Electrical equipment fault diagnosis method and system based on knowledge graph

Abstract

The invention relates to the technical field of electrical equipment fault diagnosis, in particular to an electrical equipment fault diagnosis method and system based on a knowledge graph, and the method comprises the steps: obtaining the multi-mode sensing data of electrical equipment, and constructing a fault diagnosis knowledge graph; mapping entities and relationships in the atlas to a low-dimensional continuous vector space by adopting a knowledge representation learning algorithm, and extracting semantic feature vectors; inputting the data into a graph neural network model to learn neighborhood information, executing link prediction and knowledge completion, and determining a potential fault causal chain in a preset operation environment; and matching the causal chain with the logic rule base, performing causal logic derivation on the diagnosis result in combination with a large-scale language model, and outputting a fault root cause analysis report and an operation and maintenance management decision suggestion. According to the method, multi-modal perception data and knowledge representation learning are deeply fused, implicit fault association which is not explicitly recorded in historical operation and maintenance data is mined by using a graph neural network, and interpretable analysis is provided in combination with a language model.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment fault diagnosis technology, specifically to a knowledge graph-based method and system for electrical equipment fault diagnosis. Background Technology

[0002] Against the backdrop of the rapid development of the energy internet and smart grids, the physical topology of power systems is becoming increasingly complex, and the dimensions and scale of monitoring data are growing exponentially. The stable operation of electrical equipment has become a core cornerstone for ensuring the order of modern life, and the industry has placed higher demands on asset management and high-precision predictive maintenance throughout the entire life cycle of equipment.

[0003] Early electrical equipment diagnosis primarily relied on analytical methods based on physical models and simple expert systems. While these methods performed well under single operating conditions, they showed limitations when handling multi-source heterogeneous data. Subsequently, the rise of deep learning techniques such as convolutional neural networks and long short-term memory networks improved the accuracy of feature extraction. However, purely data-driven deep learning models are essentially "black box" mechanisms, unable to elucidate the causal logic of fault occurrence and struggling to directly extract unstructured "dark knowledge" such as maintenance manuals, historical inspection logs, and expert experience. Furthermore, existing diagnostic systems often face the problem of sparse power fault data, resulting in many potential fault correlations not being explicitly recorded, making it difficult to achieve closed-loop management from fault identification to decision-making recommendations.

[0004] Existing technologies struggle to effectively integrate multi-source heterogeneous data with deep expert mechanisms, resulting in untraceable causal logic for complex faults. Furthermore, data sparsity limits prediction accuracy, making it impossible to provide highly interpretable closed-loop decision support and resource allocation recommendations for power operation and maintenance.

[0005] To address this, a knowledge graph-based method and system for diagnosing electrical equipment faults are proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a knowledge graph-based method and system for diagnosing electrical equipment faults, which achieves traceability of fault logic and predictive maintenance by deeply integrating mechanistic knowledge with multi-source heterogeneous data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A knowledge graph-based method for fault diagnosis of electrical equipment includes: Acquire multimodal sensing data for power equipment and construct a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, voiceprint data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; A knowledge representation learning algorithm is used to map the entities and relationships in the power fault diagnosis knowledge graph to a low-dimensional continuous vector space, and extract the semantic feature vectors of the entities and relationships. The semantic feature vectors are then input into a graph neural network model, which learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment. The potential fault causal chain is matched with a preset logical rule base, and the diagnostic results are deduced by combining a large-scale language model to output an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

[0008] Preferably, constructing a power fault diagnosis knowledge graph based on the multimodal perception data specifically includes the following steps: Adopting a top-down design principle, defining the entity categories, attribute sets, and relationship sets of the power fault diagnosis knowledge graph; wherein, the entity categories include equipment entities, components, fault modes, monitoring signals, and processing suggestions; the relationship sets include physical inclusion relationships, logical triggering relationships, fault-causing relationships, and operation and maintenance suggestion relationships; using a named entity recognition algorithm to identify the equipment entity names, component names, and fault phenomena from the unstructured text data, and using a relationship extraction model incorporating an attention mechanism to determine the causal logical path between the equipment entity, the component name, and the fault phenomenon, generating a first set of triples; Deep learning algorithms are used to identify external defect entities in the visual image data and abnormal noise pattern entities in the acoustic print data, and the identification results are attached to the corresponding equipment entities as fault feature entities. At the same time, the physical parameter threshold states in the sensor monitoring data are converted into monitoring signal entities in the entity category to generate a second triplet set. The first triplet set and the second triplet set are aligned and fused in terms of attributes and stored in the native graph database to form the power fault diagnosis knowledge graph.

