A thermal power plant fault diagnosis method based on a large language model

Abstract

The application discloses a kind of fire power generation equipment fault diagnosis methods based on large language model, it is related to the technical field of intelligent diagnosis of fire power generation equipment, including the following steps: generating fire power equipment multi-source state data frame;Based on working condition identification feature and historical normal working condition segment, generate working condition peeling residual sequence;Generate equipment time delay causal propagation graph;Generate fault evidence passport based on working condition peeling residual sequence;Fault evidence passport is hung to equipment time delay causal propagation graph, generates candidate fault propagation chain;Candidate fault propagation chain, fault evidence passport, counterfactual pruning result and equipment regulation knowledge are input evidence chain constraint type DeepSeek-R1 fire power fault diagnosis model, generate fault diagnosis result;According to the closed loop update of on-site review result and overhaul feedback result.The present application can improve the evidence traceability, causal consistency and diagnostic accuracy of fire power equipment fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for thermal power generation equipment, and in particular to a fault diagnosis method for thermal power generation equipment based on a large language model. Background Technology

[0002] With the development of thermal power generating units towards deep peak shaving, frequent start-stop, and multi-system coupled operation, intelligent fault diagnosis technology for boilers, turbines, pulverizing equipment, flue gas systems, feedwater systems, and environmental protection equipment has received widespread attention. Existing publicly available technologies have attempted to apply expert knowledge bases, large language models, knowledge graphs, or causal networks to power generation equipment fault diagnosis. For example, CN119293232A discloses a method for combining expert knowledge bases, prompt word templates, and large language models to generate state sets based on equipment condition monitoring time-series data and retrieve fault diagnosis results; CN112307218A discloses a method for constructing a knowledge graph knowledge base for fault diagnosis of typical equipment in smart power plants; and CN109359662A discloses a causal analysis and multi-layer Bayesian network fault diagnosis method for the non-stationary characteristics of megawatt-class ultra-supercritical units. These disclosures demonstrate that large language models, knowledge graphs, and causal analysis have been applied in the field of power generation or power plant equipment fault diagnosis.

[0003] However, in field applications at thermal power plants, DCS operating data, SIS historical data, protection interlock signals, alarm records, operating logs, and maintenance work orders suffer from inconsistent sampling frequencies, timescale drift, inconsistent equipment names, and complex measurement point attributions, making it difficult to uniformly bind fault evidence. Deep peak shaving, soot blowing, pulverizer start-up and shutdown, coal quality fluctuations, and environmental regulations generate numerous interpretable disturbances, and existing diagnostic methods easily misjudge normal operating condition disturbances as faults. Equipment anomalies also propagate with time delays along the medium flow direction, control loops, and protection interlocks. Simply relying on knowledge base retrieval or ordinary large language model reasoning makes it difficult to constrain the direction of fault propagation, propagation delay, and the credibility of evidence sources, easily leading to problems such as no root cause evidence, backpropagation explanations, or unverifiable diagnostic conclusions. Therefore, there is an urgent need for a fault diagnosis method for thermal power equipment capable of performing time-series alignment, operating condition stripping, time-delay causal constraints, evidence passport encapsulation, and evidence chain constraint reasoning on multi-source state data.

[0004] Therefore, how to provide a fault diagnosis method for thermal power generation equipment based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a fault diagnosis method for thermal power generation equipment based on a large language model. This invention fully utilizes dynamic time warping and time synchronization, equipment time delay causal propagation graph, counterfactual pruning, and the DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints. It describes in detail the process of extracting fault evidence, generating propagation chains, and binding diagnostic conclusions for thermal power equipment faults. It has the advantages of high diagnostic accuracy, strong evidence traceability, and low false alarm rate.

[0006] A fault diagnosis method for thermal power generation equipment based on a large language model according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect operating data of thermal power generating units, perform dynamic time warping and time calibration, KKS equipment coding mapping, measurement point attribution binding, anomaly marker writing and text quadruple extraction to generate multi-source status data frames for thermal power equipment;

[0008] Step 2: Extract operating condition identification features from the multi-source state data frames of thermal power equipment, match historical normal operating condition segments and calculate the equipment operating condition baseline state, peel off interpretable deviations based on standard disturbance envelopes, and generate an operating condition peeling residual sequence.

[0009] Step 3: Configure device nodes, causal edges, propagation direction, propagation delay, and edge reliability according to device connection, media flow direction, control loop, protection interlock, and historical anomaly propagation time difference to generate a device delay causal propagation graph;

[0010] Step 4: Extract abnormal evidence based on the residual sequence stripping under operating conditions, and encapsulate it according to equipment ownership, residual strength, source credibility, time delay matching degree and conflict state to generate fault evidence passport;

[0011] Step 5: Attach the fault evidence passport to the device delay causal propagation graph, perform delay constraint path search and counterfactual pruning, and generate candidate fault propagation chains;

[0012] Step 6: Input the candidate fault propagation chain, fault evidence passport, counterfactual pruning results and equipment procedure knowledge into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model. Through evidence passport credibility gating, equipment causal chain constraint reasoning, counterfactual consistency verification and diagnostic conclusion evidence binding, generate fault diagnosis results.

[0013] Step 7: Generate diagnostic deviations based on the on-site verification results and maintenance feedback results, and update the equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and evidence chain constraint type DeepSeek-R1 thermal power fault diagnosis model diagnostic constraints.

[0014] Optionally, step one specifically includes:

[0015] Collect operating data of thermal power generating units and establish a data source cache queue according to data source, acquisition channel and original acquisition time stamp;

[0016] Using the unit's AGC command change points, load change points, auxiliary machine start / stop signals, protection action signals, and major regulation change points as time calibration anchor points, dynamic time warping and time calibration are performed to generate a unified sampling time scale;

[0017] Resampling is performed on continuous measurement point data, protection events, alarm events, and text events according to a unified sampling time standard;

[0018] Establish a KKS device coding table, a device alias dictionary, a device hierarchy table, and a system affiliation table. Perform character normalization, keyword window matching, and edit distance matching on the device names in the text records to complete the KKS device coding mapping.

[0019] The measurement point affiliation binding is performed based on the measurement point number, measurement point name, physical quantity type, equipment, system, and associated equipment fields in the measurement point configuration table.

[0020] For data with empty values, exceeding the range, in a dormant state, single-point mutation, or manually entered data, write missing, over-range, dormant, low-confidence, and manually sourced data respectively;

[0021] Text quadtuples are extracted from operation logs, inspection records, and maintenance work orders to generate a quadtuple of "equipment - abnormal phenomenon - handling action - event time";

[0022] The processed data is aggregated according to a unified sampling timescale to generate a multi-source status data frame for thermal power equipment. The multi-source status data frame for thermal power equipment includes equipment identifier, measurement point identifier, system identifier, unified sampling timescale, status value, data source, data type, confidence flag, anomaly flag, text quadruple field, and time inheritance flag.

[0023] Optionally, step two specifically includes:

[0024] Operating condition identification features are extracted from the multi-source status data frames of thermal power equipment. These features include unit load, load change rate, coal mill operating combination, total air volume, main steam pressure, low calorific value of coal, soot blowing status, denitrification ammonia injection volume, and ambient temperature.

[0025] Based on the operating condition identification characteristics, the following conditions are identified: load increase condition, load decrease condition, coal mill start-up and shutdown condition, soot blowing disturbance condition, coal quality fluctuation condition, and environmental protection regulation condition.

[0026] Match historical normal operating condition segments from the historical normal operating database and remove measurement point samples with missing, over-range, dormant, and low confidence markers;

[0027] The baseline value of the measuring point is determined based on the mean or median of the corresponding measuring point within the historical normal operating condition segment. The upper limit and lower limit of normal fluctuation are determined based on the standard deviation range or quantile range of the corresponding measuring point. The continuous judgment threshold is determined in combination with the equipment response inertia to generate the equipment operating condition baseline state.

[0028] Standard disturbance envelopes are established based on load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, and environmental regulation.

[0029] The current equipment status is compared with the equipment operating condition baseline status to remove interpretable deviations, retain uninterpretable deviations, and generate an operating condition stripping residual sequence in chronological order.

[0030] Optionally, the step of comparing the current equipment state with the equipment operating condition baseline state, removing interpretable deviations, retaining uninterpretable deviations, and generating an operating condition stripping residual sequence in chronological order specifically includes:

[0031] When the load change rate exceeds the preset load change rate threshold and the AGC command changes continuously, the current time window will be marked as either a load increase condition or a load decrease condition.

[0032] When the coal mill start / stop signal is valid, mark the current time window as the coal mill start / stop condition;

[0033] When the soot blowing valve action signal is valid, mark the current time window as a soot blowing disturbance condition;

[0034] When the change in the lower heating value of adjacent coal quality testing cycles exceeds the preset coal quality fluctuation threshold, the current time window is marked as a coal quality fluctuation condition.

[0035] When the amount of ammonia injected for denitrification, the inlet parameters for desulfurization, or the environmental regulation instructions continuously deviate from the stable range, the current time window will be marked as an environmental regulation condition.

[0036] Based on historical normal disturbance events, the standard disturbance envelopes corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, and environmental regulation are determined respectively.

