Closed-loop fault diagnosis method and device based on combination of AI intelligent agent and power equipment simulation

Abstract

The invention discloses a closed-loop fault diagnosis method based on the combination of an AI intelligent agent and power equipment simulation, which is applied to the field of power system fault diagnosis, and comprises the following steps: on the basis of a preset knowledge graph basic data set and power system multi-source data, performing data cleaning, feature extraction and knowledge integration; constructing a unified power equipment fault diagnosis knowledge graph, and performing reasoning on the knowledge graph and time sequence characteristics in combination with large model fine tuning or a mixed reasoning engine of a graph neural network to generate M candidate fault hypotheses; calculating the similarity between simulation and actual measurement waveforms through a DTW algorithm and a frequency spectrum comparison method, and screening high-consistency hypotheses; based on the logic verification rule base, performing causal graph reasoning, constraint checking and anti-factual thinking on the high-consistency hypotheses, and eliminating non-logic hypotheses; feeding back the abnormality found in the verification link to the AI agent, dynamically adjusting the reasoning strategy through reinforcement learning, and generating a convergent diagnosis result; and outputting a target diagnosis conclusion based on the converged diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of power system fault diagnosis, and in particular to a closed-loop fault diagnosis method based on the combination of AI intelligent agents and power equipment simulation. Background Technology

[0002] Currently, there are four main technical approaches in the field of power equipment fault diagnosis, each with its own characteristics and limitations: Traditional model-driven diagnosis is based on standards such as IEC 61850 and relies on simulation platforms such as PSCAD / EMTDC and RTDS to reproduce fault waveforms. Although it ensures physical consistency, it has poor adaptability to new faults and relies on manual modeling and debugging; Data-driven open-loop diagnosis uses deep learning models such as CNN and LSTM and ensemble learning methods to analyze historical data. The accuracy increases with the sample size, but the diagnostic logic is difficult to interpret and has limited robustness to out-of-sample operating conditions; Knowledge-driven reasoning achieves rule-based reasoning by constructing a knowledge graph of power equipment. It has strong interpretability, but faces problems such as high knowledge base maintenance costs and reasoning efficiency constrained by rules; Digital twins and human-machine collaboration are development trends. Supported by digital twin platforms, existing products are mostly based on expert systems or simulations and have not achieved fully automatic closed-loop adaptive reasoning.

[0003] Existing technologies generally suffer from the following shortcomings: the diagnostic process is mostly open-loop, lacking a closed-loop mechanism of "hypothesis-verification-checking-feedback", making it difficult to correct errors in a timely manner; the verification mechanism is singular, relying solely on data-driven or simulation verification, lacking introspective verification of causal or physical logic; the degree of automation and intelligence is insufficient, requiring extensive expert intervention and failing to achieve end-to-end unattended real-time diagnosis; the diagnostic results lack a complete chain of evidence and have poor interpretability; and the online learning ability is weak, making it difficult to integrate real feedback into models or rule bases for continuous optimization. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A closed-loop fault diagnosis method based on the combination of AI intelligent agents and power equipment simulation is characterized by the following steps: acquiring multi-source data from the power system, including IED waveform recordings, SCADA historical data, GIS metadata, expert case databases, and a pre-set knowledge graph base dataset, wherein the pre-set knowledge graph base dataset includes structured knowledge of fault rules, equipment parameters, and case information; constructing a unified power equipment fault diagnosis knowledge graph based on the pre-set knowledge graph base dataset and the multi-source data from the power system through data cleaning, feature extraction, and knowledge integration; and using the constructed knowledge graph, combined with a hybrid inference engine of large-scale model fine-tuning or graph neural networks, to perform inference on the knowledge graph and time-series features. The system generates M candidate fault hypotheses, including fault type, location, triggering condition, and confidence level. Each candidate fault hypothesis is parameterized, and the power simulation platform is used to run the simulation, acquiring simulation waveforms. The similarity between the simulation and measured waveforms is calculated using the DTW algorithm and spectrum comparison method to select highly consistent hypotheses. Based on a logic verification rule base, causal graph reasoning, constraint checks, and counterfactual thinking are performed on the highly consistent hypotheses to eliminate illogical hypotheses. Anomalies detected in the verification process are fed back to the AI ​​agent, which dynamically adjusts the reasoning strategy through reinforcement learning to generate convergent diagnostic results. Based on the convergent diagnostic results, an interactive report containing fault causes, key parameters, and evidence chains is generated, outputting the target diagnostic conclusion.

[0005] A closed-loop fault diagnosis device based on the integration of AI agents and power equipment simulation is disclosed. The device includes: an acquisition module for acquiring multi-source data from the power system, including IED waveform recordings, SCADA historical data, GIS metadata, an expert case library, and a pre-set knowledge graph dataset, wherein the pre-set knowledge graph dataset includes structured knowledge of fault rules, equipment parameters, and case information; a processing module for constructing a unified power equipment fault diagnosis knowledge graph based on the pre-set knowledge graph dataset and the multi-source power system data, through data cleaning, feature extraction, and knowledge integration; and utilizing the constructed knowledge graph, combined with a hybrid inference engine of large-scale model fine-tuning or graph neural networks, to process the knowledge graph and time-series features. The algorithm performs inference to generate M candidate fault hypotheses, including fault type, location, triggering condition, and confidence level. Each candidate fault hypothesis is parameterized, and the power simulation platform is used to run the simulation, acquiring simulation waveforms. The similarity between the simulation and measured waveforms is calculated using the DTW algorithm and spectrum comparison method to select highly consistent hypotheses. Based on a logic verification rule base, causal graph reasoning, constraint checks, and counterfactual thinking are performed on the highly consistent hypotheses to eliminate illogical hypotheses. Anomalies detected in the verification process are fed back to the AI ​​agent, which dynamically adjusts the inference strategy through reinforcement learning to generate convergent diagnostic results. Based on the convergent diagnostic results, an interactive report containing fault causes, key parameters, and evidence chains is generated, outputting the target diagnostic conclusion.

[0006] Its beneficial effects are as follows: This invention provides a closed-loop fault diagnosis device based on the combination of AI intelligent agents and power equipment simulation. By adopting a triple verification mechanism of "hybrid reasoning + simulation verification + logic verification", multiple rounds of iterative verification can significantly reduce the risk of misdiagnosis and missed diagnosis. After generating candidate fault hypotheses through hybrid reasoning, highly consistent hypotheses are screened through simulation verification (such as DTW algorithm and spectrum comparison). Then, illogical hypotheses are further eliminated through causal graph reasoning, constraint checking, and counterfactual thinking to ensure the accuracy of the diagnostic results.

