A Multimodal Power Grid Fault Diagnosis Method and System Based on Causal Event Graph

Abstract

This invention discloses a multimodal power grid fault diagnosis method and system based on causal event graphs, belonging to the field of intelligent operation and maintenance technology for power systems. The method includes: preprocessing historical fault case data of power grid equipment to construct a causal event graph database; upon receiving a fault diagnosis request, a planning agent analyzes the request, combines power domain knowledge to generate an initial set of fault hypotheses, assigns corresponding credibility scores to the fault hypotheses, retrieves causal subgraphs as evidence in an iterative loop, and updates the credibility scores of the fault hypotheses using an inference agent; and generates a multimodal diagnosis report after a termination condition is met. This invention solves the problems of insufficient causal modeling and poor interpretability in traditional methods, significantly improving diagnostic accuracy, efficiency, and user trust.

Description

Technical Field

[0001] This invention belongs to the field of power grid fault analysis technology, and relates to a multimodal power grid fault diagnosis method and system based on causal event graphs. Background Technology

[0002] As a critical infrastructure, the safe and stable operation of the power grid is of paramount importance. Sudden failures of core equipment such as transformers, circuit breakers, and GIS switchgear can not only cause enormous economic losses but also trigger widespread power outages, threatening public safety. Through long-term operation and maintenance practices, the power industry has accumulated a vast amount of equipment failure cases, existing in the form of accident reports, maintenance records, and technical analyses—an extremely valuable knowledge base. However, this data is often unstructured text, with varying formats, complex content, and includes a large amount of important multimodal information such as images and data curves. Traditional fault diagnosis mainly relies on the personal experience of maintenance personnel and manual retrieval of historical data. This method is inefficient, difficult to pass on knowledge, and prone to misjudgment when facing new or complex faults due to insufficient experience or incomplete information retrieval.

[0003] With the development of artificial intelligence technology, especially the rise of Natural Language Processing and Retrieval Augmentation (RAG) technology, the intelligent utilization of these massive amounts of case data has become possible. Existing intelligent question answering systems typically adopt a technical path of "text vectorization - similarity retrieval - language model generation." While these methods perform reasonably well when dealing with simple and direct questions, their limitations become increasingly apparent when facing complex power grid fault diagnosis scenarios. First, they treat fault cases as independent text blocks, lacking modeling of the deep causal logic between events within the fault, making it difficult for the model to perform effective reasoning and only providing superficially related text fragments. Second, existing RAG systems typically employ one-time or fixed-depth retrieval strategies, failing to adaptively adjust the depth of information collection according to the complexity and ambiguity of the question. For complex problems requiring multi-step reasoning, insufficient contextual information often prevents in-depth "thinking," resulting in low answer accuracy.

[0004] Finally, current generated results are mostly plain text. Even when retrieved images are included, they are merely simple listings, failing to organically combine images with textual explanations. This makes it difficult to intuitively and clearly explain the evolution process and key aspects of the fault, resulting in poor interpretability and user experience. Therefore, there is an urgent need for a novel intelligent diagnostic method that combines causal logic modeling, dynamic adaptive retrieval, and multimodal visual evidence fusion. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multimodal power grid fault diagnosis method and system based on causal event graphs. This aims to address the problems of insufficient modeling of fault causal logic, rigid retrieval strategies, lack of reasoning ability, and poor interpretability of output results in existing technologies. This invention constructs an intelligent agent system that simulates the cognitive process of human experts "proposing hypotheses, seeking evidence, and verifying reasoning," achieving high-precision, high-efficiency, and highly interpretable intelligent diagnosis of power grid equipment faults. Compared with traditional methods relying on experience and static retrieval, this invention can dynamically adjust the reasoning depth in complex scenarios, significantly reducing the false positive rate and improving the traceability of diagnostic results.

[0006] To achieve the above objectives, the technical solution proposed in this invention is: a multimodal power grid fault diagnosis method based on causal event graphs, comprising the following steps:

[0007] S1: Preprocess historical fault case data of power grid equipment to construct a causal event graph database;

[0008] S2: When a fault diagnosis request is received from the user, the planning agent analyzes the fault diagnosis request, combines it with knowledge of the power field, generates an initial set of fault hypotheses, and assigns a corresponding confidence score to the fault hypothesis.

[0009] S3: In each round of the loop, select the fault hypothesis with the highest confidence score from the current set of fault hypotheses and generate a structured graph retrieval; then retrieve one or more relevant causal subgraphs as the evidence set for the current round; finally, the reasoning agent evaluates the support and contradiction of the obtained evidence for the current fault hypothesis and updates the confidence score of the fault hypothesis.