[0009] Preferably, the step of generating the first set of triples is as follows: the named entity recognition algorithm uses a long short-term memory network unit combined with dynamic convolution to capture semantic features in the unstructured text data; a convolutional neural network combined with an attention mechanism and a bidirectional long short-term memory network are introduced to construct the relation extraction model; the attention mechanism is used to automatically assign weights to keywords in the unstructured text data to identify semantic associations between the core entities; the causal logical path between the core entities is calculated and output through the relation extraction model; the causal logical path includes a containing component path describing the component inclusion relationship, and a damage-induced abnormal path describing the fault evolution; the extracted core entities are used as nodes, and the causal logical path is used as edges, and the association mapping is performed according to the triple structure to generate the first set of triples.

[0010] Preferably, a convolutional neural network is used to identify external defect entities in the visual image data; a feature extraction algorithm combined with voiceprint pattern matching is used to identify abnormal noise pattern entities in the voiceprint data; and a long short-term memory network is used to process the physical parameters in the sensor monitoring data and convert them into the monitoring signal entities.

[0011] Preferably, the step of extracting the semantic feature vector is as follows: using the TransR model as the knowledge representation learning algorithm, constructing corresponding vector representations for each entity and each relation in the power fault diagnosis knowledge graph; projecting the entity from the original entity space to the relation space to which the corresponding relation belongs, and processing one-to-many and / or many-to-many associations between entities and relations in the power equipment; calculating based on the geometric displacement constraints of the head entity vector, relation vector, and tail entity vector in the relation space, so that the interconnected entities and relations satisfy the preset translation relationship in the vector space; and iteratively training by continuously optimizing the loss function of the geometric displacement constraints to obtain the semantic feature vector that can retain the original triple logical structure.

[0012] Preferably, determining the potential fault causal chain includes: using the semantic feature vector as initial input, learning the neighborhood information of nodes in the power fault diagnosis knowledge graph using the graph neural network model, and capturing the deep nonlinear association between the entity and the relationship through multi-layer aggregation operations; using the graph neural network model to perform link prediction on the missing logical associations in the power fault diagnosis knowledge graph, and completing the implicit fault relationships not explicitly recorded in the historical operation and maintenance data; based on the completed implicit fault relationships, combined with the meteorological conditions and operating parameters in the preset operating environment, calculating the degree of similarity between the current equipment state and the known fault mode in the vector space; and connecting the interrelated entities and the completed implicit fault relationships according to the logical evolution order to determine the potential fault causal chain.

[0013] Preferably, the output of an interpretable fault root cause analysis report and operation and maintenance management decision recommendations includes: inputting the potential fault causal chain into the large-scale language model, and using the large-scale language model to perform textual parsing of the semantic relationships in the potential fault causal chain; retrieving preset physical constraints from the logical rule base, and performing logical verification and fact alignment on the results of the textual parsing; extracting entity attributes and historical cases related to the potential fault causal chain from the power fault diagnosis knowledge graph, generating the fault root cause analysis report according to the logical evolution order, and tracing and displaying the causal path of fault triggering; and matching corresponding operating procedures and resource requirements from the processing suggestion class entities of the power fault diagnosis knowledge graph according to the severity of the fault root cause analysis report, and outputting the operation and maintenance management decision recommendations.

[0014] A knowledge graph-based electrical equipment fault diagnosis system includes: Knowledge graph construction module: acquires multimodal sensing data for power equipment and constructs a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, acoustic print data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; Causal chain module: Using a knowledge representation learning algorithm, the entities and relationships in the power fault diagnosis knowledge graph are mapped to a low-dimensional continuous vector space, and semantic feature vectors of the entities and relationships are extracted; the semantic feature vectors are input into a graph neural network model, and the graph neural network model learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph, and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment; Decision output module: Matches the potential fault causal chain with a preset logical rule base, and combines a large-scale language model to perform causal logical deduction on the diagnostic results, outputting an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application utilizes convolutional neural networks to extract image features and combines them with long short-term memory networks to process time-series monitoring data, transforming unstructured expert experience into a graph semantic network. This deep integration of mechanistic knowledge and data-driven approaches marks the entry of diagnosis into the cognitive intelligence stage. Compared to purely data-driven black-box mechanisms, it significantly improves the comprehensive judgment capability of complex faults, laying a semantic foundation for subsequent accurate diagnosis.