[0037] When the current deviation falls within the corresponding standard disturbance envelope and the duration does not exceed the allowable duration, the current deviation is determined to be an interpretable deviation and is removed.

[0038] When the current deviation exceeds the normal fluctuation range of the equipment's operating condition reference state, the duration reaches the corresponding measurement point's continuous judgment threshold, and it cannot be explained by the standard disturbance envelope, the current deviation will be retained as an unexplainable deviation.

[0039] Unexplainable deviations are aggregated according to equipment identification, measurement point identification, unified sampling time scale, deviation direction, deviation amplitude, duration, and confidence level to generate a working condition stripping residual sequence.

[0040] Optionally, step three specifically includes:

[0041] Construct a set of equipment nodes based on the equipment ledger, P&ID drawings, DCS control logic, and protection interlocking table of the thermal power generating unit;

[0042] Construct a set of causal edges, and configure the propagation direction, propagation delay, and edge credibility for each causal edge;

[0043] The propagation direction is determined based on equipment connection, medium flow direction, control circuit and protection interlock, forming a pulverizing combustion propagation side, a wind-powder propagation side, a wind and smoke regulation side, a steam-water regulation side and an environmental protection coupling side;

[0044] Based on the occurrence time of the residual loss of adjacent device nodes in historical abnormal events, the propagation time difference between the start time of the abnormality of the upstream node and the start time of the abnormality of the downstream node is determined.

[0045] Statistical distribution analysis is performed on multiple propagation time differences corresponding to the same causal edge. Outlier sample elimination method based on statistical distribution is used to determine the propagation time difference distribution interval, and propagation delay is generated based on the propagation time difference distribution interval.

[0046] The edge credibility is configured based on the clarity of the equipment mechanism, the number of historical abnormal samples, the concentration of the propagation time difference distribution, and the on-site verification results.

[0047] Write device nodes, causal edges, propagation directions, propagation delays, and edge credibility into the graph structure to generate a device delay causal propagation graph.

[0048] Optionally, step four specifically includes:

[0049] Based on the residual sequence stripping under working conditions, extract abnormal time windows where the residual is not zero and the duration reaches the continuous judgment threshold of the corresponding measuring point, and merge abnormalities in the same device whose consecutive abnormal time window intervals do not exceed the preset interval threshold into the same residual abnormal event.

[0050] Extract abnormal evidence based on residual abnormal events, and associate alarm signals, SOE signals, operation logs, inspection records or maintenance work orders within the same time window with residual abnormal evidence to form a multi-source abnormal evidence set;

[0051] The residual strength is calculated based on the degree and duration of the residual amplitude exceeding the normal fluctuation range of the equipment's operating condition reference state;

[0052] The source credibility is determined based on the degree of automatic data collection, the degree of human intervention, the persistence of anomalies, and the reliability of on-site verification. The source credibility of the corresponding text evidence is reduced when the text quadruple field has a time inheritance marker.

[0053] The delay matching degree is determined based on the matching relationship between the occurrence time of abnormal evidence and the propagation delay range of adjacent device nodes in the device delay causal propagation diagram;

[0054] Conflict states are generated based on the consistency between alarms, residuals, text records, and associated measurement points.

[0055] Encapsulate multi-source anomaly evidence within the same device and the same anomaly time window to generate a fault evidence passport. The fault evidence passport includes fault evidence passport number, device affiliation, system affiliation, evidence type, anomaly start time, anomaly end time, residual strength, source credibility, time delay matching degree, conflict state, and associated measurement point set.

[0056] The fault evidence passports are sorted according to equipment ownership and anomaly start time, and an index relationship is established between the fault evidence passports and the operating condition stripping residual sequence, the multi-source state data frame of thermal power equipment, and the equipment delay causal propagation graph.

[0057] Optionally, step five specifically includes:

[0058] The fault evidence passport is attached to the corresponding device node in the device delay causal propagation graph according to the device affiliation;

[0059] Select the device node with the earliest anomaly start time and whose source credibility meets the preset conditions as the starting point for path search.

[0060] Extend downstream from the starting point of the path search along the causal edges in the device delay causal propagation graph, and perform a delay-constrained path search;

[0061] When the next device node joins the abnormal propagation path, it must simultaneously meet the following conditions: consistent propagation direction, existence of fault evidence, abnormal start time falling within the propagation delay range of the corresponding causal edge, source credibility meeting preset conditions, and residual abnormal time windows being continuously connected within the corresponding propagation delay range.

[0062] When the start time, fluctuation range, and duration of an abnormal path can be explained by the standard disturbance envelope corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, or environmental regulation, the abnormal path shall be removed.

[0063] When an anomaly is triggered only by a low-confidence measurement point and no synchronous anomaly occurs in the upstream measurement point, downstream measurement point, or similar redundant measurement point, the anomaly path is removed as a measurement point anomaly path.

[0064] When the abnormal propagation direction is opposite to the direction of equipment connection, medium flow, or control loop, the abnormal path is cut off as a path that is not valid in the reverse direction.

[0065] The propagation chain score is calculated based on source credibility, residual strength, delay matching degree, conflict state, causal edge credibility, and evidence complementarity. The evidence complementarity is determined based on the number of evidence types in the candidate fault propagation chain, such as residual abnormalities, alarm signals, SOE signals, operation logs, inspection records, or maintenance work orders.

[0066] The candidate fault propagation chains are output in descending order of propagation chain score. The candidate fault propagation chain includes root cause candidate nodes, propagation node sequence, affected nodes, associated fault evidence passport, propagation delay matching results, counterfactual pruning results, and propagation chain score.

[0067] Optionally, step six specifically includes:

[0068] The candidate fault propagation chain, fault evidence passport, counterfactual pruning results and equipment procedure knowledge are input into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model.

[0069] Equipment procedures include operating procedures, maintenance procedures, equipment manuals, protection interlock instructions, and historical case studies.

[0070] The DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints includes a procedure knowledge retrieval layer, a diagnosis task package construction layer, an evidence passport credibility gating layer, a structured diagnosis prompt layer, an equipment causal chain constraint reasoning layer, a counterfactual consistency verification layer, and a diagnosis conclusion evidence binding layer.

[0071] The procedure knowledge retrieval layer retrieves procedure knowledge fragments from equipment procedure knowledge based on the root cause candidate node, propagation node sequence, system attribution, and abnormal phenomenon fields in the candidate fault propagation chain.

[0072] The diagnostic task package construction layer generates a diagnostic task number and writes the diagnostic task number, candidate fault propagation chain number, root cause candidate node, propagation node sequence, propagation direction, propagation delay range, propagation chain score, fault evidence passport number, counterfactual pruning result, and procedure clause number into the diagnostic task package.

[0073] The evidence passport credibility gating layer generates evidence input weights as evidence weights based on source credibility, residual strength, time delay matching degree, and conflict state, and divides the faulty evidence passport into core evidence area and auxiliary evidence area;

[0074] The structured diagnostic prompt layer constructs fixed prompt templates based on candidate fault propagation chains, core evidence areas, auxiliary evidence areas, counterfactual pruning results, and equipment procedure knowledge;

[0075] The equipment causal chain constraint inference layer verifies the equipment entities and propagation relationships in the output of the evidence chain constraint DeepSeek-R1 thermal power fault diagnosis model based on the candidate fault propagation chain.

[0076] The counterfactual consistency verification layer generates a counterfactual verification matrix based on the counterfactual pruning results and verifies the output of the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model.

[0077] The diagnostic conclusion evidence binding layer binds evidence to the output that has passed the verification to generate fault diagnosis results.

[0078] Optionally, the evidence passport credibility gating layer normalizes the residual intensity and generates evidence input weights as evidence weights based on source credibility, normalized residual intensity, time delay matching degree, and conflict state.

[0079] Based on the evidence weight and the preset evidence partitioning threshold, the fault evidence passport is divided into the core evidence area, the auxiliary evidence area, and the fault evidence passport that has not entered the model inference input.

[0080] The structured diagnostic prompt layer incorporates candidate fault propagation chain area, core evidence area, auxiliary evidence area, counterfactual pruning result area, equipment procedure knowledge area, and output format area into a fixed prompt template;

[0081] The device causal chain constraint reasoning layer uses candidate fault propagation chains as root cause reasoning boundaries and adopts a streaming interception method for output verification. When the output of the evidence chain constraint DeepSeek-R1 thermal power fault diagnosis model contains non-candidate fault propagation chain devices, devices without bound fault evidence passports, reverse propagation relationships, or propagation interpretations that exceed the propagation delay range, the generation is interrupted and regenerated based on the original diagnostic task package.

[0082] The counterfactual consistency verification layer constructs a counterfactual verification matrix, which includes load disturbance, soot blowing operation, start and stop of coal mill, coal quality fluctuation, environmental protection regulation, measurement point anomaly and communication anomaly. Each item is written with whether to exclude, the basis for exclusion, the fault evidence passport number and the counterfactual pruning rule number.

[0083] When the counterfactual verification matrix contains unexcluded items or missing bound number items, the missing items will be written into the feedback prompt and the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model will be regenerated.