[0007] The dual verification mechanism (parallel "simulation physical verification" and "logical causal verification") improves the robustness of diagnosis, reduces the bias that may exist in a single verification method, increases diagnostic accuracy by 20%, and significantly reduces the false alarm rate. It enhances automation and intelligence, achieving end-to-end automated processes. From data acquisition (IED waveform recording, SCADA historical data, etc.), knowledge graph construction, and candidate fault hypothesis generation, to simulation verification, logical verification, closed-loop optimization, and report generation, it eliminates the need for extensive manual intervention by experts, enabling unattended real-time diagnosis and significantly improving efficiency by over 50%.

[0008] The closed-loop optimization layer dynamically adjusts the inference strategy (such as rule weights and model hyperparameters) through reinforcement learning and supports online updates of the knowledge graph (new devices / procedures are automatically extracted into entities and relationships through an NLP pipeline), enabling the system to have long-term online learning capabilities, continuously adapt to new fault conditions, and achieve sustained improvement in long-term accuracy. The system supports multiple simulation platforms and models, can quickly adapt to different power grid topologies and equipment models, and has strong robustness and transferability. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the application of a closed-loop fault diagnosis device based on the combination of AI intelligent agent and power equipment simulation in cloud service equipment, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a closed-loop fault diagnosis device based on the combination of AI intelligent agent and power equipment simulation, provided as an embodiment of the present invention. Detailed Implementation

[0010] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Figure 1 This application describes a conference live streaming system and its implementation method according to exemplary embodiments. In one embodiment, this application also proposes a closed-loop fault diagnosis device based on the combination of AI intelligent agents and power equipment simulation.

[0011] In this application embodiment, a closed-loop fault diagnosis method based on the combination of AI intelligent agent and power equipment simulation is described, such as... Figure 1 As shown: S101, acquire multi-source data of the power system.

[0012] In one implementation, when acquiring multi-source data from a power system, the first step is to collect IED waveform data. This data records waveform changes during power equipment faults. For example, when a short-circuit fault occurs on a line, the IED waveform recorder will record detailed current and voltage waveform data before and after the fault, including information such as sampling rate, channel ID, and timestamp, and conforms to the COMTRADE standard. Next is SCADA historical data, which covers various real-time monitoring data during normal operation and faults of the power system, such as bus voltage and line current at different substations at different times. For example, it may include data such as voltage fluctuation curves and current change trends of a substation over the past 24 hours.

[0013] GIS metadata is also an important component, containing spatial information such as the geographical location and topology of power equipment. This includes the installation locations of equipment like transformers and circuit breakers, as well as their interconnections, such as the specific geographical coordinates and route of two substations connected by a transmission line. The expert case database stores past cases of power equipment fault handling, including fault types, causes, handling methods, and results. For example, a case of CT saturation fault was documented in detail, including waveform characteristics (peak flat-top), diagnostic process, and solutions.

[0014] The pre-defined fault rules in the foundational dataset of the knowledge graph clarify the correspondence between different fault types and features. For example, arcing faults are usually accompanied by a waveform feature with high-frequency noise components >300Hz. The equipment parameters include information such as the rated voltage and current of various power equipment, such as the rated voltage of a certain type of transformer being 110kV and the rated current being 500A. The case information is a structured compilation of historical fault cases, facilitating the construction and reasoning of the knowledge graph. By collecting this multi-source data, a comprehensive and solid data foundation is laid for subsequent power equipment fault diagnosis work.

[0015] S102, based on a pre-set knowledge graph dataset and multi-source data from the power system, constructs a unified knowledge graph for power equipment fault diagnosis through data cleaning, feature extraction, and knowledge integration.

[0016] In one implementation, a pre-defined knowledge graph dataset and multi-source power system data are cleaned to remove noise, fill in missing values, and correct outliers, generating standardized raw data. The pre-defined knowledge graph dataset contains fault rules, equipment parameters (e.g., a transformer with a rated voltage of 110kV), historical cases (e.g., CT saturation fault records from 2023), and multi-source power system data, including IED waveforms (including short-circuit fault waveforms) and SCADA real-time voltage and current values ​​(e.g., a 10kV line current of 300A). Specifically, the following steps are performed: noise removal: removing spike pulse waveforms caused by sensor interference from IED waveforms; missing value filling: using interpolation to fill in 5 minutes of blank current data in the SCADA system; outlier correction: correcting disconnector equipment types incorrectly labeled as "circuit breaker" in the GIS metadata. Finally, a raw dataset containing standardized waveform data, equipment parameters, and case information is generated.

[0017] Feature extraction is performed on standardized raw data to extract temporal and static features, generating structured feature data. Temporal features include waveform amplitude, harmonic content, and envelope characteristics, while static features include equipment topology, operating procedures, and fault type. Specifically, the temporal features extracted from IED waveform recording data are directly correlated with the dynamic variation patterns of fault waveforms. Waveform amplitude (peak value 2000A): Based on the COMTRADE standard waveform recording file, the maximum current waveform value during a short-circuit fault on a 10kV line is extracted using a peak detection algorithm. This value needs to match the "typical amplitude range of short-circuit faults" node (1500A-2500A) in the knowledge graph as a quantitative indicator of fault severity. Harmonic content (3rd harmonic percentage 5%): The fault waveform is analyzed using FFT transformation to calculate the ratio of the 3rd harmonic component to the fundamental component, which meets the technical requirement in Article 1 that "spectral characteristics are used to distinguish fault types." For example, this indicator can help eliminate CT saturation faults (CT saturation is usually accompanied by a surge in higher harmonics). Envelope characteristics (maximum amplitude occurrence time 14:23:05): The waveform envelope is obtained through Hilbert transform, and the timestamp corresponding to the maximum amplitude is located. It is compared with the fault alarm time of the SCADA system (14:23:04) to verify the consistency of data time and provide a time dimension basis for subsequent causal inference.

[0018] Static features extracted from GIS metadata and knowledge graph data reflect the inherent attributes and rules of the equipment. Equipment topology (110kV busbar connected to 3 outgoing lines): Based on GIS topology data, the connection relationship between the busbar and outgoing lines is analyzed, constructing a "busbar-line" topology graph node to provide spatial constraints for fault location (e.g., faults can only occur within the range of lines connected to the busbar); Operating procedures (circuit breaker operation sequence): The operating rule of "disconnecting the line circuit breaker first, then the busbar circuit breaker" is extracted from structured procedure documents and used as a constraint condition for the logic verification layer to ensure that fault assumptions comply with maintenance specifications; Fault type (single-phase ground fault): Combining the typical feature descriptions of "single-phase ground fault" in the expert case library, it is defined as a fault type node in the knowledge graph and associated with feature tags such as "increased zero-sequence current" and "decreased phase voltage," providing static knowledge support for hybrid reasoning.