[0016] 2. This invention utilizes graph neural networks to learn and perform link prediction through neighborhood information. By mining the implicit correlations between meteorological conditions, environmental disturbances, and equipment status that are not explicitly recorded, the system can identify more complete potential fault causal chains. This enhances the system's ability to capture rare or nonlinear state evolution faults, achieving a leap from "passive maintenance" to "predictive maintenance."

[0017] 3. This invention deeply integrates a large-scale language model with a logical rule base, effectively solving the illusion problem commonly found in generative models through physical constraint verification. By tracing the causal path of fault triggering, the system provides operation and maintenance experts with a traceable diagnostic path supported by semantic logic, significantly enhancing the credibility and professionalism of the diagnostic results. Furthermore, this mechanism supports few-sample learning and prior knowledge injection, reducing dependence on massive labeled data and achieving a closed-loop decision output from diagnosis to treatment suggestions, effectively reducing the long-term operation, maintenance, and upgrade costs of the system. Attached Figure Description

[0018] Figure 1 This is a flowchart of a knowledge graph-based electrical equipment fault diagnosis method proposed in this invention. Figure 2 This is a flowchart illustrating a knowledge graph-based electrical equipment fault diagnosis method proposed in this invention. Figure 3 This is a system architecture diagram of an electrical equipment fault diagnosis system based on knowledge graphs proposed in this invention. Detailed Implementation

[0019] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1

[0021] Please see Figures 1 to 2 This invention provides a method for diagnosing electrical equipment faults based on knowledge graphs, the technical solution of which is as follows: A knowledge graph-based method for fault diagnosis of electrical equipment, such as Figures 1-2 As shown, it includes: Acquire multimodal sensing data for power equipment and construct a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, voiceprint data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; A knowledge representation learning algorithm is used to map the entities and relationships in the power fault diagnosis knowledge graph to a low-dimensional continuous vector space, and extract the semantic feature vectors of the entities and relationships. The semantic feature vectors are then input into a graph neural network model, which learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment. The potential fault causal chain is matched with a preset logical rule base, and the diagnostic results are deduced by combining a large-scale language model to output an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

[0022] Further, a power fault diagnosis knowledge graph is constructed based on the multimodal perception data, specifically including the following steps: Adopting a top-down design principle, the entity categories, attribute sets, and relationship sets of the power fault diagnosis knowledge graph are defined; wherein, the entity categories include equipment entities, components, fault modes, monitoring signals, and processing suggestions; the relationship sets include physical inclusion relationships, logical triggering relationships, fault-causing relationships, and operation and maintenance suggestion relationships; Named entity recognition algorithms are used to identify the equipment entity names, component names, and fault phenomena from the unstructured text data, and a relationship extraction model incorporating an attention mechanism is used to determine the causal logical path between the equipment entity, the component name, and the fault phenomenon, generating a first set of triples; Deep learning algorithms are used to identify external defect entities in the visual image data and abnormal noise pattern entities in the acoustic print data, and the identification results are attached to the corresponding equipment entities as fault feature entities. At the same time, the physical parameter threshold states in the sensor monitoring data are converted into monitoring signal entities in the entity category to generate a second triplet set. The first triplet set and the second triplet set are aligned and fused in terms of attributes and stored in the native graph database to form the power fault diagnosis knowledge graph.

[0023] Specifically, in the entity alignment and attribute fusion process, an entity disambiguation algorithm based on similarity calculation is adopted. By calculating the similarity of the name strings and the similarity of the topological structure of entities in different triplet sets, it is determined whether they point to the same physical device object. For the dynamic attributes in the second triplet set, an entity-attribute-value structure is adopted. The real-time monitored physical parameters are used as dynamic attribute values ​​and are mounted to the corresponding entity nodes of the power fault diagnosis knowledge graph through timestamp indexing, so as to realize the semantic association between static ledgers and dynamic operating status.

[0024] The edit distance algorithm is used to calculate the text similarity of entity names in different triples. For long words in power industry terminology, the text score is obtained by normalization, which ranges from 0 to 1.

