A method and system for extrapolating a world model of power equipment fault evolution based on causal graphs

By using a world model extrapolation method for power equipment fault evolution based on causal graphs, the problem of insufficient modeling of the future evolution process of faults in existing technologies is solved, and the proactive extrapolation of fault evolution trends and the improvement of causal consistency and interpretability of robot operation decisions are realized.

CN123070352APending Publication Date: 2026-11-10HUNAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611585113.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing power equipment fault detection methods lack the ability to model the multi-step evolution process of faults in the future. They are difficult to describe the continuous process of faults from environmental load, changes in operating status, and structural degradation to the formation of risk outcomes. Furthermore, they do not fully utilize the influence of robot actions on future observation, fault confirmation, and equipment status, resulting in insufficient reliability of the model in predicting long-tailed faults, complex coupled faults, and faults under extreme operating conditions.

Method used

A world model extrapolation method for power equipment fault evolution based on causal graphs is adopted. By acquiring multi-source input data, performing time alignment and causal constraint processing, a mapping relationship between causal nodes and variable-level latent space state variables is established. The state transition extrapolation is performed using causal constraint information to extrapolate the latent space state sequence at future moments, and the fault evolution process is combined with external operating conditions and robot action intervention conditions.

Benefits of technology

It improves the causal consistency and interpretability of fault prediction, can proactively infer the future evolution trend of faults, and enhances the accuracy and safety of robot operation decisions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070352A_ABST
    Figure CN123070352A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for extrapolating a world model of power equipment fault evolution based on causal graphs. The method includes: acquiring and preprocessing multi-source input data; standardizing and constraining the causal graph of power faults to generate causal constraint information; semantically aligning causal graph nodes with latent space state variables of the world model variable level, establishing a semantic mapping relationship; encoding the preprocessed multi-source input data, allocating it to corresponding causal nodes based on correlation relationships to obtain node-level latent state observation representations, and then updating the current-time variable-level latent space state variable set based on the semantic mapping relationship; and extrapolating state transitions based on causal constraint information to output the latent space state sequence at future times and the corresponding fault evolution results. This invention extends fault detection to fault evolution extrapolation, subjecting state transitions to the causal mechanism of power equipment faults, which is beneficial for improving the causal consistency and extrapolation reliability of fault prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault prediction technology, and in particular relates to a method and system for extrapolating a world model of power equipment fault evolution based on causal graphs. Background Technology

[0002] In recent years, with the development of online monitoring, intelligent inspection, condition assessment, and risk early warning technologies, power equipment fault detection has gradually evolved from manual experience-based judgment and single-point alarms to automated diagnosis based on multi-source data fusion and intelligent model analysis. Transformers, switchgear, GIS equipment, circuit breakers, cables, insulators, and transmission lines are susceptible to load fluctuations, environmental temperature and humidity, pollution, thunderstorms, salt spray, harmonics, voltage fluctuations, mechanical vibration, material aging, and external operations during long-term operation, gradually exhibiting fault symptoms such as insulation degradation, partial discharge, poor contact, abnormal temperature rise, mechanical loosening, and flashover risks. Actual operation and maintenance not only requires identifying whether the equipment is currently abnormal, but also determining whether the abnormality will continue to develop, whether external conditions will accelerate the fault evolution, and how future observation results and risk status will change after robots perform retesting, inspection, or handling actions. Therefore, power equipment fault analysis is evolving from current state identification to fault evolution prediction and robot operation decision-making.

[0003] Existing technologies mainly include fault detection based on monitoring or inspection data, fault prediction based on historical sequences or digital twins, fault identification based on multimodal inspection by power robots, and auxiliary analysis based on fault knowledge, expert rules, or causal relationships. Fault detection methods typically utilize operational data, infrared or visible light images, partial discharge, oil chromatography, vibration, and acoustic signals to identify, classify, or risk-scorrate current defects. However, most of these methods focus on single-moment detection and lack explicit modeling of the continuous evolution of faults. Historical sequence and digital twin methods can predict equipment status trends, health, or remaining lifespan based on historical equipment state sequences, online monitoring data, degradation indicators, health indicators, or digital twin models. However, these are usually passive predictions based on existing observations and are difficult to extrapolate from changes in external operating conditions such as humidity and load, as well as the impact of robot retesting, cleaning, fastening, or isolation actions on future states. Existing robotic inspection methods typically incorporate sensors such as visible light, infrared, ultraviolet, acoustic, partial discharge, lidar, point cloud, and positioning / navigation sensors for inspection path execution, equipment status acquisition, defect identification, instrument readings, infrared temperature measurement, and partial discharge detection. These methods primarily use the robot as a multimodal data acquisition platform, failing to adequately model the impact of robot actions on future observation quality, fault confirmation confidence, local equipment condition, and operational risks. Fault knowledge, expert rules, and causal relationships, on the other hand, typically utilize equipment structural knowledge, fault mechanisms, historical cases, expert rules, or knowledge graphs to describe the relationships between equipment components, fault symptoms, fault causes, and remedial measures. This aims to improve the interpretability of fault diagnosis and the generation of remedial recommendations. These methods are primarily used for fault explanation, case retrieval, and remedial recommendations, but rarely directly participate in the state transition process of predictive models.

[0004] In summary, while existing power equipment fault detection and prediction technologies have made some progress in fault identification, defect detection, condition scoring, trend prediction, multimodal inspection, and fault interpretation, they still have the following shortcomings in scenarios involving future evolution and robot operation decision-making. First, existing fault detection methods focus heavily on current state identification and lack the ability to model the multi-step evolution of faults, making it difficult to describe the continuous process of a fault from environmental load, changes in operating state, structural degradation, the appearance of fault symptoms, to the formation of risk outcomes. Second, existing fault prediction methods are mostly passive predictions, lacking the ability to extrapolate under external operating conditions and robot intervention, making it difficult to compare the impact of different operating condition changes or different robot actions on future fault risks and observation results. Third, existing power robot inspection methods mainly use robots as multimodal data acquisition platforms, failing to fully utilize the impact of robot actions on future observation, fault confirmation, equipment status, and operational risks, making it difficult to support proactive re-testing, risk avoidance, handling, and operational strategy selection. Fourth, existing fault knowledge or causal relationships are mostly used for diagnosis, explanation, and treatment suggestions, and have not been effectively transformed into state transition constraints within the fault prediction model. This makes the model prone to learning non-causal correlations, affecting the predictive reliability of long-tail faults, complex coupled faults, and extreme operating condition faults.

[0005] Therefore, existing technologies cannot fully meet the needs of power equipment operation and maintenance scenarios for fault future evolution prediction, external operating condition intervention, robot action intervention, multimodal future observation prediction, and fault causal explanation path output. To solve the above problems, it is necessary to propose a new power equipment fault evolution prediction method, which maps the causal relationships in the power equipment fault mechanism to the state variables in the fault prediction model, so that the model is constrained by causal relationships when making future state transitions; at the same time, external operating conditions and robot actions are used as interventionable variables to predict the future evolution process of fault state, risk state, and robot multimodal observation results under different intervention conditions. Based on this, this invention proposes a power equipment fault evolution world model prediction method and system based on causal graphs to improve the causal consistency and interpretability of fault prediction, risk assessment, and robot operation decisions. Summary of the Invention

[0006] To address the above technical problems, this invention provides a method and system for extrapolating a world model of power equipment fault evolution based on causal graphs.

[0007] The technical solution adopted by this invention to solve its technical problem is: A method for extrapolating a world model of power equipment fault evolution based on causal graphs, the method comprising the following steps: S100: Acquire multi-source input data of the target power equipment and perform time-aligned preprocessing on the multi-source input data to obtain time-aligned multi-source input data; S200: Obtain the cause-effect graph of power failure, and perform normalization and causal constraint processing on the cause-effect graph to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model. S300: Obtain the world model, which contains a predefined variable-level latent space state variable structure. Align the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model, and establish a semantic mapping relationship between the causal nodes and the variable-level latent space state variables. S400: Encode the time-aligned multi-source input data to obtain the corresponding latent variables after encoding. Based on the correlation between the multi-source input data and causal nodes, assign the latent variables after encoding to the corresponding causal nodes. Aggregate the latent variables after encoding that belong to the same causal node to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, update the corresponding variable-level latent space state variables using the node-level latent state observation representation to form the variable-level latent space state variable set at the current time. Each variable-level latent space state variable corresponds to the corresponding causal node semantics. S500: Based on causal constraint information, the set of variable-level latent space state variables at the current moment is subjected to causal constraint state transition, and the latent space state sequence at the future moment is deduced. Based on the latent space state sequence at the future moment, the future fault evolution result of the target power equipment is obtained.

[0008] Preferably, in S100, the multi-source input data includes at least one of the following: power equipment operating status data, power robot multimodal observation data, robot motion data, and external working condition data; Time alignment preprocessing includes: for high-frequency signals, extracting statistical features, frequency domain features, time-frequency features, or event features; for low-frequency data, performing interpolation, preservation, or window aggregation; and for robot multimodal observation data, associating data according to robot pose, timestamp, and device object. The time-aligned set of multi-source input data at time t is represented as follows: in, Indicates the first A collection of multi-source input data at any given time. This represents data indicating the operating status of power equipment. This represents multimodal observation data of the robot. Represents robot motion data, This represents external operating condition data.