[0010] S4: At the end of each iteration, check if there are any fault hypotheses whose credibility scores reach the credibility acceptance threshold; if the termination condition is met, exit the loop; otherwise, the reasoning agent will revise, merge or split the unverified fault hypotheses based on the current evidence set, generate a new set of fault hypotheses and recalculate the credibility scores, and then return to step S3.

[0011] S5: After the loop terminates, select the fault hypothesis with the highest credibility score that has been verified, and generate an explanatory text for fault cause analysis and handling suggestions based on the causal subgraph that supports the fault hypothesis; during the generation process, when a key event node is described, call the visual language model to calculate the explanatory relevance, and adaptively embed images or charts into the generated text to achieve multimodal evidence fusion, and finally output a fault diagnosis report containing text and images or charts corresponding to the text. The text is organized in the order of the causal event chain to make the fault diagnosis report content traceable.

[0012] Specifically, the process of constructing a causal event graph database includes: using a pre-trained language model to perform entity recognition on the original text of historical power grid equipment failure cases and extracting candidate nodes; then using a relation extraction model to identify the logical relationships described in the original text and construct causal or temporal connection edges between nodes; and simultaneously, using image recognition and content analysis technologies to automatically associate the images in the historical power grid equipment failure cases with the corresponding event nodes in the causal event graph, ultimately obtaining the causal event graph database.

[0013] Specifically, the confidence scores of all initial failure hypotheses are normalized so that the sum of the confidence scores is 1.

[0014] Specifically, in step S3, a structured graph retrieval query is generated for the fault hypothesis with the highest credibility score. The process of the reasoning agent evaluating the evidence adopts a Bayesian update framework. The credibility score of each fault hypothesis is represented by its posterior probability. Whenever new evidence is retrieved, the corresponding posterior probability is updated according to the Bayesian formula.

[0015] Specifically, when the credibility scores of all fault hypotheses are below the credibility rejection threshold and the number of iterations reaches the upper limit, it is determined that the existing information is insufficient to make a diagnosis, and a clarifying question is proactively generated for the user to obtain additional information to generate a new fault hypothesis, thereby guiding the next round of the diagnostic process.

[0016] The present invention also provides a multimodal power grid fault diagnosis system based on causal event graphs, comprising:

[0017] The data processing and database building module is used to preprocess historical fault case data of power grid equipment and build a causal event graph database.

[0018] The hypothesis generation module is used to analyze the problem through a planning agent when a fault diagnosis request is received from the user, combine knowledge in the power field, generate an initial set of fault hypotheses, and calculate an initial confidence score for each fault hypothesis.

[0019] The iterative verification module is used to select the failure hypothesis with the highest credibility score in each round of the loop, generate a structured graph retrieval query, retrieve relevant causal subgraphs from the causal event graph database as evidence, and evaluate the support and contradiction of the evidence for the hypothesis through the reasoning agent, and update the credibility score of the failure hypothesis.

[0020] The loop control module is used to check whether the credibility score of the failure hypothesis has reached the credibility acceptance threshold at the end of each iteration. If the termination condition is met, the loop will be exited. Otherwise, the inference agent will be driven to correct, merge or split the unverified failure hypothesis, generate a new set of failure hypotheses and recalculate the credibility score, and restart the iteration verification.

[0021] The multimodal report generation module is used to select the verified fault hypothesis with the highest confidence score after the loop terminates, generate explanatory text of fault cause analysis and handling suggestions based on the cause-effect subgraph supporting the hypothesis, and call a visual language model to calculate the explanatory relevance between the image or chart and the text during the generation process, adaptively embed the image or chart into the text, and generate a multimodal fault diagnosis report containing the text and the image or chart.

[0022] Specifically, the data processing and database building module includes:

[0023] The entity recognition unit is used to perform entity recognition on the original text of historical fault cases of power grid equipment using a pre-trained language model, and extract equipment, components, monitoring parameters, fault phenomena, and operating behaviors as candidate nodes.

[0024] The relation extraction unit is used to identify causal, temporal, or conditional logical relationships in the original text using a relation extraction model, and to construct connection edges between nodes.

[0025] The multimodal association unit is used to automatically associate maintenance images, operating curves, and event nodes in the causal event graph of historical fault cases of power grid equipment through image recognition and content analysis technology.

[0026] The graph storage unit is used to store the structured causal event graph in the causal event graph database to support graph retrieval and query.

[0027] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the multimodal power grid fault diagnosis method based on causal event graphs.