[0025] The Jaccard similarity coefficient is used to calculate the overlap of the adjacent node sets of two entity nodes in order to determine their positional consistency in the physical topology or logical hierarchy of the power system.

[0026] The text score and the topological overlap are weighted and summed to obtain the final comprehensive similarity index; when the index exceeds a preset judgment threshold (e.g., 0.85), it is determined that the entity points to the same physical device entity.

[0027] Different attributes that are determined to be the same entity will be deduplicated and merged. For static attributes that conflict, the existing ledger data in the original graph database will be used as the standard. For dynamic parameters, they will be appended in the order of timestamps.

[0028] This embodiment achieves structured integration of all-dimensional information about power equipment through top-down ontology design and multi-source data fusion mechanisms. Utilizing entity disambiguation algorithms and entity-attribute-value structures, it solves the problems of "information silos" and entity redundancy under multi-source heterogeneous data, ensuring the uniqueness of the diagnostic subject. By attaching real-time dynamic parameters to static ledgers via timestamp indexes, it realizes dynamic semantic association between physical equipment and digital models, significantly improving the system's ability to finely represent equipment status and providing a solid data foundation for achieving fine-grained root cause localization from the overall to the local level.

[0029] Further, the step of generating the first set of triples is as follows: the named entity recognition algorithm uses a long short-term memory network unit combined with dynamic convolution to capture semantic features in the unstructured text data; a convolutional neural network combined with an attention mechanism and a bidirectional long short-term memory network are introduced to construct the relation extraction model; the attention mechanism is used to automatically assign weights to keywords in the unstructured text data to identify semantic associations between the core entities; the causal logical path between the core entities is calculated and output through the relation extraction model; the causal logical path includes the containing component path describing the component inclusion relationship, and the damage-induced abnormal path describing the fault evolution; the extracted core entities are used as nodes, and the causal logical path is used as edges, and the association mapping is performed according to the triple structure to generate the first set of triples.

[0030] In the named entity recognition process, the dynamic convolutional layer captures local key features in the unstructured text data through convolutional kernels with adaptive weights, and uses its output vector as the input of the long short-term memory network unit to enhance the model's ability to capture long-distance dependencies of power industry terms; the relation extraction model automatically calculates the contribution of each phrase in the text to a specific causal path through an attention mechanism matrix, and inputs the high-weight semantic features into the fully connected layer for classification, thereby achieving accurate determination of causal relationships such as "damage" and "abnormality".

[0031] This embodiment significantly improves the efficiency of extracting "dark knowledge" from fragmented unstructured knowledge by introducing dynamic convolution and attention mechanisms. The dynamic convolutional layer accurately captures the local features of power industry terminology through adaptive weights, solving the problem of insufficient long-distance dependency capture in traditional models when dealing with complex syntax. Combined with the automatic calculation of the contribution of the attention mechanism matrix to causal paths, it achieves accurate determination of semantic logic such as "damage" and "abnormality," providing high-quality triplet data support for constructing a fault knowledge graph with causal tracing capabilities.

[0032] Furthermore, a convolutional neural network is used to identify external defect entities in the visual image data; a feature extraction algorithm combined with voiceprint pattern matching is used to identify abnormal noise pattern entities in the voiceprint data; and a long short-term memory network is used to process the physical parameters in the sensor monitoring data and convert them into the monitoring signal entities.

[0033] In the voiceprint recognition process, the acoustic features of the voiceprint data are first extracted using Mel frequency cepstral coefficients, and then Euclidean distance is calculated between the voiceprint data and the pre-stored abnormal sound patterns in the knowledge graph to achieve pattern matching. In the monitoring signal conversion process, the long short-term memory network performs sliding window sampling on the time-series monitoring data, identifies numerical fluctuation features that exceed the preset normal range, and encapsulates them into monitoring signal entities with specific type labels (such as: abnormal temperature, discharge intensity) as evidence nodes that trigger subsequent reasoning logic.

[0034] This solution integrates convolutional neural networks, Mel-frequency cepstral coefficients, and long short-term memory networks to achieve precise multi-dimensional perception of the "visual, auditory, and physical" status of power equipment. By utilizing Euclidean distance for acoustic pattern matching and combining it with sliding window sampling to identify numerical fluctuations in time-series data, it can accurately capture subtle abnormal sounds and nonlinear state evolutions that are difficult to detect using traditional methods. This technology efficiently transforms raw heterogeneous signals into monitoring signal entities with semantic labels, providing deterministic evidence nodes for subsequent reasoning and significantly improving the sensitivity and accuracy of fault trigger identification in complex operating environments.