[0009] Preferably, the cause-effect graph of power faults in S200 is represented as follows: in, Representing the cause-and-effect graph of power outages. Represents the set of causal nodes. Represents the set of causal edges. Represents the set of node attributes and edge attributes; Causal nodes include at least one of the following: environmental load nodes, equipment operating status nodes, equipment structural status nodes, fault symptom nodes, risk outcome nodes, robot action nodes, and robot observation nodes. Any causal edge is represented as: in, Indicates that by node Pointing to node The causal edge, Indicates the causal edge type. Indicates a time lag. Symbols indicating positive or negative action. Indicates the edge confidence or effect strength. Indicates whether the current causal edge involves a modifiable variable; The normalization process includes mapping semantically similar causal nodes to uniform variable identifiers and labeling the causal nodes with their types. The causal constraint processing includes: completing and normalizing the attributes of causal edges, labeling the modifiable variables and modifiable edges, and generating causal constraint information based on the normalized causal nodes and causal edges; the causal constraint information includes at least one of the following: variable identifier, variable type, equipment component to which the variable belongs, observation mode corresponding to the variable, set of causal parent variables, set of causal child variables, edge type, time lag, direction of action, sign of action, edge confidence, and modifiable identifier.

[0010] Preferably, the power fault causal graph in S200 is generated by extracting textual causal relationships: candidate causal event pairs are extracted from power equipment fault reports, maintenance records, defect records, robot inspection reports, standard procedures, and expert experience texts using a causal expression pattern library, and represented as follows: in, This refers to a collection of power equipment fault reports, maintenance records, defect records, robot inspection reports, or standard procedure texts. This represents a causal expression pattern library. Indicates the first One candidate cause event, Indicates the first One candidate result event, This indicates the pattern or confidence level that triggers the current causal relationship. Indicates the number of candidate causal event pairs; Candidate causal edges can be obtained by normalizing, fusing, and verifying candidate causal events. For candidate causal edges, the edge type, direction of action, time lag, action sign, action strength, edge confidence, and intervention attributes are determined by combining expert verification, historical operation data statistics, physical mechanism rules, or simulation results, forming a power fault causal graph for state transition constraints in the world model.

[0011] Preferably, the mapping from causal graph nodes to variable-level latent space state variables in S300 is represented as follows: in, This represents the mapping function from causal nodes to latent space state variables in the world model. The first in the cause-effect graph of power failure is represented by the second. A causal node, Represents nodes The corresponding number Continuous observation of data or status data, Represents nodes Aligned Time-space variables and latent space state variables Indicates the first The set of latent space state variables at time step. This represents the number of latent space state variables involved in fault evolution simulation. Each variable-level latent space state variable corresponds to variable description information, which includes at least one of the following: variable identifier, variable type, equipment component to which the variable belongs, observation mode corresponding to the variable, whether the variable can be intervened, the set of causal parent variables of the variable, and the set of causal child variables of the variable. The mapping is used to determine the correspondence between causal nodes and variable-level latent space state variables. The state value of the variable-level latent space state variable at each time step is updated by the latent variable after being encoded by S400 based on the input data of the corresponding node.

[0012] The S400 includes: S410: For power equipment operating status data, use operating status coding methods such as status feature extraction, time series modeling, or a combination of both to obtain the power equipment operating status coding representation; For multimodal observation data of electric robots, a multimodal perception coding method that integrates images, sensor signals and spatial information is adopted to obtain the multimodal observation coding representation of electric robots; For robot motion data, a motion coding method that combines discrete motion attribute embedding with continuous motion parameter standardization is adopted, and robot motion latent variables are generated by the motion encoder. For external operating condition data, a working condition coding method is adopted, which combines continuous operating condition quantity standardization, discrete operating condition event embedding, and time window feature fusion. External operating condition latent variables are generated by the working condition encoder. S420: The various encoded representations and latent variables obtained in S410 are unified as encoded latent variables of each data source. The association between multi-source input data and causal nodes is established. The encoded latent variables are assigned to the corresponding causal nodes and aggregated to generate the corresponding node-level latent state observation representation. S430: Based on the above semantic mapping relationship, using the node-level latent state observation representations corresponding to each causal node, fill or update the corresponding variable-level latent space state variables to form the variable-level latent space state variable set at the current moment. Specifically: in, Indicates the first The set of latent space state variables at the variable level that is filled or updated at each moment. Indicates the first The variable in the first... The latent space state variables at time step. , This represents the number of latent space state variables involved in fault evolution simulation.

[0013] Preferably, in S500, based on causal constraint information, a causal constraint state transition is performed on the set of variable-level latent space state variables at the current moment to deduce the latent space state sequence at future moments, including: S510: Generate a causal constraint matrix based on the causal constraint information. When there are nodes in the causal constraint information Pointing to node When there is a causal edge, let the matrix elements , representing variables Allow to variable Transmit state information; when there is no corresponding causal edge in the causal constraint information, let the matrix elements... , representing variables Direct participation in variables is not allowed Status update; S520: For any target latent space state variable Only its own historical state and satisfying The set of causal parent variables, allowed external working condition variables, and allowed robot action variables participate in the state update at the next moment; for variable pairs that do not satisfy the causal constraint relationship, during the state transition process, an information transmission control mechanism is used to suppress or shield their information transmission based on the causal constraint matrix. S530: The next-time state sequence of the target latent space state variables is represented as follows: in, Representing variables The corresponding state transition function, This represents the current state of the target variable itself. This represents the set of causal parent variables that are allowed to participate in the update of the target variable's state. Indicates the time lag of causal action. Represents external operating condition variables. Represents robot action variables; state transition function It is implemented using a data-driven time-series state transition model; S540: Traverse all latent variables Aggregate variables The complete set of latent space states for the next time step is obtained, and S520 to S530 are executed iteratively to deduce the sequence of latent space states for multiple future time steps.

[0014] Preferably, S540 further includes: under external working condition intervention or robot action intervention, performing intervention operations on the corresponding variables in the world model, and performing multi-step fault evolution deduction based on the causal constraint state transition relationship after intervention to obtain the latent space state sequence for multiple future time steps; The intervention operations include: setting conditions or replacing inputs for future external working condition sequences and robot action sequences; setting or resetting the state of the target variable and truncating the corresponding causal incoming edges for direct state intervention; and adjusting the edge weights for adjusting the causal action mechanism. Intervention operations are represented as follows: in, Represents the variable Perform intervention procedures. This indicates the set intervention state. Representing variables The set of causal parent variables, This indicates cutting off all causal parent variable pairs. Information transmission; Given the future sequence of external operating conditions and the robot's action sequence, a multi-step recursive process is performed based on a causal constrained world model to obtain the future... A sequence of latent space states at each time step: in, Indicating the intervention conditions The first deduction obtained The state of the latent space at all times. Indicates the first under causal constraint Step-world model state transition function, Indicates the external operating condition intervention sequence. Indicates the robot's action intervention sequence. Represents the constraints of the cause-effect graph of power failures.

[0015] Preferably, in S500, obtaining the future fault evolution result of the target power equipment based on the latent space state sequence at future time steps includes: inputting the latent space state sequence of multiple future time steps into the decoding module and outputting the future evolution result of the power equipment fault state; the future evolution result includes a fault evolution sequence, a risk state sequence, and a power robot multimodal observation prediction result; wherein, the fault evolution sequence is used to represent the state change process of the target power equipment in multiple future time steps, including normal state, abnormal symptom state, early fault state, fault development state, severe risk state, or fault type probability sequence; the risk state sequence is used to represent the risk probability, risk level, health score, fault propagation state, outage risk, or operational risk at each future time step; the power robot multimodal observation prediction result is used to represent the observation results obtained by the robot in the future under given robot action conditions, including one or more of the following: visible light image prediction result, infrared temperature distribution prediction result, ultraviolet discharge prediction result, acoustic anomaly prediction result, partial discharge signal prediction result, point cloud change prediction result, contact feedback prediction result, or observation confidence prediction result; The results of fault evolution, risk state, and multimodal observation and prediction of electric robots are expressed as follows: in, Indicates the first Fault evolution prediction results at time points. Indicates the first Risk state prediction results at any given time Indicates the first Multimodal observation and prediction results of the time-lapse robot , and These represent the fault state decoder, the risk state decoder, and the robot observation decoder, respectively.

[0016] Preferably, S500 further includes: S600: Based on the state change magnitude, causal edge activation intensity, variable contribution, and time sequence of each latent space state variable during the multi-step deduction process, a fault causal explanation path is generated as supplementary explanation information for future evolution results. The fault causal explanation path is used to represent the main causal chain that leads to fault evolution and risk changes, including external operating condition variables, equipment operating state variables, equipment structural state variables, fault symptom variables, risk outcome variables, and robot action variables; The cause-and-effect explanation path for the fault is represented as follows: in, Indicates from the first Time to the The set of causal explanation paths for failures at any given moment. Indicates candidate causal paths, Indicates the first Causal edge of time The activation intensity or contribution This indicates that the top scorers will be selected. Path; Also includes: S700: Based on the fault evolution sequence, risk state sequence, and robot multimodal observation and prediction results, compare the risk changes, observation information gain, action cost, and operational safety of different action schemes, and generate suggestions for the inspection, retesting, avoidance, handling, or manual takeover of the power robot.

