Multi-mode general-purpose cooperative causal thinking chain reasoning power anomaly detection method and system

Abstract

The invention provides a power anomaly detection method and system based on multi-mode special collaborative causal thinking chain reasoning, and relates to the technical field of power anomaly detection. Comprising the following steps: acquiring an embedded representation of multi-modal data in power equipment, inputting the embedded representation into a general model, and generating a framework of a causal chain; the power exclusive model is based on a framework of a causal chain, each node of the causal chain is refined layer by layer, detail information is supplemented, key factors are filled, and a causal graph is obtained; inputting the causal graph into a hypergraph neural network, mining high-order association in data through multi-layer convolution operation, and updating embedded representation of nodes in the causal graph to obtain an updated causal graph; and on the basis of the embedded representation of the nodes in the updated causal graph, calculating the potential risk influence of the nodes, and determining abnormal nodes. According to the method, a general model and an electric power exclusive model are combined, efficient and accurate abnormal causal chain reasoning is realized by integrating multi-modal data, and the accuracy and effectiveness of electric power anomaly detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power anomaly detection, and in particular to a multi-modal collaborative causal thinking chain reasoning power anomaly detection method and system. Background Art

[0002] In the daily operation of power systems, equipment abnormal events may be caused by a variety of complex factors, and these anomalies often have chain reactions and a wide range of impacts. Most existing anomaly detection and analysis systems rely on a single data source, making it difficult to fully capture the multi-dimensional characteristics of anomalies and to identify the deep causal relationships between abnormal events. With the introduction of multimodal data (such as visual, audio, and text data), power systems are able to collect multi-faceted information on equipment operation, but how to effectively integrate these multimodal data and conduct in-depth causal analysis remains a problem that needs to be solved.

[0003] In addition, the general large model has strong adaptability in general anomaly detection, but it has shortcomings in the refined analysis of specific fields. If it is applied alone to the power field, it often cannot achieve the purpose of accurate detection. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a multi-modal general-specialist collaborative causal thinking chain reasoning power anomaly detection method and system, which combines the general model with the power-specific model for general-specialist collaborative analysis, and realizes efficient and accurate abnormal causal chain reasoning by integrating multi-modal data, thereby improving the accuracy and effectiveness of power anomaly detection.

[0005] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the present invention provides a multi-modal collaborative causal thinking chain reasoning method for power anomaly detection.

[0006] The multi-modal collaborative causal thinking chain reasoning power anomaly detection method includes the following steps: Acquire multimodal data from power equipment, preprocess it and convert it into a unified embedding representation; The unified embedding representation is input into the general model, and the causal chain generation process is refined into multiple subtasks through DDCoT reasoning technology. Each subtask is advanced layer by layer, and the nodes and edges of each subtask are determined to obtain the final causal chain skeleton. The power-specific model is based on the skeleton of the causal chain generated by the general model. It refines each node of the causal chain layer by layer, supplements detailed information and fills in key factors to obtain a causal diagram. The causal graph is input into the hypergraph neural network, and the high-order associations in the data are mined through multi-layer convolution operations. The embedded representations of the nodes in the causal graph are updated to obtain the updated causal graph. Based on the embedded representation of the nodes in the updated causal graph, the nodes with abnormalities are determined.

[0007] The second aspect of the present invention provides a multi-modal collaborative causal thinking chain reasoning power anomaly detection system.

[0008] The multi-modal collaborative causal thinking chain reasoning power anomaly detection system includes: The data acquisition module is configured to: acquire multimodal data from the power equipment, perform preprocessing and convert the data into a unified embedding representation; The general model data processing module is configured to: input the unified embedding representation into the general model, refine the causal chain generation process into multiple subtasks through the DDCoT reasoning technology, advance each subtask layer by layer, determine the nodes and edges of each subtask, and obtain the final causal chain skeleton; The power-specific model data processing module is configured as follows: the power-specific model is based on the skeleton of the causal chain generated by the general model, and each node of the causal chain is refined layer by layer, and detailed information is supplemented and key factors are filled in to obtain a causal diagram; The hypergraph neural network data processing module is configured to: input the causal graph into the hypergraph neural network, mine high-order associations in the data through multi-layer convolution operations, update the embedded representations of the nodes in the causal graph, and obtain an updated causal graph; The detection module is configured to: determine the nodes with abnormalities based on the embedded representations of the nodes in the updated causal graph. The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the multi-modal collaborative causal thinking chain reasoning power anomaly detection method as described in the first aspect of the present invention.

[0009] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the multimodal collaborative causal thinking chain reasoning power anomaly detection method as described in the first aspect of the present invention are implemented.

[0010] One or more of the above technical solutions have the following beneficial effects: The present invention provides a multi-modal general-purpose collaborative causal thinking chain reasoning power anomaly detection method and system. By combining hypergraph neural network and causal reasoning technology, a causal chain analysis framework with dynamic update and risk warning is constructed. First, the system collects multi-modal data of power equipment, and after denoising and feature extraction, it is input into the model with a unified embedding representation; the general model is responsible for basic anomaly detection, forming the skeleton structure of the causal chain, and gradually filling in the key nodes and detailed information in the causal chain through the refined analysis of the exclusive model. Then, the system uses the hypergraph neural network to extract high-order causal relationships in multi-layer convolution and hierarchically display the causal chain. Finally, through the dynamic update mechanism, the system can adjust the causal chain structure in time and generate a real-time risk warning report based on the risk nodes in the chain. The method of the present invention enables the power system to achieve comprehensive monitoring, accurate analysis and effective warning of abnormal events when facing complex anomalies, greatly improving the management efficiency and safety of the power system.

[0011] The present invention proposes a general-specialized collaborative multimodal causal thinking chain reasoning method, which combines the general semantic understanding of the general model and the refined analysis of the power-specific model, and uses multimodal thinking chain technology to organically coordinate the reasoning processes of the two, gradually identifying the root causes, possible impacts and risks of abnormalities, thereby improving the recognition accuracy and response efficiency of complex abnormalities.

[0012] The present invention proposes a multimodal anomaly hierarchical causal chain analysis method based on hypergraph neural network and causal reasoning. The system captures complex high-order associations in multimodal data through hypergraph neural network, and automatically constructs causal chains in combination with causal reasoning technology, and performs layer-by-layer analysis and causal verification of key factors in the chain. This method can present the anomaly detection results in the form of a hierarchical causal chain, intuitively reveal the causal logic behind the anomaly, and provide users with detailed causal relationship analysis and reasoning results, thereby achieving transparency and refinement of anomaly detection and cause analysis.

[0013] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0015] Figure 1 This is a flow chart of the method of embodiment 1.

[0016] Figure 2 This is a schematic diagram of the cause-effect chain skeleton of Example 1. DETAILED DESCRIPTION

[0017] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0018] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0019] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0020] Embodiment 1 This embodiment discloses a multimodal general-special collaborative causal thinking chain reasoning power anomaly detection method, which combines a general model with a power-specific model to achieve comprehensive analysis of multimodal data to identify complex anomalies in the power system. The method first collects and preprocesses multimodal data such as visual, audio, and text, and converts them into embedded representations; then the general model is responsible for generalized anomaly detection, generates the skeleton structure of the causal chain, and the dedicated model performs refined analysis to gradually fill in the key nodes in the chain.

[0021] Afterwards, the hypergraph neural network is used to identify multi-level high-order causal relationships, generate a hierarchical causal chain display, and use a dynamic update mechanism to supplement newly detected anomalies in real time to generate a risk warning report.

[0022] The method of this embodiment can effectively reveal the causal logic behind the anomaly, and provides an accurate and efficient anomaly detection and risk warning solution for the power system.

[0023] like Figure 1 As shown, the multi-modal collaborative causal thinking chain reasoning power anomaly detection method includes the following steps: Acquire multimodal data from power equipment, preprocess it and convert it into a unified embedding representation; The unified embedding representation is input into the general model, and the causal chain generation process is refined into multiple subtasks through DDCoT reasoning technology. Each subtask is advanced layer by layer, and the nodes and edges of each subtask are determined to obtain the final causal chain skeleton. The power-specific model is based on the skeleton of the causal chain generated by the general model. It refines each node of the causal chain layer by layer, supplements detailed information and fills in key factors to obtain a causal diagram. The causal graph is input into the hypergraph neural network, and the high-order associations in the data are mined through multi-layer convolution operations. The embedded representations of the nodes in the causal graph are updated to obtain the updated causal graph. Based on the embedded representation of the nodes in the updated causal graph, the nodes with abnormalities are determined.

[0024] The method of this embodiment will be explained in detail below. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method provided by this embodiment may specifically include: Step 1: In multimodal causal chain analysis, data collection and preprocessing are the basis for ensuring the quality of analysis. The system collects multiple modal data (visual, audio, and text) from power equipment and performs denoising, enhancement, and feature extraction on them. This process standardizes data of different modalities and converts them into a unified embedding representation for subsequent collaborative analysis of general and exclusive models. This process establishes a high-quality data input foundation for the system, making subsequent anomaly detection and causal analysis more accurate and reliable.

[0025] Furthermore, the system first collects visual image data, mainly through cameras installed around the power equipment to monitor the equipment in real time. These visual image data can provide information about the appearance of the equipment and be used to identify physical abnormalities such as surface damage, corrosion, overheating and discoloration.