[0019] Using JSONSchema to associate and map time-series features with static features, for example: {"device_id":"Line102","timestamp":"2024-05-20T14:23:05", "temporal_features":{"amplitude_peak":2000,"harmonic_ratio_3rd":0.05,"envelope_max_time":"14:23:05"}, "static_features":{"topology":"Bus110→Line102", The structured data, defined as "operation_rule":"BreakerSequence_001","fault_type_label":"SinglePhaseGround"}}, establishes a link between dynamic waveform features and static device knowledge, providing a unified feature input format for the AI ​​agent's inference layer.

[0020] Integrate the structured feature data with the preset knowledge graph base data, map the fault rules, device parameters, and case information into the nodes and relationships of the knowledge graph, and generate an initial knowledge graph. Integrate the structured feature data with the preset knowledge graph base data. Node mapping: Define "110 kV transformer", "single-phase ground short circuit", and "peak value 2000 A" as the nodes of the knowledge graph; Relationship definition: Establish the association relationships of "fault type - waveform feature" (such as "single-phase ground → sudden increase in amplitude") and "device - parameter" (such as "transformer - rated voltage 110 kV"); Case integration: Store the causal pair of "CT saturation → flat-topped waveform" in the historical cases as a triple into the graph. Generate an initial knowledge graph containing 120 nodes and 300 relationships.

[0021] Based on the initial knowledge graph, automatically extract the entities and relationships of new devices and regulations through the NLP pipeline, call the KGAPI to insert nodes and edges, update the historical case triples, and generate a dynamically extended unified power equipment fault diagnosis knowledge graph. When the system accesses the newly commissioned "220 kV GIS equipment", automate the knowledge integration through the NLP pipeline: Use the named entity recognition (NER) model to parse the device regulation document, and extract the core entities "GIS circuit breaker" and "220 kV busbar" from the text "The 220 kV GIS equipment includes 3 groups of GIS circuit breakers, which are respectively connected to phases A, B, and C of the 220 kV busbar", and label the entity types (equipment type, topology type).

[0022] Identify the "connected to" relationship between the "GIS circuit breaker" and the "220 kV busbar" through the relationship extraction model, generate the relationship triple <GIS circuit breaker, connected to, 220 kV busbar>; Call the knowledge graph API (KGAPI) to perform the node insertion operation, assign a unique device ID (such as Dev-GIS-001) to the "GIS circuit breaker", and associate the "220 kV busbar" with the existing topology node; At the same time, insert the "connected to" relationship edge and supplement the attribute information (such as the connection method is "rigid busbar connection") to ensure that the new device topology relationship is seamlessly connected with the existing graph.

[0023] For newly occurring "arc fault" cases, updates are performed according to the specification of "case information mapped to triples". The key feature "high-frequency noise > 300Hz" is extracted from the fault recording and matched with the causal rule "arc fault → high-frequency noise > 300Hz" in the logic verification rule base to confirm the fault type label. The fault event is structured and mapped to the core triple <Event ID: E202405, Device: Bus22, Fault Type: Arc Fault>, supplementing attribute information (such as the fault occurrence timestamp "2024-05-10T09:15:30" and the feature parameter "high-frequency noise peak 350Hz"). The triple is inserted into the "case node set" of the knowledge graph via KGAPI, and a "fault occurrence" relationship edge is established between the "Bus22" device node and the "arc fault" type node. Simultaneously, the case base index is updated to ensure that the latest case can be called upon for subsequent inference.

[0024] By integrating newly added device nodes and case triples, the knowledge graph achieves dynamic expansion. Node scale expansion: 8 new device-type nodes (including GIS circuit breakers, busbar branches, etc.) and 12 relationship edges (including connection relationships, membership relationships, etc.) are added; Rule adaptation and update: The constraint condition of "220kV busbar fault impact range" in the inference rules is automatically adjusted based on the new device topology; Case timeliness assurance: The latest arc flash fault cases and historical cases form a feature comparison library, providing more comprehensive case support for subsequent inference of similar faults. The entire process requires no manual intervention, achieving real-time dynamic expansion of the knowledge graph.

[0025] S103 utilizes the constructed knowledge graph, combined with a hybrid inference engine of large model fine-tuning or graph neural network, to infer the knowledge graph and temporal features, generating M candidate fault hypotheses.

[0026] In one implementation, the constructed knowledge graph and time-series feature data are fused together, and the relationships between knowledge graph nodes and time-series feature parameters are associated to generate hybrid inference input data. When generating the hybrid inference input data, it is necessary to extract equipment topology relationships and fault rule node association information from the knowledge graph, such as static knowledge like "the connection relationship between 110kV bus 12 and outgoing lines Line 3 and Line 4" and "the causal mapping between short-circuit faults and current surge characteristics." Simultaneously, it is associated with time-series feature data, including dynamic parameters such as the current amplitude of 2000A during a Line 3 fault, the 3rd harmonic percentage of 5%, and the time of maximum amplitude occurrence of 14:23:05. The node IDs, relationship types, and time-series feature parameters of the knowledge graph are bound together using JSON format to form hybrid inference input data containing "equipment-feature-relationship" triples, ensuring that the AI ​​agent can simultaneously call upon static knowledge and dynamic features for inference.

[0027] Based on hybrid reasoning input data, parallel reasoning is performed through rule-based reasoning, case-based reasoning, and a fine-tuned large-scale power model, according to a dynamic weighted formula. The weights of each inference method are calculated to generate multi-source inference results. Based on the hybrid inference input data, a multi-source parallel inference mechanism is initiated: the rule inference module outputs preliminary conclusions according to preset fault rules (e.g., "current amplitude exceeds 1.5 times the rated value → short circuit fault"); the case inference module retrieves historical single-phase grounding fault cases from the knowledge graph (e.g., the 2023 Line 5 fault case, characterized by a sudden current increase of 2000A±500A) and calculates the feature matching degree; the fine-tuned large power model (LLM) performs semantic understanding of the input topological relationships and waveform features, and outputs natural language inference results. The results are then calculated using a dynamic weighting formula. The weights of each reasoning method are calculated, with the accuracy of rule reasoning Acc1=0.8, case reasoning Acc2=0.9, and LLM reasoning Acc3=0.95. After calculation, the weights are updated to rule reasoning 0.25, case reasoning 0.3, and LLM reasoning 0.45, and finally, a multi-source reasoning result containing the three types of reasoning conclusions is generated.

