Intelligent diagnosis method and system based on remote monitoring of power plant equipment

Abstract

The application discloses an intelligent diagnosis method and system based on power plant equipment remote monitoring, and belongs to the technical field of power plant equipment intelligent diagnosis. The method adopts an edge-cloud collaborative architecture, an edge gateway completes multi-source data preprocessing and fine-grained working condition marking, and disturbance false alarms are inhibited; a bidirectional constraint mechanism-AI hybrid diagnosis model and a working condition correlation knowledge base with dynamic credibility weight are constructed; self-adaptive alarm thresholds are generated according to working conditions to realize hierarchical early warning; chain fault reasoning tracing is driven by early warning, and an auditable reasoning link and maintenance suggestion are generated; diagnosis review is completed by using digital twin bidirectional simulation; and automatic sample classification, model and knowledge base incremental iterative updating are realized based on actual maintenance results. The application solves the problems of traditional static threshold false and missed alarms, poor diagnosis explainability, no self-learning closed loop and the like, realizes early warning and root cause positioning of power plant equipment hidden dangers, adapts to group-level remote predictive operation and maintenance, and reduces the risk of unit unplanned shutdown.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for power plant equipment, and more specifically, to an intelligent diagnostic method and system based on remote monitoring of power plant equipment. Background Technology

[0002] Power plant generator units comprise a large number of critical equipment, including steam turbines, boilers, generators, fans, and feedwater pumps. The safe operation of these equipment directly determines the stable power supply of the units. Traditional operation and maintenance models mainly rely on local DCS monitoring and regular manual inspections. Monitoring methods are limited to the local site and it is difficult to achieve remote centralized monitoring across plant areas. Fault diagnosis often uses static threshold alarms, which are prone to generating a large number of false alarms and missed alarms when operating conditions change. Multi-source heterogeneous data (vibration, temperature, pressure, oil, electrical quantities) cannot be effectively integrated. After a fault occurs, troubleshooting relies on engineers' experience, which is time-consuming and makes it difficult to achieve early warning of potential hazards and predictive maintenance.

[0003] Most existing remote diagnostic solutions only complete data uploading and simple over-limit alarms, lacking intelligent reasoning capabilities that combine mechanistic knowledge with data-driven approaches. The interpretability of diagnostic results is poor, there is a lack of complete closed-loop handling suggestions, and the models cannot adaptively update with equipment aging and changes in operating conditions, thus limiting their effectiveness in actual engineering implementation. Summary of the Invention

[0004] One technical problem to be solved The purpose of this invention is to address the problems that most existing remote diagnostic solutions only complete data uploading and simple over-limit alarms, lack the intelligent reasoning ability that combines mechanistic knowledge and data-driven approaches, have poor interpretability of diagnostic results, lack complete closed-loop handling suggestions, and the models cannot adaptively update with equipment aging and changes in operating conditions, thus limiting their effectiveness in practical engineering implementation.

[0005] Two technical solutions To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent diagnostic method based on remote monitoring of power plant equipment includes: Step S1: Multi-source data acquisition and data preprocessing: Collect multi-source operating parameters of core equipment in the power plant, and complete data cleaning, noise reduction, and normalization preprocessing through the edge gateway. Based on the processed parameters, perform fine-grained operating condition classification and labeling. Step S2: Construct a bidirectional constraint mechanism-AI hybrid diagnostic model and a working condition related knowledge base: Build a device mechanism knowledge base, establish a fault-symptom mapping relationship and fault tree logic, and bind each fault symptom to a dynamic credibility weight that is strongly correlated with the working condition type and signal duration in S1, which is different from a fixed weight knowledge base. Step S3, Adaptive Dynamic Monitoring and Hierarchical Early Warning: The cloud receives data uploaded from the edge and operating condition tags, combines the operating condition tags to generate adaptive alarm thresholds, calculates equipment health indicators, and triggers hierarchical early warnings; Step S4, Auditable Chain-like Intelligent Fault Diagnosis and Root Cause Tracing: Driven by the S3 early warning event, it integrates the S2 model algorithm and S1 operating condition data to complete fault reasoning, confidence correction, root cause location, and generate a traceable reasoning chain. Step S5, Digital Twin Two-Way Collaborative Visualization and Diagnostic Verification: It forward-embeds the diagnostic conclusions of S4 to realize visualization, while in reverse outputting simulation data to assist S4 in secondary verification, forming a diagnostic enhancement closed loop, which is different from traditional one-way visualization. Step S6: Fully automated closed-loop iteration based on inference chain backtracking: Relying on the auditable inference chain of S4, the review conclusion of S5, and the actual results of on-site maintenance, the sample screening and model update are automatically completed, and the S1 working condition filtering rules, S2 model and knowledge base parameters are optimized in reverse iteration to achieve continuous evolution of the entire system.