[0035] Further, the step of extracting the semantic feature vector is as follows: using the TransR model as the knowledge representation learning algorithm, constructing corresponding vector representations for each entity and each relation in the power fault diagnosis knowledge graph; projecting the entity from the original entity space to the relation space to which the corresponding relation belongs, and processing one-to-many and / or many-to-many associations between entities and relations in the power equipment; calculating based on the geometric displacement constraints of the head entity vector, relation vector, and tail entity vector in the relation space, so that the interconnected entities and relations satisfy the preset translation relationship in the vector space; and iteratively training by continuously optimizing the loss function of the geometric displacement constraints to obtain the semantic feature vector that can retain the original triple logical structure.

[0036] Specifically, the geometric displacement constraint follows the formula ,in and Representing the head entity vector respectively Tail entity vector Projection matrix in relation space The projection vector under the action, i.e. , ,and The vector represents the relationship. During iterative training, the loss function employs a hinge-based interval loss function. By comparing the scores of positive sample triples (real-world fault logic) and negative sample triples (error logic formed by randomly replaced entities), the distance between positive sample entities is narrowed, while the distance between negative sample entities is widened. The dimension of the low-dimensional continuous vector space is preset to 100 to 200 dimensions to ensure that complex semantic relationships are preserved while reducing computational complexity.

[0037] This scheme introduces the TransR knowledge representation learning model to transform symbolic knowledge into low-dimensional continuous vectors, enabling numerical computation of power fault logic. Utilizing a relational space projection mechanism, it effectively solves the complex one-to-many and many-to-many relationship problems commonly found in electrical equipment, exhibiting stronger semantic expressive power compared to the basic model. Iterative training using a hinge loss function based on intervals ensures significant discriminative power between positive and negative samples in the vector space, efficiently preserving the logical structure of the original triples in a low-dimensional space of 100 to 200 dimensions, providing a high-precision mathematical foundation for subsequent in-depth mining of hidden fault relationships.

[0038] Further, determining the potential fault causal chain includes: using the semantic feature vector as initial input, learning the neighborhood information of nodes in the power fault diagnosis knowledge graph using the graph neural network model, and capturing the deep nonlinear association between the entity and the relationship through multi-layer aggregation operations; using the graph neural network model to perform link prediction on the missing logical associations in the power fault diagnosis knowledge graph, and completing the implicit fault relationships not explicitly recorded in the historical operation and maintenance data; based on the completed implicit fault relationships, combined with the meteorological conditions and operating parameters in the preset operating environment, calculating the degree of similarity between the current equipment state and the known fault mode in the vector space; and connecting the interrelated entities and the completed implicit fault relationships according to the logical evolution order to determine the potential fault causal chain.

[0039] Specifically, the multi-layer aggregation operation adopts a graph attention network architecture. By calculating the contribution weights of neighboring nodes to the central node, the feature vectors of adjacent entities and their relationships are weighted and summed to update the representation vector of the central node. In the link prediction process, a scoring function is used to model the probability that a specific relationship exists between any two nodes in the graph. When the calculated probability score exceeds a preset threshold (e.g., 0.85), it is determined that the hidden fault relationship exists and automatic completion is performed. The calculation of proximity in the vector space specifically adopts a cosine similarity or Euclidean distance algorithm. The real-time feature vector that integrates meteorological conditions and operating parameters is compared with the typical fault mode vectors stored in the knowledge graph. By setting a similarity threshold, the fault mode to which the current device belongs is determined, and a directed path search is performed along the hidden fault relationship, thereby connecting and forming a complete causal chain of the potential fault.

[0040] This solution significantly enhances the system's ability to uncover hidden fault correlations by introducing graph attention networks and link prediction mechanisms. It effectively addresses the common "data sparsity" problem in power system operation and maintenance by using weighted aggregation of neighborhood information to capture deep nonlinear correlations. Furthermore, by combining meteorological and operational parameters for cosine similarity comparison, it can accurately identify and complete implicit logic not explicitly recorded in historical data, thereby determining the complete causal chain of faults and achieving advanced prediction and accurate source tracing of complex evolution paths.