[0028] The multi-source inference results are fused and analyzed. A graph neural network (GCN) is used to deeply mine the topological relationships and temporal features of the knowledge graph, filtering out candidate results with consistent fault type, location, and triggering conditions to generate a preliminary fault hypothesis set. During the fusion analysis of the multi-source inference results, the GCN first performs feature learning on the knowledge graph topology, inputting 128-dimensional node features (including equipment type and rated parameters). After two layers of GCN processing, topological association weights are output, strengthening the connection weights between Bus12 and Line3. Simultaneously, temporal convolution is performed on the temporal features to extract the abrupt change time and amplitude change trend of the fault waveform. By comparing the fault type (all pointing to single-phase grounding or short circuit), fault location (concentrated near Line3 or Bus12), and triggering conditions (load abrupt change period) of the three types of inference results, candidate results with consistent conclusions are filtered out, such as "Line3 single-phase grounding fault, triggered by load abrupt change condition" and "Bus12 insulation breakdown, accompanied by voltage drop," forming a preliminary fault hypothesis set.

[0029] The initial set of fault hypotheses is evaluated for confidence. Hypotheses whose confidence levels meet the threshold requirements are retained, generating M candidate fault hypotheses, each including fault type, location, triggering condition, and confidence level. The confidence evaluation metrics include inference path completeness (e.g., whether rule-based inference covers all protection action logic) and feature matching degree (e.g., waveform features with cosine similarity to the case library ≥ 0.85). A confidence threshold of 0.7 is set, eliminating low-confidence hypotheses such as "Line 4 open circuit fault (confidence 0.62, due to insufficient feature matching degree)". Finally, hypotheses meeting the threshold requirements are retained, generating M candidate fault hypotheses, such as: "H1: Line 3 single-phase ground fault, location approximately 1.5km from Bus 12, triggering condition 14:23 load surge, confidence 0.85; H2: Bus 12 busbar insulation breakdown, fault phase A, triggering condition long-term high-load operation, confidence 0.72", providing clear candidate directions for subsequent simulation verification.

[0030] S104 parameterizes each candidate fault hypothesis, drives the power simulation platform to run the simulation, obtains the simulation waveform, calculates the similarity between the simulation and the measured waveform using the DTW algorithm and the spectrum comparison method, and selects the hypothesis with high consistency.

[0031] In one implementation, each candidate fault hypothesis is parameterized, extracting key parameters including fault location and impedance, and encapsulating them into simulation input parameters using a unified JSON structure. Specifically, each candidate fault hypothesis is parameterized, extracting fault location (e.g., "connection between Bus12 and Line3") and impedance parameters (e.g., resistance r = 0.1Ω, reactance x = 0.5Ω), and encapsulating them into simulation input parameters using a unified JSON structure. An example is shown below: {"hypotheses":[{"id":"H1","type":"SLG","parameters": {"location":"Bus12-Line3","impedance":{"r":0.1,"x":0.5}}},{"id":"H2","type":"Three-phase short circuit","parameters":{"location":"Line3 middle section","impedance":{"r":0.05,"x":0.3}}}]}.

[0032] The parameterized simulation input parameters drive the power simulation platform to run the simulation, automatically retrieving equipment models and network topology data to generate simulation waveform data for the corresponding fault scenario. The parameterized simulation input parameters are then passed to the power simulation platform (such as PSCAD / RTDS), which automatically retrieves the corresponding equipment models (such as the conductor model of Line 3 and the busbar model of Bus 12) and network topology data (such as the connection relationship between Line 3 and Bus 12), runs the simulation, and generates waveform data for the fault scenario. For example, the simulation waveform for H1 will exhibit the characteristics of a single-phase ground fault: a sudden increase in current in the faulty phase (peak value 2000A) and an increase in voltage in the non-faulty phases.

[0033] Similarity calculations were performed on simulated and measured waveform data. Temporal similarity was calculated using the DTW algorithm, and frequency domain similarity was calculated through spectral comparison. For temporal similarity calculation, the DTW algorithm was used with a window size of 10 sampling points and Euclidean distance as the metric. The calculated DTW value for H1 was 8.5 (lower values ​​indicate greater similarity). Frequency domain similarity was calculated through spectral comparison with an FFT resolution of 0.5 Hz. The calculated spectral cosine similarity for H1 was 0.92 (≥0.9 threshold).

[0034] Based on temporal similarity and frequency domain similarity, according to the comprehensive scoring formula The hypotheses are scored, and those with a score ≥ 0.8 are selected to generate a set of highly consistent failure hypotheses. =20, the score of H1 is (0.6×(1-8.5 / 20)+0.4×0.92=0.6×0.575+0.368=0.345+0.368=0.713) (slightly lower than 0.8, requiring further verification); the calculated score of H2 is 0.85, which meets the threshold requirement. Finally, H2 with a score ≥0.8 is selected to generate a highly consistent fault hypothesis set.

[0035] S105, based on the logic verification rule base, performs causal graph reasoning, constraint checking, and counterfactual thinking on highly consistent assumptions to eliminate illogical assumptions.

[0036] In one implementation, high consistency assumptions and corresponding simulation waveforms and equipment parameter information are extracted, fault types are associated with waveform features and operational constraints, and logic verification input data is generated. The high consistency assumptions (such as "H2: Line 3 three-phase short circuit fault") and associated information are systematically extracted and associated. Specifically, this includes fault type labels, explicitly marked as "three-phase short circuit," and associated with the "three-phase short circuit" node in the knowledge graph; simulation waveform features, extracting current amplitude (sudden increase to 3000A), harmonic components (second harmonic accounting for 8%), and timestamp (14:23:06), and binding them to the corresponding features of the measured waveform; equipment parameters, including the rated voltage (110kV), real-time monitoring voltage (0.9pu), and the topology node to which Line 3 belongs (connected to Bus 12), all derived from the equipment parameter nodes in the knowledge graph; and operational constraints, based on the topology of Bus 12 in the GIS metadata, determining that it must maintain connectivity with the three outgoing lines Line 3, Line 4, and Line 5.

[0037] The above information is structured and linked using JSON format, as shown in the following example: {"hypothesis_id":"H2","fault_type":"three-phase short circuit", "waveform_features":{"current_amplitude":3000,"harmonic_2nd":0.08,"timestamp":"14:23:06"},"device_params":{"Line3_voltage_rated":110,"Line3_voltage_real":0.9,"topology":"Bus12→Line3"}, The generated logical verification input data, defined by "constraints":{"Bus12_connectivity":["Line3","Line4","Line5"]}}, can be directly called by the causal reasoning module.