[0006] Further limitations on the above technical solution: In step S1, the collected multi-source operating parameters of the core power plant equipment include equipment vibration, temperature, pressure, flow rate, electrical quantity, oil chromatography, unit operation instructions, and load parameters; the fine-grained operating condition classification and labeling specifically involves performing edge lightweight pre-diagnosis based on precise operating condition labels: when short-term parameter offsets caused by human operation, transient fluctuations, or external disturbances are identified, they are determined to be invalid anomalies, alarm uploads are suppressed, and only steady-state valid anomaly data, continuous transient anomaly data, and corresponding operating condition labels are encrypted and uploaded to the cloud, while invalid disturbance data is stored locally. Based on the parameter change rate, unit instructions, and real-time load, four categories of fine-grained operating condition labeling are completed: steady-state / start-stop transient / operational disturbance / abnormal disturbance.

[0007] Further limitations on the above technical solution: In step S2, the bidirectional constraint mechanism-AI hybrid diagnostic model has a training phase and an inference phase. In the training phase, the fault symptom logic of the mechanism knowledge base is embedded as a loss constraint term into the loss function of the AI ​​sub-model to realize the mechanism knowledge to constrain the training of the AI ​​model. In the inference phase, the AI ​​sub-model outputs a set of candidate faults and preliminary confidence levels. The mechanism inference sub-model combines the working condition label and signal duration to obtain the evidence credibility weights, performs multi-evidence time-series weighted causal inference on the candidate faults, and corrects the confidence levels of the candidate faults.

[0008] Further limitations on the above technical solution: In step S3, the specific method for generating adaptive alarm thresholds is as follows: The cloud receives real-time data and fine-grained operating condition tags uploaded by S1, calls the threshold parameters of the corresponding operating condition in the knowledge base of S2, dynamically generates alarm thresholds that are adapted to the current operating status, and calculates the overall health of the equipment in real time. The specific method for graded early warning is as follows: based on the degree of health decay, the duration of parameter exceeding the limit, and the type of operating condition, four levels of early warning are distinguished: prompt, general, important, and emergency. Abnormalities caused by transient disturbances are automatically downgraded and marked to avoid false triggering of diagnosis.

[0009] Further clarification of the above technical solution: In step S4, the specific steps for auditable chain-like intelligent fault diagnosis and root cause tracing are as follows: S401, Pre-input Dependency: Receives S3 warning trigger signal, abnormal parameters, operating condition label and abnormal duration; S402, AI Initial Inference: Call the AI ​​model trained in S2 and output the initial candidate fault set and basic confidence level; S403, Mechanism Weighted Correction: Call the dynamic confidence weight of S2, combine the working condition characteristics of S1 and the duration of abnormality, perform time-series weighted causal reasoning on the initial AI result, correct the confidence, eliminate working condition interference misjudgment, and lock the root cause of the fault. S404, Inference Chain Generation: Records the entire chain from "Operating Condition Label → AI Initial Output → Mechanism Weight Correction → Final Fault Conclusion" to generate an auditable diagnostic report and simultaneously output maintenance and handling suggestions; S405, Suspected Fault Backup: When the confidence level is insufficient, output the investigation priority based on the weighted weight to guide manual review.

[0010] Further clarification of the above technical solution: In step S5, the visualization implementation of the diagnostic conclusion of S4 is specifically achieved by receiving the root cause of the fault and component location information of S4, accurately highlighting the faulty component in the digital twin model, and dynamically displaying the changes in equipment status in combination with the real-time operation data of S1. The reverse output simulation data assists the S4 secondary verification in the following ways: based on the current S1 operating condition and the S4 fault type, a fault simulation deduction is carried out to generate standard fault simulation features, which are then sent back to the S4 diagnostic module to perform a secondary verification of the original diagnostic results and improve the diagnostic accuracy.

