Controlled LLM guided fuzzy bayesian network based power equipment fault diagnosis method
By using a controlled LLM-guided fuzzy Bayesian network, combined with multi-source data and a security constraint library, the problems of fuzzy uncertainty and security constraints in power equipment fault diagnosis are solved, achieving efficient, interpretable, and secure fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIONTO ENERGY INVESTMENT CO LTD HEBEI
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies struggle to handle fuzzy uncertainties, lack interpretability, and fail to provide proactive diagnostic planning capabilities under safety constraints in power equipment fault diagnosis, leading to false alarms, missed alarms, and unenforceable diagnostic recommendations.
A controlled LLM-guided fuzzy Bayesian network is adopted. By accessing multi-source data, a knowledge base and a security constraint base are constructed to generate a fuzzy Bayesian network for evidence reasoning and verification action planning. Combined with a large language model, a structured diagnostic report is generated, and manual confirmation and model updates are performed.
It achieves robustness to sensor noise and operating condition disturbances, improves the accuracy and interpretability of diagnosis, ensures the safety and executability of diagnostic recommendations, reduces invalid troubleshooting steps, and enhances the system's adaptive learning capability.
Smart Images
Figure CN122241432A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for power production equipment, and in particular to a method for fault diagnosis of power equipment using a controlled LLM-guided fuzzy Bayesian network. Background Technology
[0002] As the core rotating equipment in thermal power plants, the vibration state of steam turbine units directly affects the safe, stable operation and economic benefits of the unit. In actual operating conditions, steam turbine vibration faults exhibit significant strong coupling and nonlinear characteristics: a single symptom of excessive vibration amplitude may be caused by multiple root causes such as rotor mass imbalance, shaft misalignment, oil film oscillation, dynamic-static rubbing, or thermal bending; conversely, the same fault root cause may also exhibit differentiated spectral characteristics under different loads or oil temperatures. This complex mapping relationship between multiple symptoms and multiple faults presents a significant challenge to fault diagnosis.
[0003] Existing technologies have the following limitations in solving the above problems: Limitations of deterministic methods based on thresholds and rules. Traditional monitoring systems often employ fixed threshold alarms or linear scoring logic. However, industrial field data is frequently accompanied by sensor noise, and fault characteristics often fall within the threshold gray area. This rigid logic struggles to handle fuzzy uncertainties, easily leading to false alarms or missed alarms. Furthermore, expert systems built on human experience are costly to maintain, prone to rule conflicts when cross-subsystem coupling is involved, and lack the ability to generalize to new concurrent faults.
[0004] The black-box nature and sample dependence of purely data-driven methods. While data-driven methods such as deep learning perform excellently in single fault classification, they heavily rely on massive amounts of labeled samples, while high-value fault samples from large equipment such as steam turbines are extremely scarce. More importantly, end-to-end neural networks lack interpretability and cannot provide a chain of evidence consistent with physical mechanisms, making it difficult for maintenance personnel to verify and trust diagnostic conclusions.
[0005] The lack of knowledge integration and proactive planning capabilities is a significant issue. Existing diagnostic models are mostly passive monitoring, analyzing only historical data. However, complex fault diagnosis often requires proactive execution of specific verification actions to obtain crucial identification information. Current technologies lack an information-theoretic-based dynamic programming mechanism to generate the optimal minimal verification sequence. Furthermore, while Large Language Models (LLMs) possess powerful text knowledge understanding capabilities, they are susceptible to factual errors or the illusion of generating unreliable content, and struggle to directly generate logical structures that satisfy strict probabilistic constraints. Moreover, existing diagnostic systems often lack interaction with the hard safety logic of the distributed control system (DCS) when outputting recommendations, potentially leading to recommendations that violate unit safety interlock rules and lack on-site executability.
[0006] In summary, there is an urgent need for a hybrid intelligent diagnostic method that can integrate unstructured text knowledge with real-time sensor data, explicitly handle uncertainty, and proactively plan the optimal diagnostic path under safety constraints. Summary of the Invention
[0007] The purpose of this invention is to provide a power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network, which solves the problems of existing technologies, such as the inability of hard threshold determination to handle fuzzy uncertainty, the lack of interpretability and sample dependence in pure data-driven methods, and the lack of proactive diagnosis planning capabilities under security constraints.