[0084] The diagnostic conclusion evidence binding layer binds the root cause of the fault to the fault evidence passport number in the core evidence area, binds the fault level to the risk level in the protection interlock status, alarm level and equipment procedure knowledge, binds the scope of impact to the affected nodes in the candidate fault propagation chain, binds the verification measurement points to the associated measurement point set, and binds the handling suggestions to the procedure clause number.

[0085] The diagnostic confidence is calculated based on the evidence weight of the core evidence area, the score of the candidate fault propagation chain, the completeness of the counterfactual consistency check, and the degree of matching of equipment procedure knowledge, and the fault diagnosis result is generated.

[0086] Optionally, step seven specifically includes:

[0087] Collect on-site verification results and maintenance feedback results, and compare the on-site verification results, maintenance feedback results and fault diagnosis results to generate diagnostic deviations;

[0088] When the manually confirmed root cause is inconsistent with the fault root cause, a root cause device bias is generated; when the actual abnormal propagation time exceeds the propagation delay range of the corresponding causal edge, a propagation delay bias is generated; when low-confidence fault evidence affects the diagnostic conclusion, an evidence weight bias is generated; when normal disturbances are not stripped away or the real fault propagation path is mistakenly pruned by counterfactual pruning, a counterfactual pruning bias is generated.

[0089] The equipment operating condition baseline status is updated based on the diagnostic deviation. Normal operating segments confirmed on-site are added to the historical normal operating condition segment library. Historical normal operating condition segments are re-matched, and the baseline value, normal upper limit, normal lower limit, and duration threshold of the corresponding measuring point are updated using a sliding window weighted statistical method.

[0090] The propagation delay and edge confidence are updated based on the propagation delay deviation. The minimum and maximum propagation delays of the corresponding causal edges are corrected using an exponential moving average method. The edge confidence of the corresponding causal edges is increased or decreased based on the on-site confirmation results.

[0091] The evidence weights are updated based on the evidence weight bias, and the counterfactual pruning rules are updated based on the counterfactual pruning bias.

[0092] Based on the maintenance feedback results, the diagnostic constraints of the DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints are updated. The fault phenomena, root causes, fault evidence passport numbers, candidate fault propagation chain numbers, handling actions and recovery results confirmed on site are written into the thermal power fault case index library. The thermal power fault case index library serves as a source of historical handling cases in equipment procedure knowledge.

[0093] The updated equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and diagnostic constraints are written into the closed-loop update record and bound to the diagnostic task number, fault evidence passport number, and candidate fault propagation chain number.

[0094] The beneficial effects of this invention are:

[0095] (1) This application generates multi-source status data frames for thermal power equipment by using dynamic time warping and time calibration, KKS equipment code mapping, measurement point attribution binding, abnormal mark writing and text quadruple extraction. This solves the problems of inconsistent time scales, inconsistent equipment names and unclear measurement point attribution among DCS operation data, SIS historical data, protection interlock signals, alarm records, operation logs and maintenance work orders. It enables continuous measurement points, event signals and text records to be associated under a unified sampling time scale and a unified equipment code, thereby improving the consistency and traceability of fault diagnosis input data.

[0096] (2) This application calculates the equipment operating condition baseline state by matching historical normal operating condition segments with operating condition identification features, and uses standard disturbance envelopes to strip away the interpretable deviations caused by load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations and environmental regulation, and generates an operating condition stripping residual sequence. This solves the problem that normal disturbances under deep peak shaving and frequent operating condition switching are misjudged as equipment failures, so that subsequent abnormal evidence mainly comes from unexplainable deviations, reducing the false alarm rate and improving the stability of fault detection under complex operating conditions.

[0097] (3) This application generates candidate fault propagation chains by using equipment time delay causal propagation graphs, fault evidence passports, time delay constraint path search and counterfactual pruning. Then, it inputs the evidence chain constraint type DeepSeek-R1 thermal power fault diagnosis model to perform evidence passport credibility gating, equipment causal chain constraint reasoning, counterfactual consistency verification and diagnostic conclusion evidence binding. This solves the problems that ordinary large language models are prone to generating root causes without evidence, backpropagation interpretation and unverifiable conclusions. It enables fault root causes, scope of influence, verification measurement points and disposal suggestions to be bound to fault evidence passports, propagation chains and procedural knowledge, thereby improving the credibility, interpretability and on-site maintenance guidance value of thermal power equipment fault diagnosis. Attached Figure Description

[0098] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0099] Figure 1 This is a flowchart of a fault diagnosis method for thermal power generation equipment based on a large language model proposed in this invention;

[0100] Figure 2This is a schematic diagram of a fault diagnosis method for thermal power generation equipment based on a large language model proposed in this invention;

[0101] Figure 3 This is a framework diagram of the DeepSeek-R1 thermal power fault diagnosis model, which is a chain-of-evidence-constrained model, in the fault diagnosis method for thermal power equipment based on a large language model proposed in this invention. Detailed Implementation

[0102] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0103] refer to Figures 1-3 A fault diagnosis method for thermal power generation equipment based on a large language model includes the following steps:

[0104] Step 1: Collect operating data of thermal power generating units, perform dynamic time warping and time calibration, KKS equipment coding mapping, measurement point attribution binding, anomaly marker writing and text quadruple extraction to generate multi-source status data frames for thermal power equipment;

[0105] Step 2: Extract operating condition identification features from the multi-source state data frames of thermal power equipment, match historical normal operating condition segments and calculate the equipment operating condition baseline state, peel off interpretable deviations based on standard disturbance envelopes, and generate an operating condition peeling residual sequence.

[0106] Step 3: Configure device nodes, causal edges, propagation direction, propagation delay, and edge reliability according to device connection, media flow direction, control loop, protection interlock, and historical anomaly propagation time difference to generate a device delay causal propagation graph;

[0107] Step 4: Extract abnormal evidence based on the residual sequence stripping under operating conditions, and encapsulate it according to equipment ownership, residual strength, source credibility, time delay matching degree and conflict state to generate fault evidence passport;

[0108] Step 5: Attach the fault evidence passport to the device delay causal propagation graph, perform delay constraint path search and counterfactual pruning, and generate candidate fault propagation chains;

[0109] Step 6: Input the candidate fault propagation chain, fault evidence passport, counterfactual pruning results and equipment procedure knowledge into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model. Through evidence passport credibility gating, equipment causal chain constraint reasoning, counterfactual consistency verification and diagnostic conclusion evidence binding, generate fault diagnosis results.

[0110] Step 7: Generate diagnostic deviations based on the on-site verification results and maintenance feedback results, and update the equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and evidence chain constraint type DeepSeek-R1 thermal power fault diagnosis model diagnostic constraints.

[0111] In this embodiment, step one specifically includes:

[0112] Collect operating data from thermal power generating units and establish a data source cache queue according to data source, acquisition channel, and original acquisition time stamp; the operating data of thermal power generating units includes DCS operating data, SIS historical data, protection interlock signals, equipment alarm records, operation logs, inspection records, maintenance work orders, coal quality data, environmental protection operation data, and environmental data; write the original acquisition time stamp, data source identifier, data type identifier, and access batch identifier into each data entry;

[0113] Dynamic time warping is performed using the unit AGC command change point, load change point, auxiliary machine start / stop signal, protection action signal and major regulation change point as time anchor points. The event sequence in the reference data source is matched with the event sequence in the data source to be corrected. The dynamic warping path is determined according to the order of event triggering and the time distance of events. The sampling time scale of the data source to be corrected is shifted, locally stretched or locally compressed according to the dynamic warping path to generate a unified sampling time scale.

[0114] The data that has been calibrated is resampled; for continuous measurement point data, interpolation or retention processing is performed according to the unified sampling time scale; for protection events, alarm events and text events, the event status is written into the time window where the corresponding unified sampling time scale is located; for the case where there are multiple sampled values ​​for the same measurement point in the same time window, the retained value is determined according to the data source priority and sampling time distance.

[0115] Perform KKS equipment coding mapping to establish a KKS equipment coding table, equipment alias dictionary, equipment hierarchy table, and system affiliation table; perform character normalization processing on equipment names in text records, remove unit numbers, symbol intervals, and non-equipment descriptive words, and generate the equipment field to be matched;

[0116] The device field to be matched is matched with the device alias dictionary, and the overall similarity is determined by combining keyword window matching and edit distance matching. When the overall similarity reaches the mapping threshold, the device field to be matched is mapped to the corresponding KKS device code. When the overall similarity does not reach the mapping threshold, the device field is written to the low confidence tag.

[0117] Perform measurement point attribution binding, read the measurement point number, measurement point name, physical quantity type, belonging device, belonging system, and associated device fields from the measurement point configuration table, and bind the devices that have completed KKS device code mapping to the measurement points; for measurement points that directly reflect the status of a single device, bind them to the corresponding device identifier; for common measurement points, write the main attribution device and associated attribution device according to the device connection relationship, media flow direction, and measurement point purpose;

[0118] Perform the following actions: Write an anomaly flag; write a missing flag for data with empty values ​​or data exceeding the allowed missing time; write an over-range flag for data exceeding the instrument's range, engineering constraints, or physical reasonable range; write a dormant flag for measuring points where the equipment is in shutdown, standby, or maintenance status; write a low-confidence flag for data with a single-point sudden change but which has not changed synchronously with upstream, downstream, or similar redundant measuring points; and write a manual source flag for manually entered text, manually confirmed records, and manually reviewed results.