[0011] Further limitations on the above technical solution: In step S6, the specific steps of the fully automatic closed-loop iteration based on inference chain backtracking are as follows: Step S601, Data Comparison: Enter the actual fault results from on-site inspection and automatically compare them with the complete inference chain stored in S4 and the review conclusion in S5; Step S602, Automatic Sample Classification: No manual labeling required; automatically distinguishes between accurately diagnosed samples, misdiagnosed samples, and suspected faulty samples. Step S603, Closed-loop update: Use the new samples to perform incremental training, and simultaneously update the weights of the S2 mechanism knowledge base and the AI ​​model parameters, while optimizing the S1 edge case pre-filtering rules. Step S604, Version Distribution: The iterated new model, new operating condition rules, and new weight parameters are distributed to the edge and cloud to achieve the next round of monitoring and diagnostic accuracy upgrade.

[0012] An intelligent diagnostic system based on remote monitoring of power plant equipment includes, The on-site edge perception layer includes sensors, DCS interface units, and edge gateway modules. Sensors are installed at core monitoring points of various equipment in the power plant, with at least two sensors per parameter monitoring point for parameter complementarity. Sensors are used to collect operating signals such as equipment temperature, vibration, pressure, electrical, and oil parameters. The DCS interface units are used to interface with the power plant's existing control system to read unit operating conditions and parameters. The edge gateway module has a built-in edge preprocessing unit, local caching unit, and edge preliminary anomaly detection unit. It performs raw data noise reduction, outlier removal, operating condition labeling, local caching, preliminary anomaly screening, and encrypted structured data uploading to the cloud. The network transmission layer includes industrial intranets and 5G / 4G communication units, providing encrypted data transmission channels, supporting breakpoint resume, and enabling secure data interaction between the edge perception layer and the cloud platform layer; The cloud platform layer includes a database module, a knowledge base management module, an intelligent diagnostic calculation module, an early warning and response module, and a model iteration and training module; The application presentation layer includes a remote monitoring web client and a mobile operation and maintenance terminal.

[0013] Further limitations on the above technical solution: Database module: stores basic device files, real-time time-series data, historical fault data, and maintenance records; Knowledge base management module: maintains equipment mechanism knowledge base, fault tree, fault case library, and maintenance procedure library, and supports operation and maintenance personnel to visually edit and update knowledge; Intelligent diagnostic computing module: Loads the hybrid intelligent diagnostic model, performs feature extraction, fault reasoning, root cause tracing, outputs fault type, confidence level, and maintenance suggestions, performs multi-evidence fusion diagnosis, assigns weights to multi-dimensional evidence such as vibration, temperature, oil, and electrical components, comprehensively calculates fault confidence level, and avoids misjudgment based on a single parameter; Early warning and handling module: realizes adaptive threshold calculation based on operating conditions, hierarchical early warning, and completes alarm message distribution; Model Iteration Training Module: Receives real fault samples after maintenance, performs incremental training and updates on the hybrid intelligent diagnostic model, and optimizes model parameters.

[0014] Further limitations on the above technical solution: The cloud platform layer also sets up a digital twin simulation unit to construct a digital twin of the equipment and map real-time data to the twin model interface to display the component status.

[0015] Three beneficial effects Compared with existing technologies, the present invention provides an intelligent diagnostic method and system based on remote monitoring of power plant equipment, which has the following beneficial effects: 1. This invention adopts an edge-cloud collaborative architecture, where preprocessing and local caching are completed at the edge, reducing the transmission pressure on the cloud, balancing real-time performance and industrial network security, and adapting to the complex network conditions at power plant sites.

[0016] 2. The system adopts a hybrid diagnostic approach combining a mechanism knowledge base and an AI model to overcome the shortcomings of pure data-driven models, such as poor interpretability and weak small-sample fault identification. By combining fault tree causal reasoning, the system can locate the root cause of the fault, output complete maintenance and treatment suggestions, and ensure that the diagnostic results are traceable.

[0017] 3. Adopting dynamic adaptive alarm thresholds based on operating conditions, which differs from traditional static thresholds, significantly reduces false alarms and missed alarms under varying load conditions, enables early warning of potential equipment hazards, supports predictive maintenance, and reduces the risk of unplanned unit shutdowns.

[0018] 4. It has closed-loop self-learning capability, and real fault samples after maintenance are fed back to iterate the model and knowledge base, and the diagnostic performance continues to improve as the running time increases.