[0008] To achieve the above objectives, this invention provides a method for fault diagnosis of power equipment using a controlled LLM-guided fuzzy Bayesian network, comprising the following steps: S1. Access multi-source data and build a knowledge base and a security constraint base. The knowledge base includes a measurement point dictionary, a root cause dictionary, and an action dictionary. S2. Analyze multi-source data through a large language model to generate a list of nodes, a list of edges, candidate prior probabilities, and candidate conditional probability tables for the candidate fuzzy Bayesian network, and perform structural verification to obtain the fuzzy Bayesian network. S3. Map the continuous measurement point observation values corresponding to the symptom node to the fuzzy membership degree of each state through the preset membership degree function, normalize the fuzzy membership degree to obtain the soft evidence probability, correct the soft evidence probability through the evidence reliability weight to form the corrected soft evidence, and combine the corrected soft evidence of all symptom nodes into an evidence set. S4. Input the evidence set into the fuzzy Bayesian network to perform posterior inference, obtain the posterior probability distribution and confidence interval of each root cause node, sort the posterior probabilities, and output the Top-K root causes with the highest posterior probabilities as the candidate set. S5. Generate a set of candidate verification actions based on the action dictionary, define result variables and value space for each candidate verification action, calculate the value information of each candidate verification action, and calculate the comprehensive utility value in combination with cost, risk and time consumption. S6. Based on the security constraint library, perform hard constraint filtering on the candidate verification action set, output the executable action set, sort the executable action set in descending order according to the comprehensive utility value, and select a specific number of executable action sets to form a priority verification action set. S7. Calculate the contribution of each piece of evidence to the root cause distribution based on the evidence set, and construct the evidence chain based on the contribution. Organize the Top-K root causes and their confidence intervals, the evidence chain, the priority check action set and the reasons for the filtered actions into a structured diagnostic result. Provide the structured diagnostic result as a fact constraint to the big language model, and the big language model generates a natural language diagnostic report and maintenance suggestions within the scope of the fact constraint. S8. Feed back the manually confirmed results to update the fuzzy Bayesian network parameters and action dictionary.
[0009] Furthermore, in S2, the candidate fuzzy Bayesian network is represented as a directed acyclic graph. , where the set of nodes , For the set of root cause nodes, For the set of symptom nodes, It is a set of directed edges.
[0010] Furthermore, in S2, each node output by the large language model is mapped to a dictionary entity in the knowledge base, and unmapped candidate nodes are replaced or removed.
[0011] Furthermore, in S2, the structural verification includes acyclicity verification and causal edge validity verification; wherein the acyclicity verification ensures the graph... Given a directed acyclic graph, the validity check of causal edges restricts unreasonable pre-defined connection relationships; When validation fails, structural corrections are performed, including deleting edges that cause cycles, merging duplicate nodes, and replacing non-standard names with dictionary-standard names.
[0012] Further, in step S3, the continuous measurement point observations corresponding to the symptom nodes are mapped to fuzzy membership degrees of each state through a preset membership function. The fuzzy membership degrees are normalized to obtain the soft evidence probability. The soft evidence probability is then corrected using evidence reliability weights to form the corrected soft evidence, which includes the following steps: S301. A trapezoidal membership function is used to map the continuous observations of the symptom nodes to the membership degrees of each state, where the states include high state, medium state and low state; the threshold parameter of the trapezoidal membership function is given by the measurement point dictionary and configured in segments according to the load segment or speed segment. S302. Normalize the membership vector to obtain the soft evidence probability, calculated using the following formula: ; in, As a symptom node In state The probability of soft evidence; For state The corresponding membership value; For state The corresponding membership value, As a symptom node Observed values; S303, Introducing a weighting for the reliability of evidence The formula for correcting soft evidence is as follows: ; in, The corrected probability of soft evidence; As a symptom node Weight of evidence reliability As a symptom node All states, As a symptom node The total number of states.
[0013] Furthermore, in step S4, the step of inputting the evidence set into a fuzzy Bayesian network to perform posterior inference and obtain the posterior probability distribution and confidence interval of each root cause node includes: S401. Calculate the posterior probability of each root cause node through Bayesian inference. S402. Apply Beta distribution to binary root cause events. Perform parameter uncertainty modeling, where and Define the shape parameters of the Beta distribution and output the corresponding parameters. Confidence interval The significance level is indicated by .
[0014] Furthermore, in step S5, calculating the value information of each candidate verification action includes the following steps: S501. Calculate the posterior entropy of the root cause. The calculation formula is as follows: ; in, In the current set of evidence Posterior entropy of the next root cause; Root cause node The posterior probability; The total number of root cause nodes; S502, Calculate candidate verification actions The value information is calculated using the following formula: ; in, Candidate verification action Valuable information; In the current set of evidence The value of the next action result variable is The predicted probability; To observe the results of the action The posterior entropy of the root cause; Candidate verification action The value space of .
[0015] Furthermore, in step S6, the formula for calculating the comprehensive utility value is as follows: ; in, Candidate verification action Execution costs; Candidate verification action Risk level; Candidate verification action The estimated time; These are the weighting coefficients for the first, second, third, and fourth comprehensive utility values. Candidate verification action The overall utility value.
[0016] Further, in S6, the method for hard constraint filtering of the candidate verification action set based on the security constraint library is as follows: the constraint judgment result is 1 when the candidate verification action satisfies all constraints in the security constraint library, and the constraint judgment result is 0 when the candidate verification action violates any constraint in the security constraint library. The set of executable actions is defined as follows: ;in, A set of executable actions; For the set of candidate verification actions; For candidate verification actions; Candidate verification action In context The constraint determination result.