[0041] Furthermore, the system outputs an interpretable root cause analysis report and operation and maintenance management decision recommendations, including: inputting the potential fault causal chain into the large-scale language model, and using the large-scale language model to perform textual parsing of the semantic relationships in the potential fault causal chain; retrieving preset physical constraints from the logical rule base, and performing logical verification and fact alignment on the results of the textual parsing; extracting entity attributes and historical cases related to the potential fault causal chain from the power fault diagnosis knowledge graph, generating the fault root cause analysis report according to the logical evolution order, and tracing and displaying the causal path of fault triggering; and matching corresponding operating procedures and resource requirements from the processing suggestion class entities of the power fault diagnosis knowledge graph according to the severity of the fault root cause analysis report, and outputting the operation and maintenance management decision recommendations.

[0042] Specifically, in the logical verification and fact alignment process, the preliminary text generated by the large-scale language model is extracted as candidate triples, and these triples are compared with the physical constraints in the logical rule base (e.g., the causal correspondence between transformer gas protection actions and characteristic gas component concentrations) for logical consistency. If there is a conflict, the logical rule base takes precedence, and the output of the large-scale language model is corrected through prompt word engineering to eliminate illusions. When matching the corresponding operating procedures and resource requirements, a semantic similarity algorithm is used to associate the attributes of the fault mode entities in the fault root cause analysis report with the processing suggestion entities in the graph. This automatically retrieves and outputs the list of maintenance tools, spare parts models, and standard operating instructions text that match the fault mode most closely, thereby achieving closed-loop guidance from diagnosis to execution.

[0043] This embodiment effectively eliminates the "illusion" problem of generative models by deeply integrating a large-scale language model with a logical rule base and using physical constraints for logical verification. This mechanism supports tracing and displaying the causal path of faults from a knowledge graph, significantly enhancing the interpretability of diagnostic results and the trustworthiness of maintenance experts. Furthermore, by matching suggested entities using semantic similarity algorithms, the system can automatically output accurate lists of maintenance tools and standard operating procedures, achieving closed-loop management from diagnosis to execution, greatly improving the efficiency of root cause analysis and reducing mean time to repair.

[0044] This embodiment achieves a leap from a "black box model" to "cognitive intelligence" in power fault diagnosis by deeply integrating multimodal perception data and knowledge representation learning technology. First, knowledge graph construction technology unifies heterogeneous data such as visual, voiceprint, text, and sensor data, eliminating information silos and significantly improving the comprehensive perception dimension of complex fault symptoms. Second, the introduction of graph neural network link prediction and knowledge completion mechanisms enables accurate discovery of hidden causal relationships even under sparse power fault sample conditions, achieving advanced prediction of fault causal chains. Finally, the combination of a large-scale language model and logical rule base ensures that the diagnostic results possess both the readability of natural language and the rigor of power physical mechanisms. This method significantly improves the accuracy and interpretability of fault root cause analysis, provides closed-loop maintenance decision support for operation and maintenance personnel, effectively reduces the average maintenance time of critical equipment, and ensures the safe and stable operation of the power system.

[0045] Example 2

[0046] Please see Figure 3 This invention provides an electrical equipment fault diagnosis system based on knowledge graphs, the technical solution of which is as follows: A knowledge graph-based electrical equipment fault diagnosis system, such as Figure 3 As shown, it includes: Knowledge graph construction module: acquires multimodal sensing data for power equipment and constructs a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, acoustic print data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; Causal chain module: Using a knowledge representation learning algorithm, the entities and relationships in the power fault diagnosis knowledge graph are mapped to a low-dimensional continuous vector space, and semantic feature vectors of the entities and relationships are extracted; the semantic feature vectors are input into a graph neural network model, and the graph neural network model learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph, and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment; Decision output module: Matches the potential fault causal chain with a preset logical rule base, and combines a large-scale language model to perform causal logical deduction on the diagnostic results, outputting an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

[0047] This embodiment uses the diagnosis of sudden faults in an oil-immersed power transformer in a substation as a specific application scenario to describe in detail the application process of the system of the present invention: The system first acquires real-time data through multi-source monitoring equipment deployed within the station: An industrial camera captured images of obvious oil leakage on the transformer bushing, providing visual image data. An acoustic fingerprint sensor detected an internal, uneven, muffled discharge sound, providing acoustic fingerprint data. An online monitoring device showed the transformer oil temperature instantly rising from 55°C to 85°C, triggering a light gas protection activation, providing sensor monitoring data. The graph construction module used a convolutional neural network to identify oil leakage entities, used Mel-frequency cepstral coefficients to match discharge noise patterns, and used LSTM to identify temperature anomaly signals. These entities are dynamically linked to transformer graph nodes using timestamp indexes, achieving a semantic association between static records and dynamic operating status, forming a real-time triplet set.