[0019] 5. The system supports centralized remote monitoring of multiple power plants, visual display via Web and mobile terminals, and is adapted to group-level remote operation and maintenance scenarios, thereby improving the level of digitalization and intelligence in power plant operation and maintenance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a detailed flowchart of step S4 of the present invention; Figure 3 This is a detailed flowchart of step S6 of the present invention; Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0021] The following is in conjunction with the appendix Figure 1-4 The present invention will be described in further detail below.

[0022] Example 1: This example addresses the potential early leakage of the heating surface of the water-cooled wall in a thermal power plant boiler by implementing the complete process of the intelligent diagnostic method based on remote monitoring of power plant equipment according to the present invention.

[0023] Step S1: Multi-source data acquisition and data preprocessing: The edge gateway collects multi-source operating parameters of the boiler: temperature at various measuring points of the water-cooled wall, furnace pressure, flue gas temperature, feedwater flow rate, steam flow rate, exhaust gas temperature, steam drum water level, unit load, boiler operating commands, and flue gas oxygen content; the edge gateway completes data cleaning, noise reduction, and normalization processing.

[0024] Based on the parameter change rate, unit instructions, and real-time load, four categories of fine-grained operating conditions are marked. The unit is currently operating under a steady-state condition at 75% load, excluding start-up and shutdown transients, human operation disturbances, and external disturbances.

[0025] The edge gateway detected an abnormal drop in temperature at a local measuring point on the water-cooled wall, corresponding to a slight fluctuation in furnace pressure. This anomaly is a valid steady-state anomaly. Short-term disturbances are suppressed to invalidate alarms. The pre-processed time-series data with steady-state operating condition labels is encrypted and uploaded to the cloud; the original disturbance data is cached and retained locally.

[0026] Step S2: Construct a bidirectional constraint mechanism-AI hybrid diagnostic model and a working condition-related knowledge base: The mechanism knowledge base includes a built-in boiler fault tree and fault-symptom mapping relationship: the symptoms corresponding to water-cooled wall leakage include local wall temperature drop, furnace pressure fluctuation, increased feedwater-steam flow difference, and decreased flue gas temperature; each symptom is bound to a dynamic reliability weight under 75% load steady-state conditions.

[0027] Training phase: The boiler fault tree logic is embedded as a loss constraint term into the AI ​​sub-model loss function, and historical boiler fault samples are used to complete the model training, realizing the constraint of mechanism knowledge on AI training; The reasoning phase awaits input of cloud-based early warning events.

[0028] Step S3: Adaptive Dynamic Monitoring and Hierarchical Early Warning of Operating Conditions The cloud receives real-time boiler data uploaded from the edge, along with a 75% load steady-state condition tag. It then calls the adaptive alarm threshold corresponding to this load condition from the knowledge base, instead of directly using a plant-wide uniform static threshold. The overall health status of the boiler equipment is calculated in real time. If the health status shows a significant decline, and the duration of parameter exceeding the limit and steady-state operating conditions are considered, an important level warning is determined, triggering a warning event and pushing a warning signal, abnormal parameters, operating condition tags, and abnormal duration to the downstream S4.

[0029] Step S4: Auditable chain-based intelligent fault diagnosis and root cause tracing: S401, Pre-input Dependency: Receives S3 important early warning signals, abnormal parameters such as water-cooled wall temperature, furnace pressure, and flow difference, 75% load steady-state condition label, and abnormal duration of 12 minutes. S402, AI Initial Inference: Call the trained AI sub-model and output the candidate fault set: ① Water-cooled wall heating surface leakage (basic confidence level 0.78); ② Local coking in the furnace (basic confidence level 0.41); ③ Temperature measurement point fault (basic confidence level 0.22). S403, Mechanism Weighted Correction: Retrieve the dynamic credibility weights of each symptom under steady-state conditions from the S2 knowledge base, and combine the duration of the anomaly to perform multi-evidence time-series weighted causal reasoning: The increased feedwater-steam flow difference and furnace pressure fluctuations match multiple pieces of evidence of water-cooled wall leakage, increasing the confidence of this fault to 0.89; the coking fault lacks corresponding evidence of flue gas oxygen content, and the measuring point fault lacks evidence of synchronous anomalies at multiple measuring points, so the confidence is reduced to 0.27 and 0.13 respectively, locking the root cause to the early leakage potential of the boiler water-cooled wall heating surface; S404, Inference Chain Generation: Complete record of the chain: [75% load steady-state condition label → AI initial output candidate fault set → multi-evidence weighted correction from mechanism knowledge base → final conclusion: early leakage of water-cooled wall], generate an auditable diagnostic report, and output maintenance and handling suggestions: reduce unit load, focus on checking the corresponding area of ​​water-cooled wall tubes, and arrange for shutdown to check for leaks; S405. The overall confidence level of this case is sufficient, and there is no need to proceed with the fallback procedure for suspected faults.