[0038] Based on the logic verification input data, the causal chain completeness of fault type and waveform characteristics is checked through the causal graph inference rule set, generating causal verification results. Based on the logic verification input data, a deep verification is performed by calling the preset causal graph inference rule set: the causal mapping rule "three-phase short circuit → current surge + increased proportion of low-order harmonics" is retrieved from the logic verification rule library (derived from the feature summarization of 100+ similar faults in the expert case library); it is checked whether the "current surge of 3000A" in the simulation waveform is consistent with the physical cause of the three-phase short circuit (direct phase-to-phase conduction leading to a sudden drop in impedance), confirming that there is no contradiction.

[0039] Verify whether it contains all the typical characteristics of a three-phase short circuit (such as no high-frequency noise and similar current amplitudes in each phase). The waveform characteristics of H2 are all satisfied, with no missing items. The final causality verification result is generated: "The fault type and waveform characteristics of H2 form a complete causal chain, which conforms to physical laws and passes the causality verification."

[0040] The assumptions that pass the causal verification are subjected to constraint checks to verify whether the equipment operating parameters meet the safety thresholds and topological constraints, generating constraint verification results. The safety thresholds are voltage ≤ 1.2 pu and current ≤ 1.5 pu. For H2, which passes the causal verification, a double check is performed based on the equipment constraint rules in the knowledge graph: Equipment operating parameter threshold verification: Referring to the "Safety Operating Thresholds for Power Equipment" node in the knowledge graph, Line3's real-time voltage 0.9 pu ≤ 1.2 pu (voltage threshold) and real-time current 1.3 pu ≤ 1.5 pu (current threshold) are both within the safety range. Topology constraint verification: Based on the post-fault topology simulation results, the connectivity matrix of Bus12 and Line3 shows that no electrical islands have been formed (connectivity index ≥ 0.95), which meets the topology constraint of "Bus12 needs to maintain multiple outgoing line connections" in the GIS metadata. The constraint verification result is generated: "The equipment operating parameters and topology of H2 meet the constraint conditions and pass the constraint verification."

[0041] Based on the assumption that the constraint verification has passed, we engage in counterfactual thinking, following formal logic. Verify the exclusivity of the hypothesis and generate a counterfactual verification result. Construct a counterfactual proposition: if H2 "three-phase short circuit" is true, then the waveform should not exhibit "high-frequency noise > 300Hz" (a typical characteristic of arc faults, derived from the causal rule base). Perform spectral analysis on the simulated waveform (FFT resolution 0.5Hz). The results show that the energy proportion of the high-frequency band (>300Hz) is <0.01, with no high-frequency noise components. The characteristics of the simulated waveform are consistent with the conclusion of the counterfactual proposition, and no contradictory features are found. The generated counterfactual verification result is: "H2 passes counterfactual verification, exclusivity is established, and no other fault type characteristic interference was found."

[0042] Based on the combined results of causal verification, constraint verification, and counterfactual verification, hypotheses with logical contradictions or those that do not meet constraints are eliminated, generating a logically consistent set of fault hypotheses. The final screening is performed by combining the results of the three verifications, as follows: H2 passed causal verification (complete causal chain), constraint verification (parameters and topology are compliant), and counterfactual verification (no contradictory features). Compared with other hypotheses: H1 (single-phase grounding) was eliminated due to failure in counterfactual verification (arc fault characteristics appeared), and H3 (CT saturation) was eliminated due to a broken causal chain (no flat-top waveform). The hypothesis set is finalized: only H2 passed all verifications, with no logical contradictions or constraint violations. The final logically consistent set of fault hypotheses is generated: {H2: Line 3 three-phase short-circuit fault, confidence level 0.85}, providing reliable input for subsequent closed-loop optimization.

[0043] S106, the anomalies found in the verification process are fed back to the AI ​​agent, and the reasoning strategy is dynamically adjusted through reinforcement learning to generate convergent diagnostic results.

[0044] In one implementation, abnormal information detected in the logic verification process is extracted, associated with corresponding fault hypotheses, simulation scores, and verification flags, and feedback input data is generated. For example, the hypothesis "H1: Line 3 single-phase ground fault" is removed due to counterfactual verification failure (high-frequency noise >300Hz, contradicting the characteristics of a single-phase ground fault); H1 has a simulation score of 0.713 (below the 0.8 threshold) and a verification flag of "counterfactual not established." H2 passes all verifications, with a simulation score of 0.85 and a verification flag of "passed." The above information is integrated into feedback input data, as shown in the example below: {"abnormal_hypotheses":[{"id":"H1","simulation_score":0.713,"logic_flag":"The counterfactual does not hold"}], "valid_hypotheses":[{"id":"H2","simulation_score":0.85,"logic_flag":"pass"}]}.

[0045] Based on the feedback input data, a reinforcement learning state space S = {number of candidate hypotheses, simulation score, logic verification flag} is constructed, and an action space A = {increase rule priority, decrease rule priority, retrain LLM} is defined to generate reinforcement learning parameters. Specifically, the state space S: {number of candidate hypotheses = 2, simulation score = [0.713, 0.85], logic flags = ["counterfact not true", "pass"]}; the action space A: {increase rule priority (e.g., "arc fault high-frequency feature verification rule"), decrease rule priority (e.g., "single-phase grounding harmonic threshold rule"), retrain LLM}; the reward function parameter uses R = 0.5 × AccuracyReward + 0.3 × SpeedReward − 0.2 × FalseAlarmPenalty}, where H2's AccuracyReward = 1 (correct) and H1's FalseAlarmPenalty = 1 (false positive). These reinforcement learning parameters provide a basis for policy adjustment.

[0046] Using reinforcement learning parameters, based on the reward function R=0.5×AccuracyReward+0.3×SpeedReward−0.2×FalseAlarmPenalty, for the cases of H1 misjudgment (failed counterfactual verification) and H2 passing verification: AccuracyReward=1 (H2 correctly identified), SpeedReward=1 (inference time 2.3 seconds, meeting the ≤3-second threshold), FalseAlarmPenalty=1 (H1 misjudged); the calculated reward value R=0.5×1+0.3×1-0.2×1=0.6, which reflects that there is room for optimization in the current policy.