[0017] A system for extrapolating a world model of power equipment fault evolution based on causal graphs, used to execute a method for extrapolating a world model of power equipment fault evolution based on causal graphs, including: The data acquisition module is used to acquire multi-source input data of the target power equipment and perform time-aligned preprocessing on the multi-source input data to obtain time-aligned multi-source input data. The causal graph construction module is used to obtain the causal graph of power failure, and to perform normalization and causal constraint processing on the causal graph to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model. The mapping relationship establishment module is used to obtain the world model, which contains a predefined variable-level latent space state variable structure. It aligns the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model and establishes the semantic mapping relationship between the causal nodes and the variable-level latent space state variables. The encoding and mapping module is used to encode the time-aligned multi-source input data to obtain the corresponding encoded latent variables. Based on the correlation between the multi-source input data and causal nodes, the encoded latent variables are assigned to the corresponding causal nodes. The encoded latent variables belonging to the same causal node are aggregated to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, the corresponding variable-level latent space state variables are updated using the node-level latent state observation representation to form the set of variable-level latent space state variables at the current time. Each variable-level latent space state variable corresponds to the semantics of the corresponding causal node. The causal constraint deduction module is used to perform causal constraint state transitions on the set of variable-level latent space state variables at the current moment based on causal constraint information, and to deduce the latent space state sequence at future moments. The results output module is used to obtain the future fault evolution results of the target power equipment based on the potential space state sequence at future time.

[0018] This invention encodes multi-source input data into a set of latent space state variables with causal semantics. Through multi-step deduction via causal constraint state transitions, it can deduce the latent space state sequence of future moments from the current state, extending fault detection from simply determining whether a fault is currently abnormal to actively predicting its future evolution trend. By acquiring a causal graph of power faults and generating causal constraint information, the information transfer between variables is constrained during the state transition process of the world model. This ensures that the state transition follows the physical mechanism of equipment faults rather than statistical correlations, avoiding erroneous deductions caused by non-causal correlations and improving the causal consistency of prediction results. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for extrapolating a world model of power equipment fault evolution based on causal graphs in one embodiment of the present invention; Figure 2 This is a schematic diagram of a world model deduction framework for power equipment fault evolution based on causal graphs in one embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] This invention provides a method and system for extrapolating a world model of power equipment fault evolution based on causal graphs. The flowchart of the method is as follows: Figure 1 As shown in the diagram, the deduction framework is illustrated below. Figure 2 As shown.

[0022] (a) S100: Acquire multi-source input data of the target power equipment and perform time alignment preprocessing on the multi-source input data to obtain time-aligned multi-source input data.

[0023] In one embodiment, in S100, the multi-source input data includes at least one of the following: power equipment operating status data, power robot multimodal observation data, robot motion data, and external working condition data; Power equipment operating status data includes electrical operating data, thermal status data, mechanical status data, insulation status data, switch and protection status data, alarm data, and maintenance record data. Specifically, it can include one or more of the following: voltage, current, load, active power, reactive power, equipment temperature, oil temperature, winding temperature, partial discharge signals, vibration signals, acoustic signals, oil chromatography data, insulation resistance, switch status, protection action records, alarm records, historical defect records, and maintenance records. It can also include other data that can characterize the operating status or status changes of the target power equipment. Multimodal observation data for power robots includes one or more of the following: visible light images, infrared images, ultraviolet images, acoustic detection data, partial discharge detection data, point cloud data, lidar data, force data, tactile data, and robot pose data. Robot motion data includes one or more of the following: movement, approach, circling, dwelling, zoom shooting, changing perspective, infrared re-measurement, ultraviolet re-measurement, partial discharge detection, acoustic detection, contact detection, cleaning, fastening, isolation, avoidance, and stopping / waiting. Robot motion data may also include the time of motion occurrence, duration of motion, object of motion, spatial location of motion, motion execution parameters, and motion safety constraints. External operating condition data includes one or more of the following: ambient temperature, humidity, wind speed, rainfall, lightning, pollution level, salt spray, altitude, load fluctuation, voltage fluctuation, harmonic disturbance, short-term overload, seasonal changes, and operation mode switching.

[0024] Because the data sources, sampling frequencies, and data formats are different, this invention aligns the data from different sources according to a unified time window.

[0025] Time alignment preprocessing includes: for high-frequency signals, extracting statistical features, frequency domain features, time-frequency features, or event features; for low-frequency data, performing interpolation, preservation, or window aggregation; and for robot multimodal observation data, associating data according to robot pose, timestamp, and device object. The time-aligned first The multi-source input data set at time t is represented as: in, Indicates the first A collection of multi-source input data at any given time. This represents data indicating the operating status of power equipment. This represents multimodal observation data of the robot. Represents robot motion data, This represents external operating condition data.

[0026] Through S100, the system obtains a multi-source data foundation to represent the current operating status of the equipment, on-site robot observations, robot operation behavior, and external environmental conditions, providing input for subsequent causal graph processing, variable-level state coding, and fault evolution inference.

[0027] (ii) S200: Obtain the cause-effect graph of power failure, and perform normalization and causal constraint processing on the cause-effect graph of power failure to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model.

[0028] Specifically, a power fault causal graph is used to characterize the causal relationships between environmental loads, equipment operating status, equipment structural status, fault symptoms, risk outcomes, robot actions, and robot observations. It can be a pre-built graph, or a graph retrieved from external knowledge bases, expert systems, equipment fault knowledge bases, digital twin systems, operation and maintenance knowledge bases, or historical fault case databases. It can also be a graph generated through text causal relationship extraction, expert rule modeling, historical data causal discovery, simulation rule generation, or a combination of the above methods.

[0029] It should be noted that this invention does not limit the specific source and construction method of the power fault causal graph. As long as it can characterize the causal relationship between environmental loads, equipment operating status, equipment structural status, fault symptoms, risk outcomes, robot actions, and robot observations, and can further generate causal constraint information for constraining the state transition of the world model, the purpose of this invention can be achieved. The power fault causal graph can be constructed internally by this system or provided by an external system. The core of this invention lies in: performing causal constraint processing on the acquired power fault causal graph, and using the processed causal constraint information in the latent space state transition process of the world model.

[0030] In one embodiment, the power fault cause-effect graph in S200 is represented as follows: in, Representing the cause-and-effect graph of power outages. Represents the set of causal nodes. Represents the set of causal edges. Represents the set of node attributes and edge attributes; Causal nodes include at least one of the following: environmental load nodes, equipment operating status nodes, equipment structural status nodes, fault symptom nodes, risk outcome nodes, robot action nodes, and robot observation nodes. Environmental load nodes represent external operating conditions such as high temperature, high humidity, thunderstorms, pollution, salt spray, heavy load, and harmonic disturbances. Equipment operating status nodes represent operating states such as voltage, current, load, temperature, partial discharge intensity, vibration amplitude, and oil chromatography status. Equipment structural status nodes represent structural states such as insulation status, contact status, sealing status, mechanical loosening status, and cooling status. Fault symptom nodes represent fault manifestations such as partial discharge, abnormal temperature rise, abnormal noise, abnormal vibration, abnormal oil chromatography, surface cracks, flashover symptoms, and poor contact symptoms. Risk outcome nodes represent fault type, fault level, risk probability, risk level, downtime risk, or operational risk. Robot action nodes represent observation actions, detection actions, contact actions, handling actions, or avoidance actions. Robot observation nodes represent multimodal observation results obtained or predicted after the robot performs actions.

[0031] Any causal edge is represented as: in, Indicates that by node Pointing to node The causal edge, Indicates the causal edge type. Indicates a time lag. Symbols indicating positive or negative action. Indicates the edge confidence or effect strength. This indicates whether the current causal edge involves a modifiable variable.

[0032] After obtaining the causal graph of power faults, this invention performs the following processing on the graph. First, the causal nodes are normalized. Normalization includes mapping semantically similar causal nodes to unified variable identifiers and labeling the causal nodes by type; for example, high humidity, humid air, and high-humidity environment are normalized to high-humidity environment nodes; insulation degradation, reduced insulation performance, and insulation deterioration are normalized to insulation state degradation nodes; and partial discharge enhancement, increased partial discharge, and abnormal discharge pulse are normalized to partial discharge anomaly nodes. Based on the node semantics and equipment objects, the nodes are labeled as environmental load nodes, equipment operating status nodes, equipment structural status nodes, fault symptom nodes, risk result nodes, robot action nodes, or robot observation nodes. Node types are used for subsequent grouping of variable-level latent space state variables and state update constraints.