[0026] During the acquisition process, the system automatically synchronizes the image timestamps to ensure that all image data is time-consistent with data from other modalities, such as audio and text. The frequency of visual image acquisition depends on the task requirements, which can be either periodic image capture or high-frequency acquisition when a specific event is detected. For example, when the equipment produces abnormal vibrations, the system increases the image acquisition frequency to capture detailed information about the abnormal process.

[0027] Specifically, the feature embedding of the visual image data V is processed by Vision Transformer (ViT), and the output feature representation is:

[0028] Among them, v is the visual feature and V is the visual image data.

[0029] Furthermore, audio data is collected in real time by a high-sensitivity microphone installed near the equipment. Audio recordings can reflect the mechanical operation status inside the equipment, such as whether there is abnormal noise caused by wear and looseness of equipment parts.

[0030] Since environmental noise may interfere with the accuracy of audio data, the system performs real-time denoising on the audio. During the acquisition process, a low-pass filter is used to remove high-frequency noise, and a band-pass filter is used to extract signals in a specific frequency range related to the device. These preprocessing steps ensure that the audio data accurately reflects the operating status of the device and keeps pace with the visual image data, providing support for multimodal anomaly detection.

[0031] Specifically, the audio data A is processed by Audio Transformer, which takes the spectrogram or waveform of the audio data as input, obtains the timing characteristics and frequency changes of the audio through the self-attention mechanism, and generates a high-dimensional audio embedding:

[0032] Where a is the audio feature and A is the audio data.

[0033] Furthermore, text data includes log information of the equipment, mainly by extracting operation logs, fault reports and maintenance records from the equipment monitoring system. These text logs record the operation process of the equipment, the feedback of the operators and the abnormal events of the equipment, and are an important data source for analyzing historical abnormalities and equipment status in the system.

[0034] The system uses advanced large language models such as GPT-4 to generate embeddings and extract deep features for text data. These large models are pre-trained on a rich corpus and can capture the deep semantics and contextual information of the text. The text embedding generation process is as follows:

[0035] Where t is the text feature and T is the text data.

[0036] Furthermore, for longer texts (such as device logs), the system generates global semantic embeddings by segmenting and aggregating them. In addition, the multi-layer embedding structure of the large language model can provide the system with hierarchical semantic information. The low-level embedding is used to extract basic semantic features (such as device status), while the high-level embedding captures more complex causal logic and abnormal associations.

[0037] Furthermore, the extracted visual features, audio features, and text features are standardized and normalized and fused into a unified embedding representation.

[0038] All data is standardized into an embedded vector format that can be directly used by general models to achieve multi-modal collaborative analysis. Through these embedded representations, the system can integrate information from different modalities, identify anomalies, and conduct in-depth causal chain analysis.

[0039] After normalizing the data of each modality, they are fused into a unified multimodal embedding matrix E to ensure that data of different modalities can be collaboratively analyzed in the same space:

[0040] The above step 1 implements data preprocessing and multimodal embedding generation. Overall: Visual, audio, and text data are collected and denoised to ensure high-quality input of each modality. Visual data is extracted through Vision Transformer (ViT) to obtain visual features, audio data is extracted through Audio Transformer to obtain audio features, and text data is generated through text embedding using a large language model (such as GPT-4).

[0041] The above three features are standardized and normalized and fused into a unified multimodal embedding matrix E to ensure that data from different modalities are analyzed collaboratively in the same space.

[0042] Step 2: The generation of a multimodal causal chain of thought that is collaborative and specialized is the core of the system's identification and analysis of anomalies. This process combines the general model and the power-specific model, and through the Duty-Distinct Chain-of-Thought (DDCoT) reasoning technology, the processing of multimodal data is completed step by step to form a sophisticated causal chain. The general model first performs a general analysis of the multimodal data to generate the skeleton of the causal chain; then, the power-specific model conducts an in-depth analysis of specific nodes, supplements detailed information, and gradually fills in key factors. This collaborative reasoning can comprehensively and efficiently identify and resolve complex anomalies, and improve the accuracy and reliability of causal analysis.

[0043] In this embodiment, in general: General models refer to machine learning models that can be widely used in various anomaly detection scenarios. They are usually models that have been pre-trained on large-scale data, such as GPT-4, ViT (Visual Transformer) and other large models. These models have the ability to perform preliminary causal analysis on multimodal data (such as vision, audio, and text). The function of the general model is to provide general anomaly detection and generate the basic skeleton of the causal chain, and then the power-specific model will perform more detailed power industry-specific analysis.

[0044] The power-specific model is a customized model that combines the power industry knowledge base and the experience of domain experts, and is specifically used for anomaly detection and analysis of power systems. By learning the historical data of power equipment and industry-specific characteristics (such as equipment load, failure mode, etc.), the model further refines the causal chain generated by the general model, gradually fills in the key node information in the causal chain, and identifies the deep-seated causes of power anomalies.

[0045] Furthermore, the definitions, functions, and technical characteristics of the general model and the power-specific model are explained respectively, and examples are given as follows: General Model: Definition: The general model is a multimodal data processing model for generalized anomaly detection and analysis, suitable for preliminary anomaly identification in various scenarios. It learns general patterns through pre-training on large-scale data and can identify signs of generalized anomalies.

[0046] Function: The general model is responsible for performing preliminary analysis of multimodal data (such as vision, audio, and text) to generate the basic skeleton of the causal chain, providing a rough framework for subsequent in-depth analysis.

[0047] Technical features: Universal models are usually based on large-scale pre-trained deep learning models, such as Visual Transformer (ViT) for visual feature extraction, large language models (such as GPT-4) for text analysis, and AudioTransformer for audio feature analysis. The design of universal models enables them to have good generalization capabilities, able to handle multiple data types and perform generalized reasoning.

[0048] For example, in a power system, when a general model analyzes data from a substation, it can preliminarily identify possible damage in images, abnormal noise in audio, and key warning statements in logs, but it does not have the detailed analysis capabilities unique to power.

[0049] Power-Specific Model: Definition: The power-specific model is a professional model designed for the power system. It combines the knowledge base of the power field and historical data unique to the field, and focuses on in-depth analysis of specific anomalies in the power system.

[0050] Function: The power-specific model conducts a detailed analysis of the basic skeleton of the causal chain generated by the general model, gradually fills in the detailed information and associated features of each key node, and identifies and analyzes complex abnormal events in the power system.

[0051] Technical features: The power-specific model uses databases unique to the power industry, such as historical fault records, equipment operating characteristics, operating procedures, etc., and conducts in-depth correlation analysis of abnormal causes through feature extraction methods related to the power system (such as hypergraph neural networks or adaptive attention mechanisms), and gradually builds a complete structure of the cause-effect chain.

[0052] For example, after detecting that a transformer is overloaded, the power-specific model will conduct an in-depth analysis of the specific cause of the anomaly based on historical load data, equipment type, and environmental factors, such as identifying whether the performance is degraded due to excessive ambient temperature or due to aging of internal components. In addition, it will analyze whether the anomaly will affect other downstream equipment.

[0053] As the first-level analysis tool, the general model undertakes the basic detection task of multimodal anomalies. It is responsible for processing the collected multimodal data. In this embodiment, after the unified embedding representation is input into the general model, the DDCoT method is used to refine the generation process of the causal chain into a series of subtasks. For example, when a device anomaly is detected, the general model will decompose the problem into several steps, including detecting the initial anomaly point, identifying potential direct causes, and preliminarily evaluating the impact on surrounding devices.

[0054] After that, the DDCoT method simulates the step-by-step reasoning of humans and advances each subtask layer by layer. At this stage, the general model not only generates the basic skeleton of the causal chain, but also forms a preliminary judgment on the cause of the anomaly through its extensive semantic understanding and zero-sample reasoning capabilities, providing a basic framework for the causal chain.

[0055] Furthermore, in the process of breaking down the causal chain generation process into multiple subtasks through DDCoT reasoning technology, a multi-level recursive subtask generation method is adopted to obtain the final causal chain skeleton. The specific process includes: According to the type of multimodal data, the anomaly detection of each modality data is processed independently; The anomaly detection of each modal data is divided into a current level task, and the current level task is further divided into a plurality of subtasks according to a recursive formula, wherein the plurality of subtasks include detecting an initial anomaly point, identifying a potential direct cause, and evaluating the impact on surrounding devices; Recursively analyze multiple subtasks in the current level task to determine the causal chain skeleton of the current level task; Based on the recursive analysis results of the current-level task, the identification of potential direct causes and the evaluation of the impact on peripheral devices in multiple subtasks of the current-level task are divided into the next-level task, and the next-level task is further divided into multiple subtasks according to the recursive method; Recursively analyze multiple subtasks in the next-level task to determine the causal chain skeleton of the next-level task; Iterate the above multi-level recursive subtask generation and analysis process to obtain the final causal chain skeleton.

[0056] It should be noted that the nodes in the causal graph represent events or anomalies (such as equipment failures, abnormal temperatures in the power system, etc.), and these nodes are connected to each other through causal relationships. In the causal chain, the connection between nodes represents the causal path of the event, marking the transmission from a certain anomaly (node) to the potential cause or subsequent impact.

[0057] The edges in the causal graph are determined based on certain logical relationships (such as subtasks at each level), indicating the causal relationship between nodes. The edges in each causal graph connect different nodes, reflecting the transmission path from one abnormal event to another (or from an abnormality to its direct cause, impact, etc.). These edges are usually inferred layer by layer in the causal reasoning process through the results of recursive task generation and subtask analysis.