[0017] Furthermore, in step S7, the step of calculating the contribution of each piece of evidence to the root cause distribution based on the evidence set, and constructing an evidence chain based on the contribution, includes: S701. Calculate the contribution of each piece of evidence to the root cause distribution based on the evidence set. The calculation formula is as follows: ; in, For the first The contribution of each piece of evidence; Let KL divergence be a metric. For evidence set The posterior distribution of the root causes; To remove the first Root cause posterior distribution after 1 piece of evidence; To obtain evidence Remove the first A subset following one piece of evidence; S702. Evidence with a contribution rate higher than the preset evidence screening threshold and whose corresponding symptoms can be mapped to the standard measurement point name or standard fault symptom term in the measurement point dictionary is organized into an evidence chain according to time sequence and causal logic.
[0018] Therefore, the power equipment fault diagnosis method using the above-mentioned controlled LLM-guided fuzzy Bayesian network has the following beneficial effects: (1) By mapping the continuous measurement point observations to fuzzy membership soft evidence and introducing the evidence reliability weight mechanism, the problem of misjudgment in the fuzzy transition region of traditional hard threshold judgment is effectively solved, the robustness to sensor noise and operating condition disturbance is enhanced, and the diagnostic system can still maintain a stable diagnostic accuracy under low signal-to-noise ratio conditions. (2) Using fuzzy Bayesian network as the verifiable probabilistic reasoning kernel, all diagnostic conclusions can be traced back to specific test point evidence, conditional probability dependencies and posterior reasoning paths. The large language model only organizes language under the factual constraints of the structured reasoning results and does not participate in the probabilistic reasoning process, thus achieving strong interpretability and auditability. (3) Based on the value information-driven verification planning mechanism, the verification action that can reduce the root cause uncertainty is selected first in each step of diagnosis, and the comprehensive utility evaluation is carried out in combination with cost, risk and time consumption, which effectively reduces the invalid screening steps and achieves efficient diagnosis with fewer steps. (4) Introduce a safety constraint library to perform hard constraint filtering on the candidate action set, covering constraint types such as protection interlocks, regulations and tickets and working condition boundaries, to ensure that all diagnostic suggestions output by the system meet the on-site safety system requirements, and avoid the problem of the intelligent diagnostic system giving inoperable or dangerous suggestions; (5) By updating the fuzzy Bayesian network parameters and action dictionary parameters by manually confirming and updating the maintenance conclusions, the diagnostic system can achieve continuous adaptive learning, enabling the system to gradually adapt to the operating characteristics of specific units and continuously improve diagnostic accuracy and planning quality in long-term operation.
[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0020] Figure 1 This is a flowchart of the power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to the present invention. Figure 2 This is a comparison chart of the average number of diagnostic steps for different diagnostic methods of the present invention; Figure 3 This is a comparison chart of the diagnostic accuracy of various diagnostic methods under different noise levels according to the present invention; Figure 4 This is a comparison chart of the safety constraint verification results of different diagnostic strategies of the present invention; Figure 5 This is a comparison chart showing the convergence effect of the VOI-based verification sequence of the present invention on the root cause posterior uncertainty. Detailed Implementation
[0021] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0022] It should be noted that the main technical terms involved in this invention are defined as follows: Value of Information (VOI) refers to the expected reduction in decision uncertainty after performing an information acquisition action; Conditional Probability Table (CPT) refers to a data structure in a Bayesian network that describes the conditional probability distribution of a child node given the state of its parent node; Soft evidence refers to the fact that in Bayesian network inference, the observed node is not definitively in a certain state, but rather represents the degree to which it belongs to each state in the form of a probability distribution; Fuzzy transition region refers to the uncertain region where the observed value of a measurement point is between the normal value and the alarm threshold, and it is difficult to make a deterministic state determination in this region using a hard threshold; Feedback refers to feeding back the manually confirmed information of the diagnostic results and the maintenance conclusions to the system to update the model parameters for continuous learning.
[0023] This embodiment addresses the vibration scenario of a steam turbine in a thermal power unit. Please refer to [link / reference]. Figure 1 A method for fault diagnosis of power equipment using controlled LLM-guided fuzzy Bayesian networks includes the following steps: S1. Integrate multi-source data to build a knowledge base and safety constraint library for turbine vibration diagnosis; Step S1 is used to solidify the multi-source data and operation and maintenance knowledge involved in turbine vibration diagnosis into configurable, indexable, and executable engineering assets, providing a unified input for subsequent large language model candidate generation, fuzzy Bayesian network inference, and value information planning. Specifically, it includes the following steps: S101, Data Sources and Standardization; It integrates vibration monitoring data, distributed control system operating condition data, oil system data, and event and alarm data, and simultaneously integrates text documents such as procedures, work orders, defect reports, and accident bulletins. Using equipment tag numbers as the primary key, it establishes a mapping relationship between measurement points, components, operating conditions, events, and work orders, and normalizes measurement point aliases, abbreviations, and synonyms to form standardized entities that can be consistently invoked for enhanced retrieval and reasoning.