[0030] Step S5: Digital Twin Two-Way Collaborative Visualization and Diagnostic Verification: Positive visualization: Receive the root cause of the S4 fault and the component location, highlight the corresponding water-cooled wall tube section in the boiler digital twin, and display the time-series change curves of wall temperature, flow rate and furnace pressure in real time. Reverse simulation verification: Based on the steady-state condition of 75% load, the early leakage fault of the water-cooled wall is simulated, and the simulation feature dataset is generated and sent back to the S4 diagnostic module; the simulation features are compared with the abnormal features measured on site to verify and confirm the validity of the diagnostic conclusion, eliminate the misjudgment of single-point faults of sensors, and further confirm the authenticity of the fault.

[0031] Step S6: Fully automated closed-loop iteration based on inference chain backtracking: S601, Data Comparison: After on-site maintenance is completed, the actual maintenance results are entered: It is confirmed that there is a micro-leakage in the local tube wall of the water-cooled wall; The system automatically compares the actual maintenance results with the complete inference chain stored in S4 and the twin simulation verification conclusion in S5. S602, Automatic Sample Classification: Automatically identifies samples as accurate diagnostic samples; S603, Closed-loop update: Add the fault sample to the sample library and perform incremental training; update the dynamic credibility weights of each symptom of water-cooled wall leakage in the boiler mechanism knowledge base, and optimize the pre-filtering rules for boiler operating conditions at the edge end. S604, Version Distribution: The updated AI model, knowledge base weights, and operating condition filtering rules are distributed to the edge gateway and cloud platform for subsequent boiler equipment monitoring and diagnosis.

[0032] Maintenance personnel can view early warning information, diagnostic reports, and virtual reality screens, and receive maintenance work order prompts through web clients and mobile maintenance terminals.

[0033] Example 2: This example addresses the potential oil film oscillation fault in the #2 bearing of a steam turbine generator set by applying the complete process of the intelligent diagnostic method of this invention.

[0034] Step S1: Multi-source data acquisition and data preprocessing: The edge gateway collects multi-source parameters of the steam turbine: vibration amplitude of bearings #1-#3, vibration spectrum, bearing metal temperature, lubricating oil inlet temperature, lubricating oil pressure, unit speed, unit active load, and control valve opening command; the edge gateway completes noise reduction, elimination of outlier values, and normalization preprocessing.

[0035] The unit is currently operating under rated load steady-state conditions. We eliminate start-up speed-up transients and valve operation disturbances, and perform fine-grained operating condition labeling.

[0036] A significant increase in the low-frequency component of the #2 bearing shaft vibration was detected, along with a slight increase in bearing temperature. The duration of the anomaly continued to accumulate, which was determined to be a valid steady-state anomaly. The processed time-series data, along with the rated load steady-state condition label, was encrypted and uploaded to the cloud. Instantaneous interference data was cached locally.

[0037] Step S2: Construct a bidirectional constraint mechanism-AI hybrid diagnostic model and a working condition-related knowledge base: The mechanism knowledge base includes a turbine fault tree and fault-symptom mapping: typical symptoms of oil film oscillation include: bearings exhibiting near-half-frequency vibration components, significant increases in shaft vibration, and lubricating oil pressure / temperature deviating from the optimal range; each symptom is bound to a dynamic reliability weight under rated load steady-state conditions.

[0038] Training phase: The turbine failure mechanism and oil film oscillation fault tree logic are embedded into the AI ​​sub-model as constraint loss, and training is completed using historical turbine bearing failure samples; Inference phase: waiting for warning input.

[0039] Step S3: Adaptive Dynamic Monitoring and Hierarchical Early Warning of Operating Conditions The cloud receives uploaded vibration, temperature, and oil pressure parameters, as well as the rated load steady-state operating condition label; it retrieves the adaptive alarm threshold under this operating condition, which is different from the relaxed threshold during the start-up process. The turbine's overall health index is calculated, and the health index is rapidly declining. Combined with the vibration exceeding the limit amplitude and duration, it is determined to be an important warning, triggering a warning event and outputting a warning signal, abnormal parameters, operating condition label, and abnormal duration to S4.