[0058] More specifically, the general model generates the "skeleton of the causal chain" based on the "unified embedding representation". The specific data processing process includes: Generation of preliminary causal chains: Using the Duty-Distinct Chain-of-Thought (DDCoT) reasoning technique: The unified embedding matrix E is input into the general model, and the model applies the DDCoT reasoning technology to decompose the causal chain generation process into multiple subtasks. The characteristic of DDCoT technology is that it advances the reasoning process step by step, simulates the human way of thinking, and thus generates a preliminary causal chain skeleton.

[0059] Subtask definition: In the DDCoT reasoning framework, we first detect the initial anomaly, then identify the potential direct cause, and then evaluate the impact on the surrounding devices or systems. Each subtask completes a specific reasoning step independently, gradually filling in the nodes of the causal chain skeleton.

[0060] DDCoT technology embeds and matches visual, audio, and text data into corresponding subtasks based on the semantic information of different modal features. For example, visual data may correspond to the detection of physical anomalies (such as surface damage and discoloration), while audio data may correspond to the analysis of changes in operating noise, and text data is used to supplement device log information.

[0061] Node construction and linking of causal relationships: After completing the preliminary causal chain, the system gradually establishes the connection relationship of the chain by analyzing the modal feature similarity and time association between each node. For example, by associating visual features with the timestamp of text logs, the association between abnormal devices and related events in the causal chain can be identified.

[0062] The output results of each subtask will be combined to form the basic skeleton of a causal graph, in which feature nodes of different modalities represent direct or indirect causal relationships through edges.

[0063] Iterative optimization of skeleton structure: After the initial causal chain skeleton is generated, the general model continues to optimize the hierarchical structure of the causal chain by updating the embedding vectors of each node. This process is based on the calculation of the similarity and correlation between the modal features in the chain. Through multiple iterations, the unclear causal relationship is gradually clarified and supplemented.

[0064] Output the causal chain skeleton: After completing the above steps, the general model generates a preliminary causal chain skeleton containing multimodal data, which is used for further detailed analysis by the power-specific model.

[0065] Furthermore, recursive analysis is performed on multiple subtasks in the current level task to determine the causal chain skeleton of the current level task, including: Detect initial outliers: The embedded representation of each modal data in the current level is used to perform generalized anomaly detection through a general model, identify the anomalies, and use the identified anomalies as the first-class nodes in the current level task, with the corresponding embedded representation; Identify potential direct causes: The first-class node in the current hierarchical task is taken as the analysis object, combined with the corresponding embedding representation, the data of the power equipment related to the first-class node is detected, the potential direct cause of the first-class node is identified and verified, and after the verification is completed, the potential direct cause is taken as the second-class node, connected to the first node through an edge, and the embedding representation of the first-class node and the second-class node is updated; Evaluate the impact on surrounding devices: Take the first-class nodes and second-class nodes in the current hierarchical task as analysis objects, identify the impact of the initial abnormal point and its potential direct cause on other devices or systems, determine the conduction path, and take the identified other affected devices or systems as the third-class nodes. Connect the third-class nodes with the second-class nodes and / or the first-class nodes based on the conduction path, and update the embedded representations of the first-class nodes, the second-class nodes, and the third-class nodes; Based on the above-mentioned first-type nodes, second-type nodes, third-type nodes and the edges between them, the causal chain skeleton of the current level task is determined.

[0066] The embedded representation of the second type of node is obtained in the process of identifying the potential direct cause of the first type of node; the embedded representation of the third type of node is obtained in the process of determining other devices or systems affected by the first type of node and the second type of node. It can be understood that in order to facilitate the mapping with the nodes in the subsequent hypergraph, the embedded representations of the first type of node, the second type of node, and the third type of node in this embodiment all contain content of multiple corresponding modes.

[0067] Further: In the above technical solution, each type of data anomaly detection is processed independently. For example, visual anomaly detection, audio anomaly detection, and text anomaly detection are all processed independently, and anomalies are identified in the corresponding modality respectively.

[0068] The purpose of initial anomaly detection is to determine the starting point of the causal chain, that is, the initial abnormal event that may exist in the power system. In the initial subtask, the embedding representation of each modality is generalized through the general model to detect anomalies and identify the anomalies. The identification results of the anomalies will generate the initial causal chain nodes. Once the anomaly is identified, it will be used as the first node of the causal chain with the corresponding embedding representation, which contains preliminary abnormal characteristics.

[0069] In specific implementation, the system first inputs multimodal data (such as vision, audio, text, etc.) into a general model to identify abnormal features in the data. For example, the damage features of substation equipment can be identified in visual data, abnormal noise can be detected in audio data, and alarm information can be found in text logs.

[0070] The purpose of potential direct cause identification is to determine the direct cause of the initial abnormal point and provide a preliminary causal relationship for the causal chain. After detecting the abnormal point, DDCoT will conduct further cause analysis on the node based on the multimodal features of the abnormal point. For example, the system will check the load records, environmental data, etc. of the power equipment related to the abnormal point. The system gradually verifies the potential direct cause based on the correlation of multimodal data (such as abnormal temperature, excessive load, etc.). DDCoT technology adds features layer by layer through each subtask and updates the node representation. After confirming the direct cause, the direct cause is used as a new node and connected to the initial abnormal point through an edge to form the basic edge in the causal chain.

[0071] In specific implementation, after the initial abnormal point is determined, the general model will further analyze the factors directly related to the abnormal point. For example, equipment damage in visual data may be related to overload, and abnormal noise in audio data may be related to mechanical failure. DDCoT divides these potential direct causes into multiple subtasks at the next level, obtains data associated with the potential direct causes, and continues to the next level of analysis.

[0072] After the direct cause is identified, DDCoT will generate subtasks for impact assessment layer by layer, further analyze the correlation of space, time and function, generate new nodes and edges, gradually add the potential impact and conduction path of peripheral devices to the causal chain, and judge the impact of the initial abnormality and direct cause on other devices or systems. For example, the detection of abnormal temperature of a device may cause overload or other hidden problems of peripheral devices. The purpose of the preliminary assessment of the impact on peripheral devices is to evaluate the potential impact of the initial abnormal point and its direct cause on the peripheral devices in the system and determine the conduction path of the causal chain.

[0073] For each new node, the system updates the node embedding representation to preserve its feature changes during the conduction process.

[0074] In specific implementation, based on the initially identified anomalies and direct causes, DDCoT will gradually analyze the devices or nodes that may be affected by the anomaly by gradually analyzing the operating data and anomaly records of surrounding devices and evaluating the transmission range of the initial anomaly.

[0075] Through the above layer-by-layer subtask advancement and node addition, the system can form a preliminary judgment on the cause of the anomaly. For example, through multi-layer causal reasoning, DDCoT will identify the possible root cause and use it as a preliminary judgment of the abnormal event. After the preliminary judgment, the system uses each abnormal cause and its related events as nodes, and the direct cause and conduction path as edges to generate a causal chain skeleton. This skeleton reflects the complete causal path from the initial anomaly to the direct cause and then to the conduction impact.

[0076] In this causal chain structure, each node contains its associated abnormal feature information, indicating the cause and impact of the abnormality. For example, when analyzing the failure of substation equipment, DDCoT technology first identifies the appearance damage of the equipment in the image as the initial abnormal point (subtask 1). Next, the load data and equipment history records are analyzed to identify overload as the direct cause (subtask 2). Then, the conduction risk of overload to adjacent equipment is evaluated, such as whether it will cause temperature rise or load fluctuation of other equipment (subtask 3). Through the step-by-step processing of these subtasks, the skeleton structure of the causal chain is generated, providing a basis for subsequent detailed analysis. The general model uses the generalized semantic understanding ability to generate the basic skeleton of the causal chain and preliminarily judge the cause of the abnormality; then, the subsequent power-specific model analyzes and refines the specific factors of the key nodes layer by layer.

[0077] Specifically, consider the spatial, temporal and functional correlation between the initial abnormal point and its potential direct cause and other devices or systems, conduct spatial correlation analysis, functional correlation analysis and temporal correlation analysis between the initial abnormal point and its potential direct cause and other devices or systems, identify the impact of the initial abnormal point and its potential direct cause on other devices or systems, and determine the transmission range and path: (1) Spatial relationship: If the anomaly identified in the visual modality data affects neighboring devices, a spatial correlation analysis task with these devices is generated.

[0078] (2) Functional relevance: For example, if a mechanical failure is causing an audio anomaly, analyze whether it may have a chain effect on other functionally connected devices.

[0079] It can be understood that this embodiment adopts a multi-level recursive subtask generation method to generate a multi-layer structure of the causal chain through recursive iteration. In each layer of the structure, the connection relationship between the nodes and the edges between the nodes is determined, so as to integrate all the layers and finally construct the causal chain skeleton. In specific implementation, after the preliminary evaluation is completed, the DDCoT technology will generate recursive subtasks based on the analysis results to extend the causal chain to a deeper level. The system will check the results of each subtask and continue to generate the next level of tasks, filling in the nodes of the causal chain layer by layer until the causal chain skeleton structure is formed.

[0080] At each level of recursion, it is checked whether the conduction of the initial anomaly is likely to trigger a subsequent device failure or alarm event, generating further causal nodes.

[0081] It can be understood that by iterating the above multi-level recursive subtask generation process and analysis process, the final causal chain skeleton is obtained, which is specifically: Based on the causal chain skeleton of multiple levels of tasks after iteration, the final causal chain skeleton is obtained. The final causal chain skeleton reflects the complete causal path from the initial anomaly to the direct cause and then to the transmitted impact.