[0119] The system performs text quadruple extraction, segments operation logs, inspection records, and maintenance work orders into sentences, identifies equipment fields based on an equipment dictionary, anomaly fields based on an anomaly lexicon, handling fields based on a handling action lexicon, and event time fields based on time expression rules. When equipment, anomaly, and handling fields exist in the same sentence, a quadruple "equipment—anomaly—handling action—event time" is generated. When no explicit event time appears in the text, the text record creation time or the time of the adjacent context is used as the event time, and a time inheritance tag is written in it.

[0120] After completing dynamic time warping and calibration, KKS equipment encoding mapping, measurement point attribution binding, anomaly marker writing, and text quadruple extraction, the data is collected according to a unified sampling time scale to generate a multi-source status data frame for thermal power equipment. The multi-source status data frame for thermal power equipment includes equipment identifier, measurement point identifier, system identifier, unified sampling time scale, status value, data source, data type, confidence marker, anomaly marker, text quadruple field, and time inheritance marker.

[0121] In this embodiment, step two specifically includes:

[0122] Operating condition identification features are extracted from multi-source status data frames of thermal power equipment. These features include unit load, load change rate, coal mill operating combination, total air volume, main steam pressure, coal quality low calorific value, soot blowing status, denitrification ammonia injection amount, and ambient temperature. Data with missing, over-range, or low confidence markers are not used as the primary criterion but only as auxiliary references.

[0123] The current operating condition is identified based on the operating condition identification characteristics; when the load change rate exceeds 2MW / min and the AGC command changes continuously, the current time window is marked as either a load increase or load decrease condition. The 2MW / min is determined based on the unit's normal AGC adjustment rate and the historical upper limit of stable load fluctuation; when the coal mill start-up and shutdown signal is valid, the current time window is marked as a coal mill start-up and shutdown condition.

[0124] When the soot blowing valve action signal is valid, the current time window is marked as a soot blowing disturbance condition; when the change in the lower heating value of coal in adjacent coal quality testing cycles exceeds 1.5 MJ / kg, the current time window is marked as a coal quality fluctuation condition, and 1.5 MJ / kg is determined based on the historical test fluctuation range of the coal fed into the furnace and the combustion adjustment record; when the amount of ammonia injected for denitrification, the desulfurization inlet parameters, or the environmental protection adjustment command continuously deviates from the stable range, the current time window is marked as an environmental protection adjustment condition.

[0125] Match historical normal operating condition segments from the historical normal operating database. The historical normal operating database consists of historical segments in which no protection action occurred, no valid fault work order existed, no continuous alarm occurred, and the operating record was confirmed as normal.

[0126] The matching conditions for similar operating conditions are set as follows: load deviation not exceeding 30MW, load change rate deviation not exceeding 2MW / min, coal mill operation combination difference not exceeding one unit, ambient temperature deviation not exceeding 5℃, and coal quality lower heating value deviation not exceeding 1.5MJ / kg. It is also required that the soot blowing state and environmental regulation state are in the same or similar disturbance window. The 30MW is determined based on the statistical particle size of the load segment of the 1,000 kW unit, the 5℃ is determined based on the seasonal operation data of the circulating cooling and flue gas system, and the coal mill combination difference not exceeding one unit is determined based on the influence boundary of the pulverizing system commissioning combination on the combustion state.

[0127] The historical normal operating condition segments obtained through matching are subjected to a reliability screening process, and measurement point samples with missing, over-range, dormant, and low-reliability markers are removed. When the number of valid samples meets the quantity requirements, the benchmark value of the measurement point is determined based on the mean or median of the corresponding measurement points within similar operating condition segments. The upper limit and lower limit of normal fluctuation are determined based on the standard deviation range or quantile range of the corresponding measurement points, and the continuous judgment threshold is determined in combination with the equipment response inertia. The continuous judgment threshold for coal mill current is set to 60 to 120 seconds, and the continuous judgment threshold for main steam pressure is set to 120 to 300 seconds.

[0128] Standard disturbance envelopes are established based on load changes, soot blowing operations, pulverizer start-up and shutdown, coal quality fluctuations, and environmental regulation. The standard disturbance envelopes are obtained from historical normal disturbance events and are used to represent the allowable deviation range and allowable duration of the corresponding disturbance on relevant measuring points within a time window. The duration of the soot blowing disturbance envelope is set to 5 to 15 minutes after the start of the soot blowing operation, based on the time distribution of furnace negative pressure, flue gas temperature, and oxygen content returning to the stable range in historical soot blowing records. The duration of the pulverizer start-up and shutdown disturbance envelope is set to 10 to 30 minutes after the pulverizer start-up and shutdown, based on the historical statistical results of the pulverizing system output and combustion parameters returning to stability.

[0129] The current equipment status is compared with the equipment operating condition baseline status. When the current deviation falls within the corresponding standard disturbance envelope and the duration does not exceed the allowable duration corresponding to the standard disturbance envelope, the deviation is determined as an interpretable deviation and is removed. When the current deviation exceeds the normal fluctuation range of the equipment operating condition baseline status, the duration reaches the corresponding measurement point's duration determination threshold, and cannot be explained by the standard disturbance envelope corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, or environmental regulation, the deviation is retained as an unexplainable deviation.

[0130] Unexplainable deviations are collected according to equipment identification, measurement point identification, unified sampling time scale, deviation direction, deviation amplitude, duration and confidence level, and a working condition stripping residual sequence is generated in chronological order.

[0131] In this embodiment, step three specifically includes:

[0132] Based on the equipment ledger, P&ID drawings, DCS control logic and protection interlock table of the thermal power generating unit, a set of equipment nodes is constructed. The equipment nodes are divided into boiler, steam turbine, generator, pulverizing system, flue gas system, water supply system, condensate system and environmental protection system, and further refined to equipment units that can be independently collected, alarmed or repaired.

[0133] Construct a set of causal edges. Directed causal edges are used to represent the abnormal influence relationship between one device node and another device node. Each causal edge is configured with propagation direction, propagation delay and edge credibility.

[0134] The propagation direction is determined based on equipment connections, media flow direction, control circuits, and protection interlocks: the coal feeder to the coal mill and the coal mill to the furnace are configured as the pulverizing and combustion propagation side; the primary air fan to the coal mill and the coal mill to the furnace are configured as the air-coal propagation side; the induced draft fan to the furnace negative pressure is configured as the air-flue gas regulation side; the feedwater pump to the feedwater flow rate and the feedwater flow rate to the main steam parameters are configured as the steam-water regulation side; and the denitrification ammonia injection regulation to the air preheater differential pressure is configured as the environmental protection coupling side.

[0135] Propagation delay is determined based on the historical anomaly propagation time difference: the occurrence time of the residual of the working condition stripping of adjacent equipment nodes in the historical anomaly event is read, the time difference between the start time of the upstream node anomaly and the start time of the downstream node anomaly is taken as the propagation time difference, and samples corresponding to manual operation, equipment shutdown, dormant measurement points and low confidence measurement points are removed.

[0136] Statistical distribution analysis is performed on multiple propagation time differences corresponding to the same causal edge to generate the propagation time difference distribution interval; outlier propagation time difference samples are removed by using the interquartile range, and the 5th percentile of the propagation time difference distribution is taken as the minimum propagation delay, and the 95th percentile is taken as the maximum propagation delay, thereby generating the propagation delay range of the causal edge.

[0137] The propagation delay from abnormal coal feeder feed rate to abnormal coal mill current is set to 5 to 30 seconds, based on the coal feed disturbance transmission record of the pulverizing system; the propagation delay from coal mill abnormality to furnace negative pressure fluctuation is set to 20 to 60 seconds, based on the furnace negative pressure response time distribution after coal mill output disturbance; the propagation delay from furnace combustion disturbance to main steam pressure fluctuation is set to 60 to 180 seconds, based on the historical propagation time difference distribution under the response inertia of boiler heat storage and steam-water system.

[0138] The edge confidence level is configured based on the clarity of the equipment mechanism, the number of historical anomaly samples, the concentration of propagation time difference distribution, and the results of on-site verification. For causal edges that are jointly confirmed by equipment connections, control loops, and protection interlocks and have a concentrated propagation time difference distribution, a higher edge confidence level is set. For causal edges that are only supported by historical anomaly propagation time differences and have a discrete propagation time difference distribution, a lower edge confidence level is set.

[0139] Write device nodes, causal edges, propagation directions, propagation delays, and edge credibility into the graph structure to generate a device delay causal propagation graph.

[0140] In this embodiment, step four specifically includes:

[0141] Read the residual sequence of the working condition stripping, and extract the abnormal time window where the residual is not zero and the duration reaches the continuous judgment threshold of the corresponding measurement point according to the equipment identifier, measurement point identifier and unified sampling time scale; according to the DCS sampling cycle and short-term jitter filtering requirements, merge the abnormalities with a continuous abnormal time window interval of no more than 30 seconds in the same equipment into the same residual abnormal event.