[0028] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned multimodal power grid fault diagnosis method based on causal event graphs.

[0029] This invention achieves several improvements in methodology and system architecture: by combining causal event graphs with multimodal data, it breaks through the traditional diagnostic methods that rely solely on textual information, significantly enhancing reasoning capabilities in complex scenarios; through adaptive retrieval and iterative verification mechanisms, it simulates the dynamic thinking process of human-like experts—"proposing hypotheses—finding evidence—verifying reasoning"—resulting in significant improvements in accuracy and robustness; by intelligently and selectively embedding images and charts into diagnostic reports, it not only enhances the interpretability and intuitiveness of conclusions but also improves the user's understanding experience; by introducing an interactive clarification mechanism, the system can proactively guide users to supplement key clues when information is insufficient, thereby avoiding erroneous or incomplete judgments due to a lack of evidence. In summary, this invention can achieve more efficient, accurate, and interpretable results in the fault diagnosis of various key power grid equipment such as transformers, circuit breakers, and GIS equipment, demonstrating good versatility and application value. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the framework flow of the method of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in detail below.

[0032] Reference Figure 1 A multimodal power grid fault diagnosis method based on causal event graphs includes the following steps:

[0033] S1. Data processing and database construction steps: Preprocess the historical fault case data of power grid equipment, which includes unstructured information such as text descriptions, operating curves, sensor data, maintenance images and operation records, and construct a causal event graph database.

[0034] S2. Hypothesis generation step: When a fault diagnosis request is received from the user, the planning agent analyzes the fault diagnosis request, combines it with knowledge of the power field, generates an initial set of fault hypotheses, and assigns a corresponding confidence score to the fault hypothesis.

[0035] S3. Iterative Hypothesis Verification Steps: In each round of the loop, select the fault hypothesis with the highest confidence score from the current set of fault hypotheses and generate a structured graph retrieval; then retrieve one or more relevant causal subgraphs as the evidence set for the current round; finally, the reasoning agent evaluates the support and contradiction of the obtained evidence for the current fault hypothesis and updates the confidence score of the fault hypothesis.

[0036] S4. Loop Control and Termination Steps: At the end of each iteration, check if there are any fault hypotheses whose credibility scores reach the credibility acceptance threshold; if the termination condition is met, exit the loop; otherwise, the reasoning agent will revise, merge or split the unverified fault hypotheses based on the current evidence set, generate a new set of fault hypotheses and recalculate the credibility scores, and then return to step S3.

[0037] S5. Multimodal Fault Diagnosis Report Generation Steps: After the loop terminates, select the fault hypothesis with the highest confidence score that has been verified, and generate an explanatory text with fault cause analysis and handling suggestions based on the causal subgraph that supports the fault hypothesis; during the generation process, when a key event node is described, call the visual language model to calculate the explanatory relevance, and adaptively embed images or charts into the generated text, thereby realizing multimodal evidence fusion, and finally generating a multimodal fault diagnosis report containing the text and the images or charts.

[0038] Step S1 of this method preprocesses the massive amount of historical fault case data of power grid equipment. Instead of simply segmenting it into text blocks, it uses advanced natural language processing and computer vision technologies to parse each historical fault case of power grid equipment into a causal event graph. This causal event graph not only includes textual information, but also automatically associates multimodal data such as key images, charts, and operation records with corresponding event nodes, enabling the knowledge base to have cross-modal reasoning capabilities.

[0039] In this causal event graph, each node represents a specific physical event, equipment status, or abnormal monitoring data, such as "partial discharge signal of phase A winding exceeds the standard" or "operation of circuit breaker tripping." The directed edges between nodes precisely represent the causal or temporal relationships between them, formalized as follows:

[0040] ;

[0041] in, This represents the i-th node. This represents the j-th node. This represents the m-th node. This represents the nth node.

[0042] Furthermore, key images from historical power grid equipment failure cases, such as photos of equipment burn marks and data trend charts from dissolved gas analysis (DGA), are bound as attribute information to corresponding event nodes in the causal event graph, forming a multimodal knowledge network. This design ensures that the diagnostic process involves not only abstract logical reasoning but also intuitive visual evidence, thereby enhancing the understanding of experts and users.