[0047] Based on the reward signal, the inference rules and model parameters are dynamically updated. The weight of the "single-phase grounding harmonic threshold rule" is reduced from 0.3 to 0.2 (due to the misjudgment in H1 being related to the leniency of this rule), while the weight of the "high-frequency noise counterfactual verification rule" is increased from 0.2 to 0.3 (to strengthen the exclusiveness verification). The Large Power Model (LLM) is marked as needing retraining because it failed to recognize the contradiction between "high-frequency noise > 300Hz" and single-phase grounding in H1. The training data is supplemented with 100+ feature comparison samples of arc faults and single-phase grounding. The strategy optimization results are generated, including a rule weight update table (e.g., "Rule ID: R002, old weight 0.3 → new weight 0.2") and an optimization scheme for the LLM retraining parameters (learning rate 5e-5, Epoch=50).

[0048] Based on the policy optimization results, the AI ​​agent reconstructs the inference logic, regenerates candidate fault hypotheses, and iteratively executes multi-source information fusion, simulation verification, and logic verification steps until the diagnostic results converge, generating the final converged fault diagnosis result. The candidate hypotheses are regenerated, and combined with the updated rule weights, the risk of misjudgment in class H1 is eliminated. A new hypothesis, "H3: Line 3 arc fault" (matching high-frequency noise features), is added, expanding the candidate hypothesis set to {H2, H3}.

[0049] The verification process is as follows: Multi-source information fusion: H3 is associated with the rule "arc fault → high-frequency noise > 300Hz" in the knowledge graph, and the high-frequency component of its simulation waveform accounts for 12% of the time-series features; Simulation verification: H3's simulation score is calculated to be 0.79 (DTW value 9.2, spectral similarity 0.88), and it enters the logic verification because it is <0.8 threshold; Logic verification: H3's causal chain is broken (arc faults should have continuous arc characteristics, but the simulation waveform has no continuous high-frequency components), and it is removed. Convergence judgment: In the new round of verification, H2's causal verification integrity score increased from 0.8 to 0.85 due to rule weight optimization, and the simulation score stabilized at 0.85. With no other valid assumptions, it meets the convergence condition of "consistent results in two consecutive iterations".

[0050] Generate the final result and output the converged diagnostic result: {H2: Line 3 three-phase short circuit fault, confidence level 0.88 (0.03 higher than the initial level)}, and record the strategy optimization log (e.g., "Adjustment of rule R002 reduced the false positive rate by 15%").

[0051] S107, based on the converged diagnostic results, generates an interactive report containing the cause of the fault, key parameters, and evidence chain, and outputs the target diagnostic conclusion.

[0052] In one implementation, information is extracted from the converged diagnostic results, including the final fault cause, key simulation parameters, inference path, and verification nodes. The simulation waveform comparison chart and logic verification log are correlated to generate the original report data. A comprehensive information extraction is performed on the converged diagnostic result (H2: Line 3 three-phase short-circuit fault). Specifically, the final fault cause is clearly a three-phase short circuit in the middle section of Line 3, triggered by a load surge at 14:23 (consistent with the load fluctuation timestamp of 14:23:04 recorded by the SCADA system). Key simulation parameters, including the fault location (approximately 2km from Bus 12), impedance parameters (r=0.05Ω, x=0.3Ω), and simulation score (0.85, calculated based on the comprehensive scoring formula), are extracted from the original output of the simulation verification layer.

[0053] The inference path meticulously records the weight allocation of the hybrid inference stage (rule-based inference 0.25, case-based inference 0.3, LLM inference 0.45), core simulation verification metrics (DTW=7.2, spectral similarity 0.91), and the pass status of logic verification (complete causal chain, satisfied constraints, counterfactual validity), and is associated with the strategy adjustment log of the closed-loop optimization layer. Verification nodes include causal verification pass flags (matching the "three-phase short circuit → current surge" rule), constraint parameters (voltage 0.9pu≤1.2pu, current 1.3pu≤1.5pu), and counterfactual verification logs (high-frequency noise <300Hz, conforming to exclusive logic). Associated files include simulation and measured waveform comparison graphs (peak difference <5%) and a complete logic verification log (including detailed reasons why H1 was removed due to counterfactual contradictions). All information is stored in JSON format for key parameters, and the storage paths of the associated waveform graphs and log files are linked to ensure traceability.

[0054] Based on the original data in the report, the fault diagnosis conclusions, key parameters, and evidence chain elements are integrated according to a preset template to generate the structured content of the report. Fault diagnosis conclusions: The fault type, specific location (middle section), and triggering cause (load change) of the three-phase short circuit in Line 3 are clearly identified. The origin of the confidence level of 0.88 is explained—the initial confidence level of 0.85 was improved after optimization by the reinforcement learning strategy (rule weight adjustment), which corresponds to the reward function calculation result of the closed-loop optimization layer.

[0055] Key parameters are summarized below: Simulation parameters: fault location, impedance value, DTW value, spectral similarity, and comprehensive score; Equipment parameters: Line3 rated voltage 110kV, real-time voltage 0.9pu, and real-time current 1.3pu; Verification indicators: causal verification pass status, constraint threshold compliance status, and counterfactual logic verification results; Evidence chain elements: The entire process of "hybrid reasoning → simulation verification → logic verification" is sorted out in chronological order, and the key criteria for each step are marked (such as simulation score ≥ 0.8, constraint parameters within the safety threshold), and associated with the rule nodes in the knowledge graph.

[0056] Briefly explain why H1 (single-phase ground fault) was removed due to failure of counterfactual verification. Its simulation waveform contains high-frequency noise >300Hz, contradicting the characteristics of a single-phase ground fault, and conforms to the formal expression of counterfactual logic. The structured content adopts a chapter-based layout to ensure the logical coherence of the fault cause, parameters, and evidence chain.

[0057] The structured content is visualized to generate a comparison chart of simulated and measured waveforms, a diagram of the inference path, and a logic verification node map. An interactive viewing interface is provided to generate an interactive draft report. The comparison chart of simulated and measured waveforms adopts an overlay display mode, and the peak points, DTW values ​​(7.2), and similarity curves are marked according to the requirements of Article 1. It supports zooming to view waveform details before and after 14:23:05 (corresponding to the time when the maximum amplitude value appears).

[0058] The flowchart illustrates the flow of "hybrid reasoning → simulation verification → logic verification," using different colors to indicate the weight of each step (e.g., dark blue for LLM reasoning, indicating 0.45) and its pass / fail status (green for pass), corresponding to the parallel reasoning architecture of the AI ​​agent's reasoning layer. The knowledge graph presents the relationships between "fault type - feature - constraint," such as "three-phase short circuit → current surge (3000A) → voltage ≤ 1.2 pu," with node size reflecting the strength of the association, matching the causal graph reasoning rule set of the logic verification layer.