[0082] After all subtasks are executed, this embodiment combines the results of all subtasks to generate the basic structure of the cause-effect chain, providing a basis for in-depth analysis of the power-specific model. The output includes the cause-effect relationship of each node, the potential direct cause and its transmission path to the surrounding equipment.

[0083] like Figure 2 As shown, the following example is given: Initial node detection: Assuming that a device overheating is detected in the image (initial abnormal point), the system generates node A.

[0084] Direct cause subtask: Check the equipment load record, confirm that overload is the direct cause, generate node B, and add edge A→B.

[0085] Impact assessment subtask: Detect that the overload affects the operation of surrounding equipment, generate nodes C and D, and add edges B→C and B→D.

[0086] Causal chain skeleton generation: After forming a preliminary cause judgment, nodes A, B, C, and D constitute the skeleton of the causal chain. The node embedding representation is updated in each layer of subtasks, and the edges represent the causal relationship between nodes.

[0087] Above Figure 2 The nodes represent abnormal events, direct causes, and potential influencing factors in multimodal data; the edges represent causal relationships, and the direction of the edges shows how the abnormality is transmitted and how the direct cause leads to subsequent impacts; the hierarchical structure of the causal chain is divided into multiple levels of initial abnormality, direct cause, and transmission impact. The initial abnormality node (A) represents the detected initial abnormal event (for example, abnormal device temperature) as the first node of the causal chain; the direct cause node (B) indicates that in the direct cause identification stage of DDCoT reasoning, the direct cause of the temperature abnormality is confirmed to be the device overload, and this node is connected to the initial abnormality node through a causal edge; the transmission impact nodes (C, D, E) indicate that through further analysis of multimodal data, the potential impact of abnormal events on surrounding devices (such as temperature transmission to other devices) or possible related factors (such as excessively high ambient temperature increases the load on the cooling system) are confirmed.

[0088] Furthermore, the power-specific model is based on the skeleton of the causal chain generated by the general model, and refines each node of the causal chain layer by layer, supplements detailed information and fills in key factors to obtain a causal diagram.

[0089] After the general model completes the basic analysis, the power-specific model is connected to the analysis process to conduct a refined analysis of the causal chain skeleton for the power industry. The dedicated model uses its rich knowledge in the power field to deeply identify details that the general model may not fully capture. For example, after the general model identifies that the equipment is overloaded, the dedicated model will further analyze the historical load of the equipment, the specific equipment type and its location in the power system. This step allows the dedicated model to refine each node of the causal chain layer by layer, combining the equipment status with industry knowledge to identify deep potential causes and risks. Through this layer-by-layer analysis, the causal chain gradually develops from the initial framework to a detailed causal network.

[0090] In this embodiment, causal reasoning constructs a causal graph, which represents the causal relationship between different multimodal feature nodes through edges. Assume that the causal graph G consists of nodes X and edges E, and the edges represent the causal relationship. Based on Pearl's causal reasoning framework, the causal effect of the system can be represented by estimating the causal influence between different nodes:

[0091] Among them, X, Y, Z: These symbols usually represent different events or variables. X represents an externally intervened event or variable, usually an event that we control or influence. Y represents an outcome event, that is, the target event we are concerned about, which is usually affected by X. Z represents an intermediate variable or covariate, that is, other factors that may affect the relationship between X and Y in the causal chain. P(Y|do(X)): represents the probability of event Y occurring in the case of external intervention X. Here, do(X) represents an operation of "human intervention" or "influence". It is used to eliminate confounding factors to evaluate the causal effect of Y caused purely by X. P(Z=z): represents the probability that event Z takes the value z. Z can be a random variable, and its probability distribution P(Z=z) represents the probability of occurrence of each possible value. ∑zP(Y|X,Z=z): represents the sum of the conditional probabilities of all possible values ​​of Z.

[0092] This part calculates the sum of the probabilities of Y occurring given X at different values ​​of Z. This gives us an expected value that can be used to analyze the overall impact of X on Y under various possible scenarios.

[0093] Furthermore, based on the skeleton of the causal chain generated by the general model, the power-specific model refines each node of the causal chain layer by layer, supplements the detailed information and fills in the key factors to obtain a causal diagram, which includes: Additional details: For each node in the final causal chain skeleton, the power-specific model associates the embedded representation of each node with the knowledge base of the power industry, and combines the background data of equipment operation to generate an embedded representation that is more in line with the characteristics of the power field; the power knowledge base contains power equipment operation data and failure modes, and the background data of equipment operation includes the historical failure records of the equipment, equipment type and location; Filling key factors: For each node in the final causal chain skeleton, the power-specific model further introduces multi-dimensional data of the node. Through the correlation analysis of multi-dimensional data, the key influencing factors of the node are determined and filled in, and the deep potential causes of each node are identified; Generate a cause-effect diagram: After the final causal chain skeleton of the general model is supplemented with detailed information and filled with key factors, the final causal chain skeleton is expanded into a causal graph.

[0094] Furthermore, the above technical solution specifically includes: Enhanced and embedded optimization of power industry features: First, the power-specific model starts with the causal chain skeleton generated by the general model and refines each node in the chain. The system associates the embedded representation of each node with the specific knowledge base of the power industry, and combines the background data of the equipment operation (such as the historical fault records of the equipment, the type and location of the equipment, etc.) to generate an embedded representation that is more in line with the characteristics of the power field.

[0095] The power knowledge base contains a wealth of power equipment operation data and failure modes to support the model to identify common types of problems with specific equipment. For example, the load history data of the transformer is used to determine the cause of overload, while the temperature curve of the cable helps detect insulation aging.

[0096] Dynamic filling of key factors: For the key nodes of the causal chain, the exclusive model automatically screens the key factors closely related to the node through the "dynamic feature selection mechanism". For example, after detecting the abnormal temperature of a certain device, the system will call the historical temperature data and environmental data of the device to supplement the temperature characteristics of the node to reveal the source and impact of the abnormal temperature.

[0097] The dynamic feature selection mechanism uses an adaptive attention mechanism to adaptively assign weights to each feature based on the current state of the chain and the newly detected data. The adaptive attention mechanism can adjust the focus in real time, focusing on the features most relevant to the current anomaly, ensuring that the detailed information of each node is accurate and closely related to the actual situation. By dynamically adjusting the attention weights, the model can prioritize important features in multimodal data to achieve accurate supplementation and in-depth analysis of abnormal nodes.

[0098] Correlation analysis of node multi-dimensional data: For each node in the causal chain, the dedicated model further introduces multi-dimensional data of the node, including environmental factors (such as temperature and humidity) and equipment status data (such as current, voltage, and load). Through multimodal association analysis, the model can identify the deep potential causes of each node. For example, in a high humidity environment, the possibility of insulation failure increases, and the model will refine the "humidity" factor as a key influencing factor of the node.

[0099] To achieve the above analysis, the dedicated model uses a context-aware module to extract features from the associated data of each node layer by layer, ensuring that key detailed information of the power industry is retained during the refinement process.

[0100] Hierarchical refinement of causal chain nodes: The dedicated model analyzes the potential impact of different nodes on the overall system layer by layer according to their position in the causal chain. For example, for the initial abnormal node, the model will focus on analyzing its direct cause, while for the node close to the end of the causal chain, the model will evaluate its conductive impact on the system to ensure the rationality of the overall causal relationship of the chain.

[0101] In terms of specific implementation, the model uses a multi-layer hypergraph neural network (HNN) structure to recursively refine node information. In each layer of hypergraph convolution, the model aggregates multimodal information related to the current node through hyperedges to generate an updated node embedding representation. As the layers progress, the hypergraph neural network extracts detailed information layer by layer (such as peak points or key environmental conditions in historical load data) and gradually adds it to the node feature representation, allowing the node to integrate high-order information from other related nodes and achieve multi-level causal chain refinement.

[0102] Constructing a multimodal causal graph: After each node is refined and key factors are supplemented, the dedicated model will expand the causal chain skeleton into a multimodal causal graph. By connecting the causal relationships between different nodes through the Graph Neural Network (GNN), the model can comprehensively consider the multimodal features of each node (visual, audio, text, etc.) to achieve efficient causal relationship identification and causal graph update.

[0103] The causal graph will be adjusted in real time according to the operating status of the equipment to capture the dynamic correlation between devices, thereby further improving the causal chain structure of power anomaly detection.

[0104] Furthermore, when processing causal chain nodes, the power-specific model needs to obtain data on various power equipment related to the current node to ensure the accuracy of abnormal cause identification: First, it calls up the device's operating data, including real-time and historical current, voltage, power, frequency, phase, temperature and other information, to analyze whether the device is overloaded, overheated or has abnormal current; Environmental data is also a key component, such as ambient temperature, humidity, wind speed and rainfall, especially high temperature and high humidity, which may affect the insulation or heat dissipation of the equipment; In addition, the model also calls on the equipment's condition monitoring data, such as vibration frequency, noise level, oil level, and oil temperature, which can help detect whether the equipment has internal wear, abnormal vibration, or poor heat dissipation; At the same time, maintenance and repair records are also an important source of information for power-specific model analysis. Historical failures, component replacements, maintenance cycles, and inspection data can provide clues to whether equipment is at risk of aging. To better understand the conduction path of the anomaly, the model also analyzes the topology of the power grid and the connection relationship between adjacent devices to identify whether the anomaly may have a chain reaction on other devices; Finally, historical anomalies and external factors (such as natural disasters or construction disturbances) may also provide further context to the model, allowing for a more comprehensive identification of potential causes and possible impacts of current anomalies.