[0043] The process of constructing a causal event graph database includes: using a pre-trained language model to perform entity recognition on the original text of historical power grid equipment fault case data, extracting equipment, components, monitoring parameters, fault phenomena, and operational behaviors as candidate nodes; then using a relation extraction model to identify logical relationships such as "caused by... leading to..." and "first...then..." described in the original text, constructing causal or temporal connection edges between nodes; simultaneously, through image recognition and content analysis techniques, automatically associating images in historical power grid equipment fault cases with corresponding event nodes in the causal event graph. The final causal event graph database is denoted as: ,in: This represents a set of nodes, corresponding to device status, fault phenomena, or operational events. Represents the nth node; Denotes the set of edges, edges Used to characterize causal, temporal, or conditional relationships; This represents the i-th node. This represents the j-th node. This represents the edge connecting the i-th node and the j-th node; For a multimodal dataset, each node Multimodal evidence, including state images, can be attached. Operating curve Or inspection form ; Represents a node Multimodal evidence.

[0044] More specifically, the scanned historical fault case reports of power grid equipment are first processed using Optical Character Recognition (OCR) technology to obtain the original text. Next, a natural language processing pipeline based on fine-tuning of a large language model processes the text. This pipeline includes: 1) Named Entity Recognition (NER), used to accurately identify specialized entities such as "500kV Shenwei Substation No. 1 Main Transformer," "SF6 Gas Pressure Anomaly," and "Differential Protection"; 2) Relation Extraction (RE), which focuses on identifying causal, temporal, and conditional logical relationships contained in the text. For example, from the sentence "Due to the overheating of the connector, the contacts melted, ultimately causing a phase-to-phase short circuit," a causal chain such as "connector overheating" -> "contact melting" -> "phase-to-phase short circuit" is extracted. These extracted entities and relationships together constitute the nodes and edges of the causal event graph. Simultaneously, a computer vision module can be set up to process the images and charts in the case study. Through image classification and content recognition, it associates images (such as equipment appearance images and infrared thermal images) with corresponding event nodes in the causal event graph (such as "bushing cracks" and "severe overheating of phase B current transformer"). All this structured information is ultimately stored in a high-performance causal event graph database, such as Neo4j, for complex graph pattern graph retrieval queries.

[0045] In step S2 of this embodiment, when the system receives a fault diagnosis request from a user, such as "A 220kV transformer tripped, accompanied by gas protection activation. What could be the cause?", a dedicated planning agent module is activated. The fault diagnosis request is treated as a problem. First, deep semantic analysis is performed on the problem, and combined with domain knowledge, this open-ended problem is transformed into a set of specific fault hypotheses that can be subsequently verified, such as: "Fault Hypothesis 1: Inter-turn short circuit occurred in the internal winding," "Fault Hypothesis 2: A through-type external short circuit caused the protection to malfunction," and "Fault Hypothesis 3: The on-load tap changer malfunctioned." In this way, the present invention can quickly generate multiple candidate explanations with limited input information, improving the comprehensiveness of the diagnosis. For fault diagnosis requests... The initial set of fault assumptions is represented as follows: , This represents the m-th fault hypothesis, with the superscript indicating the cycle number and 0 indicating the initial state; using Let represent the initial confidence score of the i-th failure hypothesis. ,satisfy: , indicating that under the single root cause hypothesis, the confidence score sums to 1.

[0046] In practice, when a user inputs a fault diagnosis request through an interactive interface, the Planner Agent begins its work. This Planner Agent is essentially a large language model guided by specific instructions. After receiving the user's question, its internal instructions guide it to "role-play," simulating a power grid fault analysis expert to conduct a preliminary analysis and decompose the fault diagnosis request into several mutually exclusive or parallel core assumptions. For example, for a fault diagnosis request "GIS equipment frequently exhibits partial discharge signals, but routine testing has not revealed any obvious abnormalities. Please analyze the cause," the planning agent might output a structured list of initial fault hypotheses: { "Initial Fault Hypothesis Set": [{ "Hypothesis ID": "H1", "Description": "Interior contains free metal particles", "Confidence Score": 0.9}, { "Hypothesis ID": "H2", "Description": "Insulator surface has deposits or is damp", "Confidence Score": 0.6}, { "Hypothesis ID": "H3", "Description": "Sensor itself is faulty or there is signal interference", "Confidence Score": 0.5} ]}, "Normalized Confidence": [{ "H1": 0.45}, { "H2": 0.30}, { "H3": 0.25}], with each fault hypothesis accompanied by an initial confidence score.

[0047] In step S3, the system enters an iterative "hypothesis-verification" loop. In each loop, the executing AI selects the most probable fault hypothesis and generates a highly targeted graph search query around it. This graph search query is no longer a vague keyword but may be a structured graph pattern, such as "finding a causal chain directly connected to the 'gas protection action' node, with an upstream path containing nodes of 'winding discharge' or 'rapid oil temperature rise'." The system uses this graph search query to efficiently search the causal event graph database, finding all matching causal subgraphs. These causal subgraphs provide direct evidence supporting or refuting the current fault hypothesis. This search method effectively avoids interference from irrelevant information, making diagnosis more accurate and efficient.