[0048] The causal chain module utilizes pre-constructed semantic feature vectors for deep reasoning: The system uses the TransR model to project the identified entities such as oil seepage, discharge sound, and high oil temperature into the relation space, processes the one-to-many association between entities and relations, and extracts semantic feature vectors.

[0049] Graph neural network models aggregate transformer neighborhood information to uncover hidden fault relationships not explicitly recorded in the historical operation and maintenance data of the equipment under sparse power fault samples.

[0050] The system combines the current high ambient humidity meteorological conditions and uses cosine similarity in vector space to calculate the degree of similarity between the current state and the known fault modes, and connects them in the logical evolution order to form a complete causal chain of potential faults.

[0051] The decision output module transforms the above logical chain into operation and maintenance suggestions: After parsing the causal chain using a large-scale language model, the logic rule base is retrieved to confirm the causal correspondence between the transformer gas protection action and the characteristic gas concentration. The output is then corrected through prompt word engineering to eliminate illusions and ensure factual alignment.

[0052] The system extracts entity attributes and historical cases from the graph, generates a root cause analysis report, and traces the causal path from oil seepage leading to insulation dampness to the final temperature rise.

[0053] Based on the severity of the report, the system automatically matches and outputs the corresponding list of maintenance tools, spare parts models, and standard operating procedure texts from the processing suggestion entities, thus achieving closed-loop guidance from diagnosis to execution.

[0054] By applying this embodiment, maintenance personnel can obtain logically supported professional diagnostic results and operational suggestions after a fault occurs, which significantly improves the efficiency of root cause analysis and ensures the safe and stable operation of the power system.

[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for fault diagnosis of electrical equipment based on knowledge graphs, characterized in that, include: Acquire multimodal sensing data for power equipment and construct a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, voiceprint data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; A knowledge representation learning algorithm is used to map the entities and relationships in the power fault diagnosis knowledge graph to a low-dimensional continuous vector space, and extract the semantic feature vectors of the entities and relationships. The semantic feature vectors are then input into a graph neural network model, which learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment. The potential fault causal chain is matched with a preset logical rule base, and the diagnostic results are deduced by combining a large-scale language model to output an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

2. The method for diagnosing electrical equipment faults based on knowledge graphs according to claim 1, characterized in that, Constructing a power fault diagnosis knowledge graph based on the multimodal perception data specifically includes the following steps: Adopting a top-down design principle, defining the entity categories, attribute sets, and relationship sets of the power fault diagnosis knowledge graph; wherein, the entity categories include equipment entities, components, fault modes, monitoring signals, and processing suggestions; the relationship sets include physical inclusion relationships, logical triggering relationships, fault-causing relationships, and operation and maintenance suggestion relationships; using a named entity recognition algorithm to identify the equipment entity names, component names, and fault phenomena from the unstructured text data, and using a relationship extraction model incorporating an attention mechanism to determine the causal logical path between the equipment entity, the component name, and the fault phenomenon, generating a first set of triples; Deep learning algorithms are used to identify external defect entities in the visual image data and abnormal noise pattern entities in the acoustic print data, and the identification results are attached to the corresponding equipment entities as fault feature entities. At the same time, the physical parameter threshold states in the sensor monitoring data are converted into monitoring signal entities in the entity category to generate a second triplet set. The first triplet set and the second triplet set are aligned and fused in terms of attributes and stored in the native graph database to form the power fault diagnosis knowledge graph.

3. The method for diagnosing electrical equipment faults based on knowledge graphs according to claim 2, characterized in that, The steps for generating the first set of triples are as follows: the named entity recognition algorithm uses a long short-term memory network unit combined with dynamic convolution to capture semantic features in the unstructured text data; a convolutional neural network combined with an attention mechanism and a bidirectional long short-term memory network are introduced to construct the relation extraction model; the attention mechanism is used to automatically assign weights to keywords in the unstructured text data to identify semantic associations between the core entities; the relation extraction model is used to calculate and output the causal logical path between the core entities; the causal logical path includes the containing component path describing the component inclusion relationship, and the damage-induced anomaly path describing the fault evolution; The extracted core entities are used as nodes, and the causal logic paths are used as edges. The association mapping is performed according to the triplet structure to generate the first triplet set.