[0033] Causal constraint processing includes: attribute completion and normalization of causal edges, annotation of modifiable variables and modifiable edges, and generation of causal constraint information based on the normalized causal nodes and edges. Specifically, for each causal edge, its direction of action, edge type, time lag, action sign, action strength, edge confidence, and applicable operating conditions are determined. Edge types include one or more of the following: environmental causal edge, operational evolution edge, structural degradation edge, symptom generation edge, risk transformation edge, action observation edge, and action handling edge. For external operating condition nodes and robot action nodes, their modifiable attributes are annotated. For example, external operating condition nodes such as high humidity, high temperature, heavy load, thunderstorms, and increased pollution can be used as external operating condition intervention variables; robot action nodes such as infrared re-measurement, partial discharge detection, cleaning, fastening, isolation, and avoidance can be used as robot action intervention variables.

[0034] Causal constraint information includes one or more of the following: variable identifier, variable type, equipment component to which the variable belongs, observation mode corresponding to the variable, set of causal parent variables, set of causal child variables, edge type, time lag, direction of action, sign of action, edge confidence, and intervention indicator.

[0035] In one embodiment, the power fault causal graph is generated by extracting textual causal relationships: candidate causal event pairs are extracted from power equipment fault reports, maintenance records, defect records, robot inspection reports, standard procedures, and expert experience texts using a causal expression pattern library, and represented as follows: in, This refers to a collection of power equipment fault reports, maintenance records, defect records, robot inspection reports, or standard procedure texts. This represents a causal expression pattern library. Indicates the first One candidate cause event, Indicates the first One candidate result event, This indicates the pattern or confidence level that triggers the current causal relationship. Indicates the number of candidate causal event pairs; Candidate causal edges can be obtained by normalizing, fusing, and verifying candidate causal events. For candidate causal edges, the edge type, direction of action, time lag, action sign, action strength, edge confidence, and intervention attributes are determined by combining expert verification, historical operation data statistics, physical mechanism rules, or simulation results, forming a power fault causal graph for state transition constraints in the world model.

[0036] Specifically, the causal expression pattern library includes causal trigger expressions such as "due to...leading to...", "because...therefore...", "...cause...", "...cause...", "...inducing...", "...causing...", "...originating from...", etc.

[0037] Through step S200, this invention processes power fault causal maps from different sources and with different forms of expression into causal constraint information that can be used by the world model. This causal constraint information is not simply used for fault interpretation, but rather for constraints on information transmission direction, state update relationship, and intervention propagation path during subsequent variable-level latent space state transition processes.

[0038] In S500, causal constraint information is specifically used to generate a causal constraint matrix. Based on the edge type, time lag, action sign, and edge confidence or action strength of the causal edge, the state information transmitted from the causal parent variable to the target variable is filtered, transformed, or weighted, so that the causal edge attributes generated by S200 can be actually used in the subsequent state transition process.

[0039] (iii) S300: Obtain the world model, which contains a predefined variable-level latent space state variable structure. Align the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model, and establish a semantic mapping relationship between the causal nodes and the variable-level latent space state variables.

[0040] Specifically, existing world models typically use a single overall latent state to represent the environmental state. In contrast, this invention organizes the latent space states of the world model into a set of variable-level latent space state variables, giving each latent space state variable a clear causal semantic. For example, the environmental humidity node corresponds to the environmental load latent space state variable, the load node corresponds to the equipment operating state latent space state variable, the insulation state node corresponds to the equipment structural state latent space state variable, the partial discharge node corresponds to the fault symptom latent space state variable, the risk level node corresponds to the risk state latent space state variable, the robot infrared re-measurement action node corresponds to the robot action latent space state variable, and the infrared observation result node corresponds to the robot observation latent space state variable.

[0041] In one embodiment, the mapping from causal graph nodes to variable-level latent space state variables in S300 is represented as follows: in, This represents the mapping function from causal nodes to latent space state variables in the world model. The first in the cause-effect graph of power failure is represented by the second. A causal node, Represents nodes The corresponding number Continuous observation of data or status data, Represents nodes Aligned Time-space variables and latent space state variables; To facilitate subsequent state transition calculations, each variable-level latent space state variable is associated with variable description information. This description information includes at least one of the following: variable identifier, variable type, equipment component to which the variable belongs, corresponding observation mode, whether the variable can be interfered with, the variable's causal parent variable set, and the variable's causal child variable set. The causal parent variable set refers to the set of variables in the power fault causality graph that are allowed to directly influence the target variable's state update; the causal child variable set refers to the set of variables in the power fault causality graph that are directly influenced by the target variable. Mapping is used to determine the correspondence between causal nodes and variable-level latent space state variables. The state value of the variable-level latent space state variable at each time step is updated by the latent variable encoded by S400 based on the input data of the corresponding node.

[0042] Through S300, this invention transforms the external causal graph of power faults into a variable organization structure within the world model. The causal graph is no longer merely used for retrieval or interpretation, but rather determines which latent space state variables in the world model can participate in the state update of the target variable, which variables cannot directly transmit information, and which variables can be intervened by external operating conditions or robot actions.

[0043] It is understandable that the alignment method between causal graph nodes and world model state variables can be replaced. In one implementation, a one-to-one correspondence between causal nodes and latent space state variables can be used; in other implementations, one-to-many, many-to-one, hierarchical alignment, semantic embedding alignment, or dynamic alignment methods can also be used. For example, multiple external operating condition nodes can be jointly mapped to environmental load latent variables, and a device structural state node can also be split into multiple structural state sub-variables. As long as the nodes or node types in the causal graph can be transformed into variable-level latent space state variables and used for subsequent state transition constraints, they all belong to the alternative implementation methods of this invention.

[0044] (iv) S400: Encode the time-aligned multi-source input data to obtain the corresponding latent variables after encoding. Based on the correlation between the multi-source input data and causal nodes, assign the latent variables after encoding to the corresponding causal nodes. Aggregate the latent variables after encoding that belong to the same causal node to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, update the corresponding variable-level latent space state variables using the node-level latent state observation representation to form the set of variable-level latent space state variables at the current time. Each variable-level latent space state variable corresponds to the corresponding causal node semantics.

[0045] In one embodiment, S400 includes: S410: For power equipment operating status data, an operating status coding method using state feature extraction, time-series modeling, or a combination of both is employed to obtain a coded representation of the power equipment operating status. Specifically, for continuous numerical or low-frequency time-series data such as voltage, current, load, active power, reactive power, equipment temperature, oil temperature, winding temperature, oil chromatography data, and insulation resistance, normalization, statistical feature extraction, window aggregation, recurrent neural networks, time-series Transformers, state-space models, or combinations thereof are used for coding. For high-frequency signal data such as partial discharge, vibration, and acoustics, statistical feature extraction, frequency domain feature extraction, and time-series modeling are employed. Encoding methods include frequency transformation, event feature extraction, one-dimensional convolutional networks, recurrent neural networks, temporal Transformers, or combinations thereof. For discrete state or event data such as switch status, protection action records, and alarm records, one-hot encoding, tag encoding, embedding encoding, event feature extraction, or event sequence encoding are used. For structured or text-based data such as historical defect records and maintenance records, field encoding, event encoding, text feature extraction, text encoders, or combinations thereof are used. For other unlisted power equipment operating status data, the corresponding encoding method is adopted according to its data format to obtain the corresponding input data encoding representation. It should be noted that the specific data mentioned above are illustrative examples. S400 encodes the power equipment operating status data that was actually acquired and completed in S100 and used for this fault evolution simulation. For data not listed in the above examples, the corresponding encoding method is adopted according to its data form. After encoding, the data is associated with the corresponding causal nodes based on the equipment object, equipment component, timestamp, observation mode, or variable semantics. Data that cannot be associated with causal nodes does not participate in the filling or updating of variable-level latent space state variables.

[0046] For multimodal observation data of electric robots, a multimodal perception coding method that integrates images, sensor signals and spatial information is adopted to obtain a multimodal observation coding representation of electric robots. Specifically, a visual encoder, a multimodal encoder, a target detection network, a semantic segmentation network, a point cloud encoder or a combination thereof can be used for coding. For robot motion data, a motion encoding method that combines discrete motion attribute embedding with continuous motion parameter standardization is adopted, and robot motion latent variables are generated by motion encoder. Specifically, for robot motion data, discrete attributes such as motion type, motion object, and motion action parts are converted into discrete codes or embedded representations, and continuous attributes such as motion spatial position, robot pose, motion duration, and motion execution parameters are normalized. Then, the discrete codes are concatenated or fused with the normalized continuous attributes, and robot motion latent variables are obtained through robot motion encoder. For external operating condition data, an operating condition coding method that combines continuous operating condition quantity standardization, discrete operating condition event embedding, and time window feature fusion is adopted to generate external operating condition latent variables through an operating condition encoder. Specifically, for external operating condition data, continuous operating condition variables such as temperature, humidity, wind speed, pollution, load disturbance, and voltage disturbance are normalized, and discrete operating condition events such as rainfall, lightning, and operation mode switching are converted into discrete codes or embedded representations. At least one of the mean, maximum, minimum, rate of change, or cumulative amount is extracted according to a preset time window, and external operating condition latent variables are obtained through an external operating condition encoder. S420: The various encoded representations obtained in S410 are unified with the latent variables as encoded latent variables of each data source. The association between multi-source input data and causal nodes is established based on at least one of equipment object, equipment component, timestamp, observation mode and variable semantics. The encoded latent variables are assigned to the corresponding causal nodes and aggregated to generate the corresponding node-level latent state observation representation. S430: Based on the above semantic mapping relationship, using the node-level latent state observation representations corresponding to each causal node, fill or update the corresponding variable-level latent space state variables to form the variable-level latent space state variable set at the current moment. Specifically: in, Indicates the first The set of latent space state variables at the variable level that is filled or updated at each moment. Indicates the first The variable in the first... The latent space state variables at time step. , This represents the number of latent space state variables involved in fault evolution simulation.