[0048] Next, an intelligence agent responsible for reasoning rigorously evaluates the retrieved evidence. It analyzes the logical consistency, completeness, and contradictions between these causal subgraphs and the current hypothesis. Through this process, the agent dynamically updates the credibility score of each hypothesis. If a hypothesis is supported by a large amount of strong, uncontradictory evidence, its credibility score will significantly increase; conversely, if the evidence contradicts it, its credibility score will decrease. If the evidence retrieved in this round is insufficient to make a clear judgment, or if new, unexpected phenomena are discovered, the reasoning agent will proactively revise existing hypotheses or derive new, more refined sub-hypotheses, and then initiate a new round of retrieval-verification cycle. This process closely resembles the dynamic thinking process of an experienced engineer facing a complex problem, constantly proposing conjectures, searching for clues, eliminating errors, and gradually approaching the truth. Thus, this invention can maintain continuous exploration and optimization when facing complex, ambiguous, or information-scarce situations, rather than hastily drawing conclusions.

[0049] For example, in step S3, control flow is transferred to the execution and reasoning agent, and the iterative verification loop begins. In the first round, the system selects the most credible fault hypothesis H1: "Free metal particles exist internally." Based on this fault hypothesis, the execution agent automatically generates a standard graph retrieval query statement, following the standard Cypher graph database path pattern syntax, for example:

[0050] In this statement, the keyword and These represent "match" and "return" respectively; parentheses Square brackets are used to represent nodes. Used to indicate relationships between nodes; curly braces The inner part represents the attribute constraints of the node; the colon... The "event" that follows indicates the node label. Expression In the middle, letters "Causes" is a relation variable name used to refer to the set of matched relations; "causes" indicates the relation type. This is a path length qualifier in Cypher syntax, indicating that edges of the "cause" relation type can be repeated consecutively 1 to 5 times in the path, meaning that 1 to 5 "cause" relations are allowed between the starting node and the target node. The purpose of this graph retrieval query is to find all paths in the causal event graph database that originate from the "free metal particle" event, proceed through no more than 5 causal steps, and ultimately lead to the "partial discharge signal." After executing the graph retrieval query, the causal event graph database returns several matching causal subgraphs.

[0051] The Reasoner Agent receives these causal subgraphs as evidence. It is also a large language model guided by specific reasoning instructions. It analyzes each causal subgraph one by one and evaluates its fit with the H1 fault hypothesis. For example, a causal subgraph showing that "metal particles" cause "surface electric field distortion of the basin insulator," leading to "partial discharge," strongly supports the H1 fault hypothesis. Another causal subgraph might show that similar discharge phenomena are caused by "insulator aging," which, while not direct evidence, provides another possibility. The Reasoner Agent synthesizes all the evidence and updates the credibility score of the H1 fault hypothesis according to predefined evaluation rules (or a smaller evaluation model). Simultaneously, it evaluates other hypotheses, potentially lowering the credibility scores of the H2 and H3 fault hypotheses due to a lack of evidence. The loop terminates if, after one round, the credibility score of the H1 fault hypothesis exceeds the credibility acceptance threshold. If the condition is not met, but evidence frequently mentions "pot-type insulators," the reasoning agent may revise the H2 fault hypothesis to "discharge caused by particles or moisture on the surface of the pot-type insulator," and initiate a new round of more targeted searches.

[0052] For the fault hypothesis with the highest credibility score Generate structured graph retrieval queries t represents the number of iterations, where the reasoning agent evaluates evidence using a Bayesian update framework, and the credibility score of each failure hypothesis H represents its posterior probability. , where E is the accumulated set of evidence;

[0053] Whenever new evidence is retrieved, the corresponding posterior probability is updated according to Bayes' theorem:

[0054] ,in The conditional probability of the i-th piece of evidence The value is estimated by the natural language reasoning ability of a large language model, representing the degree of conformity of the i-th piece of evidence under the existing fault hypothesis and evidence. Let i represent the posterior probability of the i-th piece of evidence. Indicates the i-th piece of evidence. This represents the updated cumulative set of evidence. It represents the union of sets.

[0055] In step S4, the termination condition of the iteration loop is: when the credibility score of a certain fault hypothesis exceeds a preset credibility acceptance threshold, the system considers that the root cause has been found; or, when all fault hypotheses are falsified, or there is still no clear conclusion after a certain number of iterations, the system will judge that the information is insufficient and may ask the user a guiding question to obtain more clues.