[0142] Extract abnormal evidence based on residual abnormal events: For residual abnormalities whose duration reaches the corresponding measurement point's continuous judgment threshold, write them into residual abnormal evidence; For alarm signals, SOE signals, operation logs, inspection records, or maintenance work orders that exist within the same time window, associate them with residual abnormal evidence by time window to form a multi-source abnormal evidence set under the same equipment.

[0143] Calculate residual strength: Compare the current residual amplitude with the normal fluctuation range in the equipment operating condition baseline state. The greater the residual amplitude exceeds the normal fluctuation range and the longer the duration, the higher the residual strength. When the residual amplitude only slightly exceeds the normal fluctuation range and the duration is close to the continuous judgment threshold, the residual strength is marked as low. When the residual amplitude significantly exceeds the normal fluctuation range and continuously exceeds the continuous judgment threshold, the residual strength is marked as medium or high.

[0144] Determine the source credibility: Based on the degree of automatic data collection, degree of manual intervention, persistence of anomalies, and reliability of on-site verification, the source credibility of protection interlock signals is set to 0.95, the source credibility of continuous residual anomalies is set to 0.85, the source credibility of equipment alarm signals is set to 0.60, the source credibility of structured text records is set to 0.40, and the source credibility of single-point instantaneous alarms is set to 0.20; when structured text records have time inheritance markers, the source credibility of the corresponding text evidence is reduced;

[0145] Determine the time delay matching degree: Compare the occurrence time of the abnormal evidence with the propagation time delay range of adjacent device nodes in the device time delay causal propagation graph. When the occurrence time of the abnormal evidence falls within the propagation time delay range of the corresponding causal edge, the time delay matching degree is marked as a match; when the abnormal evidence is earlier than the minimum propagation time delay or later than the maximum propagation time delay, the time delay matching degree is marked as a mismatch; when the abnormal evidence lacks comparable upstream or downstream evidence, the time delay matching degree is marked as pending confirmation.

[0146] Conflict state generation: When an equipment alarm is triggered but no operating condition stripping residual appears at the corresponding associated measurement point, the evidence is written into the measurement point conflict state; when the text record shows that the equipment is normal but continuous residual abnormalities exist, the evidence is written into the text conflict state; when a single point anomaly is not supported by upstream measurement points, downstream measurement points, or similar redundant measurement points, the evidence is written into the single point conflict state.

[0147] Multi-source anomaly evidence within the same device and the same anomaly time window is encapsulated to generate a fault evidence passport. The fault evidence passport includes the fault evidence passport number, device affiliation, system affiliation, evidence type, anomaly start time, anomaly end time, residual strength, source credibility, time delay matching degree, conflict state, and associated measurement point set.

[0148] When multiple pieces of evidence point to the same device and the abnormal time windows overlap, they are merged into the same failure evidence passport, and the overall credibility is improved; when there is a conflict between multiple pieces of evidence, the conflict is retained and the overall weight of the failure evidence passport is reduced.

[0149] The generated fault evidence passports are sorted according to equipment ownership and anomaly start time, and an index relationship is established between the fault evidence passports and the operating condition stripping residual sequence, the multi-source state data frame of thermal power equipment, and the equipment time delay causal propagation graph for subsequent candidate fault propagation chain generation.

[0150] In this embodiment, step five specifically includes:

[0151] Read the fault evidence passport and the device delay causal propagation graph, and attach each fault evidence passport to the corresponding device node according to the device affiliation; when there are multiple fault evidence passports for the same device node, sort them according to the anomaly start time, source credibility and residual strength, and retain the evidence group with overlapping time windows as the anomaly evidence set of the device node.

[0152] Determine the starting point for path search: Select the device node with the earliest anomaly start time and the source credibility meets the conditions as the starting point; based on the lowest available credibility of the device alarm signal and continuous residual anomalies, set the starting point source credibility threshold to 0.60; when multiple device nodes meet the conditions at the same time, prioritize the device node with higher residual intensity and fewer conflict states.

[0153] Execution delay constraint path search: From the starting point, extend downstream along the causal edge in the device delay causal propagation graph. When the next device node joins the anomaly propagation path, it must simultaneously satisfy the following conditions: consistent propagation direction, existence of fault evidence passport and source credibility meets the conditions, anomaly start time falls within the propagation delay range of the corresponding causal edge, credibility of evidence source meets the conditions, and residual change has continuity.

[0154] Tolerance processing for propagation delay: Based on the unified sampling time scale and DCS resampling cycle, the delay matching tolerance is set to 5 to 10 seconds; when the deviation between the next device node's abnormal start time and the causal edge propagation delay boundary does not exceed this tolerance, it is still judged as a delay match; when the deviation exceeds this tolerance, the corresponding path is marked as a delay mismatch and propagation stops.

[0155] Evidence verification of abnormal propagation paths: If there is only a single point instantaneous alarm in the path, or if the subsequent device nodes in the path lack residual abnormalities, alarm signals, SOE signals or structured text support, the path expansion shall be stopped; if the path contains at least two types of evidence in continuous residual abnormalities and alarm signals, SOE signals or text records, the path shall be retained as a valid abnormal propagation path.

[0156] Perform counterfactual pruning: When the start time, fluctuation range, and duration of an abnormal path can be explained by the standard disturbance envelope corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, or environmental regulation in step two, the path is pruned; when the anomaly is triggered only by a low-confidence measuring point, and no synchronous anomalies occur at upstream measuring points, downstream measuring points, or similar redundant measuring points, the path is pruned as an abnormal measuring point path; when the direction of anomaly propagation is opposite to the direction of equipment connection, medium flow, or control loop, the path is pruned as a reverse invalid path.

[0157] Calculate the propagation chain score: Based on the source credibility, residual strength, time delay matching degree, conflict state, causal edge credibility, and evidence complementarity of the fault evidence passports in the path, the abnormal propagation path that has not been pruned is evaluated by weighted comprehensive evaluation; the higher the source credibility, the higher the residual strength, the more complete the time delay matching, the fewer the conflict state, the higher the edge credibility, and the more diverse the evidence types, the higher the propagation chain score.

[0158] The candidate fault propagation chains are output in descending order of propagation chain score. The candidate fault propagation chain includes root cause candidate nodes, propagation node sequence, affected nodes, associated fault evidence passport, propagation delay matching results, counterfactual pruning results, and propagation chain score.

[0159] In this embodiment, step six specifically includes:

[0160] The candidate fault propagation chain, fault evidence passport, counterfactual pruning results, and equipment procedure knowledge are input into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model to generate fault diagnosis results. The equipment procedure knowledge includes operating procedures, maintenance procedures, equipment manuals, protection interlock instructions, and historical handling cases. This model includes a procedure knowledge retrieval layer, a diagnostic task package construction layer, an evidence passport credibility gating layer, a structured diagnostic prompt layer, an equipment causal chain constraint reasoning layer, a counterfactual consistency verification layer, and a diagnostic conclusion binding layer.

[0161] The procedure knowledge retrieval layer receives the root cause candidate nodes, propagation node sequence, system attribution, and abnormal phenomenon fields in the candidate fault propagation chain. It establishes a local vector knowledge base for equipment procedure knowledge according to equipment KKS code, system attribution, procedure clause number, fault phenomenon, verification measurement point, handling action, and risk level in equipment procedure knowledge. During retrieval, it uses HNSW graph index to perform approximate nearest neighbor retrieval and recalls 5 to 15 procedure knowledge fragments according to cosine similarity.

[0162] The diagnostic task package construction layer generates a diagnostic task number and writes the diagnostic task number, candidate fault propagation chain number, root cause candidate node, propagation node sequence, propagation direction, propagation delay range, propagation chain score, fault evidence passport number, counterfactual pruning result and procedure clause number into the diagnostic task package to form the input context of this model.

[0163] The evidence passport credibility gating layer traverses the fault evidence passport, reading source credibility, residual strength, time delay matching degree, and conflict status. The residual strength is normalized using the Sigmoid function, with the normalization benchmark determined based on the distribution of real abnormal residual strength in historical correct diagnosis samples. The source credibility, the normalized residual strength value, and the time delay matching degree are multiplied to obtain the evidence input weight. When the conflict status is a measurement point conflict, text conflict, or single point conflict, the evidence input weight is reduced according to a preset attenuation coefficient.

[0164] Based on the weight distribution of highly credible evidence in historical correct diagnostic samples, the core evidence threshold is set to 0.70; fault evidence passports with an input weight greater than 0.70 are written into the core evidence area, fault evidence passports with an input weight between 0.20 and 0.70 are written into the auxiliary evidence area, and fault evidence passports with an input weight less than 0.20 are not included in the model inference input.

[0165] The structured diagnostic prompt layer constructs a fixed prompt template, which includes a candidate fault propagation chain area, a core evidence area, an auxiliary evidence area, a counterfactual pruning result area, an equipment procedure knowledge area, and an output format area. The candidate fault propagation chain area contains the equipment node sequence, causal edge direction, and propagation delay range. The core evidence area contains the fault evidence passport number, equipment attribution, residual strength, source credibility, delay matching degree, and associated measurement points. The counterfactual pruning result area contains the pruning results for load disturbance, soot blowing operation, start and stop of coal mill, coal quality fluctuation, environmental regulation, measurement point anomaly, and communication anomaly.