[0105] By combining these multimodal data, the power-specific model can comprehensively and accurately analyze the abnormal conditions of each node, providing support for the refinement of the causal chain and the construction of the abnormal transmission path.

[0106] Furthermore, the power-specific model identifies the specific process of "deep underlying causes", including: When analyzing the power equipment data associated with the current node, the power-specific model will gradually deepen its understanding by combining industry knowledge to systematically identify the underlying potential causes of equipment anomalies. The following is the detailed data processing flow of the power-specific model: First, the power-specific model extracts real-time current, voltage, temperature and other equipment operation data from the current node, and compares it with the normal operating status to preliminarily determine whether the equipment exceeds the normal operating range. The preliminary analysis of these data provides a starting point for anomaly detection. Next, the model will load information about equipment load, environmental impact, etc. from the industry knowledge base, and through query and correlation analysis, determine whether the current operating status meets the safety limit or optimal operating range of the equipment. If the current or voltage value is found to be high, the model will further trace the historical data related to the load to determine whether overload operation is the cause of the equipment abnormality.

[0107] After the initial analysis of the operating data is completed, the power-specific model will conduct a deeper analysis in combination with environmental data (such as temperature and humidity). The model correlates the internal and external temperatures of the equipment and refers to the environmental impact standards in industry knowledge to determine whether the anomaly is affected by environmental factors. For example, after detecting that the equipment temperature is too high, the model will call historical ambient temperature data and combine it with the equipment behavior model under similar environmental conditions to analyze whether it is caused by ambient temperature. If the model determines that the environment is a potential factor, it will be used as a supplementary node in the causal chain.

[0108] Furthermore, the model will call status monitoring data and maintenance records. By analyzing the equipment's vibration frequency, noise level, oil level, oil temperature and other data, the model can identify whether there are hidden problems such as component wear and aging. For example, when the model detects that the vibration frequency is abnormal and the noise is too high, it will combine the fault feature library in the industry knowledge to match the current equipment features with the known equipment failure mode to determine whether it is potential component aging or mechanical loss. If the match is successful, the fault feature is marked as the underlying cause of the current node.

[0109] Finally, the model will refer to historical similar abnormal events and the equipment topology to verify the potential cause of the abnormality. In this process, the model uses the topological relationship to determine whether the abnormality of the current node may be related to the operating status of the upstream device or adjacent equipment. For example, if there is a high voltage output in the upstream device, the power-specific model will analyze its conduction effect on the downstream device. Combined with the chain reaction data of the industry knowledge base, the model can identify the conduction link of the abnormality in the system and ultimately determine the underlying potential cause.

[0110] Through the above-mentioned in-depth analysis, the power-specific model can not only identify the direct cause of the current node abnormality, but also combine the gradual association of industry knowledge and data characteristics to extract the deep-seated potential causes, providing a reliable basis for the further refinement of the causal chain.

[0111] Furthermore, after the power-specific model is gradually refined and filled, the system performs integrity checks on each node in the causal chain.

[0112] The causal relationship of each node needs to be verified in collaboration with the general model and the dedicated model to ensure the logic and accuracy of the causal chain. This step not only verifies the rationality of multimodal analysis, but also strengthens the hierarchical structure of the causal chain. When the causal chain is finally formed, it includes a detailed analysis from direct causes to deeper risks. The system can generate a panoramic view of abnormal events based on this chain, providing an accurate basis for the management and decision-making of the power system.

[0113] Step 3: The causal graph is input into the hypergraph neural network, and the high-order associations in the data are mined through multi-layer convolution operations. The embedded representation of the nodes in the causal graph is updated to obtain the updated causal graph, which includes: The causal graph is input into the hypergraph neural network, the nodes of the causal graph are extracted as the nodes of the hypergraph, and the weights are calculated according to the similarity and correlation between the modal features, so as to determine the hyperedges between the hypergraph nodes and complete the construction of the hypergraph; Through multi-layer convolution operations, multi-level causal reasoning is performed on the constructed hypergraph structure to extract high-order features in the data layer by layer: in each layer of convolution operations, the embedded representation of the hypergraph node is updated, and the information of the surrounding hypergraph nodes is gradually aggregated, so that each hypergraph node obtains relevant other modal features; Based on the embedded representations of the updated hypergraph nodes, the embedded representations of the nodes in the causal graph are updated to obtain an updated causal graph.

[0114] The Hypergraph Neural Network (HNN) is used to model high-order associations of multimodal data, thereby realizing the identification and hierarchical display of complex causal relationships. The introduction of the hypergraph neural network enables the system to aggregate data features of different modalities, and through multi-layer convolution operations, deeply mine high-order associations in the data to generate a clear causal chain. Finally, the system displays the causal chain in a hierarchical form, analyzing it layer by layer from direct causes to potential impacts, providing users with a comprehensive causal analysis perspective.

[0115] The causal graph is input into the hypergraph neural network, and the nodes of the causal graph are extracted as the nodes of the hypergraph. The causal graph contains first-category nodes, second-category nodes and third-category nodes, and the embedded representation of each type of node contains multimodal features.

[0116] Furthermore, in the hypergraph construction phase, the system converts multimodal data into a hypergraph structure in order to capture high-order relationships in the data. The nodes in the hypergraph represent data features of different modalities, such as the appearance features of visual images, the acoustic features of audio recordings, and the semantic features of text logs. In order to better represent the interactions and associations between multimodal data, the system uses hyperedges to connect multiple nodes, which can capture the complex relationships between multiple modal features. For each hyperedge in the hypergraph, the system calculates the weight based on the similarity and correlation between modal features to ensure that the connections between different modal data are fully displayed. Through this high-order structured modeling, the hypergraph neural network can effectively aggregate contextual information in multimodal data and build the basic structure of the causal chain.

[0117] Furthermore, weights are calculated based on the similarity and correlation between modal features, and the hyperedges determined based on the weights include: When the causal graph is input into the hypergraph neural network, the key step in constructing the hypergraph is to calculate the weights based on the similarity and correlation between modal features to determine the hyperedges between nodes.

[0118] It can be understood that the nodes in the hypergraph correspond to the multimodal features of the same power facility. When constructing hyperedges, the similarity and correlation of these multimodal data need to be considered.

[0119] The following is the specific calculation method: Modal feature similarity calculation: Feature extraction: First, extract multimodal features of each node, such as visual features, audio features, and text features, and generate embedding vectors for each modality. Visual data may be embedded using Vision Transformer, audio data may be embedded using Audio Transformer, and text data may be embedded using a large language model (such as GPT).

[0120] Similarity measurement: For the similarity between different modalities, cosine similarity or Euclidean distance can be used to measure the similarity between the embedding vectors of nodes. Set a similarity threshold. If the similarity is greater than the threshold, it means that these nodes have a strong correlation.

[0121] Multimodal similarity fusion: After obtaining multimodal similarity metrics such as vision, audio, and text, the similarity scores of these modalities are fused through weighted average or attention mechanism to comprehensively evaluate the similarity between nodes. For example:

[0122] Among them, α, β, and γ are weighting coefficients, which can be adjusted according to the importance of the mode.

[0123] Modal feature correlation calculation: Spatiotemporal correlation: If the nodes in the causal graph have temporal or spatial dependencies (for example, the spatial locations of adjacent devices are close, or the time of operation events is close), the spatiotemporal correlation can be calculated for these nodes. The correlation between nodes is measured by time difference and spatial distance.

[0124] Topological relationship: In the power system, the topological structure and physical connection relationship of the equipment are very important. Use the equipment topology diagram to determine the topological distance of the nodes in the cause-effect diagram, or quantify the association through the physical connection situation. Nodes with higher topological association can be given higher association weights.

[0125] Relevance fusion: Considering the spatiotemporal and topological relationships comprehensively, a relevance measure is generated. For example, the relevance measure is combined with the topological relationship according to the weight:

[0126] Among them, δ, η, and ζ are weight coefficients, which are set according to the application scenario.

[0127] Weight calculation and hyperedge determination: Weight calculation: After obtaining the similarity and association measures, the final hyperedge weight is determined by the combination of the two. For example, the comprehensive weight can be generated by weighted summation or product method:

[0128] Where λ and μ are adjustable parameters used to balance the contribution of similarity and relevance.

[0129] Hyperedge determination: A hyperedge is constructed for each pair of nodes based on the calculated weights. A weight threshold is set, and only pairs of nodes with weights above the threshold are connected as hyperedges. For multiple nodes with high similarity and association, their embedded representations are aggregated through hyperedges to enhance high-order associations.

[0130] Example of constructing a hypergraph: For example, nodes A, B, and C have high modal similarity and spatiotemporal correlation, so a hyperedge connecting A, B, and C is created in the hypergraph to aggregate their multimodal features and complete high-order correlation modeling in the causal chain. This approach allows the hypergraph neural network to further explore the complex correlations in the data in multi-layer convolution operations and dynamically update the node embeddings to generate an updated causal graph.

[0131] The high-order association in this embodiment refers to the node participating in information exchange with multiple associated nodes through hyperedges. It is not just a simple adjacency relationship, but an overall information fusion through the associated features of a group of nodes. This high-order association modeling enables each node to capture more complex causal relationships (for example, direct and indirect abnormal conduction paths, cross-device linkage effects) after multi-layer convolution updates, thereby generating a progressive causal chain structure. Ultimately, the embedded representation of each node can contain its multimodal association features, as well as causal information from direct to indirect, thus forming a clear causal chain.