[0056] Therefore, when the confidence scores of all failure hypotheses are low and the maximum number of iterations has been reached, If the system determines that the existing information is insufficient to make a diagnosis, it will proactively generate a clarifying question for the user. This is used to acquire additional information to generate new, more accurate fault hypotheses, thereby guiding the next round of the diagnostic process. Loop control can be represented as:

[0057] ;

[0058] Here, Stop represents the loop termination condition. This represents the i-th fault hypothesis. This represents the maximum current confidence score among all failure hypotheses. As a threshold for credibility, The confidence threshold is used to reject the candidate. The confidence score for the i-th fault hypothesis is its posterior probability.

[0059] Once the loop terminates and the final cause of the failure is determined, the multimodal answer generation step begins. Based on the validated failure hypothesis and the complete causal chain evidence supporting it, a natural language-based failure diagnosis report is organized and generated. This report goes beyond simply providing a conclusion; it systematically explains, step by step, how the failure evolved from its initial symptoms, through a series of intermediate stages, to its current outcome, following the causal chain. During the text generation process, when a key event node is reached, such as describing "partial discharge signal," the system intelligently invokes a built-in Visual Language Model (VLM) to determine the value of the associated "discharge signal waveform" in interpreting the current text. If the VLM determines that the waveform significantly enhances persuasiveness and clarity, the system adaptively embeds the waveform into the appropriate position within the text paragraph, possibly accompanied by a brief caption. Ultimately, the user is presented with a logically clear, well-supported, and visually rich professional diagnosis report, enabling them not only to know what happened but also why.

[0060] The process by which the visual language model calculates the explanatory relevance is as follows:

[0061] ;

[0062] The process of adaptively deciding whether to embed multimodal data is based on a utility function:

[0063] ;

[0064] in It is an image. It is the currently generated text fragment; It is the image-to-text explanatory relevance score calculated by the visual language model, representing whether the image can clearly explain the text content; For visual language model functions; It is the embedding utility score; It is the cost of the "cognitive complexity" of the image; only Embedding is only performed at specific times to ensure high relevance and understandability of the embedded data. The utility threshold; for The weights; for The weight.

[0065] Meanwhile, the fault diagnosis report has good traceability, and users can clearly see the relationship between reasoning path and evidence, which greatly improves trust and practicality.

[0066] In this embodiment, it is assumed that the H1 fault hypothesis is finally verified after two rounds of iteration. The system then activates the answer generation module. This module obtains the most complete causal chain supporting H1 and "translates" it into fluent natural language. The generator will describe it as follows: "After analysis, the root cause of the frequent partial discharge signals in this GIS equipment is most likely the presence of free metal particles inside it. The evolution path of this process is as follows: First, metal particles generated during long-term operation or installation of the equipment move randomly within the cavity under the influence of an electric field. When these particles move to high field strength regions, such as the surface of a basin insulator, they cause severe distortion of the local electric field." When describing the "surface of a basin insulator," the system will retrieve an "electric field simulation diagram" associated with this node. At this time, the Visual Language Model (VLM) is invoked, which receives the image and the current text "causes severe distortion of the local electric field" as input and outputs a high relevance score. Since the score exceeds the embedding threshold, the system inserts this simulation diagram below the current paragraph and automatically generates a caption: "Simulation diagram of electric field distortion of metal particles on the surface of the insulator." The entire generation process continues until the complete causal chain is explained and corresponding processing suggestions are attached, ultimately forming a complete, well-reasoned, and illustrated diagnostic report presented to the user.

[0067] The method of the present invention was compared with other methods, and the results are shown in Table 1.

[0068] Table 1: Performance comparison of this invention with other methods under various evaluation metrics

[0069]

[0070] Among them, BM25 is a classic sparse retrieval function that retrieves and ranks documents based on a ranking formula of term frequency and inverse document frequency; SimCSE is a dense sentence vector model obtained through unsupervised contrastive learning; StaticRAG is an end-to-end baseline method that relies solely on static retrieval and generation; CausalGraph is a retrieval and inference method based on causal graph structure modeling; AgentRAG is an improved method that introduces an intelligent agent mechanism for adaptive multi-round retrieval and inference; MultiRet is a cascaded retrieval method that supports multi-hop evidence chains; Recall@5 represents the proportion of successfully retrieved relevant items in the first 5 returned results; Precision@3 represents the proportion of truly relevant items in the first 3 returned results; NDCG@10 is the normalized depreciation cumulative gain, used to measure the ranking quality of the first 10 returned results. The method of this invention realizes intelligent fault diagnosis of power grid equipment based on causal modeling and multimodal fusion, and its performance under various evaluation indicators is superior to the comparative methods, exhibiting higher accuracy and practicality.