[0047] The above encoding results are not simply concatenated into a single latent vector. Each encoded variable is mapped to its corresponding latent space state variable slot according to the semantic mapping relationship of S300. For example, the highest temperature, area of ​​the temperature rise region, and temperature gradient in the infrared image can be entered into the temperature rise anomaly variable; the number of pulses, discharge amplitude, and spectral characteristics of the partial discharge signal can be entered into the partial discharge variable; the dirty area, cracked area, or damaged area in the visible light image can be entered into the surface state variable or structural state variable; and the robot action type and action parameters can be entered into the action variable.

[0048] Through S400, this invention enables multi-source heterogeneous data to be integrated not in an unstructured manner, but to enter the corresponding latent space state variables according to the variable semantics in the power fault causal graph, providing structured state input for subsequent causal constraint state transitions.

[0049] (v) S500: Based on causal constraint information, perform causal constraint state transition on the variable-level latent space state variable set at the current moment, deduce the latent space state sequence at the future moment, and obtain the future fault evolution result of the target power equipment based on the latent space state sequence at the future moment.

[0050] Specifically, in obtaining the first The set of latent space state variables at time step Subsequently, this invention constructs a latent space state transition process for a world model based on causal constraint information.

[0051] In one embodiment, S500 performs causal constraint state transitions on the current set of variable-level latent space state variables based on causal constraint information to deduce a sequence of latent space states for future times, including: S510: Generate a causal constraint matrix based on the causal constraint information. When there are nodes in the causal constraint information Pointing to node When there is a causal edge, let the matrix elements , representing variables Allow to variable Transmit state information; when there is no corresponding causal edge in the causal constraint information, let the matrix elements... , representing variables Direct participation in variables is not allowed Status update; S520: For any target latent space state variable Only its own historical state and satisfying The set of causal parent variables, allowed external working condition variables, and allowed robot action variables participate in the state update at the next moment. For variable pairs that do not satisfy the causal constraint relationship, during the state transition process, information transmission control mechanisms are used to suppress or shield their information transmission based on the causal constraint matrix. For example, information transmission can be suppressed or shielded by state transition matrix, attention mask, adjacency matrix mask, graph message passing gating, gradient channel constraint, or state update gating. S530: The next-time state sequence of the target latent space state variables is represented as follows: in, Representing variables The corresponding state transition function, This represents the current state of the target variable itself. This represents the set of causal parent variables that are allowed to participate in the update of the target variable's state. Indicates the time lag of causal action. Represents external operating condition variables. Represents robot action variables; state transition function This invention employs a data-driven temporal state transition model, such as recurrent neural networks, gated recurrent units, long short-term memory networks, graph neural networks, graph attention networks, Transformers, state-space models, neural differential equations, probabilistic state-space models, or combinations thereof. The invention does not limit the specific network structure; its core lies in defining the direction of information transfer and the state update relationship between variables based on causal constraints during the latent space state transition process.

[0052] S540: Traverse all latent variables Aggregate variables The complete set of latent space states for the next time step is obtained, and S520 to S530 are executed iteratively to deduce the sequence of latent space states for multiple future time steps. Through S500, this invention transforms the causal graph of power failure into state transition structural constraints in the world model, enabling the world model to follow the causal mechanism of power equipment failure when extrapolating future states, rather than relying entirely on data correlation to freely learn state transitions.

[0053] It is understandable that the specific model structure for causal constraint state transitions can be replaced. The state transition function of the world model can be implemented using graph neural networks, graph attention networks, recurrent neural networks, long short-term memory networks, Transformers, state-space models, neural ordinary differential equations, probabilistic state-space models, Bayesian networks, or hybrid models combining physical mechanism models and neural networks. Causal constraints can be implemented using hard masks, soft gating, attention masks, causal message passing matrices, edge type parameters, causal consistency loss, or rule post-processing. As long as the model limits the information transmission direction and state update relationship between latent space state variables based on causal constraint information, the purpose of this invention can be achieved.

[0054] In one embodiment, S540 further includes: under external working condition intervention or robot action intervention, performing intervention operations on the corresponding variables in the world model, and performing multi-step fault evolution deduction based on the causal constraint state transition relationship after intervention to obtain the latent space state sequence for multiple future time steps; external working condition intervention includes changes in working conditions such as high temperature, high humidity, thunderstorms, increased pollution, enhanced salt spray, increased load, voltage fluctuations, enhanced harmonics, short-term overload, and switching of operating modes. Robot action intervention includes actions such as robot approach observation, changing perspective, zooming for shooting, infrared re-measurement, ultraviolet re-measurement, partial discharge detection, acoustic detection, contact detection, cleaning, fastening, isolation, avoidance, and requesting manual takeover.

[0055] This invention categorizes robot actions into observational actions, detectional actions, remedial actions, and risky actions. Observational actions primarily affect future observation results or observation confidence levels, such as approaching the target area, changing the viewing angle, zooming in for a shot, and infrared retesting. Detectional actions primarily affect the observability of specific fault symptoms, such as partial discharge detection, acoustic detection, and ultraviolet detection. Remedial actions can affect the local state or risky state of equipment, such as cleaning, securing, and isolating. Risky actions may introduce additional operational risks, such as approaching energized areas, contact detection, or operating in complex spaces.

[0056] Given a future sequence of external working conditions or a sequence of robot actions, the sequence can be input into a causal constraint world model by setting conditions or replacing inputs. When it is necessary to simulate a scenario where the target variable is directly set externally, the target variable can be state set or reset and its original causal incoming edges can be truncated. This type of processing is called direct state intervention. When it is necessary to simulate changes in the causal action mechanism, the edge weights or state transition parameters of the corresponding causal edges can be adjusted. This type of processing is called mechanism adjustment.

[0057] Robot action variables do not participate in the update of all variable-level latent space state variables by default. Instead, the range of variables they can influence is determined based on the causal edges between the robot action nodes and the target nodes. Observation-type actions such as approaching for observation, changing the viewpoint, zooming for shooting, and infrared re-measurement mainly affect the robot's observation results or observation confidence. Detection-type actions such as partial discharge detection, acoustic detection, and ultraviolet detection mainly affect the observability of corresponding fault symptoms or detection results. Handling actions such as cleaning, fastening, and isolation mainly affect the structural state, fault symptoms, or risk state of the corresponding equipment. Risk-type actions such as approaching energized areas, contact detection, or operating in complex spaces can affect the operational risk state.

[0058] In one implementation, when intervention is performed on a variable, the variable undergoes state setting, state reset, input replacement, causal edge truncation, or edge weight adjustment. For example, when simulating a future increase in humidity, the humidity variable is set as the target intervention state, and its impact on insulation status, partial discharge symptoms, and risk level is calculated in subsequent simulations; when simulating a robot performing infrared retesting, the robot action variable is set as the infrared retesting action, and the impact of this action on future infrared observation results, the confidence level of temperature rise anomaly confirmation, and risk assessment results is simulated; when simulating a robot performing cleaning actions, the robot action variable is set as the cleaning action, and its impact on contamination status, leakage current, partial discharge risk, and fault risk level is simulated.

[0059] Intervention operations are represented as follows: in, Represents the variable Perform intervention procedures. This indicates the set intervention state. Representing variables The set of causal parent variables, This indicates cutting off all causal parent variable pairs. The information transmission; by setting or adjusting the causal inbound edge of the variable to be intervened, the variable can be no longer determined by the original parent variable, but directly set by the intervention condition.

[0060] Given the future sequence of external operating conditions and the robot's action sequence, a multi-step recursive process is performed based on a causal constrained world model to obtain the future... The sequence of latent space states at each time step: in, Indicating the intervention conditions The first deduction obtained The state of the latent space at all times. Indicates the first under causal constraint Step-world model state transition function, Indicates the external operating condition intervention sequence. Indicates the robot's action intervention sequence. Represents the constraints of the cause-effect graph of power failures. Indicates the step size of the deduction.

[0061] Using the methods described above, the system can compare the differences in fault evolution under different intervention schemes. For example, it can compare the fault evolution trajectory and risk state changes under different conditions such as no intervention, future high humidity conditions, robot infrared re-measurement, robot partial discharge detection, robot cleaning, and robot obstacle avoidance.

[0062] It is understandable that external working condition intervention, robot motion intervention, and output methods can be interchangeable. External working condition intervention can be achieved through single variable setting, future working condition sequence setting, working condition scenario label setting, edge weight adjustment, or counterfactual inference. Robot motion intervention can be modeled according to motion type, motion parameters, or motion sequence, and can also be classified into observation-type motions, detection-type motions, handling-type motions, and risk-type motions. Output results can be one or more of the following: fault evolution sequence, risk score sequence, risk level sequence, robot multimodal observation and prediction results, structured observation and prediction results, risk differences before and after intervention, fault causal explanation path, or robot operation suggestions.