[0059] Buttons such as "View Raw Data," "Download Waveform File," and "Backtrack Reasoning Process" have been added, supporting navigation to the original records of each stage (e.g., clicking "Simulation Verification" allows viewing the original output log of the PSCAD simulation). The initial draft report is in web format, compatible with both desktop and mobile devices.

[0060] The initial draft of the interactive report undergoes a completeness check, necessary explanatory information is added, and a final interactive report containing complete diagnostic evidence is generated, outputting the target fault diagnosis conclusion. Specifically, the following supplementary explanations are provided: a new "Impact of Strategy Optimization" section is added, explaining the specific reason for the confidence level increasing from 0.85 to 0.88—the weight of the "high-frequency noise counterfactual verification rule" in reinforcement learning increases from 0.2 to 0.3, improving the verification strictness, consistent with the rule weight update table of the closed-loop optimization layer.

[0061] Outlier values ​​are marked as follows: In the waveform graph, indicate the reason for the 10% deviation between the measured value and the simulated value at a certain moment—sensor sampling delay, and indicate the impact of this deviation on the comprehensive score (calculated to not affect the final result); The compliance check is as follows: Confirm that all parameter units (Ω, pu, Hz) and time format (ISO8601) comply with industry standards (such as the COMTRADE format's requirements for timestamps).

[0062] The final output is an interactive report containing complete diagnostic data, showing the target diagnostic conclusion: "Three-phase short circuit fault in the middle section of Line 3. It is recommended to immediately isolate the faulty line and inspect the conductor insulation." The report allows maintenance personnel to trace the original data and verification logic of any link through the interactive interface.

[0063] like Figure 2As shown, a closed-loop fault diagnosis device based on the combination of AI intelligent agents and power equipment simulation includes: an acquisition module for acquiring multi-source data of the power system, including IED waveform recordings, SCADA historical data, GIS metadata, expert case databases, and a preset knowledge graph basic dataset, wherein the preset knowledge graph basic dataset includes structured knowledge of fault rules, equipment parameters, and case information; a processing module for constructing a unified power equipment fault diagnosis knowledge graph based on the preset knowledge graph basic dataset and the multi-source data of the power system through data cleaning, feature extraction, and knowledge integration; and utilizing the constructed knowledge graph, combined with a hybrid inference engine of large model fine-tuning or graph neural networks, to process the knowledge graph and time series data. The system uses features to infer M candidate fault hypotheses, including fault type, location, triggering condition, and confidence level. Each candidate fault hypothesis is parameterized, and the power simulation platform is used to run the simulation, acquiring simulation waveforms. The similarity between the simulation and measured waveforms is calculated using the DTW algorithm and spectrum comparison method to filter highly consistent hypotheses. Based on a logic verification rule base, causal graph reasoning, constraint checks, and counterfactual reasoning are performed on the highly consistent hypotheses to eliminate illogical hypotheses. Anomalies detected in the verification process are fed back to the AI ​​agent, which dynamically adjusts the inference strategy through reinforcement learning to generate convergent diagnostic results. Based on the convergent diagnostic results, an interactive report containing fault causes, key parameters, and evidence chains is generated, outputting the target diagnostic conclusion.

[0064] A computing device includes a memory for storing computer program instructions and a processor for executing the computer program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute any one of the closed-loop fault diagnosis devices based on the combination of AI intelligent agent and power equipment simulation.

[0065] The methods and / or embodiments in this application can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processing unit, it performs the functions defined in the methods of this application.

[0066] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0067] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0068] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within this application.

Claims

1. A closed-loop fault diagnosis method based on the combination of AI intelligent agent and power equipment simulation, characterized in that, include: Acquire multi-source data from the power system, including IED waveform recording, SCADA historical data, GIS metadata, expert case database, and a pre-set knowledge graph base dataset. The pre-set knowledge graph base dataset includes structured knowledge of fault rules, equipment parameters, and case information. Based on a pre-defined knowledge graph dataset and multi-source data from the power system, a unified knowledge graph for power equipment fault diagnosis is constructed through data cleaning, feature extraction, and knowledge integration. By utilizing the constructed knowledge graph and combining it with a hybrid inference engine of large model fine-tuning or graph neural network, reasoning is performed on the knowledge graph and temporal features to generate M candidate fault hypotheses, including fault type, location, triggering condition and confidence level. Each candidate fault hypothesis is parameterized, and the power simulation platform is driven to run the simulation to obtain the simulation waveform. The similarity between the simulation and the measured waveform is calculated by the DTW algorithm and the spectrum comparison method, and the hypothesis with high consistency is selected. Based on the logic verification rule base, cause-effect graph reasoning, constraint checking and counterfactual thinking are performed on high consistency assumptions to eliminate illogical assumptions. Anomalies detected during the verification process are fed back to the AI ​​agent, which dynamically adjusts the inference strategy through reinforcement learning to generate convergent diagnostic results. Based on the converged diagnostic results, an interactive report containing the cause of the fault, key parameters, and evidence chain is generated, and the target diagnostic conclusion is output.

2. The method according to claim 1, characterized in that, Based on a pre-defined knowledge graph dataset and multi-source power system data, a unified knowledge graph for power equipment fault diagnosis is constructed through data cleaning, feature extraction, and knowledge integration, including: The pre-defined knowledge graph base dataset and multi-source data of the power system are cleaned to remove noise, fill in missing values, correct outliers, and generate standardized raw data. Feature extraction is performed on standardized raw data to extract time-series and static features, generating structured feature data. The time-series features include waveform amplitude, harmonic content, and envelope features, while the static features include equipment topology, operating procedures, and fault types. The structured feature data and the pre-set knowledge graph base data are integrated to map fault rules, equipment parameters and case information into nodes and relationships of the knowledge graph, thereby generating an initial knowledge graph; Based on the initial knowledge graph, the entities and relationships of newly added equipment and procedures are automatically extracted through the NLP pipeline, nodes and edges are inserted by calling KGAPI, historical case triples are updated, and a dynamically expanded unified knowledge graph for power equipment fault diagnosis is generated.