[0040] Step S4: Auditable chain-based intelligent fault diagnosis and root cause tracing: S401, Pre-input Dependency: Receive S3 warning, #2 bearing vibration spectrum, shaft amplitude value, lubricating oil temperature and pressure, rated load steady-state condition label, and abnormal duration of 18 minutes; S402, AI Initial Inference: The AI ​​model outputs the following candidate fault sets: ① Bearing oil film oscillation (basic confidence level 0.74); ② Bearing rotor imbalance (basic confidence level 0.46); ③ Bearing housing looseness (basic confidence level 0.31). S403, Mechanism-Weighted Correction: The dynamic credibility weights of the knowledge base are invoked, and multi-evidence causal reasoning is performed based on operating conditions and the duration of the anomaly: Significant half-frequency vibration characteristics are observed in the actual measurements, and the lubricating oil temperature is slightly higher than normal, matching the core symptom of oil film oscillation; the confidence level is corrected to 0.86. Evidence such as a significantly increased proportion of first harmonic frequency in rotor imbalance and a lack of base vibration characteristics in bearing housing loosening are also present; the confidence levels for these two are lowered to 0.33 and 0.21, respectively. The root cause is identified as a potential oil film oscillation fault in turbine #2 bearing. S404, Reasoning Chain Generation: Record the complete audit chain: Rated load steady-state condition label → AI output candidate fault set → Mechanism fault tree + dynamic weighted correction → Final fault: #2 bearing oil film oscillation, output diagnostic report; Maintenance and handling suggestions: Prioritize adjusting the lubricating oil supply temperature and oil pressure, closely monitor bearing vibration, if the vibration continues to worsen, reduce the load and stop the machine to check the bearing clearance when appropriate. S405. This case has sufficient confidence and does not trigger any suspected fallback procedures.

[0041] Step S5: Digital Twin Two-Way Collaborative Visualization and Diagnostic Verification Forward visualization: The #2 bearing component is highlighted in the digital twin model of the steam turbine to display the real-time curves of shaft vibration spectrum, bearing temperature, and lubricating oil parameters; Reverse simulation verification: Based on the current rated load conditions, conduct oil film oscillation fault simulation, generate simulated vibration spectrum characteristics and send them back to the S4 diagnostic module; compare the simulation characteristics with the actual field spectrum for verification, confirm that the fault characteristics match, and eliminate sensor fault interference.

[0042] Step S6: Fully automated closed-loop iteration based on inference chain backtracking: S601, Data Comparison: After on-site handling and maintenance, the actual results are entered: It is confirmed that the clearance of #2 bearing is offset and oil film oscillation occurs; the system automatically compares the actual maintenance results, S4 inference chain, and S5 twin verification conclusion; S602, Automatic Sample Classification: Mark samples as precise diagnostic samples; S603, Closed-loop update: Use this sample to incrementally train the hybrid model, update the dynamic credibility weights of each symptom corresponding to the oil film oscillation in the turbine knowledge base, and optimize the turbine operating condition disturbance filtering rules at the edge end. S604, Version Distribution: The updated model, knowledge base parameters, and operating condition rules are distributed to the cloud and edge gateway to iteratively upgrade the subsequent turbine fault diagnosis capabilities.

[0043] Operations and maintenance personnel can view alerts, diagnostic reports, and virtual machine status through web clients and mobile operations and maintenance terminals, and receive operations and maintenance handling reminders.