[0063] In one embodiment, S500 obtains the future fault evolution result of the target power equipment based on the potential space state sequence at future times, including: The latent space state sequence for multiple future time steps is input into the decoding module, which outputs the future evolution results of the power equipment fault state. The future evolution results include the fault evolution sequence, the risk state sequence, and the multimodal observation and prediction results of the power robot. Among them, the fault evolution sequence is used to represent the state change process of the target power equipment in multiple future time steps, including normal state, abnormal symptom state, early fault state, fault development state, severe risk state, or fault type probability sequence; the risk state sequence is used to represent the risk probability, risk level, health score, fault propagation state, outage risk, or operational risk at each future time; the multimodal observation and prediction results of the power robot are used to represent the observation results obtained by the robot in the future under given robot action conditions, including visible light image prediction results, infrared temperature distribution prediction results, ultraviolet discharge prediction results, acoustic anomaly prediction results, partial discharge signal prediction results, point cloud change prediction results, contact feedback prediction results, or observation confidence prediction results. The results of fault evolution, risk state, and robot multimodal observation and prediction are expressed as follows: in, Indicates the first Fault evolution prediction results at time points. Indicates the first Risk state prediction results at any given time Indicates the first Multimodal observation and prediction results of the time-lapse robot , and These represent the fault evolution decoder, the risk state decoder, and the robot observation decoder, respectively.

[0064] In practical engineering implementation, the results of robot multimodal observation and prediction can be complete images, complete signals, or structured predictions. For example, for infrared observation, it can output the future maximum temperature, the area of ​​the temperature rise region, the temperature gradient, or the confidence level of infrared anomalies; for partial discharge detection, it can output the future number of discharge pulses, discharge amplitude, probability of discharge type, or partial discharge risk score; for visible light observation, it can output the location of the defect region, the defect category, and the defect visibility score. This can reduce the difficulty of complete multimodal reconstruction and improve engineering feasibility.

[0065] In one embodiment, S500 is followed by: S600: Based on the state change amplitude, causal edge activation intensity, variable contribution, and time sequence of each latent space state variable during the multi-step deduction process, a fault causal explanation path is generated as supplementary explanation information for future evolution results. The fault causal explanation path is used to represent the main causal chain leading to fault evolution and risk changes, including external operating condition variables, equipment operating state variables, equipment structural state variables, fault symptom variables, risk outcome variables, and robot action variables. The cause-and-effect explanation path for the fault is represented as follows: in, Indicates from the first Time to the The set of causal explanation paths for failures at any given moment. Indicates candidate causal paths, Indicates the first Causal edge of time The activation intensity or contribution This indicates that the top scorers will be selected. Path; For example, under high humidity and polluted conditions, the system can output the following causal explanation path: increased ambient humidity → enhanced conductivity of surface contaminants → increased leakage current → enhanced partial discharge → abnormal infrared temperature rise → increased risk level.

[0066] For example, under the condition of robot cleaning action, the system can output the following causal explanation path: robot cleaning action → reduced surface dirt state → reduced leakage current → reduced risk of partial discharge → reduced risk level.

[0067] Through S600, this invention not only outputs the fault prediction result, but also outputs the temporally ordered causal path that leads to the prediction result, which makes it easier for maintenance personnel to understand the reasoning basis and also makes it easier for the power robot to select the next operation action.

[0068] In one embodiment, it also includes: S700: After obtaining fault evolution sequences, risk state sequences and robot multimodal observation and prediction results under different external working conditions and robot action conditions, compare the risk changes, observation information gain, action cost and operation safety of different action schemes, and generate suggestions for inspection, retesting, avoidance, handling or manual takeover of power robots.

[0069] For example, when the simulation results show that a certain device has an increased risk of partial discharge under high humidity and heavy load conditions, the system can suggest that the robot perform partial discharge retesting or infrared retesting; when the simulation results show that a certain contact detection action may cause an increase in operational risk, the system can suggest that the robot avoid the area or request manual intervention; when the simulation results show that cleaning or tightening actions can reduce the risk of subsequent failures, the system can output them as candidate actions.

[0070] It should be noted that S700 can be used as an optional embodiment of the present invention. The core of the present invention lies in the fault evolution world model deduction based on causal graph constraints, and the robot operation decision is a further application based on the deduction results.

[0071] Next, the specific implementation process of the present invention will be explained using the evolution of pollution discharge risk in substation insulators as an example.

[0072] First, the system acquires operational status data, external operating condition data, and robot inspection data for the target insulator. Operational status data includes load, voltage, current, and historical defect records; external operating condition data includes humidity, temperature, rainfall, and pollution level; robot observation data includes visible light images, infrared images, and partial discharge detection data; robot motion data includes proximity observation, infrared re-measurement, partial discharge detection, and cleaning actions.

[0073] Then, the system acquires a pre-constructed causal graph of insulator pollution discharge risk and performs causal constraint processing on the graph. This causal graph includes nodes such as humidity, pollution state, surface leakage current, insulation state, partial discharge, infrared temperature rise anomaly, risk level, robot infrared re-measurement, robot partial discharge detection, and robot cleaning. Causal edges include humidity pointing to pollution conductivity, pollution conductivity pointing to leakage current, leakage current pointing to partial discharge, partial discharge pointing to infrared temperature rise anomaly, infrared temperature rise anomaly pointing to risk level, and robot infrared re-measurement pointing to infrared observation confidence level, and robot cleaning pointing to pollution state, etc.

[0074] Subsequently, the system maps the aforementioned causal nodes to variable-level latent space state variables in the world model, for example... , , , , , , , The robot's visible light and infrared images are processed by a visual encoder and then mapped to dirt state variables and infrared temperature rise anomaly variables; the partial discharge signal is processed by a signal encoder and then mapped to a partial discharge variable; the humidity and rainfall data are processed by a working condition encoder and then mapped to an environmental load variable.

[0075] When multiple observations, such as visible light, infrared, and partial discharge, are simultaneously associated with the same causal node, the system aggregates the corresponding encoded representations according to observation quality, time freshness, or detection confidence, and uses the node-level latent state observation representation obtained from the aggregation to update the current state of the corresponding variable. When there is a lack of effective observations at a certain moment, the system maintains or updates the current state based on the state of the variable at the previous moment and the state of its causal parent variable.

[0076] In the current state, if the system detects high ambient humidity, heavy surface contamination, and a slight temperature rise in the infrared image, the causal-constrained world model will extrapolate the future based on the current latent space state and causal graph. Without robotic intervention, the model can deduce a fault evolution sequence of increased leakage current, enhanced partial discharge, abnormally increased temperature rise, and elevated risk level, and output the corresponding causal explanation path: increased humidity → enhanced conductivity of contaminants → increased leakage current → enhanced partial discharge → elevated risk level.

[0077] Under the intervention condition of the robot performing infrared retesting, model simulation shows that the physical risk status of the equipment may not decrease immediately, but the confidence level of infrared observation and the confidence level of abnormal temperature rise confirmation increase. Based on this, the system can recommend that the robot perform infrared retesting to improve the reliability of fault confirmation. Under the intervention condition of the robot performing cleaning, model simulation shows that the level of contamination decreases, leakage current decreases, partial discharge risk decreases, and risk level decreases. Based on this, the system can consider cleaning as a candidate action and provide the corresponding risk reduction trend and causal explanation path.

[0078] Through the above implementation methods, the present invention can compare the future failure evolution trajectories under different external working conditions and different robot actions in the same causal constraint world model, and output the failure evolution sequence, risk state sequence, robot future observation prediction results and failure causal explanation path.

[0079] Compared with existing methods for power equipment fault detection, fault prediction, knowledge graph reasoning, digital twin diagnosis, and power robot inspection, this invention has the following advantages: (1) It can extend from current fault detection to future fault evolution prediction. Existing power equipment fault detection methods typically output the current fault category, defect level, risk score, or anomaly alarm based on current operating data, online monitoring data, inspection images, or sensor signals, mainly addressing whether an anomaly is currently present and what type of fault it is. These methods generally lack the ability to model the continuous evolution of a fault from environmental loads, changes in equipment operating status, structural degradation, the manifestation of fault symptoms, to the formation of a risk outcome. This invention encodes multi-source data into a set of variable-level latent space state variables and performs multi-step latent space state transitions through a world model. This enables the deduction of the evolution of power equipment fault states at multiple future moments, outputting fault evolution sequences and risk state sequences. Therefore, this invention can not only determine the current equipment state but also further determine whether a minor anomaly is likely to continue developing, whether future risks may escalate, and along what path the fault might evolve.

[0080] (2) It can reduce erroneous inferences caused by non-causal correlations. Existing data-driven fault prediction methods typically rely on statistical correlations in historical data for modeling. In power equipment operation and robotic inspection scenarios, inspection time, observation angle, equipment location, environmental conditions, data acquisition frequency, or certain occasional sensor fluctuations may be correlated with fault labels, but they are not necessarily the true causes of fault evolution. If the model directly learns these correlations, it is prone to non-causal state inferences and erroneous risk predictions. This invention acquires a causal graph of power faults and performs causal constraint processing on it, generating causal constraint information including causal direction, edge type, time lag, action sign, edge confidence, and interventionist attributes. Then, the causal nodes in the causal graph are aligned with the variable-level latent space state variables in the world model, so that the update of any target latent space state variable in the world model is constrained by its own historical state, the set of causal parent variables, allowed external operating condition variables, and allowed robot action variables. For variable pairs that do not satisfy the causal constraint relationship, information transmission can be suppressed or shielded through the causal constraint matrix using an information transmission control mechanism.