4. The method for diagnosing electrical equipment faults based on knowledge graphs according to claim 2, characterized in that, A convolutional neural network is used to identify external defect entities in the visual image data; a feature extraction algorithm combined with voiceprint pattern matching is used to identify abnormal noise pattern entities in the voiceprint data; and a long short-term memory network is used to process the physical parameters in the sensor monitoring data and convert them into the monitoring signal entities.

5. The method for diagnosing electrical equipment faults based on knowledge graphs according to claim 1, characterized in that, The steps for extracting the semantic feature vectors are as follows: using the TransR model as the knowledge representation learning algorithm, constructing corresponding vector representations for each entity and each relation in the power fault diagnosis knowledge graph; projecting the entity from the original entity space to the relation space to which the corresponding relation belongs, and processing one-to-many and / or many-to-many associations between entities and relations in the power equipment; The geometric displacement constraints of the head entity vector, relation vector, and tail entity vector in the relation space are calculated to ensure that the related entities and relations satisfy the preset translation relationship in the vector space. The semantic feature vector that can retain the original triple logical structure is obtained by iteratively training by continuously optimizing the loss function of the geometric displacement constraints.

6. The method for diagnosing electrical equipment faults based on knowledge graphs according to claim 1, characterized in that, Determining the potential fault causal chain includes: using the semantic feature vector as initial input, learning the neighborhood information of nodes in the power fault diagnosis knowledge graph using the graph neural network model, and capturing the deep nonlinear association between the entity and the relationship through multi-layer aggregation operations; using the graph neural network model to perform link prediction on the missing logical associations in the power fault diagnosis knowledge graph, and completing the implicit fault relationships not explicitly recorded in the historical operation and maintenance data; based on the completed implicit fault relationships, combined with the meteorological conditions and operating parameters in the preset operating environment, calculating the degree of similarity between the current equipment state and the known fault mode in the vector space; and connecting the interrelated entities and the completed implicit fault relationships according to the logical evolution order to determine the potential fault causal chain.

7. A knowledge graph-based method for diagnosing electrical equipment faults according to claim 1, characterized in that, The system outputs an interpretable root cause analysis report and operation and maintenance management decision recommendations, including: inputting the potential fault causal chain into the large-scale language model, and using the large-scale language model to perform textual parsing of the semantic relationships in the potential fault causal chain; retrieving preset physical constraints from the logical rule base, and performing logical verification and fact alignment on the results of the textual parsing; extracting entity attributes and historical cases related to the potential fault causal chain from the power fault diagnosis knowledge graph, generating the fault root cause analysis report according to the logical evolution order, and tracing and displaying the causal path of fault triggering; and matching corresponding operating procedures and resource requirements from the processing suggestion class entities of the power fault diagnosis knowledge graph according to the severity of the fault root cause analysis report, and outputting the operation and maintenance management decision recommendations.

8. A knowledge graph-based electrical equipment fault diagnosis system, characterized in that, include: Knowledge graph construction module: acquires multimodal sensing data for power equipment and constructs a power fault diagnosis knowledge graph based on the multimodal sensing data; the multimodal sensing data includes visual image data describing external defects of the equipment, acoustic print data describing the internal state of the equipment, unstructured text data describing fault mechanisms and expert experience, and sensor monitoring data describing physical parameters; Causal chain module: Using a knowledge representation learning algorithm, the entities and relationships in the power fault diagnosis knowledge graph are mapped to a low-dimensional continuous vector space, and semantic feature vectors of the entities and relationships are extracted; the semantic feature vectors are input into a graph neural network model, and the graph neural network model learns the neighborhood information of the nodes in the power fault diagnosis knowledge graph, and performs link prediction and knowledge completion to determine the potential fault causal chain of the power equipment under a preset operating environment; Decision output module: Matches the potential fault causal chain with a preset logical rule base, and combines a large-scale language model to perform causal logical deduction on the diagnostic results, outputting an interpretable fault root cause analysis report and operation and maintenance management decision suggestions.

Images