Claims

1. An intelligent diagnostic method based on remote monitoring of power plant equipment, characterized in that, include: Step S1: Multi-source data acquisition and data preprocessing: Collect multi-source operating parameters of core equipment in the power plant, and complete data cleaning, noise reduction, and normalization preprocessing through the edge gateway. Based on the processed parameters, perform fine-grained operating condition classification and labeling. Step S2: Construct a bidirectional constraint mechanism-AI hybrid diagnostic model and a working condition related knowledge base: Build a device mechanism knowledge base, establish a fault-symptom mapping relationship and fault tree logic, and bind each fault symptom to a dynamic credibility weight that is strongly correlated with the working condition type and signal duration in S1, which is different from a fixed weight knowledge base. Step S3, Adaptive Dynamic Monitoring and Hierarchical Early Warning: The cloud receives data uploaded from the edge and operating condition tags, combines the operating condition tags to generate adaptive alarm thresholds, calculates equipment health indicators, and triggers hierarchical early warnings; Step S4, Auditable Chain-like Intelligent Fault Diagnosis and Root Cause Tracing: Driven by the S3 early warning event, it integrates the S2 model algorithm and S1 operating condition data to complete fault reasoning, confidence correction, root cause location, and generate a traceable reasoning chain. Step S5, Digital Twin Two-Way Collaborative Visualization and Diagnostic Verification: It forward-embeds the diagnostic conclusions of S4 to realize visualization, while in reverse outputting simulation data to assist S4 in secondary verification, forming a diagnostic enhancement closed loop, which is different from traditional one-way visualization. Step S6: Fully automated closed-loop iteration based on inference chain backtracking: Relying on the auditable inference chain of S4, the review conclusion of S5, and the actual results of on-site maintenance, the sample screening and model update are automatically completed, and the S1 working condition filtering rules, S2 model and knowledge base parameters are optimized in reverse iteration to achieve continuous evolution of the entire system.

2. The intelligent diagnostic method based on remote monitoring of power plant equipment according to claim 1, characterized in that, In step S1, the collected multi-source operating parameters of the core power plant equipment include equipment vibration, temperature, pressure, flow rate, electrical quantities, oil chromatography, unit operation commands, and load parameters. The fine-grained operating condition classification and labeling specifically involves performing edge lightweight pre-diagnosis based on precise operating condition labels: when short-term parameter offsets caused by human operation, transient fluctuations, or external disturbances are identified, they are determined to be invalid anomalies, alarm uploads are suppressed, and only steady-state valid anomaly data, continuous transient anomaly data, and corresponding operating condition labels are encrypted and uploaded to the cloud. Invalid disturbance data are retained locally. Based on the parameter change rate, unit commands, and real-time load, four categories of fine-grained operating condition labeling are completed: steady-state / start-stop transient / operational disturbance / abnormal disturbance.

3. The intelligent diagnostic method based on remote monitoring of power plant equipment according to claim 2, characterized in that, In step S2, the bidirectional constraint mechanism-AI hybrid diagnostic model has a training phase and an inference phase. In the training phase, the fault symptom logic of the mechanism knowledge base is embedded as a loss constraint term into the loss function of the AI ​​sub-model to realize the mechanism knowledge to constrain the training of the AI ​​model. In the inference phase, the AI ​​sub-model outputs a set of candidate faults and preliminary confidence. The mechanism inference sub-model combines the working condition label and signal duration to obtain the evidence credibility weight, performs multi-evidence time-series weighted causal inference on the candidate faults, and corrects the confidence of the candidate faults.

4. The intelligent diagnostic method based on remote monitoring of power plant equipment according to claim 3, characterized in that, In step S3, the specific method for generating adaptive alarm thresholds is as follows: the cloud receives real-time data and fine-grained operating condition tags uploaded by S1, calls the threshold parameters of the corresponding operating condition in the knowledge base of S2, dynamically generates alarm thresholds that adapt to the current operating status, and calculates the overall health of the equipment in real time. The specific method for graded early warning is as follows: based on the degree of health decay, the duration of parameter exceeding the limit, and the type of operating condition, four levels of early warning are distinguished: prompt, general, important, and emergency. Abnormalities caused by transient disturbances are automatically downgraded and marked to avoid false triggering of diagnosis.

5. A smart diagnostic method based on remote monitoring of power plant equipment according to claim 4, characterized in that, In step S4, the specific steps for auditable chain-like intelligent fault diagnosis and root cause tracing are as follows: S401, Pre-input Dependency: Receives S3 warning trigger signal, abnormal parameters, operating condition label and abnormal duration; S402, AI Initial Inference: Call the AI ​​model trained in S2 and output the initial candidate fault set and basic confidence level; S403, Mechanism Weighted Correction: Call the dynamic confidence weight of S2, combine the working condition characteristics of S1 and the duration of abnormality, perform time-series weighted causal reasoning on the initial AI result, correct the confidence, eliminate working condition interference misjudgment, and lock the root cause of the fault. S404, Inference Chain Generation: Records the entire chain from "Operating Condition Label → AI Initial Output → Mechanism Weight Correction → Final Fault Conclusion" to generate an auditable diagnostic report and simultaneously output maintenance and handling suggestions; S405, Suspected Fault Backup: When the confidence level is insufficient, output the investigation priority based on the weighted weight to guide manual review.