[0071] Another embodiment of the present invention provides a multimodal power grid fault diagnosis system based on causal event graphs, comprising:

[0072] The data processing and database building module is used to preprocess historical fault case data of power grid equipment and build a causal event graph database.

[0073] The hypothesis generation module is used to analyze the problem through a planning agent when a fault diagnosis request is received from the user, combine knowledge in the power field, generate an initial set of fault hypotheses, and calculate an initial confidence score for each fault hypothesis.

[0074] The iterative verification module is used to select the failure hypothesis with the highest credibility score in each round of the loop, generate a structured graph retrieval query, retrieve relevant causal subgraphs from the causal event graph database as evidence, and evaluate the support and contradiction of the evidence for the hypothesis through the reasoning agent, and update the credibility score of the failure hypothesis.

[0075] The loop control module is used to check whether the credibility score of the failure hypothesis has reached the credibility acceptance threshold at the end of each iteration. If the termination condition is met, the loop will be exited. Otherwise, the inference agent will be driven to correct, merge or split the unverified failure hypothesis, generate a new set of failure hypotheses and recalculate the credibility score, and restart the iteration verification.

[0076] The multimodal report generation module is used to select the verified fault hypothesis with the highest confidence score after the loop terminates, generate explanatory text of fault cause analysis and handling suggestions based on the cause-effect subgraph supporting the hypothesis, and call a visual language model to calculate the explanatory relevance between the image or chart and the text during the generation process, adaptively embed the image or chart into the text, and finally generate a multimodal fault diagnosis report containing the text and the image or chart.

[0077] Specifically, the data processing and database building module includes:

[0078] The entity recognition unit is used to perform entity recognition on the original text of historical fault cases of power grid equipment using a pre-trained language model, and extract equipment, components, monitoring parameters, fault phenomena, and operating behaviors as candidate nodes.

[0079] The relation extraction unit is used to identify causal, temporal, or conditional logical relationships in the original text using a relation extraction model, and to construct connection edges between nodes.

[0080] The multimodal association unit is used to automatically associate maintenance images, operating curves, and event nodes in the causal event graph of historical fault cases of power grid equipment through image recognition and content analysis technology.

[0081] The graph storage unit is used to store the structured causal event graph in the causal event graph database to support graph retrieval and query.

[0082] Building upon the above embodiments, a user interaction module is also included. This module receives fault diagnosis requests input by the user and outputs a final diagnosis report. When evidence is insufficient, it generates clarifying questions to the user to obtain additional information. The user interaction module provides a graphical interface that allows users to input fault diagnosis requests in natural language and displays evidence tracing and multimodal embedded content in the diagnosis report through highlights, hyperlinks, or floating windows.

[0083] Based on the above embodiments, the system also includes a case incremental update module, which is used to periodically or triggerally integrate new power grid equipment fault cases into the causal event graph database through the data processing and database building module, so as to realize the continuous evolution of the knowledge base.

[0084] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned multimodal power grid fault diagnosis method based on causal event graphs.

[0085] Another embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the aforementioned multimodal power grid fault diagnosis method based on causal event graphs.

[0086] The above description merely illustrates preferred embodiments of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make modifications or alterations to the above-disclosed content to create equivalent embodiments. However, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for multi-modal power grid fault diagnosis based on a causal event graph, characterized in that, The method comprises the following steps: S1: preprocessing the historical fault case data of power grid equipment, and constructing a causal event graph database; S2: when receiving a user input fault diagnosis request, an intelligent agent analyzes the fault diagnosis request, generates an initial fault hypothesis set in combination with power field knowledge, and calculates an initial credibility score for each fault hypothesis; S3: in each round of iteration, a fault hypothesis with the highest credibility score is selected from the current fault hypothesis set, and a structured graph retrieval query is generated; Then one or more related causal subgraphs are retrieved as the evidence set of the current round; finally, the reasoning intelligent agent evaluates the support degree and contradiction degree of the obtained evidence to the current fault hypothesis, and updates the credibility score of the fault hypothesis; S4: at the end of each iteration, it is checked whether the credibility score of the fault hypothesis reaches a credibility acceptance threshold; if the termination condition is met, the loop is exited; otherwise, the reasoning intelligent agent modifies, combines or splits the unverified fault hypotheses based on the current evidence set, generates a new fault hypothesis set and recalculates the credibility score, and then returns to step S3; S5: after the loop terminates, the fault hypothesis with the highest credibility score and the verified fault hypothesis is selected, and an explanatory text of fault cause analysis and processing suggestion is generated according to the causal subgraph supporting the fault hypothesis; during the generation process, when a key event node is described, a visual language model is called to calculate the correlation between the image or chart and the explanatory text, and when the correlation is greater than a preset threshold, the image or chart is inserted into the corresponding text paragraph, so as to realize multi-modal evidence fusion, and output a fault diagnosis report containing text and at least one image or chart corresponding to the text.