3. The method according to claim 2, characterized in that, Using the constructed knowledge graph, combined with a hybrid inference engine of large model fine-tuning or graph neural networks, reasoning is performed on the knowledge graph and temporal features to generate M candidate fault hypotheses, including fault type, location, triggering condition, and confidence level, including: The constructed knowledge graph and temporal feature data are fused together, and the relationships between knowledge graph nodes and temporal feature parameters are associated to generate hybrid reasoning input data. Based on hybrid reasoning input data, parallel reasoning is performed through rule-based reasoning, case-based reasoning, and a fine-tuned large-scale power model, according to a dynamic weighted formula. Calculate the weights of each reasoning method and generate multi-source reasoning results; The results of multi-source reasoning are fused and analyzed. Combined with graph neural networks for in-depth mining of knowledge graph topological relationships and temporal features, candidate results with consistent fault type, location and triggering conditions are selected to generate a preliminary fault hypothesis set. The initial set of fault hypotheses is evaluated for confidence, and the hypotheses whose confidence meets the threshold requirements are retained. M candidate fault hypotheses are generated, which include fault type, location, triggering condition and confidence level.

4. The method according to claim 1, characterized in that, Each candidate fault hypothesis is parameterized, and the power simulation platform is driven to run the simulation to obtain simulation waveforms. The similarity between the simulation and measured waveforms is calculated using the DTW algorithm and spectrum comparison method to screen high consistency hypotheses, including: Each candidate fault hypothesis is parameterized, and key parameters including fault location and impedance are extracted and encapsulated into simulation input parameters in a unified JSON structure. The parameterized simulation input parameters drive the power simulation platform to run the simulation, automatically retrieve equipment models and grid data, and generate simulation waveform data for the corresponding fault scenarios. Similarity calculations are performed between simulated waveform data and measured waveform data. The temporal similarity is calculated using the DTW algorithm, and the frequency domain similarity is calculated by spectrum comparison. Based on temporal similarity and frequency domain similarity, according to the comprehensive scoring formula The hypotheses are scored, and those with a score ≥ 0.8 are selected to generate a set of highly consistent failure hypotheses.

5. The method according to claim 1, characterized in that, Based on a logic verification rule base, causal graph reasoning, constraint checks, and counterfactual analysis are performed on high consistency assumptions to eliminate illogical assumptions, including: Extract the high consistency assumptions and corresponding simulation waveforms and equipment parameter information, associate fault types with waveform characteristics and operating constraints, and generate logic verification input data; Based on the logical verification input data, the causal chain completeness of fault type and waveform characteristics is checked through the causal graph inference rule set, and causal verification results are generated. Constraint checks are performed on the assumptions that pass the causality verification to verify whether the equipment operating parameters meet the safety thresholds and topology constraints, and constraint verification results are generated. The safety thresholds are voltage ≤ 1.2 pu and current ≤ 1.5 pu. Based on the assumption that the constraint verification has passed, we conduct counterfactual thinking, verify the exclusivity of the assumption using formal logic, and generate counterfactual verification results. By combining the results of causal verification, constraint verification, and counterfactual verification, hypotheses that contain logical contradictions or do not meet constraints are eliminated, and a set of logically consistent fault hypotheses is generated.

6. The method according to claim 1, characterized in that, Anomalies detected during the verification process are fed back to the AI ​​agent, which dynamically adjusts the inference strategy through reinforcement learning to generate convergent diagnostic results, including: Extract abnormal information found in the logic verification process, associate it with corresponding fault assumptions, simulation scores and verification flags, and generate feedback input data; Based on the feedback input data, construct the reinforcement learning state space S={number of candidate hypotheses, simulation score, logic verification flag}, define the action space A={increase rule priority, decrease rule priority, retrain LLM}, and generate reinforcement learning parameters. Using reinforcement learning parameters, according to the reward function The reward for calculating the strategy is adjusted, the inference rule weights and model hyperparameters are dynamically updated, and the strategy optimization results are generated. Based on the policy optimization results, the AI ​​agent reconstructs the reasoning logic, regenerates candidate fault hypotheses, and iteratively executes multi-source information fusion, simulation verification, and logic verification steps until the diagnostic results converge, generating the final converged fault diagnosis result.

7. The method according to claim 6, characterized in that, Based on the converged diagnostic results, an interactive report is generated that includes the cause of the fault, key parameters, and the chain of evidence, outputting the target diagnostic conclusion, including: Information is extracted from the converged diagnostic results, including the final cause of the failure, key simulation parameters, inference path and verification nodes, and the simulation waveform comparison chart and logic verification log are correlated to generate the original report data; Based on the original data in the report, the fault diagnosis conclusions, key parameters and evidence chain elements are integrated according to the preset template to generate the structured content of the report; The structured content is visualized to generate a comparison chart of simulation and measured waveforms, a diagram of inference paths, and a graph of logic verification nodes. An interactive viewing interface is then linked to generate an interactive draft report. The initial draft of the interactive report is validated for completeness, necessary explanatory information is added, and finally an interactive report containing complete diagnostic evidence is generated, outputting the target fault diagnosis conclusion.

8. A closed-loop fault diagnosis device based on the combination of AI intelligent agent and power equipment simulation, characterized in that, The device includes: The acquisition module is used to acquire multi-source data of the power system, including IED waveform recording, SCADA historical data, GIS metadata, expert case library and preset knowledge graph basic dataset. The preset knowledge graph basic dataset includes structured knowledge of fault rules, equipment parameters and case information. The processing module is used to construct a unified knowledge graph for power equipment fault diagnosis based on a pre-set knowledge graph dataset and multi-source data from the power system. This is achieved through data cleaning, feature extraction, and knowledge integration. Using this knowledge graph, a hybrid inference engine combining large-scale model fine-tuning or graph neural networks is employed to infer the knowledge graph and temporal features, generating M candidate fault hypotheses, each including fault type, location, triggering condition, and confidence level. Each candidate fault hypothesis is parameterized, driving the power simulation platform to run simulations and acquire simulation waveforms. The similarity between the simulation and measured waveforms is calculated using the DTW algorithm and spectrum comparison method to filter highly consistent hypotheses. Based on a logic verification rule base, causal graph reasoning, constraint checks, and counterfactual reasoning are performed on the highly consistent hypotheses to eliminate illogical hypotheses. Anomalies detected during the verification process are fed back to the AI ​​agent, which dynamically adjusts the inference strategy through reinforcement learning to generate convergent diagnostic results. Based on the convergent diagnostic results, an interactive report containing fault causes, key parameters, and evidence chains is generated, outputting the target diagnostic conclusion.

9. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the closed-loop fault diagnosis device based on the combination of AI intelligent agent and power equipment simulation as described in any one of claims 1 to 7 by executing the executable instructions.

10. A computing device, the device comprising a memory for storing computer program instructions and a processor for executing the computer program instructions, wherein, When the computer program instructions are executed by the processor, the device is triggered to execute the closed-loop fault diagnosis device based on the combination of AI intelligent agent and power equipment simulation as described in any one of claims 1 to 7.

Images