows a structural schematic of a fault detection device for an electrical device according to one embodiment of the present invention. Specifically, it includes a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21. When the processor 21 executes the computer program, it implements the steps described in the above embodiment of a fault diagnosis method for an electrical device, for example... Figure 1 The steps S1-S6 described above; or, when the processor 21 executes the computer program, it implements the functions of each module in the above-described device embodiments, such as the data acquisition module 11. For example, the computer program can be divided into one or more modules, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the fault diagnosis device 20 of the electrical equipment.

[0051] In another embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the device containing the computer-readable storage medium executes the computer program, it implements a fault detection method for electrical equipment as described above. Specifically, those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0052] Accordingly, embodiments of the present invention provide a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform steps in a fault detection method for electrical equipment as described in the above embodiments, for example... Figure 1 Steps S1 to S6 as described above.

[0053] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: (1) The technical solution of this application firstly reduces the sample bias caused by differences in data acquisition environment, equipment model and working conditions by performing data augmentation processing on the sample data through the adversarial network layer and semantic mask layer, so as to maintain the accuracy stability during subsequent fault detection. (2) Based on the sample augmentation data, the fault detection model is iteratively optimized to achieve accurate learning of the characteristics of deviation-sensitive faults; entity-relationship mapping is performed on the operational fault data and multi-dimensional fault data to update the fault database; (3) Based on the updated fault database, a fault knowledge graph was constructed, which enabled fault detection during the real-time operation of electrical equipment, as well as comprehensive and timely detection of the operating status of electrical equipment.

[0054] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A fault detection method for electrical equipment, characterized in that, include: During the real-time operation of the target electrical equipment, real-time operating data of the target electrical equipment is acquired. Fault analysis is then performed on the real-time operating data based on a constructed fault knowledge graph to obtain the fault detection results of the target electrical equipment. The construction process of the fault knowledge graph includes: Obtain sample data for all the target electrical devices, wherein the sample data includes first fault sample data and second normal sample data; Sample augmentation data is obtained, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic mask layer; Based on the sample augmentation data, the initial fault detection model is trained to obtain an optimized fault detection model. The historical operating data of each of the target electrical devices is input into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data. Extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database. Based on multimodal mapping, realize the entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data to obtain integrated operational fault data. Update the initial fault database based on the integrated operational fault data to obtain the updated fault database. The updated fault database is subjected to knowledge structuring processing to obtain the fault knowledge graph.

2. The fault detection method for electrical equipment as described in claim 1, characterized in that, The obtained sample augmentation data includes: A gradient penalty condition is introduced into the adversarial network layer to constrain the first fault sample data, thereby obtaining the third sample data; A regression model is introduced into the semantic mask layer to perform semantic imputation on the second normal sample data to obtain the fourth sample data. The third sample data and the fourth sample data are integrated to obtain the sample augmentation data.

3. The fault detection method for electrical equipment as described in claim 1, characterized in that, The initial fault detection model is trained based on the sample augmentation data to obtain an optimized fault detection model. Historical operating data of each target electrical device is input into the optimized fault detection model to obtain operational fault data, including: The initial fault detection model is constructed based on a gradient boosting residual network. The sample augmentation data is input into the initial fault detection model for training to obtain the optimized fault detection model; Obtain the historical operating data of each of the target electrical devices; The historical operation data is filtered and denoised to obtain the historical operation data to be detected; Based on the optimized fault detection model, fault detection is performed on the historical operating data to be detected to obtain the operating fault data.

4. The fault detection method for electrical equipment as described in claim 1, characterized in that, The multimodal mapping method is used to map the entity-relationship data of all operational fault data to all multidimensional fault data, resulting in integrated operational fault data, including: Extract the multi-dimensional fault data corresponding to each target electrical device from the initial fault database, wherein the multi-dimensional fault data consists of at least the electrical device model, fault code, maintenance record, inspection report and test indicators; The operational fault data is subjected to semantic vector transformation to obtain operational fault data to be mapped; The multi-dimensional fault data is transformed into semantic vectors to obtain the multi-dimensional fault data to be mapped. Determine the cosine similarity between the operational fault data to be mapped and the multi-dimensional fault data to be mapped; Based on the cosine similarity and sequence labeling model, entity-relationship mapping is performed on the operational fault data and the multi-dimensional fault data to obtain integrated operational fault data.

5. The fault detection method for electrical equipment as described in claim 1, characterized in that, The process of updating the initial fault database based on the integrated operational fault data to obtain an updated fault database includes: The hash values ​​of the integrated data of each operational failure are obtained based on the hash algorithm; Based on all the hash values, the corresponding duplicate integrated data of the operation failures are removed to obtain integrated data of operation failures without redundancy. The initial fault database is updated based on the non-redundant operational fault integration data to obtain the updated fault database.

6. The fault detection method for electrical equipment as described in claim 1, characterized in that, The step of performing knowledge structuring processing on the updated fault database to obtain the fault knowledge graph includes: Extract the entities from the updated fault database and convert the entities into node information; Extract the relationships from the updated fault database and convert the relationships into edge information; Based on the graph database, an index is established for the node information and the edge information, and a uniqueness constraint is applied to the node information and the edge information to obtain the fault knowledge graph.

7. The fault detection method for electrical equipment as described in claim 1, characterized in that, After obtaining the fault detection results of the target electrical equipment, the method further includes: Based on the causal paths of the fault knowledge graph, the cause of the fault, the location of the fault, and the repair plan for the target electrical equipment are determined. The cause of the fault, the location of the fault, and the repair plan are uploaded to the user terminal, and the cause-and-effect path is visualized on the user terminal.

8. A fault detection device for electrical equipment, characterized in that, include: A fault detection module is used to acquire real-time operating data of the target electrical equipment during its real-time operation, perform fault analysis on the real-time operating data based on a constructed fault knowledge graph, and obtain fault detection results for the target electrical equipment. The construction process of the fault knowledge graph includes: The data acquisition module is used to acquire sample data of all the target electrical equipment, wherein the sample data includes first fault sample data and second normal sample data; A data augmentation module is used to obtain sample augmentation data, wherein the sample augmentation data is jointly determined by the first faulty sample data processed by the adversarial network layer and the second normal sample data processed by the semantic mask layer; The model optimization module is used to train the initial fault detection model based on the sample augmentation data to obtain an optimized fault detection model, and input the historical operating data of each of the target electrical devices into the optimized fault detection model to obtain operating fault data, wherein the historical operating data is determined by the sample data; The data optimization module is used to extract multi-dimensional fault data corresponding to each target electrical device in the initial fault database, realize entity-relationship mapping between all the operational fault data and all the multi-dimensional fault data based on multi-modal mapping, obtain integrated operational fault data, and update the initial fault database based on the integrated operational fault data to obtain the updated fault database. The graph construction module is used to perform knowledge structuring processing on the updated fault database to obtain the fault knowledge graph.

9. A fault detection device for electrical equipment, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a fault detection method for an electrical device as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements a fault detection method for electrical equipment as described in any one of claims 1 to 7.

Images