2. The method of claim 1, wherein, The process of constructing the causal event graph database comprises: using a pre-trained language model to perform entity recognition on the original text of the historical fault case of the power grid equipment, and extracting candidate nodes; then using a relation extraction model to identify the logical relationship described in the original text, and constructing causal or time sequence connection edges between nodes; at the same time, through image recognition and content analysis technology, the pictures in the historical fault case of the power grid equipment are automatically associated with the corresponding event nodes in the causal event graph, and finally the causal event graph database is obtained.

3. The method of claim 1, wherein, The credibility scores of all initial fault hypotheses are normalized so that the sum of the credibility scores is 1.

4. The method of claim 1, wherein, In step S3, a structured graph retrieval query is generated for the fault hypothesis with the highest credibility score, and the reasoning agent evaluates the evidence using a Bayesian update framework. The credibility score of each fault hypothesis represents its posterior probability. Whenever new evidence is retrieved, the corresponding posterior probability is updated according to the Bayesian formula.

5. The method of claim 1, wherein, When the credibility scores of all fault hypotheses are lower than the credibility rejection threshold, and the number of iterations reaches the upper limit, it is determined that the existing information is insufficient for diagnosis, and a clarification question is generated to the user to obtain additional information, so as to generate new fault hypotheses and guide the next round of diagnosis process.

6. A multi-modal power grid fault diagnosis system based on a causal event graph, characterized in that, The method comprises the following steps: A data processing and database building module is configured to preprocess historical fault case data of power grid equipment and build a causal event graph database; An assumption generation module is configured to generate an initial fault hypothesis set and assign an initial credibility score to each fault hypothesis by analyzing the problem through a planning agent and combining power field knowledge when receiving a user input fault diagnosis request; An iterative verification module is configured to select the fault hypothesis with the highest credibility score in each round of iteration, generate a structured graph retrieval query, retrieve relevant causal subgraphs as evidence from the causal event graph database, and evaluate the support and contradiction of the evidence to the hypothesis through an inference agent to update the credibility score of the fault hypothesis; A loop control module is configured to check whether the credibility score of the fault hypothesis reaches a credibility acceptance threshold at the end of each round of iteration, exit the loop if the termination condition is met, or drive the inference agent to modify, merge or split the unverified fault hypothesis, generate a new fault hypothesis set and recalculate the credibility score, and restart the iteration verification; A multi-modal report generation module is configured to select the verified fault hypothesis with the highest credibility score after the loop terminates, generate explanatory text of fault cause analysis and processing suggestions based on the causal subgraphs supporting the hypothesis, and call a visual language model to calculate the explanatory relevance of images or charts and text during the generation process, adaptively embed images or charts corresponding to the text in the text, and generate a multi-modal fault diagnosis report containing the text and the images or charts.

7. The system of claim 6, wherein, A user interaction module is also included to receive user input fault diagnosis requests and output final diagnosis reports, and generate clarification questions to obtain additional information when the evidence is insufficient.

8. The system of claim 7, wherein, The data processing and database building module specifically includes: An entity recognition unit is configured to use a pre-trained language model to perform entity recognition on the original text of the historical fault case of the power grid equipment, and extract the equipment, components, monitoring parameters, fault phenomena and operation behaviors as a candidate set of event nodes; A relation extraction unit is configured to use a relation extraction model to identify causal, temporal or conditional logical relationships in the original text, and build connection edges between event nodes; A multi-modal association unit is configured to automatically associate maintenance images, operation curves in the historical fault case of the power grid equipment with corresponding event nodes in the causal event graph through image recognition and content analysis technology; A graph storage unit is configured to store the structured causal event graph in the causal event graph database to support subsequent graph retrieval queries.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the multi-modal power grid fault diagnosis method based on the causal event graph according to any one of claims 1-5.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the multi-modal power grid fault diagnosis method based on the causal event graph according to any one of claims 1-5.

Images