[0132] The role of hyperedges in the clarity of causal chains The construction of hyperedges directly determines the clarity of higher-order correlations in the causal chain. Specifically: Aggregate related nodes: Aggregate nodes with high similarity and correlation through hyperedges, so that each node can gather key information from multiple related nodes when embedding and updating. The efficient information interaction between nodes on each hyperedge helps the model identify and weight the causal relationship between nodes.

[0133] Information expansion in multi-layer convolution: In multi-layer convolution operations, each layer propagates information along hyperedges, allowing nodes to obtain information about surrounding related nodes layer by layer. The calculation of hyperedge weights (based on similarity and relevance) helps the model effectively distinguish between direct and indirect effects during information expansion, thereby forming a hierarchical display of the causal chain from direct causes to potential effects.

[0134] Causal chain path determination: Hyperedges aggregate information from multiple nodes, allowing the model to automatically infer multi-level causal relationships between nodes when updating each layer. Ultimately, the updated embedding representation of the node will include its causal role in the chain, ensuring that the direct cause, transmission path, and indirect impact can be clearly presented in the causal chain display.

[0135] Furthermore, after the hypergraph is constructed, multi-level causal reasoning is performed on the constructed hypergraph structure. The multi-layer convolution operation in the hypergraph neural network can extract high-order features from the data layer by layer to identify complex causal relationships. In each layer of convolution operation, the hypergraph neural network updates the embedded representation of the node and gradually aggregates the information of the surrounding nodes so that each node can obtain other modal features related to it. This feature aggregation process is equivalent to the system adding details to the causal chain between different modalities. Through multi-level iterative updates, the system gradually captures the deep associations in the causal chain.

[0136] For example, when analyzing a device failure, the system will gradually aggregate the device's visual features (such as surface damage), audio features (such as abnormal sounds), and text logs (such as fault reports), and identify possible root causes and impact paths through multi-layer causal reasoning. This reasoning process enables the system to accurately parse complex multimodal causal chains and retain a large amount of detailed information in the chain.

[0137] In the construction of causal chains of multimodal data, Hypergraph Neural Network (HNN) is used to model high-order associations. The nodes of the hypergraph represent multimodal features, and the hyperedges e connect multiple nodes, representing the high-order relationships between modalities. For a hypergraph G=(V,E), where V is a set of nodes and E is a set of hyperedges, each hyperedge e connects multiple nodes:

[0138] in, They represent node 1, node 2, ... node k respectively.

[0139] HNN extracts high-order features from multimodal data through multi-layer convolution operations. The calculation formula for each layer of convolution is:

[0140] in, represents the node features of the lth layer, is the degree matrix of the hyperedge, is the weight matrix of the lth layer, and σ is the activation function.

[0141] Furthermore, after the causal chain is generated, the system displays the chain in a hierarchical manner, allowing users to view the causal relationship layer by layer through the interface.

[0142] The hierarchical display of the causal chain is presented step by step from direct causes to potential impacts. For example, users can first view the direct causes of equipment failure (such as component damage), and then further explore deeper related factors (such as long-term overload or lack of regular maintenance). The system displays the key causal relationships at each level of the node, and supports users to interactively query each node in the chain. By clicking on a node, users can view other modal features related to the node and further understand the causal logic behind the anomaly.

[0143] It can be understood that in this embodiment, after the hypergraph is constructed, the hypergraph neural network performs multi-level causal reasoning on the hypergraph structure through multi-layer convolution operations to identify more complex causal relationships in the causal graph. In fact, the causal graph is part of the hypergraph. The construction and reasoning process of the hypergraph directly provides support for high-order features and multimodal information for the causal graph. The two are highly related as a whole.

[0144] Reasoning process of association hypergraph and causal graph: Causal graph as the underlying structure of hypergraph: In the hypergraph, the nodes of the causal graph (i.e., abnormal events in the system, device status, etc.) are mapped to nodes of the hypergraph. The embedding representation of these nodes carries preliminary information about the causal relationship.

[0145] The hyperedges in the hypergraph form high-order associations by aggregating multiple nodes, which are used to expand the node information in the causal graph so that the nodes of the causal graph can obtain multimodal features and contextual information.

[0146] Layer-by-layer reasoning of high-order causal relationships: The convolution operation of each layer of the hypergraph neural network will update the embedded representation of the hypergraph nodes, and aggregate the modal features and association information of the surrounding nodes into the causal graph nodes layer by layer.

[0147] In this process, the causal graph nodes gradually update their representations and obtain richer multimodal context features that contain high-order causal relationships between device states and anomalies.

[0148] Causal chain refinement in hypergraph convolution: Each layer of convolution is equivalent to the layer-by-layer refinement of the causal chain. By aggregating the feature information of other nodes, the node relationship in the causal graph gradually forms a chain-like causal path. As the multi-layer convolution operation proceeds, the relationship between the nodes in the causal chain becomes clearer, showing direct causes, indirect effects, and upstream and downstream associations.

[0149] This process of layer-by-layer aggregation and updating is equivalent to adding details to the causal chain, so that the deep connections of the causal chain can be displayed in a hierarchical manner.

[0150] The updated causal graph generated is: After multi-level reasoning and updating, the embedding representation of each causal graph node contains the characteristics of multimodal data and high-order correlation information of the causal chain, forming the final causal graph. In this causal graph, the relationship between nodes is deeply mined, with a clearer causal chain structure, which can be displayed layer by layer from direct causes to potential impacts.

[0151] Through this multi-level causal reasoning, the hypergraph and causal graph form an organic whole: The nodes of the causal graph are modeled with high-order associations in the hypergraph and updated layer by layer, eventually forming a deep causal chain containing multimodal features. The convolution operation of the hypergraph directly supports the refinement of the causal graph, making the hierarchical structure of the causal chain clearly displayed.

[0152] Based on the causal graph updated by the hypergraph neural network, the specific process of detecting anomalies in the power system is as follows: Anomaly detection based on the updated causal graph: Abnormal risk score calculation: For each node in the updated causal graph, its embedding representation hi contains the aggregated information of multimodal features and can be used to determine the abnormal state of the node.

[0153] The abnormal risk score function R(i) is used to calculate the abnormal degree of the node:

[0154] Where f is a risk scoring function (e.g., a scoring network based on an MLP or attention mechanism), and θ is a model parameter. The risk scoring function analyzes whether the node embedding contains abnormal feature patterns.

[0155] Calculation of abnormal probability: Based on the risk score, the abnormal probability P(i) of the node is further calculated:

[0156] Among them, σ is the Sigmoid function, which maps the risk score to the range of [0,1] to obtain the abnormal probability of the node.

[0157] If the abnormal probability P(i) of a node exceeds the preset threshold τ, the node is marked as an abnormal node. It can be expressed by the following formula:

[0158] Here, τ is the set threshold (for example, 0.8), which means that when the abnormal probability is higher than 80%, the node is marked as abnormal.

[0159] Anomaly propagation and chain reaction detection: In order to detect whether there is a chain reaction or abnormal transmission in the causal chain, risk propagation analysis is performed on the upstream and downstream nodes of the abnormal node (i.e., the directly related adjacent nodes).

[0160] Use the formula:

[0161] Where R(j) is the risk score of node j adjacent to node i, and α is the transmission coefficient, which is used to control the degree of propagation of anomalies in the causal chain. If R(j) exceeds the threshold, node j will also be marked as a potential anomaly.

[0162] Through the above steps, the system can identify abnormal nodes in the updated causal graph and detect whether there is a conduction impact on its upstream and downstream nodes, forming a comprehensive detection of power system anomalies.

[0163] Furthermore, it also includes: New anomaly detection and feature extraction: When a new type of anomaly is discovered in the monitoring system, the data of the abnormal event (such as new visual, audio or text features) is first input into the feature extraction model. The embedded representation of the extracted new anomaly is used as a new node feature to further update the causal graph.

[0164] The feature extraction step should fully consider the feature differences of new anomalies and generate a unified embedding vector so that it is consistent with other node representations in the existing causal graph.

[0165] Dynamic addition of new nodes and edges in the causal graph: The new anomaly is added as a new node in the causal graph, and hyperedges are dynamically generated based on the relationship between the new anomaly and other nodes (such as physical location, time association, or historical data association). The new hyperedge is used to connect the new anomaly node with other related nodes to form a new causal path.

[0166] When connecting new anomaly nodes, the set of nodes to which they are connected can be determined based on algorithms such as multimodal similarity and causal reasoning to accurately reflect the potential impact path of the new anomaly in the causal chain.

[0167] Hypergraph neural network’s secondary update of the causal graph: The updated causal graph is input into the hypergraph neural network again. The HNN recalculates the embeddings of all nodes, especially the nodes connected to the new anomaly nodes, whose embedding representation aggregates the information from the new nodes to reflect the impact of the new anomaly.

[0168] Through multi-layer hypergraph convolution operations, the causal chain structure is updated layer by layer, so that all node embedding representations in the system can contain the latest anomaly information, thereby accurately capturing the potential impact path of new anomalies.

[0169] Risk propagation analysis and high-risk node identification: Using the updated node embedding representation, we conduct risk propagation analysis on new abnormal nodes and their associated nodes and calculate the risk transmission path. Based on the structure of the causal chain, we can track how new abnormalities spread to downstream nodes and identify high-risk nodes that may be affected.