[0081] Therefore, this invention can reduce the interference of irrelevant variables, accidental related variables, and non-causal observation factors on the fault evolution inference results, making the inference process more consistent with the fault mechanism of power equipment and improving the causal consistency of fault prediction and risk assessment. For long-tail faults, complex coupled faults, and extreme operating condition faults, this invention can also use the fault mechanism relationship in the causal graph to assist in the inference, alleviating the problem that a purely data-driven model is prone to instability when there are insufficient fault samples.

[0082] (3) It can support active simulation under external working conditions and robot motion intervention. Existing fault prediction methods are mostly passive, predicting future trends based on historical state sequences or existing observations. They often struggle to answer interventionist questions such as how equipment risk will change if humidity increases, load increases, thunderstorms occur, or pollution levels worsen, or how future observations and risk states will change if a robot performs infrared re-measurement, partial discharge detection, cleaning, tightening, or avoidance actions. This invention incorporates external operating conditions and robot actions as interventionist variables into the world model. Given external operating condition interventions or robot action interventions, the system can perform state setting, state reset, input replacement, causal edge truncation, or edge weight adjustment on the corresponding latent space variables, and propagate the impact downstream to equipment status, fault symptoms, and risk outcomes along the power fault causal graph.

[0083] For example, under high humidity intervention, the system can deduce the continuous impact of increased humidity on the conductivity of contaminants, leakage current, partial discharge, abnormal infrared temperature rise, and risk level; under robot infrared retest intervention, the system can deduce its impact on the infrared observation results and the confidence level of abnormal temperature rise confirmation; under robot cleaning action intervention, the system can deduce its impact on the state of contaminants, leakage current, partial discharge risk, and risk level.

[0084] Therefore, this invention can be extended from the traditional prediction of the future based on historical trends to extrapolation of the future under given working conditions or action assumptions, thereby supporting the active retesting, risk avoidance, handling, and operation strategy comparison of power robots.

[0085] (4) It can jointly output fault evolution, risk changes, future observations and causal explanation paths. Existing fault detection or prediction methods typically only output fault category, risk level, abnormal alarm, or health score, resulting in relatively simple outputs. For power robot operational decision-making, simply knowing the future risk level is insufficient; it is also necessary to determine what observations the robot might obtain after performing a certain action, whether these observations are helpful in confirming the fault, and the main reasons behind the risk changes. This invention simultaneously models the equipment fault latent state, risk state, and robot observation state in a world model. By decoding the future latent space state sequence, it can output the fault evolution sequence, risk state sequence, and multimodal observation prediction results for the power robot.

[0086] Simultaneously, this invention generates a time-ordered causal explanation path for faults based on the activation strength of causal edges, variable contribution, state change amplitude, and temporal sequence during multi-step deduction. For example, under high humidity and contamination conditions, the system can output a causal explanation path of increased ambient humidity → enhanced conductivity of surface contamination → increased leakage current → enhanced partial discharge → abnormal infrared temperature rise → increased risk level.

[0087] Therefore, this invention can not only explain whether the risk will increase, but also explain why the risk increases, along what path it increases, what the robot can observe next, and what actions should be taken, thereby improving the interpretability and feasibility of fault warning, robot operation decision-making, and operation and maintenance.

[0088] In one embodiment, a world model deduction system for power equipment fault evolution based on causal graphs is also provided, for executing a method for deducing a world model for power equipment fault evolution based on causal graphs, including: The data acquisition module is used to acquire multi-source input data of the target power equipment and perform time-aligned preprocessing on the multi-source input data to obtain time-aligned multi-source input data. The causal graph construction module is used to obtain the causal graph of power failure, and to perform normalization and causal constraint processing on the causal graph to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model. The mapping relationship establishment module is used to obtain the world model, which contains a predefined variable-level latent space state variable structure. It aligns the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model and establishes the semantic mapping relationship between the causal nodes and the variable-level latent space state variables. The encoding and mapping module is used to encode the time-aligned multi-source input data to obtain the corresponding encoded latent variables. Based on the correlation between the multi-source input data and causal nodes, the encoded latent variables are assigned to the corresponding causal nodes. The encoded latent variables belonging to the same causal node are aggregated to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, the corresponding variable-level latent space state variables are updated using the node-level latent state observation representation to form the set of variable-level latent space state variables at the current time. Each variable-level latent space state variable corresponds to the semantics of the corresponding causal node. The causal constraint deduction module is used to perform causal constraint state transitions on the set of variable-level latent space state variables at the current moment based on causal constraint information, and to deduce the latent space state sequence at future moments. The results output module is used to obtain the future fault evolution results of the target power equipment based on the potential space state sequence at future time.

[0089] Specific limitations regarding the causal graph-based world model deduction system for power equipment fault evolution can be found in the limitations of the causal graph-based world model deduction method for power equipment fault evolution described above, and will not be repeated here. Each module in the aforementioned causal graph-based world model deduction system for power equipment fault evolution can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0090] The foregoing has provided a detailed description of the method and system for extrapolating the world model of power equipment fault evolution based on causal graphs, as provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention; the descriptions of these embodiments are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A method for extrapolating a world model of power equipment fault evolution based on causal graphs, characterized in that, The method includes the following steps: S100: Acquire multi-source input data of the target power equipment and perform time-aligned preprocessing on the multi-source input data to obtain time-aligned multi-source input data; S200: Obtain the cause-effect graph of power failure, and perform normalization and causal constraint processing on the cause-effect graph to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model. S300: Obtain the world model, which contains a predefined variable-level latent space state variable structure. Align the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model, and establish a semantic mapping relationship between the causal nodes and the variable-level latent space state variables. S400: Encode the time-aligned multi-source input data to obtain the corresponding latent variables after encoding. Based on the correlation between the multi-source input data and causal nodes, assign the latent variables after encoding to the corresponding causal nodes. Aggregate the latent variables after encoding that belong to the same causal node to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, update the corresponding variable-level latent space state variables using the node-level latent state observation representation to form the variable-level latent space state variable set at the current time. Each variable-level latent space state variable corresponds to the corresponding causal node semantics. S500: Based on causal constraint information, the set of variable-level latent space state variables at the current moment is subjected to causal constraint state transition, and the latent space state sequence at the future moment is deduced. Based on the latent space state sequence at the future moment, the future fault evolution result of the target power equipment is obtained.

2. The method according to claim 1, characterized in that, In S100, the multi-source input data includes at least one of the following: power equipment operating status data, power robot multimodal observation data, robot motion data, and external working condition data; Time alignment preprocessing includes: for high-frequency signals, extracting statistical features, frequency domain features, time-frequency features, or event features; for low-frequency data, performing interpolation, preservation, or window aggregation; and for robot multimodal observation data, associating data according to robot pose, timestamp, and device object. The time-aligned first The multi-source input data set at time t is represented as: in, Indicates the first A collection of multi-source input data at any given time. This represents data indicating the operating status of power equipment. This represents multimodal observation data of the robot. Represents robot motion data, This represents external operating condition data.

3. The method according to claim 2, characterized in that, The cause-effect graph of power failure in S200 is represented as follows: in, Representing the cause-and-effect graph of power outages. Represents the set of causal nodes. Represents the set of causal edges. Represents the set of node attributes and edge attributes; Causal nodes include at least one of the following: environmental load nodes, equipment operating status nodes, equipment structural status nodes, fault symptom nodes, risk outcome nodes, robot action nodes, and robot observation nodes. Any causal edge is represented as: in, Indicates that by node Pointing to node The causal edge, Indicates the causal edge type. Indicates a time lag. Symbols indicating positive or negative action. Indicates the edge confidence or effect strength. Indicates whether the current causal edge involves a modifiable variable; The normalization process includes mapping semantically similar causal nodes to uniform variable identifiers and labeling the causal nodes with their types. The causal constraint processing includes: completing and normalizing the attributes of causal edges, labeling the modifiable variables and modifiable edges, and generating causal constraint information based on the normalized causal nodes and causal edges; the causal constraint information includes at least one of the following: variable identifier, variable type, equipment component to which the variable belongs, observation mode corresponding to the variable, set of causal parent variables, set of causal child variables, edge type, time lag, direction of action, sign of action, edge confidence, and modifiable identifier.

4. The method according to claim 3, characterized in that, The S200 power fault causal graph is generated by extracting causal relationships from text: candidate causal event pairs are extracted from power equipment fault reports, maintenance records, defect records, robot inspection reports, standard procedures, and expert experience texts using a causal expression pattern library, and represented as follows: in, This refers to a collection of power equipment fault reports, maintenance records, defect records, robot inspection reports, or standard procedure texts. This represents a causal expression pattern library. Indicates the first One candidate cause event, Indicates the first One candidate result event, This indicates the pattern or confidence level that triggers the current causal relationship. Indicates the number of candidate causal event pairs; Candidate causal edges can be obtained by normalizing, fusing, and verifying candidate causal events. For candidate causal edges, the edge type, direction of action, time lag, action sign, action strength, edge confidence, and intervention attributes are determined by combining expert verification, historical operation data statistics, physical mechanism rules, or simulation results, forming a power fault causal graph for state transition constraints in the world model.