[0024] S102. Construct a measurement point dictionary, a root cause dictionary, and an action dictionary; The measurement point dictionary defines measurement point types, locations, units, alarm thresholds, effective ranges, sampling rates, and computable feature sets. It also allows configuration of load segment thresholds and initial values for evidence reliability. The root cause dictionary defines mechanistic summaries of typical root causes, prior probability ranges, distinguishable symptom combinations, and recommended treatment points. The action dictionary defines observable symptoms, action outcome variables, costs, time consumption, risk levels, preconditions, and access requirements for executable verification actions, enabling value information planning to possess a computable action model.
[0025] S103. Construct a security constraint library.
[0026] In embodiments of this invention, a safety constraint library is used to solidify real-world power plant constraints as hard rules, covering the protection / interlocking layer, the procedure / ticket layer, and the operating condition boundary layer. The prohibited actions library must explicitly list action types that the system must not recommend under any circumstances, including protection activation / deactivation, interlocking bypass, unauthorized crossing of critical speeds, rapid adjustment of large valves under over-limit conditions, and disassembly / removal or inspection under load. For actions requiring approval, the safety constraint library filters based on access permissions and ticket status as triggering conditions.
[0027] S2. Based on the enhanced retrieval context, the large language model is used to parse the text to generate candidate fuzzy Bayesian network structures and parameter candidates, and the structure is verified. The goal of step S2 is to enable the large language model to generate candidate structures and parameters for a fuzzy Bayesian network under evidence-based conditions, and to obtain a reasonable fuzzy Bayesian network through structure verification. Specifically, this includes the following steps: S201, Enhanced Search; By using procedures and work order fragments that are manually determined to be related to the current alarm, component, and similar historical cases as context input to the large language model, the risk of illusion and cross-domain confusion can be reduced.
[0028] S202, Formal definition of candidate fuzzy Bayesian network structure; The fuzzy Bayesian network can be represented as a directed acyclic graph: ; in, For the graph structure of fuzzy Bayesian networks; A set of nodes; It is a set of directed edges.
[0029] Node set From the root cause node set and symptom node set composition: ; ; ; in, ( ) is the first One root cause node, The total number of root cause nodes; ( ) is the first A symptom node, The total number of symptom nodes; directed edges usually point from root cause nodes to symptom nodes, representing the causal relationship between root cause and symptom.
[0030] For any node Its conditional probability is given by CPT (conditional probability table or soft evidence probability obtained from membership mapping): ;in, For nodes The set of parent nodes; For nodes The conditional probability distribution given the state of its parent node.
[0031] S203, Output constraints of large language models; In this invention, the large language model is defined as a generator of candidate structures and parameter candidates. Its output adopts a structured format, including a list of nodes, a list of edges, root cause prior probability candidates, and CPT candidates or probability ranges. It can also include a confidence label for each edge to support subsequent linear mappings of strong and weak associations. Each node output by the large language model must be able to be mapped to the dictionary entity constructed in step S1; otherwise, it is considered an unsuitable candidate and triggers replacement or elimination.
[0032] S204, Structural verification.
[0033] Structural verification includes at least the following two checks: First, acyclicity verification, ensuring the diagram... First, it is a directed acyclic graph, meaning there is no path from any node back to itself via a directed edge. Second, the legality of causal edges is checked to restrict unreasonable connections, including that an action node cannot be the parent node of the root cause, that no reverse edge pointing to the root cause can be introduced without permission, and that cross-system coupling edges must meet the whitelist rules.
[0034] When the verification fails, a structural correction strategy is executed, including deleting edges that cause loops, merging duplicate nodes, and replacing non-standard names with dictionary standard names, until a fuzzy Bayesian network and its parameter candidates that satisfy the inference conditions are output.
[0035] S3. Map the continuous measurement point observations into fuzzy membership degree evidence to form the soft evidence input of the fuzzy Bayesian network; Step S3 aims to transform the continuous measurement points of turbine vibration into probability distributions belonging to various states, inputting them into a fuzzy Bayesian network as soft evidence to avoid misjudgments and jumps caused by hard thresholding. Specifically, it includes the following steps: S301. Define the membership function; With symptom nodes Continuous observations For example, As a symptom node The observed values, whose state set is presupposed to be ,in As a symptom node The total number of states. A trapezoidal membership function is used to describe the degree to which an observation belongs to a certain state. In embodiments of this invention, states include: high states, medium states, and low states. (The last sentence appears to be incomplete and possibly refers to a different concept.) Taking the membership function as an example, the expression is as follows: ; in, For observations Belongs to state Membership degree; The four threshold parameters of the trapezoidal membership function satisfy... The parameters are given by the measurement point dictionary and can be configured in segments according to load range or speed range. For symptom nodes... For each of the states, a corresponding membership function is constructed to cover its complete state space.