[0166] The device causal chain constraint inference layer uses candidate fault propagation chains as root cause inference boundaries, receives the content generated by the model in a streaming output manner, writes the output text into a sliding buffer, and identifies device entities through device dictionary matching, KKS encoding regular expression matching, and device alias reverse lookup; when the output content contains devices that are not in the candidate fault propagation chain, devices that are not bound to fault evidence passports, reverse propagation relationships, or propagation interpretations that exceed the propagation time delay range, generation is interrupted, and regeneration is performed based on the original diagnostic task package;

[0167] The counterfactual consistency verification layer constructs a counterfactual verification matrix, which includes load disturbance, soot blowing operation, start-up and shutdown of coal mill, coal quality fluctuation, environmental regulation, measurement point anomaly, and communication anomaly. Each item is written with whether it is excluded, the basis for exclusion, the binding fault evidence passport number, and the binding counterfactual pruning rule number. When there are unexcluded items or missing bound number items, the missing items are written into the feedback prompt and a regeneration is triggered. The number of regenerations is set to 1 to 3 times according to the real-time requirements of on-site diagnosis. If the verification still fails after exceeding the number of regenerations, the corresponding result is marked as pending manual review.

[0168] The diagnostic conclusion binding layer binds evidence to the output that has passed the verification. The root cause of the fault is bound to the fault evidence passport number in the core evidence area. The fault level is bound to the risk level in the protection interlock status, alarm level and equipment procedure knowledge. The scope of impact is bound to the affected nodes in the candidate fault propagation chain. The verification measurement point is bound to the set of associated measurement points. The handling suggestion is bound to the procedure clause number.

[0169] The diagnostic confidence is calculated based on the evidence input weight in the core evidence area, the candidate fault propagation chain score, the completeness of the counterfactual consistency check, and the degree of matching with equipment procedure knowledge, and the fault diagnosis result is generated. The fault diagnosis result includes the root cause of the fault, the fault level, the scope of impact, the diagnostic basis, the counterfactual exclusion items, the verification measurement points, the handling suggestions, the diagnostic confidence, and the diagnostic task number. The counterfactual exclusion items are generated by the counterfactual pruning results and the counterfactual consistency check results.

[0170] In this embodiment, step seven specifically includes:

[0171] Collect on-site verification results and maintenance feedback results. On-site verification results include verification measurement points, verification time, manually confirmed root causes, manually confirmed impact range, and manually confirmed disturbance type. Maintenance feedback results include maintenance work orders, defect elimination records, handling actions, parameter recovery time, and defect archiving results.

[0172] The on-site verification results, maintenance feedback results and fault diagnosis results are compared to generate diagnostic bias; when the manually confirmed root cause is inconsistent with the fault root cause, a root cause equipment bias is generated; when the actual abnormal propagation time exceeds the propagation delay range of the corresponding causal edge, a propagation delay bias is generated; when low-confidence fault evidence affects the diagnostic conclusion, an evidence weight bias is generated; when normal disturbances are not stripped or real faults are mistakenly pruned, a counterfactual pruning bias is generated.

[0173] The equipment operating condition baseline status is updated based on the diagnostic deviation; the normal operating segments confirmed on-site are added to the historical normal operating condition segment library, similar operating condition segments are re-matched, and the baseline value, normal upper limit, normal lower limit, and duration threshold of the corresponding measuring points are updated using a sliding window weighted statistical method; samples with low confidence markers, dormant markers, and over-range markers are not included in the update;

[0174] The propagation delay and edge credibility are updated based on the propagation delay deviation: the start time of the upstream and downstream equipment abnormalities confirmed on-site are read, the actual propagation time difference is calculated, and the minimum and maximum propagation delays of the corresponding causal edges are corrected using an exponential moving average method; when a causal edge is confirmed on-site to participate in the propagation of a real fault, the edge credibility is increased; when a causal edge is repeatedly pruned by counterfactual methods or excluded on-site, the edge credibility is decreased.

[0175] The evidence weight is updated based on the evidence weight deviation: when a single-point instantaneous alarm, conflicting text records, or low-confidence measurement points lead to incorrect diagnosis, the basic source credibility of the corresponding evidence type under the same equipment and the same fault scenario is reduced; when SOE protection interlock signals, continuous residual abnormalities, or maintenance work orders are repeatedly confirmed to be valid on-site, the evidence input weight of the corresponding evidence type is increased.

[0176] Update the counterfactual pruning rules based on counterfactual pruning deviations: When load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, or environmental regulation are identified as normal disturbances on-site but not pruned, add exclusion conditions for the corresponding disturbance envelope; when a real fault path is misjudged as a normal disturbance path, narrow the corresponding standard disturbance envelope and reduce the corresponding pruning intensity.

[0177] The diagnostic constraints of the DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints are updated based on the maintenance feedback results; the fault phenomena, root causes, fault evidence passport numbers, candidate fault propagation chain numbers, handling actions, and recovery results confirmed on site are written into the thermal power fault case index library; the thermal power fault case index library serves as the source of historical handling cases in equipment procedure knowledge, and generates new procedure knowledge matching tags, counterfactual pruning rules, and handling rules;

[0178] The updated equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and diagnostic constraints are written into the closed-loop update record and bound to the diagnostic task number, fault evidence passport number, and candidate fault propagation chain number for use in the next fault diagnosis.

[0179] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the fault diagnosis scenario of the pulverizing system, flue gas system, water supply system and environmental protection system of a 1000MW ultra-supercritical thermal power generating unit.

[0180] The unit experiences common disturbances such as AGC load increases / decreases, soot blowing, mill start / stop, coal quality fluctuations, and denitrification adjustments. 90 consecutive days of operational data were selected for verification. DCS operational data was resampled every 5 seconds, SIS historical data every 10 seconds, and protection interlock signals were written into time windows according to event timestamps. A KKS equipment coding table covering 1260 equipment nodes and 8650 measuring points was established. The load change rate threshold was set to 2 MW / min, the coal lower heating value fluctuation threshold was set to 1.5 MJ / kg, the soot blowing disturbance envelope was set to 5 to 15 minutes, the mill start / stop disturbance envelope was set to 10 to 30 minutes, the starting point source credibility threshold was set to 0.60, and the core evidence threshold was set to 0.70. After reviewing on-site maintenance records, the sample included 32 actual equipment failures and 46 normal operating condition disturbances.

[0181] During implementation, dynamic time warping and time synchronization, KKS equipment coding mapping, measurement point attribution binding, anomaly marker writing, and text quadruple extraction are first performed on DCS operation data, SIS historical data, protection interlock signals, equipment alarm records, operation logs, and maintenance work orders to generate multi-source status data frames for thermal power equipment. Then, the operating condition identification features are used to match historical normal operating condition segments to calculate the equipment operating condition baseline state. The interpretable deviation is stripped through standard perturbation envelope to obtain the operating condition stripping residual sequence.

[0182] Subsequently, based on equipment connections, media flow direction, control loops, protection interlocks, and historical anomaly propagation time differences, a time delay causal propagation diagram of the equipment is generated. Residual anomalies, alarms, SOE signals, and text records are encapsulated as fault evidence passports. Time delay constraint path search and counterfactual pruning are performed. Finally, the evidence chain constraint type DeepSeek-R1 thermal power fault diagnosis model is input, and the fault root cause, scope of impact, verification measurement points, handling suggestions, and diagnostic confidence level are output.

[0183] Table 1 Comparative Verification Data of Fault Diagnosis for Thermal Power Generation Equipment

[0184] Method A 30 24 62.5 1.6 57.7 Method B 25 11 72.0 5.8 67.9 Method C 27 10 78.1 4.5 74.4 Method D 28 8 81.3 4.8 79.5 Method of the present invention 31 4 93.8 3.6 92.3

[0185] Method A is a traditional threshold alarm method, Method B is a knowledge base rule diagnosis method, Method C is a general large language model diagnosis method, and Method D is a DeepSeek-R1 diagnosis method without evidence chain constraints.

[0186] As shown in Table 1, Method A has the fastest average diagnostic time of 1.6 minutes, and it identifies 30 actual faults, indicating that the traditional threshold alarm method has a fast response capability to sudden limit exceedances, protection actions, and obvious parameter anomalies. However, it has 24 false alarms due to normal disturbances. The main reason is that this method cannot distinguish between explainable disturbances such as soot blowing, starting and stopping of coal mills, and rapid load changes, resulting in a detection and re-verification consistency rate of only 57.7%.

[0187] Method B reduced the number of false alarms for normal disturbances to 11, which is better than Method A, indicating that the knowledge base rule diagnosis method has a good rule filtering capability for typical faults and typical operational disturbances; however, it only identified 25 actual faults and the average diagnosis time was 5.8 minutes, indicating that the fixed rules are insufficient in complex scenarios such as coal quality fluctuations superimposed with environmental regulation and air and flue gas system coupling, and the manual rule matching process is relatively slow.