5. The method according to claim 4, characterized in that, The mapping from causal graph nodes to variable-level latent space state variables in S300 is represented as follows: in, This represents the mapping function from causal nodes to latent space state variables in the world model. The first in the cause-effect graph of power failure is represented by the second. A causal node, Represents nodes The corresponding number Continuous observation of data or status data, Represents nodes Aligned Time-space variables and latent space state variables; Each variable-level latent space state variable corresponds to variable description information, which includes at least one of the following: variable identifier, variable type, equipment component to which the variable belongs, observation mode corresponding to the variable, whether the variable can be intervened, the set of causal parent variables of the variable, and the set of causal child variables of the variable. The mapping is used to determine the correspondence between causal nodes and variable-level latent space state variables. The state value of the variable-level latent space state variable at each time step is updated by the latent variable after being encoded by S400 based on the input data of the corresponding node.

6. The method according to claim 5, characterized in that, The S400 includes: S410: For power equipment operating status data, use operating status coding methods such as status feature extraction, time series modeling, or a combination of both to obtain the power equipment operating status coding representation; For multimodal observation data of electric robots, a multimodal perception coding method that integrates images, sensor signals and spatial information is adopted to obtain the multimodal observation coding representation of electric robots; For robot motion data, a motion encoding method that combines discrete motion attribute embedding with continuous motion parameter standardization is adopted, and robot motion latent variables are generated by the motion encoder. For external operating condition data, a working condition coding method is adopted, which combines continuous operating condition quantity standardization, discrete operating condition event embedding, and time window feature fusion. External operating condition latent variables are generated by the working condition encoder. S420: The various encoded representations and latent variables obtained in S410 are unified as encoded latent variables of each data source. The association between multi-source input data and causal nodes is established. The encoded latent variables are assigned to the corresponding causal nodes and aggregated to generate the corresponding node-level latent state observation representation. S430: Based on the above semantic mapping relationship, using the node-level latent state observation representations corresponding to each causal node, fill or update the corresponding variable-level latent space state variables to form the variable-level latent space state variable set at the current moment. Specifically: in, Indicates the first The set of latent space state variables at the variable level that is filled or updated at each moment. Indicates the first The variable in the first... The latent space state variables at time step. , This represents the number of latent space state variables involved in fault evolution simulation.

7. The method according to claim 6, characterized in that, Based on causal constraint information, S500 performs causal constraint state transitions on the current set of variable-level latent space state variables to deduce the latent space state sequence for future times, including: S510: Generate a causal constraint matrix based on the causal constraint information. When there are nodes in the causal constraint information Pointing to node When there is a causal edge, let the matrix elements , representing variables Allow to variable Transmit state information; when there is no corresponding causal edge in the causal constraint information, let the matrix elements... , representing variables Direct participation in variables is not allowed Status update; S520: For any target latent space state variable Only its own historical state and satisfying The set of causal parent variables, allowed external working condition variables, and allowed robot action variables participate in the state update at the next moment; for variable pairs that do not satisfy the causal constraint relationship, during the state transition process, an information transmission control mechanism is used to suppress or shield their information transmission based on the causal constraint matrix. S530: The next-time state sequence of the target latent space state variables is represented as follows: in, Representing variables The corresponding state transition function, This represents the current state of the target variable itself. This represents the set of causal parent variables that are allowed to participate in the update of the target variable's state. Indicates the time lag of causal action. Represents external operating condition variables. Represents robot action variables; state transition function It is implemented using a data-driven time-series state transition model; S540: Traverse all latent variables Aggregate variables The complete set of latent space states for the next time step is obtained, and S520 to S530 are executed iteratively to deduce the sequence of latent space states for multiple future time steps.

8. The method according to claim 7, characterized in that, S540 also includes: under external working condition intervention or robot action intervention, performing intervention operations on the corresponding variables in the world model, and performing multi-step fault evolution deduction based on the causal constraint state transition relationship after intervention to obtain the latent space state sequence for multiple future time steps; The intervention operations include: setting conditions or replacing inputs for future external working condition sequences and robot action sequences; setting or resetting the state of the target variable and truncating the corresponding causal incoming edges for direct state intervention; and adjusting the edge weights for adjusting the causal action mechanism. Intervention operations are represented as follows: in, Represents the variable Perform intervention procedures. This indicates the set intervention state. Representing variables The set of causal parent variables, This indicates cutting off all causal parent variable pairs. Information transmission; Given the future sequence of external operating conditions and the robot's action sequence, a multi-step recursive process is performed based on a causal constrained world model to obtain the future... The sequence of latent space states at each time step: in, Indicating the intervention conditions The first deduction obtained The state of the latent space at all times. Indicates the first under causal constraint Step-world model state transition function, Indicates the external operating condition intervention sequence. Indicates the robot's action intervention sequence. Represents the constraints of the cause-effect graph of power failures.

9. The method according to claim 8, characterized in that, S500 obtains the future fault evolution results of the target power equipment based on the potential space state sequence at future time points, including: Input the latent space state sequence of multiple future time steps into the decoding module, and output the future evolution result of the fault state of the power equipment; The future evolution results include fault evolution sequences, risk state sequences, and multimodal observation and prediction results of the power robot. The fault evolution sequence represents the state changes of the target power equipment over multiple future time steps, including normal state, abnormal symptom state, early fault state, fault development state, severe risk state, or a fault type probability sequence. The risk state sequence represents the risk probability, risk level, health score, fault propagation state, outage risk, or operational risk at each future time. The multimodal observation and prediction results of the power robot represent the observation results obtained by the robot in the future under given robot action conditions, including visible light image prediction results, infrared temperature distribution prediction results, ultraviolet discharge prediction results, acoustic anomaly prediction results, partial discharge signal prediction results, point cloud change prediction results, contact feedback prediction results, or observation confidence prediction results. The results of fault evolution, risk state, and robot multimodal observation prediction are expressed as follows: in, Indicates the first Fault evolution prediction results at time points. Indicates the first Risk state prediction results at any given time Indicates the first Multimodal observation and prediction results of the time-lapse robot , and These represent the fault evolution decoder, the risk state decoder, and the robot observation decoder, respectively.

10. The method according to claim 9, characterized in that, Following the S500 are: S600: Based on the state change magnitude, causal edge activation intensity, variable contribution, and time sequence of each latent space state variable during the multi-step deduction process, a fault causal explanation path is generated as supplementary explanation information for future evolution results. The fault causal explanation path is used to represent the main causal chain that leads to fault evolution and risk changes, including external operating condition variables, equipment operating state variables, equipment structural state variables, fault symptom variables, risk outcome variables, and robot action variables; The cause-and-effect explanation path for the fault is represented as follows: in, Indicates from the first Time to the The set of causal explanation paths for failures at any given moment. Indicates candidate causal paths, Indicates the first Causal edge of time The activation intensity or contribution This indicates that the top scorers will be selected. Path; Also includes: S700: Based on the fault evolution sequence, risk state sequence, and robot multimodal observation and prediction results, compare the risk changes, observation information gain, action cost, and operational safety of different action schemes, and generate suggestions for the inspection, retesting, avoidance, handling, or manual takeover of the power robot.

11. A world model deduction system for power equipment fault evolution based on causal graphs, used to perform the steps of the method as described in any one of claims 1 to 10, characterized in that, include: The data acquisition module is used to acquire multi-source input data of the target power equipment and perform time-aligned preprocessing on the multi-source input data to obtain time-aligned multi-source input data. The causal graph construction module is used to obtain the causal graph of power failure, and to perform normalization and causal constraint processing on the causal graph to generate causal constraint information, which is used to constrain the state transition process of the latent space of the world model. The mapping relationship establishment module is used to obtain the world model, which contains a predefined variable-level latent space state variable structure. It aligns the causal nodes in the power fault causal graph with the variable-level latent space state variables in the world model and establishes the semantic mapping relationship between the causal nodes and the variable-level latent space state variables. The encoding and mapping module is used to encode the time-aligned multi-source input data to obtain the corresponding encoded latent variables. Based on the correlation between the multi-source input data and causal nodes, the encoded latent variables are assigned to the corresponding causal nodes. The encoded latent variables belonging to the same causal node are aggregated to generate the corresponding node-level latent state observation representation. Based on the semantic mapping relationship, the corresponding variable-level latent space state variables are updated using the node-level latent state observation representation to form the variable-level latent space state variable set at the current time. Each variable-level latent space state variable corresponds to the semantics of the corresponding causal node. The causal constraint deduction module is used to perform causal constraint state transitions on the set of variable-level latent space state variables at the current moment based on causal constraint information, and to deduce the latent space state sequence at future moments. The results output module is used to obtain the future fault evolution results of the target power equipment based on the potential space state sequence at future time.