[0036] S302, Convert membership degree into soft evidence probability; Normalizing the membership vector yields the soft evidence probability vector: ; in, As a symptom node In state The probability of soft evidence; For state The corresponding membership value; For state The corresponding membership value, As a symptom node All states, As a symptom node The summation of all state membership degrees is used for normalization to make the sum of soft evidence probabilities equal to 1.
[0037] S303, Evidence reliability weighting.
[0038] To reflect sensor health and data integrity, an evidence reliability weight is introduced. Degrading soft evidence to a uniform distribution to enhance robustness: ; in, The corrected probability of soft evidence; As a symptom node Weighting of the reliability of evidence; As a symptom node The total number of states; This is based on no prior information. The threshold is lowered when measurement points are suspected of drifting or data is missing. This causes soft evidence to degenerate into a uniform distribution, thereby reducing the impact of erroneous evidence on the reasoning outcome; when When you completely trust the observation data, At that point, the evidence degenerates into a state of no information.
[0039] S4. Perform posterior inference based on fuzzy Bayesian networks to obtain the posterior probability distribution and confidence interval of the root cause; specifically including the following steps: S401, Posterior reasoning; The modified soft evidence from all the symptom nodes is combined into an evidence set. Input fuzzy Bayesian network, root cause nodes The posterior probability is calculated using Bayesian inference: ; in, Root cause node In a given set of evidence The posterior probability is given below; Let be a normalization constant, such that the sum of the posterior probabilities of all root cause nodes is 1; Indicates the exclusion Sum the states of all nodes outside the boundary (marginalization); For nodes The conditional probability; As a symptom node The soft evidence likelihood term is derived from the modified soft evidence probability. The corresponding state probabilities are constructed. The reasoning process can employ node trees or belief propagation to meet the timeliness requirements of online diagnosis.
[0040] S402, Top-K root cause output; Based on posterior probability All root cause nodes are sorted in descending order, and the Top-K root causes with the highest posterior probabilities are output as the candidate set. The posterior probability values of each root cause are retained to support subsequent evidence chain interpretation and diagnostic planning iteration.
[0041] S403, Confidence Interval Estimation.
[0042] To avoid overly definitive diagnostic conclusions, parameter uncertainty is modeled, preferably using a Dirichlet / Beta prior and obtaining intervals through approximation or sampling. For binary root cause events, a Beta distribution is used for approximation: ; in, and The shape parameter of the Beta distribution is updated by historical samples and refilled samples.
[0043] Correspondingly The confidence interval is: ; in, For confidence level of The confidence interval; is the quantile function of the Beta distribution; The significance level is indicated. By using interval-based output, it can suggest the need for further verification in scenarios where the posterior probabilities of multiple root causes are similar, rather than drawing a conclusion out of thin air.
[0044] S5. Construct a set of candidate verification actions, select the next verification action based on value information, and introduce a trade-off between cost, risk, and time consumption; specifically including the following steps: S501, Candidate Action Set and Outcome Variables; From the action dictionary constructed in step S1, a candidate set of verification actions is generated based on the current Top-K root cause set, the device location corresponding to the alarm object, and the applicable root cause, applicable device location, preconditions, permission requirements, and result variable definitions for each action item. And for each candidate verification action ( Define its result variable and value space This is used to describe the possible observations obtained after an action is performed, thus making the "action-information gain" computable. This represents the total number of candidate verification actions.
[0045] S502, Calculate the posterior entropy of the root cause; The level of uncertainty in the current diagnosis is measured by the root cause posterior entropy: ; in, In the current set of evidence The posterior entropy of the next root cause; Root cause node The posterior probability; This represents the total number of root cause nodes. A higher posterior entropy indicates that it is more difficult to distinguish the root cause, and the higher the diagnostic uncertainty.
[0046] S503, Calculate value information; Candidate verification action The value information is defined as the expected decrease in root cause uncertainty before and after performing the action: ; in, Candidate verification action Valuable information; In the current set of evidence The value of the next action result variable is The predicted probability; To observe the results of the action The posterior entropy of the root cause; Candidate verification action The result variable's value space. This definition enables the system to prioritize the checking action that best distinguishes candidate root causes in turbine vibration problems with similar symptoms and multiple root causes, thereby reducing invalid troubleshooting steps.
[0047] S504. Construct the action utility function and calculate the overall utility value.