[0188] Method C achieved a root cause localization accuracy of 78.1% and a repair and verification consistency rate of 74.4%, indicating that the general large language model diagnostic method can improve fault interpretation capabilities by utilizing textual information in operation logs, inspection records, and maintenance work orders. However, it lacks equipment delay causal propagation diagrams and fault evidence passport constraints, which still easily lead to problems such as unstable conclusion basis and downstream phenomena being interpreted as root causes.

[0189] Method D further improved the number of real faults identified and the root cause localization accuracy compared to Method C, indicating that DeepSeek-R1 inference models have stronger reasoning ability for complex fault chains; however, due to the lack of evidence passport credibility gating, equipment causal chain constraint reasoning and counterfactual consistency verification, there were still 8 false alarms due to normal disturbances, and the inspection and repair consistency rate was 79.5%.

[0190] The method of this invention identified 31 real faults, missing only one minor feedwater regulation anomaly; the number of false alarms due to normal disturbances was reduced to 4, the root cause localization accuracy reached 93.8%, and the consistency rate of inspection, repair, and verification reached 92.3%. These results demonstrate that the residual sequence stripping under operating conditions can reduce false alarms due to normal disturbances, the equipment delay causal propagation graph and delay constraint path search can constrain the direction of anomaly propagation, and the fault evidence passport and evidence chain constraint-type DeepSeek-R1 thermal power fault diagnosis model can improve the traceability of fault root causes, impact range, verification measurement points, and remediation recommendations.

[0191] As can be seen from the embodiments, the present invention achieves a good balance between diagnostic speed, false alarm control, root cause localization and consistency of on-site verification, and can provide verifiable, traceable and closed-loop updated fault diagnosis results for thermal power generating units under conditions of deep peak shaving, multi-disturbance coupling and multi-source data coexistence.

[0192] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for thermal power generation equipment based on a large language model, characterized in that, Includes the following steps: Step 1: Collect operating data of thermal power generating units, perform dynamic time warping and time calibration, KKS equipment coding mapping, measurement point attribution binding, anomaly marker writing and text quadruple extraction to generate multi-source status data frames for thermal power equipment; Step 2: Extract operating condition identification features from the multi-source state data frames of thermal power equipment, match historical normal operating condition segments and calculate the equipment operating condition baseline state, peel off interpretable deviations based on standard disturbance envelopes, and generate an operating condition peeling residual sequence. Step 3: Configure device nodes, causal edges, propagation direction, propagation delay, and edge reliability according to device connection, media flow direction, control loop, protection interlock, and historical anomaly propagation time difference to generate a device delay causal propagation graph; Step 4: Extract abnormal evidence based on the residual sequence stripping under operating conditions, and encapsulate it according to equipment ownership, residual strength, source credibility, time delay matching degree and conflict state to generate fault evidence passport; Step 5: Attach the fault evidence passport to the device delay causal propagation graph, perform delay constraint path search and counterfactual pruning, and generate candidate fault propagation chains; Step 6: Input the candidate fault propagation chain, fault evidence passport, counterfactual pruning results and equipment procedure knowledge into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model. Through evidence passport credibility gating, equipment causal chain constraint reasoning, counterfactual consistency verification and diagnostic conclusion evidence binding, generate fault diagnosis results. Step 7: Generate diagnostic deviations based on the on-site verification results and maintenance feedback results, and update the equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and evidence chain constraint type DeepSeek-R1 thermal power fault diagnosis model diagnostic constraints.

2. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step one specifically includes: Collect operating data of thermal power generating units and establish a data source cache queue according to data source, acquisition channel and original acquisition time stamp; Using the unit's AGC command change points, load change points, auxiliary machine start / stop signals, protection action signals, and major regulation change points as time calibration anchor points, dynamic time warping and time calibration are performed to generate a unified sampling time scale; Resampling is performed on continuous measurement point data, protection events, alarm events, and text events according to a unified sampling time standard; Establish a KKS device coding table, a device alias dictionary, a device hierarchy table, and a system affiliation table. Perform character normalization, keyword window matching, and edit distance matching on the device names in the text records to complete the KKS device coding mapping. The measurement point affiliation binding is performed based on the measurement point number, measurement point name, physical quantity type, equipment, system, and associated equipment fields in the measurement point configuration table. For data with empty values, exceeding the range, in a dormant state, single-point mutation, or manually entered data, write missing, over-range, dormant, low-confidence, and manually sourced data respectively; Text quadtuples are extracted from operation logs, inspection records, and maintenance work orders to generate quadtuples of "equipment - abnormal phenomenon - handling action - event time"; The processed data is aggregated according to a unified sampling timescale to generate a multi-source status data frame for thermal power equipment. The multi-source status data frame for thermal power equipment includes equipment identifier, measurement point identifier, system identifier, unified sampling timescale, status value, data source, data type, confidence flag, anomaly flag, text quadruple field, and time inheritance flag.

3. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step two specifically includes: Operating condition identification features are extracted from the multi-source status data frames of thermal power equipment. These features include unit load, load change rate, coal mill operating combination, total air volume, main steam pressure, low calorific value of coal, soot blowing status, denitrification ammonia injection volume, and ambient temperature. Based on the operating condition identification characteristics, the following conditions are identified: load increase condition, load decrease condition, coal mill start-up and shutdown condition, soot blowing disturbance condition, coal quality fluctuation condition, and environmental protection regulation condition. Match historical normal operating condition segments from the historical normal operating database and remove measurement point samples with missing, over-range, dormant, and low confidence markers; The baseline value of the measuring point is determined based on the mean or median of the corresponding measuring point within the historical normal operating condition segment. The upper limit and lower limit of normal fluctuation are determined based on the standard deviation range or quantile range of the corresponding measuring point. The continuous judgment threshold is determined in combination with the equipment response inertia to generate the equipment operating condition baseline state. Standard disturbance envelopes are established based on load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, and environmental regulation. The current equipment status is compared with the equipment operating condition baseline status to remove interpretable deviations, retain uninterpretable deviations, and generate an operating condition stripping residual sequence in chronological order.

4. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 3, characterized in that, The process of comparing the current equipment state with the equipment operating condition baseline state, removing interpretable deviations, retaining uninterpretable deviations, and generating an operating condition stripping residual sequence in chronological order specifically includes: When the load change rate exceeds the preset load change rate threshold and the AGC command changes continuously, the current time window will be marked as either a load increase condition or a load decrease condition. When the coal mill start / stop signal is valid, mark the current time window as the coal mill start / stop condition; When the soot blowing valve action signal is valid, mark the current time window as a soot blowing disturbance condition; When the change in the lower heating value of adjacent coal quality testing cycles exceeds the preset coal quality fluctuation threshold, the current time window is marked as a coal quality fluctuation condition. When the amount of ammonia injected for denitrification, the inlet parameters for desulfurization, or the environmental regulation instructions deviate continuously from the stable range, the current time window will be marked as an environmental regulation condition. Based on historical normal disturbance events, the standard disturbance envelopes corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, and environmental regulation are determined respectively. When the current deviation falls within the corresponding standard disturbance envelope and the duration does not exceed the allowable duration, the current deviation is determined to be an interpretable deviation and is removed. When the current deviation exceeds the normal fluctuation range of the equipment's operating condition reference state, the duration reaches the corresponding measurement point's continuous judgment threshold, and it cannot be explained by the standard disturbance envelope, the current deviation will be retained as an unexplainable deviation. Unexplainable deviations are aggregated according to equipment identification, measurement point identification, unified sampling time scale, deviation direction, deviation amplitude, duration, and confidence level to generate a working condition stripping residual sequence.

5. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step three specifically includes: Construct a set of equipment nodes based on the equipment ledger, P&ID drawings, DCS control logic, and protection interlocking table of the thermal power generating unit; Construct a set of causal edges, and configure the propagation direction, propagation delay, and edge credibility for each causal edge; The propagation direction is determined based on equipment connection, medium flow direction, control circuit and protection interlock, forming a pulverizing combustion propagation side, a wind-powder propagation side, a wind and smoke regulation side, a steam-water regulation side and an environmental protection coupling side; Based on the occurrence time of the residual loss of adjacent device nodes in historical abnormal events, the propagation time difference between the start time of the abnormality of the upstream node and the start time of the abnormality of the downstream node is determined. Statistical distribution analysis is performed on multiple propagation time differences corresponding to the same causal edge. Outlier sample elimination method based on statistical distribution is used to determine the propagation time difference distribution interval, and propagation delay is generated based on the propagation time difference distribution interval. The edge credibility is configured based on the clarity of the equipment mechanism, the number of historical abnormal samples, the concentration of the propagation time difference distribution, and the on-site verification results. Write device nodes, causal edges, propagation directions, propagation delays, and edge credibility into the graph structure to generate a device delay causal propagation graph.

6. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step four specifically includes: Based on the residual sequence stripping under working conditions, extract abnormal time windows where the residual is not zero and the duration reaches the continuous judgment threshold of the corresponding measuring point, and merge abnormalities in the same device whose consecutive abnormal time window intervals do not exceed the preset interval threshold into the same residual abnormal event. Extract abnormal evidence based on residual abnormal events, and associate alarm signals, SOE signals, operation logs, inspection records or maintenance work orders within the same time window with residual abnormal evidence to form a multi-source abnormal evidence set; The residual strength is calculated based on the degree and duration of the residual amplitude exceeding the normal fluctuation range of the equipment's operating condition reference state; The source credibility is determined based on the degree of automatic data collection, the degree of human intervention, the persistence of anomalies, and the reliability of on-site verification. The source credibility of the corresponding text evidence is reduced when the text quadruple field has a time inheritance marker. The delay matching degree is determined based on the matching relationship between the occurrence time of abnormal evidence and the propagation delay range of adjacent device nodes in the device delay causal propagation diagram; Conflict states are generated based on the consistency between alarms, residuals, text records, and associated measurement points. Encapsulate multi-source anomaly evidence within the same device and the same anomaly time window to generate a fault evidence passport. The fault evidence passport includes fault evidence passport number, device affiliation, system affiliation, evidence type, anomaly start time, anomaly end time, residual strength, source credibility, time delay matching degree, conflict state, and associated measurement point set. The fault evidence passports are sorted according to equipment ownership and anomaly start time, and an index relationship is established between the fault evidence passports and the operating condition stripping residual sequence, the multi-source state data frame of thermal power equipment, and the equipment delay causal propagation graph.

7. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step five specifically includes: The fault evidence passport is attached to the corresponding device node in the device delay causal propagation graph according to the device affiliation; Select the device node with the earliest anomaly start time and whose source credibility meets the preset conditions as the starting point for path search. Extend downstream from the starting point of the path search along the causal edges in the device delay causal propagation graph, and perform a delay-constrained path search; When the next device node joins the abnormal propagation path, it must simultaneously meet the following conditions: consistent propagation direction, existence of fault evidence, abnormal start time falling within the propagation delay range of the corresponding causal edge, source credibility meeting preset conditions, and residual abnormal time windows being continuously connected within the corresponding propagation delay range. When the start time, fluctuation range, and duration of an abnormal path can be explained by the standard disturbance envelope corresponding to load changes, soot blowing operations, start-up and shutdown of coal mills, coal quality fluctuations, or environmental regulation, the abnormal path shall be removed. When an anomaly is triggered only by a low-confidence measurement point and no synchronous anomaly occurs in the upstream measurement point, downstream measurement point, or similar redundant measurement point, the anomaly path is removed as a measurement point anomaly path. When the abnormal propagation direction is opposite to the direction of equipment connection, medium flow, or control loop, the abnormal path is cut off as a path that is not valid in the reverse direction. The propagation chain score is calculated based on source credibility, residual strength, delay matching degree, conflict state, causal edge credibility, and evidence complementarity. The evidence complementarity is determined based on the number of evidence types in the candidate fault propagation chain, such as residual abnormalities, alarm signals, SOE signals, operation logs, inspection records, or maintenance work orders. The candidate fault propagation chains are output in descending order of propagation chain score. The candidate fault propagation chain includes root cause candidate nodes, propagation node sequence, affected nodes, associated fault evidence passport, propagation delay matching results, counterfactual pruning results, and propagation chain score.

8. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step six specifically includes: The candidate fault propagation chain, fault evidence passport, counterfactual pruning results and equipment procedure knowledge are input into the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model. Equipment procedures include operating procedures, maintenance procedures, equipment manuals, protection interlock instructions, and historical case studies. The DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints includes a procedure knowledge retrieval layer, a diagnosis task package construction layer, an evidence passport credibility gating layer, a structured diagnosis prompt layer, an equipment causal chain constraint reasoning layer, a counterfactual consistency verification layer, and a diagnosis conclusion evidence binding layer. The procedure knowledge retrieval layer retrieves procedure knowledge fragments from equipment procedure knowledge based on the root cause candidate node, propagation node sequence, system attribution, and abnormal phenomenon fields in the candidate fault propagation chain. The diagnostic task package construction layer generates a diagnostic task number and writes the diagnostic task number, candidate fault propagation chain number, root cause candidate node, propagation node sequence, propagation direction, propagation delay range, propagation chain score, fault evidence passport number, counterfactual pruning result, and procedure clause number into the diagnostic task package. The evidence passport credibility gating layer generates evidence input weights as evidence weights based on source credibility, residual strength, time delay matching degree, and conflict state, and divides the faulty evidence passport into core evidence area and auxiliary evidence area; The structured diagnostic prompt layer constructs fixed prompt templates based on candidate fault propagation chains, core evidence areas, auxiliary evidence areas, counterfactual pruning results, and equipment procedure knowledge; The equipment causal chain constraint inference layer verifies the equipment entities and propagation relationships in the output of the evidence chain constraint DeepSeek-R1 thermal power fault diagnosis model based on the candidate fault propagation chain. The counterfactual consistency verification layer generates a counterfactual verification matrix based on the counterfactual pruning results and verifies the output of the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model. The diagnostic conclusion evidence binding layer binds evidence to the output that has passed the verification to generate fault diagnosis results.

9. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 8, characterized in that, The evidence passport credibility gating layer normalizes the residual intensity and generates evidence input weights as evidence weights based on source credibility, normalized residual intensity, time delay matching degree, and conflict state. Based on the evidence weight and the preset evidence partitioning threshold, the fault evidence passport is divided into the core evidence area, the auxiliary evidence area, and the fault evidence passport that has not entered the model inference input. The structured diagnostic prompt layer incorporates candidate fault propagation chain area, core evidence area, auxiliary evidence area, counterfactual pruning result area, equipment procedure knowledge area, and output format area into a fixed prompt template; The device causal chain constraint reasoning layer uses candidate fault propagation chains as root cause reasoning boundaries and adopts a streaming interception method for output verification. When the output of the evidence chain constraint DeepSeek-R1 thermal power fault diagnosis model contains non-candidate fault propagation chain devices, devices without bound fault evidence passports, reverse propagation relationships, or propagation interpretations that exceed the propagation delay range, the generation is interrupted and regenerated based on the original diagnostic task package. The counterfactual consistency verification layer constructs a counterfactual verification matrix, which includes load disturbance, soot blowing operation, start and stop of coal mill, coal quality fluctuation, environmental protection regulation, measurement point anomaly and communication anomaly. Each item is written with whether to exclude, the basis for exclusion, the fault evidence passport number and the counterfactual pruning rule number. When the counterfactual verification matrix contains unexcluded items or missing bound number items, the missing items will be written into the feedback prompt and the evidence chain-constrained DeepSeek-R1 thermal power fault diagnosis model will be regenerated. The diagnostic conclusion evidence binding layer binds the root cause of the fault to the fault evidence passport number in the core evidence area, binds the fault level to the risk level in the protection interlock status, alarm level and equipment procedure knowledge, binds the scope of impact to the affected nodes in the candidate fault propagation chain, binds the verification measurement points to the associated measurement point set, and binds the handling suggestions to the procedure clause number. The diagnostic confidence is calculated based on the evidence weight of the core evidence area, the score of the candidate fault propagation chain, the completeness of the counterfactual consistency check, and the degree of matching of equipment procedure knowledge, and the fault diagnosis result is generated.

10. The method for fault diagnosis of thermal power generation equipment based on a large language model according to claim 1, characterized in that, Step seven specifically includes: Collect on-site verification results and maintenance feedback results, and compare the on-site verification results, maintenance feedback results and fault diagnosis results to generate diagnostic deviations; When the manually confirmed root cause is inconsistent with the fault root cause, a root cause device bias is generated; when the actual abnormal propagation time exceeds the propagation delay range of the corresponding causal edge, a propagation delay bias is generated; when low-confidence fault evidence affects the diagnostic conclusion, an evidence weight bias is generated; when normal disturbances are not stripped away or the real fault propagation path is mistakenly pruned by counterfactual pruning, a counterfactual pruning bias is generated. The equipment operating condition baseline status is updated based on the diagnostic deviation. Normal operating segments confirmed on-site are added to the historical normal operating condition segment library. Historical normal operating condition segments are re-matched, and the baseline value, normal upper limit, normal lower limit, and duration threshold of the corresponding measuring point are updated using a sliding window weighted statistical method. The propagation delay and edge confidence are updated based on the propagation delay deviation. The minimum and maximum propagation delays of the corresponding causal edges are corrected using an exponential moving average method. The edge confidence of the corresponding causal edges is increased or decreased based on the on-site confirmation results. The evidence weights are updated based on the evidence weight bias, and the counterfactual pruning rules are updated based on the counterfactual pruning bias. Based on the maintenance feedback results, the diagnostic constraints of the DeepSeek-R1 thermal power fault diagnosis model with evidence chain constraints are updated. The fault phenomena, root causes, fault evidence passport numbers, candidate fault propagation chain numbers, handling actions and recovery results confirmed on site are written into the thermal power fault case index library. The thermal power fault case index library serves as a source of historical handling cases in equipment procedure knowledge. The updated equipment operating condition baseline state, propagation delay, edge credibility, evidence weight, counterfactual pruning rules, and diagnostic constraints are written into the closed-loop update record and bound to the diagnostic task number, fault evidence passport number, and candidate fault propagation chain number.

Images