[0170] In the risk propagation analysis, the risk scores of nodes affected by the new anomaly are calculated through weighted risk scores (e.g., the weights are the severity of the new anomaly and the strength of the association between nodes).

[0171] Step 4: Ensure the accuracy and timeliness of causal analysis results by continuously monitoring new data and automatically updating the causal chain.

[0172] It also includes updating the causal chain, specifically: Analyze the multimodal data of power equipment acquired in real time to obtain new anomalies; Based on the structure of the current causal graph, a hypergraph neural network is used to perform high-order correlation modeling between new anomalies and nodes in existing data, accurately capture new causal relationships, infer the possible causes and impacts of new anomalies, and dynamically adjust the nodes and connections of the current causal graph.

[0173] When a new anomaly is detected, the system automatically identifies it and adds it to the existing causal chain structure. At the same time, based on the updated causal chain, the system generates a real-time risk warning report and visualizes the causal path of the abnormal event, helping users quickly understand the root cause of the problem and potential risks, and providing support for management and decision-making.

[0174] Furthermore, during the dynamic causal chain update process, the system first analyzes the input multimodal data in real time and detects any possible new anomalies. When a new anomaly is identified, the system infers the possible causes and effects of the abnormal event based on the current causal chain structure and dynamically adjusts the nodes and connections of the causal chain. This process uses the characteristics of hypergraph neural networks to model high-order associations between new anomalies and nodes in existing data to accurately capture new causal relationships.

[0175] For example, when a new mechanical failure occurs in a piece of power equipment, the system compares the characteristics of the failure with the historical data of related equipment to identify whether there is a potential associated causal path. After identifying the new relationship, the system automatically updates the chain structure and connects the new anomaly with the existing nodes to ensure that the causal chain can reflect the latest equipment status and anomaly characteristics.

[0176] The system uses dynamic causal reasoning to achieve real-time updates of the chain, especially when new abnormal data When input, the system adjusts the existing causal chain structure according to the new data. Assuming that the current causal chain G needs to be updated to G' according to the new input data, this process can be expressed as:

[0177] Among them, ΔG is the causal chain update caused by new abnormal data.

[0178] Based on the updated causal chain, the system generates risk warning signals and calculates the potential risk impact:

[0179] in: αi: represents the weight of each abnormal event, which is usually used to measure the impact of the event on the overall risk. This parameter can be understood as the weight related to the importance of the abnormal node in the causal graph. The weight of the node can be determined by the strength of the abnormal feature in the node embedding, or it can be set according to the position and influence range of the node in the causal chain.

[0180] P(Ei|G′): represents the probability of event Ei occurring in the updated causal graph G′. This probability is usually calculated based on the risk score or abnormal probability of the node. The system analyzes the abnormal features embedded in the node in the causal graph (such as the risk score R(i)) to obtain the abnormal probability P(Ei|G′) of the node, that is, the probability of a certain abnormal node actually occurring.

[0181] How to get P(Ei|G′) from the causal graph: Each node in the updated causal graph contains an embedding representation hi, which aggregates the multimodal data features of the node, and under the multi-layer convolution operations of the hypergraph neural network, these embeddings have reflected the abnormal characteristics of the nodes.

[0182] The node risk score R(i) calculated in the anomaly detection step can be converted into anomaly probability P(Ei|G′)=σ(R(i)), where σ is the Sigmoid function mapped to the range of [0,1].

[0183] If R(i) is high enough, then P(Ei|G′) will also be high, reflecting the possibility of abnormal events at this node.

[0184] The relationship between the above parameters and the causal graph structure: The node weight αi can be assigned different weights based on the node's level, importance, or range of transmission influence in the causal chain. For example, an initial abnormal node or a key transmission node can be assigned a higher weight.

[0185] Edge influence: If the system requires chain reactions or abnormal transmission, the propagation effect of the edges in the causal graph (i.e., risk transmission between adjacent nodes) can be considered. In this case, the edges in the causal graph can contain transmission coefficients to control the degree of abnormal influence between nodes, further affecting the final abnormal probability P(Ei|G′).

[0186] Furthermore, the system will generate risk warning reports in real time to help users prevent potential safety hazards. The risk warning module uses the updated causal chain to automatically identify and analyze the potential risks of abnormal events.

[0187] For example, when the system finds that an abnormality in one device may cause a failure in downstream equipment, the system will generate an early warning signal based on the risk nodes in the causal chain. The risk early warning report displays the causal path of abnormal events in an intuitive visual form, helping users identify the specific path from the current abnormality to the potential risk. The system displays the causal relationship of key nodes on an interactive interface, and users can quickly understand the root cause of the abnormal event, the scope of impact, and its possible harm to other parts of the system by viewing the risk points on the chain.

[0188] Furthermore, during the risk warning process, the system supports multiple forms of notification mechanisms, such as pop-up alarms, email notifications, or SMS reminders, to ensure that users can obtain warning information in a timely manner and take necessary countermeasures. The system also provides detailed causal relationship descriptions, allowing users to understand the background and evolution of the current anomaly based on logical reasoning in the causal chain. This dynamic update and risk warning function not only improves the system's response speed to sudden abnormal events, but also enhances the risk management capabilities of the power system, helping users to foresee and prevent potential threats before problems occur, thereby ensuring the safe operation of the power system.

[0189] Embodiment 2 This embodiment discloses a multi-modal collaborative causal thinking chain reasoning power anomaly detection system.

[0190] The multi-modal collaborative causal thinking chain reasoning power anomaly detection system includes: The data acquisition module is configured to: acquire multimodal data from the power equipment, perform preprocessing and convert the data into a unified embedding representation; The general model data processing module is configured to: input the unified embedding representation into the general model, refine the causal chain generation process into multiple subtasks through the DDCoT reasoning technology, advance each subtask layer by layer, determine the nodes and edges of each subtask, and obtain the final causal chain skeleton; The power-specific model data processing module is configured as follows: the power-specific model is based on the skeleton of the causal chain generated by the general model, and each node of the causal chain is refined layer by layer, and detailed information is supplemented and key factors are filled in to obtain a causal diagram; The hypergraph neural network data processing module is configured to: input the causal graph into the hypergraph neural network, mine high-order associations in the data through multi-layer convolution operations, update the embedded representations of the nodes in the causal graph, and obtain an updated causal graph; The detection module is configured to: determine the nodes with abnormalities based on the embedded representations of the nodes in the updated causal graph.

[0191] Embodiment 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0192] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multi-modal collaborative causal thinking chain reasoning power anomaly detection method as described in Example 1 of the present disclosure.

[0193] Embodiment 4 The purpose of this embodiment is to provide an electronic device.

[0194] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the multi-modal collaborative causal thinking chain reasoning power anomaly detection method as described in Example 1 of the present disclosure are implemented.

[0195] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0196] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0197] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A multi-modal collaborative causal thinking chain reasoning method for power anomaly detection, characterized in that: The following steps are involved: Acquire multimodal data from power equipment, preprocess it and convert it into a unified embedding representation; The unified embedding representation is input into the general model, and the causal chain generation process is refined into multiple subtasks through DDCoT reasoning technology. Each subtask is advanced layer by layer, and the nodes and edges of each subtask are determined to obtain the final causal chain skeleton. The power-specific model is based on the skeleton of the causal chain generated by the general model. It refines each node of the causal chain layer by layer, supplements detailed information and fills in key factors to obtain a causal diagram. The causal graph is input into the hypergraph neural network, and the high-order associations in the data are mined through multi-layer convolution operations. The embedded representations of the nodes in the causal graph are updated to obtain the updated causal graph. Based on the embedded representation of the nodes in the updated causal graph, the nodes with abnormalities are determined.

2. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: Acquire multimodal data from power equipment, preprocess and convert it into a unified embedding representation, including: Collect visual image data of power equipment, extract features from the visual image data based on the Vision Transformer network model, and obtain visual features v; Collect audio data from power equipment, perform denoising and filtering, and extract audio features of the audio data based on Audio Transformer; Collect text data from the power equipment monitoring system and extract text features t of the text data based on the large language model; Standardize and normalize the visual features v, audio features a, and text features t and fuse them into a unified embedding representation .

3. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: The unified embedding representation is input into the general model, and the final causal chain skeleton is obtained by using a multi-level recursive subtask generation method. The specific process includes: According to the type of multimodal data, the anomaly detection of each modality data is processed independently; The anomaly detection of each modal data is divided into a current level task, and the current level task is further divided into a plurality of subtasks according to a recursive formula, wherein the plurality of subtasks include detecting an initial anomaly point, identifying a potential direct cause, and evaluating the impact on surrounding devices; Recursively analyze multiple subtasks in the current level task to determine the causal chain skeleton of the current level task; Based on the recursive analysis results of the current-level task, the identification of potential direct causes and the evaluation of the impact on peripheral devices in multiple subtasks of the current-level task are divided into the next-level task, and the next-level task is further divided into multiple subtasks according to the recursive method; Recursively analyze multiple subtasks in the next-level task to determine the causal chain skeleton of the next-level task; Iterate the above multi-level recursive subtask generation and analysis process to obtain the final causal chain skeleton; or, Recursively analyze multiple subtasks in the current level task to determine the causal chain skeleton of the current level task, including: Detect initial outliers: The embedded representation of each modal data in the current level is used to perform generalized anomaly detection through a general model, identify the anomalies, and use the identified anomalies as the first-class nodes in the current level task, with the corresponding embedded representation; Identify potential direct causes: The first-class node in the current hierarchical task is taken as the analysis object, combined with the corresponding embedding representation, the data of the power equipment related to the first-class node is detected, the potential direct cause of the first-class node is identified and verified, and after the verification is completed, the potential direct cause is taken as the second-class node, connected to the first node through an edge, and the embedding representation of the first-class node and the second-class node is updated; Evaluate the impact on surrounding devices: Take the first-class nodes and second-class nodes in the current hierarchical task as analysis objects, identify the impact of the initial abnormal point and its potential direct cause on other devices or systems, determine the conduction path, and take the identified other affected devices or systems as the third-class nodes. Connect the third-class nodes with the second-class nodes and / or the first-class nodes based on the conduction path, and update the embedded representations of the first-class nodes, the second-class nodes, and the third-class nodes; Based on the above first-class nodes, second-class nodes, third-class nodes and the edges between them, determine the causal chain skeleton of the current level task; or, Consider the spatial, temporal and functional correlation between the initial abnormal point and its potential direct cause and other equipment or systems, identify the impact of the initial abnormal point and its potential direct cause on other equipment or systems, and determine the transmission range and path; or, Iterate the above multi-level recursive subtask generation and analysis process to obtain the final causal chain skeleton, which is as follows: Based on the causal chain skeleton of multiple levels of tasks after iteration, the final causal chain skeleton is obtained. The final causal chain skeleton reflects the complete causal path from the initial anomaly to the direct cause and then to the transmitted impact.

4. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: The power-specific model is based on the skeleton of the causal chain generated by the general model. It refines each node of the causal chain layer by layer, supplements detailed information and fills in key factors to obtain a causal diagram, which includes: Additional details: For each node in the final causal chain skeleton, the power-specific model associates the embedded representation of each node with the knowledge base of the power industry, and combines the background data of equipment operation to generate an embedded representation that is more in line with the characteristics of the power field; the power knowledge base contains power equipment operation data and failure modes, and the background data of equipment operation includes the historical failure records of the equipment, equipment type and location; Filling key factors: For each node in the final causal chain skeleton, the power-specific model further introduces multi-dimensional data of the node. Through the correlation analysis of multi-dimensional data, the key influencing factors of the node are determined and filled in, and the deep potential causes of each node are identified; Generate a cause-effect diagram: After the final causal chain skeleton of the general model is supplemented with detailed information and filled with key factors, the final causal chain skeleton is expanded into a causal graph; or, The determining and filling of key influencing factors of the node specifically includes: Acquire multi-dimensional data of the node, the multi-dimensional data including equipment operation data, environmental data, status monitoring data, maintenance and repair record data, historical abnormal events and external factor data; The power-specific model uses a dynamic feature selection mechanism to select multi-dimensional data closely related to the current node as key influencing factors, and fills the key influencing factors into the embedded representation of the current node; The dynamic feature selection mechanism adopts an adaptive attention mechanism to adaptively assign weights to the multi-dimensional data of the current node, adjust the focus in real time, and focus on the key influencing factors of the current node; or, The equipment operation data includes real-time and historical current, voltage, power, frequency, phase and temperature information; The environmental data includes environmental temperature, humidity, wind speed and rainfall information; The condition monitoring data includes vibration frequency, noise level, oil level and oil temperature information; The maintenance and repair record data includes historical failures, parts replacements, maintenance cycles and inspection data; The historical abnormal events and external factors data include natural disasters or construction disturbances; or, The identification of the deep-level potential causes of each node includes: Based on the equipment operation data, compare it with the normal operating status to determine whether the equipment exceeds the normal working range; and combine the industry knowledge base information, through query and correlation analysis, determine whether the current operating status meets the safety limit or optimal operating range of the equipment; Based on environmental data and referring to environmental impact standards in industry knowledge, determine whether the anomaly is affected by environmental factors; Through condition monitoring data, maintenance and repair record data, identify whether there are hidden problems such as component wear and aging; Verify the potential causes of abnormalities by referring to historical abnormal events and external factor data as well as the device topology.

5. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: The causal graph is input into the hypergraph neural network, and the high-order associations in the data are mined through multi-layer convolution operations. The embedded representation of the nodes in the causal graph is updated to obtain the updated causal graph, which includes: The causal graph is input into the hypergraph neural network, the nodes of the causal graph are extracted as the nodes of the hypergraph, and the weights are calculated according to the similarity and correlation between the modal features, so as to determine the hyperedges between the hypergraph nodes and complete the construction of the hypergraph; Through multi-layer convolution operations, multi-level causal reasoning is performed on the constructed hypergraph structure to extract high-order features in the data layer by layer: in each layer of convolution operations, the embedded representation of the hypergraph node is updated, and the information of the surrounding hypergraph nodes is gradually aggregated, so that each hypergraph node obtains relevant other modal features; Based on the embedded representation of the updated hypergraph nodes, the embedded representation of the nodes in the causal graph is updated to obtain an updated causal graph; or, The weights are calculated based on the similarity and correlation between modal features to determine the hyperedges between hypergraph nodes, including: When calculating the similarity between modal features: Extract multimodal features of each node; Calculate the similarity between the modal features corresponding to any two nodes respectively, and obtain the similarity score of each corresponding modality; The similarity scores of each corresponding modality are weighted or fused to obtain a similarity measure; When calculating the correlation between modal features: If there are temporal or spatial dependencies between nodes in the causal graph, the temporal and spatial correlations between these nodes are calculated based on the time difference and spatial distance; Use the device topology graph to determine the topological distance of nodes in the causal graph, or quantify the association through physical connections to obtain an association metric; The obtained similarity measure and correlation measure are weightedly summed or multiplied again to obtain a comprehensive weight; When the comprehensive weight is greater than the set threshold, a hyperedge is connected between the two nodes; or, In each layer of convolution operation, the embedded representation of the hypergraph node is updated, and the information of the surrounding hypergraph nodes is gradually aggregated so that each hypergraph node obtains relevant other modal features, specifically: Aggregate multiple nodes based on the hyperedges in the hypergraph to form high-order associations between multiple nodes; Based on the high-order associations between multiple nodes, the nodes of the causal graph can obtain multimodal features and contextual information, thereby updating the embedding representation of the hypergraph nodes; By aggregating the feature information of other nodes, the node relationships in the causal graph gradually form a chain-like causal path. With the advancement of multi-layer convolution operations, the relationship between nodes in the causal chain gradually becomes clear, showing direct causes, indirect effects, and upstream and downstream associations.

6. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: Based on the embedded representation of the nodes in the updated causal graph, the abnormal nodes are determined, including: Based on the embedded representation of the nodes in the updated causal graph, an abnormal risk score is calculated using an abnormal risk score function; Map the abnormal risk score to the range of [0,1] and calculate the abnormal probability P(i) of the node; When the abnormal probability of a node exceeds a preset threshold, the current node is determined to be an abnormal node; or, It also includes risk propagation analysis of the upstream and downstream nodes of abnormal nodes: ； Where R(j) is the risk score of node j adjacent to node i; α is the transmission coefficient; f is the risk score function, θ is the model parameter, and h is the embedded representation of the node in the updated causal graph; If R(j) exceeds the threshold, node j will also be marked as a potential anomaly.

7. The multi-modal collaborative causal thinking chain reasoning power anomaly detection method according to claim 1 is characterized in that: It also includes updating the causal chain, specifically: Analyze the multimodal data of power equipment acquired in real time to obtain new anomalies; Add the new anomaly as a new node to the causal graph, and extract the embedding representation of the new anomaly as the feature of the new node in the causal graph; Dynamically generate new hyperedges based on the relationship between the new anomaly and other nodes. The new hyperedges are used to connect the new node with other related nodes to form a new causal path. The updated causal graph is input into the hypergraph neural network again to recalculate the embedding representations of all nodes, especially the nodes connected to the new node. The embedding representations will aggregate the information from the new node, thus reflecting the influence of the new node. Through multi-layer hypergraph convolution operations, the causal chain structure is updated layer by layer, so that all node embedding representations in the system can contain the latest anomaly information, thereby accurately capturing the potential impact path of new anomalies.

8. A multi-modal collaborative causal thinking chain reasoning power anomaly detection system, characterized by: include: The data acquisition module is configured to: acquire multimodal data from the power equipment, perform preprocessing and convert the data into a unified embedding representation; The general model data processing module is configured to: input the unified embedding representation into the general model, refine the causal chain generation process into multiple subtasks through the DDCoT reasoning technology, advance each subtask layer by layer, determine the nodes and edges of each subtask, and obtain the final causal chain skeleton; The power-specific model data processing module is configured as follows: the power-specific model is based on the skeleton of the causal chain generated by the general model, and each node of the causal chain is refined layer by layer, and detailed information is supplemented and key factors are filled in to obtain a causal diagram; The hypergraph neural network data processing module is configured to: input the causal graph into the hypergraph neural network, mine high-order associations in the data through multi-layer convolution operations, update the embedded representations of the nodes in the causal graph, and obtain an updated causal graph; The detection module is configured to: determine the nodes with abnormalities based on the embedded representations of the nodes in the updated causal graph.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the multi-modal collaborative causal thinking chain reasoning power anomaly detection method as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps in the multi-modal collaborative causal thinking chain reasoning power anomaly detection method as described in any one of claims 1-7.

Images