6. A smart diagnostic method based on remote monitoring of power plant equipment according to claim 5, characterized in that, In step S5, the positive implementation of the diagnostic conclusions of S4 to achieve visualization is specifically as follows: receiving the root cause of the fault and component location information of S4, accurately highlighting the faulty component in the digital twin model, and dynamically displaying the changes in equipment status in combination with the real-time operation data of S1. The reverse output simulation data assists the S4 secondary verification in the following ways: based on the current S1 operating condition and S4 fault type, fault simulation is carried out to generate standard fault simulation features, which are then sent back to the S4 diagnostic module to perform a secondary verification of the original diagnostic results and improve the diagnostic accuracy.

7. A smart diagnostic method based on remote monitoring of power plant equipment according to claim 6, characterized in that, In step S6, the specific steps of the fully automated closed-loop iteration based on inference chain backtracking are as follows: Step S601, Data Comparison: Enter the actual fault results from on-site inspection and automatically compare them with the complete inference chain stored in S4 and the review conclusion in S5; Step S602, Automatic Sample Classification: No manual labeling required; automatically distinguishes between accurately diagnosed samples, misdiagnosed samples, and suspected faulty samples. Step S603, Closed-loop update: Use the new samples to perform incremental training, and simultaneously update the weights of the S2 mechanism knowledge base and the AI ​​model parameters, while optimizing the S1 edge case pre-filtering rules. Step S604, Version Distribution: The iterated new model, new operating condition rules, and new weight parameters are distributed to the edge and cloud to achieve the next round of monitoring and diagnostic accuracy upgrade.

8. An intelligent diagnostic system based on remote monitoring of power plant equipment, characterized in that, include, The on-site edge perception layer includes sensors, DCS interface units, and edge gateway modules. Sensors are installed at core monitoring points of various equipment in the power plant, with at least two sensors per parameter monitoring point for parameter complementarity. Sensors are used to collect operating signals such as equipment temperature, vibration, pressure, electrical, and oil parameters. The DCS interface units are used to interface with the power plant's existing control system to read unit operating conditions and parameters. The edge gateway module has a built-in edge preprocessing unit, local caching unit, and edge preliminary anomaly detection unit. It performs raw data noise reduction, outlier removal, operating condition labeling, local caching, preliminary anomaly screening, and encrypted structured data uploading to the cloud. The network transmission layer includes industrial intranets and 5G / 4G communication units, providing encrypted data transmission channels, supporting breakpoint resume, and enabling secure data interaction between the edge perception layer and the cloud platform layer; The cloud platform layer includes a database module, a knowledge base management module, an intelligent diagnostic calculation module, an early warning and handling module, and a model iteration and training module; The application presentation layer includes a remote monitoring web client and a mobile operation and maintenance terminal.

9. An intelligent diagnostic system based on remote monitoring of power plant equipment according to claim 8, characterized in that, Database module: Stores basic device profiles, real-time time-series data, historical fault data, and maintenance records; Knowledge base management module: maintains equipment mechanism knowledge base, fault tree, fault case library, and maintenance procedure library, and supports operation and maintenance personnel to visually edit and update knowledge; Intelligent diagnostic computing module: Loads the hybrid intelligent diagnostic model, performs feature extraction, fault reasoning, root cause tracing, outputs fault type, confidence level, and maintenance suggestions, performs multi-evidence fusion diagnosis, assigns weights to multi-dimensional evidence such as vibration, temperature, oil, and electrical components, comprehensively calculates fault confidence level, and avoids misjudgment based on a single parameter; Early warning and handling module: realizes adaptive threshold calculation based on operating conditions, hierarchical early warning, and completes alarm message distribution; Model Iteration Training Module: Receives real fault samples after maintenance, performs incremental training and updates on the hybrid intelligent diagnostic model, and optimizes model parameters.

10. An intelligent diagnostic system based on remote monitoring of power plant equipment according to claim 8, characterized in that, The cloud platform layer also includes a digital twin simulation unit to build a digital twin of the equipment and map real-time data to the twin model interface to display the component status.

Images