[0048] Unify value information and engineering costs into a utility value: ; in, Candidate verification action The overall utility value; Candidate verification action Valuable information; Candidate verification action Execution costs; Candidate verification action Risk level; Candidate verification action The estimated time; These are configurable weighting coefficients for the first, second, third, and fourth comprehensive utility values. Adjusting these weighting coefficients can automatically increase the priority of low-risk actions in risk-sensitive situations. (Based on utility value) All candidate verification actions are sorted in descending order.
[0049] S6. Based on the security constraint library, perform hard constraint filtering on the candidate verification action set and output the executable action set; specifically, this includes the following steps: S601, Hard Constraint Filtering; Let the current working condition and permission context be... Define constraint decision functions : ; in, Candidate verification action In context The constraint determination result is as follows; This includes information such as the current operating status of the unit, the operator's authority level, the status of ticket approval, and the equipment operating status.
[0050] The set of executable actions is defined as follows: ; in, A set of executable actions; This is a set of candidate verification actions. The system in... Sort the data in descending order of overall utility value, select a specific number of executable action combinations with the highest overall utility value to generate a priority check action set, and ensure that the output priority check action set is executable on site.
[0051] S602, Constraint Type.
[0052] The constraint determination also covers the following situations: (1) Direct filtering of actions that match the action library must be prohibited; (2) Controlled adjustment actions that exceed the safety boundary of the working condition are filtered out when the preconditions are not met; (3) Actions requiring approval should be filtered when permissions and documents do not meet the requirements; (4) When the equipment is in a specific protection state, related operations are prohibited.
[0053] For candidate verification actions that are filtered out due to unmet constraints, the system records the specific reasons for their filtering for reference in subsequent reports.
[0054] S7, Top-K root causes, evidence chains, and priority check sets are used by a large language model to generate a natural language diagnostic report and troubleshooting recommendations; specifically, the following steps are included: S701, Evidence Chain Construction and Contribution Measurement; To avoid outputting only uninterpretable probability values, the contribution of each piece of evidence to the root cause distribution is calculated. The difference in the root cause posterior distribution before and after removing a single piece of evidence is used as a measure: ; in, For the first The contribution of each piece of evidence; KL divergence is used to measure the degree of difference between two probability distributions. For evidence set The posterior distribution of the root causes; To remove the first Root cause posterior distribution after 1 piece of evidence; To obtain evidence Remove the first A subset following each piece of evidence. The larger the value, the more crucial the evidence is to the current diagnostic conclusion. Let the evidence screening threshold be... When the first The evidence satisfies If the symptoms corresponding to the evidence can be mapped to the standard measurement point name or standard fault symptom term in the measurement point dictionary, the evidence is determined to be key evidence; then, the key evidence is organized into an evidence chain according to the time sequence and causal logic, and its corresponding measurement point and soft evidence strength are marked.
[0055] S702, Structured Results Organization; The inference and planning outputs are organized into structured diagnostic results, which include at least the Top-K root causes and their confidence intervals, key evidence chains and their source measurement point mappings, the set of priority verification actions and their comprehensive utility values, and explanations of the reasons for the filtered candidate verification actions. This structured diagnostic result ensures the traceability and auditability of the diagnostic output and serves as the sole source of fact for subsequent natural language generation.
[0056] S703, Controlled Natural Language Generation.
[0057] The structured diagnostic results are provided to the large language model as immutable factual constraints. The large language model only organizes language, cites evidence, and standardizes terminology within the scope of the facts, and must not generate new root causes, new test points, or new operational suggestions beyond the structured diagnostic results. The system inputs the Top-K root causes, confidence intervals, evidence chains, priority check action sets, filtering reasons, and retrieved similar fragments of procedures and work orders into the large language model, and limits the output through templates and field constraints. This ensures that the natural language report can express the most likely root cause and its uncertainty, the reasons why key evidence supports the conclusion, the priority basis and feasibility of the next check action, prohibited actions and their prohibition reasons, and maintenance suggestions for corresponding procedure clauses or similar work order cases. This ensures readability while maintaining consistency and auditability with the inference results.
[0058] S8. Reinforce the manual confirmation and maintenance conclusions to update the fuzzy Bayesian network parameters and action dictionary; this includes the following steps: S801, Fuzzy Bayesian network parameter update; The root causes confirmed by maintenance, the results of the measurement point verification, and the work order conclusions are fed back into the database to update the parameters of the prior probability and conditional probability tables. The conditional probability table is updated using Dirichlet priors. ; in, This is the updated Dirichlet distribution hyperparameter vector; This is the Dirichlet distribution hyperparameter vector before the update; This is the new sample count vector for this reinjection. This update method gradually adapts to the characteristics of the plant's units without compromising the stability of existing parameters. For binary root cause events, a Beta prior update is used to synchronously output the confidence interval.
[0059] S802, Action Rewards Updated; The actual cost, time consumption, and risk event feedback of action execution are used to update the action dictionary parameters. An exponential smoothing method is preferred to achieve stable convergence. Taking cost update as an example, the calculation formula is as follows: ; in, The updated action cost; This represents the actual cost of executing this action; Cost of actions performed before the update; This is the smoothing coefficient. Time consumed. Risk level The same exponential smoothing method is used for updating, so that the value information planning gradually reflects the actual implementation cost.
[0060] It should be noted that when new alarm expressions, new measurement point aliases, changes to procedures or clauses, or new situations requiring prohibited actions occur, these are fed back into the measurement point dictionary, root cause dictionary, action dictionary, and safety constraint library, and the retrieval index is updated simultaneously. This ensures that the system knowledge and field procedures remain consistent in the long term, maintaining a closed-loop stable operation of candidate generation, verification reasoning, planning filtering, interpretation output, and feedback updates.
[0061] To verify the effectiveness of the method of the present invention, simulation experiments were conducted.
[0062] like Figure 2 As shown, under various fault scenarios such as rotor imbalance, shaft misalignment, oil film oscillation, dynamic-static rubbing, and thermal bending, the method of this invention has the lowest average number of diagnostic steps overall, and maintains a stable advantage even in difficult scenarios, indicating that it can complete root cause localization with fewer checks. The scatter plot and error bar show that under disturbances such as load, oil temperature, and incomplete observations, the results of the method of this invention are more concentrated, the diagnostic process is more stable, and the repeatability is better. In contrast, the rule expert system and the CNN baseline have a higher number of checks and greater fluctuations. These results verify the technical advantages of this invention in reducing invalid checks and improving diagnostic efficiency through "soft evidence robust modeling + information value-driven check planning".
[0063] like Figure 3As shown, the accuracy of all methods decreases with decreasing signal-to-noise ratio and increased noise interference. However, the method of this invention maintains a higher and more stable accuracy across the entire range, and its performance degradation under low signal-to-noise ratio conditions is more gradual, demonstrating stronger noise robustness. This is because the invention introduces fuzzy soft evidence mapping and a reliability weighting mechanism, which can suppress evidence distortion caused by noise and enhance reasoning stability. At the same time, the more concentrated error bars indicate that it has better repeatability and engineering applicability under common noise and disturbances in industrial settings.
[0064] like Figure 4 As shown, purely data-driven strategies carry certain safety violation risks, which, while reduced by rule expert systems, cannot be completely eliminated. The method of this invention achieves "zero violations," indicating that its hard constraint library and action filtering mechanism effectively block dangerous operations that do not meet protection / interlock / ticket / condition boundaries. Simultaneously, the method of this invention exhibits a certain rate of manual escalation, reflecting its conservative strategy of transferring manual intervention when constraints conflict or information is insufficient, thereby ensuring industrial site safety compliance in an "executable and auditable" manner.
[0065] like Figure 5 As shown, with the increase in the number of verification steps, the posterior entropy of the root cause gradually decreases for all three strategies. However, the method of this invention (VOI planning) decreases faster in the early stage and enters the stable convergence interval earlier in the later stage, indicating that it can prioritize verification actions with "higher information gain" and eliminate root cause uncertainty more efficiently. In contrast, the rule expert system converges more slowly, and the random strategy decreases the slowest and has greater residual uncertainty. This result shows that the VOI planning mechanism of this invention can achieve faster and more stable root cause localization with a limited number of verification steps.
[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for fault diagnosis of power equipment using a controlled LLM-guided fuzzy Bayesian network, characterized in that, Includes the following steps: S1. Access multi-source data and build a knowledge base and a security constraint base. The knowledge base includes a measurement point dictionary, a root cause dictionary, and an action dictionary. S2. Analyze multi-source data through a large language model to generate a list of nodes, a list of edges, candidate prior probabilities, and candidate conditional probability tables for the candidate fuzzy Bayesian network, and perform structural verification to obtain the fuzzy Bayesian network. S3. Map the continuous measurement point observation values corresponding to the symptom node to the fuzzy membership degree of each state through the preset membership degree function, normalize the fuzzy membership degree to obtain the soft evidence probability, correct the soft evidence probability through the evidence reliability weight to form the corrected soft evidence, and combine the corrected soft evidence of all symptom nodes into an evidence set. S4. Input the evidence set into the fuzzy Bayesian network to perform posterior inference, obtain the posterior probability distribution and confidence interval of each root cause node, sort the posterior probabilities, and output the Top-K root causes with the highest posterior probabilities as the candidate set. S5. Generate a set of candidate verification actions based on the action dictionary, define result variables and value space for each candidate verification action, calculate the value information of each candidate verification action, and calculate the comprehensive utility value in combination with cost, risk and time consumption. S6. Based on the security constraint library, perform hard constraint filtering on the candidate verification action set, output the executable action set, sort the executable action set in descending order according to the comprehensive utility value, and select a specific number of executable action sets to form a priority verification action set. S7. Calculate the contribution of each piece of evidence to the root cause distribution based on the evidence set, and construct the evidence chain based on the contribution. Organize the Top-K root causes and their confidence intervals, the evidence chain, the priority check action set and the reasons for the filtered actions into a structured diagnostic result. Provide the structured diagnostic result as a fact constraint to the big language model, and the big language model generates a natural language diagnostic report and maintenance suggestions within the scope of the fact constraint. S8. Feed back the manually confirmed results to update the fuzzy Bayesian network parameters and action dictionary.
2. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 1, characterized in that, In S2, the candidate fuzzy Bayesian network is represented as a directed acyclic graph. , where the set of nodes , For the set of root cause nodes, For the set of symptom nodes, It is a set of directed edges.
3. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 2, characterized in that, In S2, each node output by the large language model is mapped to a dictionary entity in the knowledge base, and unmapped candidate nodes are replaced or removed.
4. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 3, characterized in that, In S2, the structural verification includes acyclicity verification and causal edge validity verification; The acyclicity check is used to ensure the graph... Given a directed acyclic graph, the validity check of causal edges restricts unreasonable pre-defined connection relationships; When validation fails, structural corrections are performed, including deleting edges that cause cycles, merging duplicate nodes, and replacing non-standard names with dictionary-standard names.
5. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 4, characterized in that, In step S3, the continuous measurement point observations corresponding to the symptom nodes are mapped to fuzzy membership degrees of each state through a preset membership function. The fuzzy membership degrees are normalized to obtain the soft evidence probability. The soft evidence probability is then corrected using evidence reliability weights to form the corrected soft evidence, which includes the following steps: S301. A trapezoidal membership function is used to map the continuous observations of the symptom nodes to the membership degrees of each state, where the states include high state, medium state and low state; the threshold parameter of the trapezoidal membership function is given by the measurement point dictionary and configured in segments according to the load segment or speed segment. S302. Normalize the membership vector to obtain the soft evidence probability, calculated using the following formula: ； in, As a symptom node In state The probability of soft evidence; For state The corresponding membership value; For state The corresponding membership value, As a symptom node Observed values; S303, Introducing a weighting for the reliability of evidence The formula for correcting soft evidence is as follows: ； in, The corrected probability of soft evidence; As a symptom node Weight of evidence reliability As a symptom node All states, As a symptom node The total number of states.
6. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 5, characterized in that, In step S4, the step of inputting the evidence set into a fuzzy Bayesian network to perform posterior inference and obtain the posterior probability distribution and confidence interval of each root cause node includes: S401. Calculate the posterior probability of each root cause node using Bayesian inference. S402. Apply Beta distribution to binary root cause events. Perform parameter uncertainty modeling, where and Define the shape parameters of the Beta distribution and output the corresponding parameters. Confidence interval The significance level is indicated by .
7. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 6, characterized in that, In step S5, calculating the value information of each candidate verification action includes the following steps: S501. Calculate the posterior entropy of the root cause. The calculation formula is as follows: ； in, In the current set of evidence Posterior entropy of the next root cause; Root cause node The posterior probability; The total number of root cause nodes; S502, Calculate candidate verification actions The value information is calculated using the following formula: ； in, Candidate verification action Valuable information; In the current set of evidence The value of the next action result variable is The predicted probability; To observe the results of the action The posterior entropy of the root cause; Candidate verification action The value space of .
8. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 7, characterized in that, In step S6, the formula for calculating the overall utility value is as follows: ； in, Candidate verification action Execution costs; Candidate verification action Risk level; Candidate verification action The estimated time; These are the weighting coefficients for the first, second, third, and fourth comprehensive utility values. Candidate verification action The overall utility value.
9. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 8, characterized in that, In step S6, the method for hard constraint filtering of the candidate verification action set based on the security constraint library is as follows: the constraint judgment result is 1 when the candidate verification action satisfies all constraints in the security constraint library, and the constraint judgment result is 0 when the candidate verification action violates any constraint in the security constraint library. The set of executable actions is defined as follows: ;in, A set of executable actions; For the set of candidate verification actions; For candidate verification actions; Candidate verification action In context The constraint determination result.
10. The power equipment fault diagnosis method using a controlled LLM-guided fuzzy Bayesian network according to claim 9, characterized in that, In step S7, the steps of calculating the contribution of each piece of evidence to the root cause distribution based on the evidence set and constructing an evidence chain based on the contribution include: S701. Calculate the contribution of each piece of evidence to the root cause distribution based on the evidence set. The calculation formula is as follows: ； in, For the first The contribution of each piece of evidence; Let KL divergence be a metric. For evidence set The posterior distribution of the root causes; To remove the first Root cause posterior distribution after 1 piece of evidence; To obtain evidence Remove the first A subset following one piece of evidence; S702. Evidence with a contribution rate higher than the preset evidence screening threshold and whose corresponding symptoms can be mapped to the standard measurement point name or standard fault symptom term in the measurement point dictionary is organized into an evidence chain in chronological order and causal logic.