Intelligent early warning and fault diagnosis system for thermal power plant

Abstract

The invention relates to the technical field of state detection and fault diagnosis, in particular to an intelligent early warning and fault diagnosis system for a thermal power plant, which comprises a multi-source data acquisition module for acquiring data in real time; the edge computing node is used for performing noise filtering and abnormal value correction on the acquired data; the digital twin modeling unit is used for constructing a dynamic simulation model of the equipment based on a physical model and historical data; the hybrid analysis engine is used for positioning early abnormal detection and fault sources; the visual early warning interface is used for dynamically displaying the health state and the fault probability of the equipment and generating a graded alarm signal; according to the invention, the multi-source data acquisition module acquires equipment multi-dimensional signals in real time, after edge computing node filtering and denoising, a digital twin modeling unit constructs a precise simulation model, a hybrid analysis engine fuses LSTM and a Bayesian algorithm, fault features are deeply mined, data weights are optimized, and the fault detection accuracy is improved. According to the system, the accuracy and timeliness of diagnosis are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of state detection and fault diagnosis, and particularly relates to an intelligent early warning and fault diagnosis system for a thermal power plant. BACKGROUND

[0002] As the core energy supply unit of the power system, the complexity of the equipment and the harshness of the operating environment of the thermal power plant put high requirements on the fault early warning and diagnosis technology. At present, the industry generally adopts a monitoring system based on threshold alarm and a regular maintenance strategy. The key parameters are collected by sensors and compared with the preset safety range to realize the preliminary identification of abnormal states. Some advanced systems introduce machine learning algorithms to predict potential equipment failures by training models based on historical data, but still mainly rely on centralized data processing architecture, and the diagnosis results are obviously lagging.

[0003] Due to the diversity of thermal power plant equipment signal types (such as vibration, temperature, pressure, etc.), and the large differences in sampling frequency, precision and reliability of different sensors, the traditional system usually adopts a simple weighted or fixed threshold fusion strategy, which is difficult to dynamically adapt to the changes in working conditions. During the period of frequent load adjustment, the coupling relationship between the boiler tube wall temperature and the steam pressure is nonlinearly enhanced. If the system fails to adjust the parameter correlation weight in real time, it may misjudge the normal working condition fluctuations as leakage risks, triggering unnecessary shutdown maintenance and causing huge economic losses. In addition, the existing methods lack sensitivity in detecting early weak fault features, and the best maintenance window has been missed when the fault is obvious.

[0004] Therefore, in order to solve the problem of high false alarm rate caused by insufficient multi-source data fusion, the present application proposes an intelligent early warning and fault diagnosis system for a thermal power plant, which dynamically optimizes the sensor data confidence by using an adaptive filtering algorithm, and realizes early and accurate positioning of faults by combining digital twin and knowledge graph technology, thereby significantly reducing the false alarm rate and improving the diagnosis timeliness. SUMMARY

[0005] In order to overcome the problem of high false alarm rate caused by insufficient multi-source data fusion of the existing system, the present application proposes an intelligent early warning and fault diagnosis system for a thermal power plant.

[0006] The technical scheme of the present application is as follows: an intelligent early warning and fault diagnosis system for a thermal power plant, comprising: a multi-source data acquisition module for real-time acquisition of temperature, pressure, vibration and current signals of the boiler, steam turbine, generator and auxiliary system; an edge computing node for performing noise filtering and outlier correction on the collected data; a digital twin modeling unit for constructing a dynamic simulation model of the equipment based on a physical model and historical data; The hybrid analysis engine includes a rapid fault screening module and a multivariate time series correlation analysis module, integrates an LSTM neural network and a Bayesian inference algorithm, and is used for early abnormality detection and fault root cause positioning. The visual early warning interface dynamically displays the equipment health status and fault probability, and generates a hierarchical alarm signal.

[0007] As preferred, the edge computing node uses a sliding window adaptive Kalman filtering algorithm to adjust the confidence weight of the sensor data, first sets an initial covariance matrix according to the historical error rate of the sensor, calculates the signal-to-noise ratio of the current data segment in real time through a sliding window mechanism, reduces the weight of the sensor data when it is detected that the sensor data continuously deviates from the predicted value, introduces a temperature drift compensation module to eliminate the influence of environmental factors on the pressure sensor, and finally outputs the optimized standardized data stream to the digital twin modeling unit.

[0008] As preferred, the digital twin modeling unit generates a device performance degradation curve and predicts the remaining useful life by comparing the simulation data with the actual operation data in real time, first establishes a multi-field coupled simulation model of high-temperature creep of the boiler pipe wall, uses the actually collected metal temperature and flue gas composition as boundary conditions, calculates the damage accumulation of the pipe microstructure every day, triggers the material aging early warning when the deviation between the simulation stress value and the measured vibration frequency spectrum characteristics exceeds 5%, extrapolates the remaining useful life by combining the Arrhenius equation, and the result is displayed in hours on the visual early warning interface.

[0009] As preferred, the hybrid analysis engine includes a two-level diagnosis structure: a first level rapid fault screening module based on frequency domain analysis of vibration signals, which uses wavelet packet decomposition to extract the energy entropy features of the turbine bearing in the 1-3 kHz frequency band, and matches them with a pre-set fault template library of gear wear and shaft misalignment; a second level multivariate time series correlation analysis module constructs a causal graph model of temperature-flow-insulation resistance of the generator cooling system, and when it is detected that the cooling water flow decreases before the winding temperature rises, the "heat exchanger blockage" fault mode is marked, and the two-level results are output after Bayesian network fusion to output the comprehensive diagnosis conclusion.

[0010] As preferred, the fault knowledge graph is stored in a Neo4j graph database and includes three types of entity nodes: device components, fault modes and maintenance measures, and the nodes are connected through "cause", "need to handle" and "prevention measures" relationship edges. When the hybrid analysis engine detects an anomaly, the knowledge graph automatically expands a three-hop association query, returns the disposal records and corresponding OEM manual chapters of similar historical cases, and highlights the preventive maintenance items not performed in the current system.

[0011] As preferred, the system dynamically updates the LSTM neural network weight by using the newly occurred fault data through an online incremental learning mechanism.

[0012] As preferred, the trigger logic of the hierarchical alarm signal includes: the first level alarm corresponds to a single parameter exceeding the static threshold but not reaching the protection value, and the system automatically records the deviation duration; the second level alarm requires two or more associated parameter abnormalities to be met, triggers a voice prompt and unlocks the relevant operation screen; the third level alarm directly sends a hard-wired trip instruction to the ETS system when detecting a protection refusal risk, and starts the accident memory function to record the full plant data for 5 minutes before the trip.

[0013] As preferred, the system outputs fault diagnosis, which includes four nested fields in JSON format: the fault location field is accurate to the device ID and component number, the probability field shows the confidence calculated by the Bayesian network, the influence range field lists the possible related systems, and the maintenance suggestion field provides the spare parts inventory status and the estimated man-hours, and is signed by the SHA-256 algorithm to ensure that the data cannot be tampered with during transmission to the MIS system.

[0014] As preferred, the hardware configuration of the edge computing node uses an explosion-proof industrial computer, which is deployed in the adjacent cabinet of the power plant DCS system, the shell meets the IP54 protection level and is built-in with a positive pressure ventilation system, and the automatic restart of the hardware watchdog circuit is realized when the computer crashes.

[0015] As preferred, the digital twin modeling unit supports linkage with the SCADA system, and through receiving real-time load instructions provided by the SCADA, the device state change is deduced in the simulation environment 60 seconds in advance, when it is predicted that the induced draft fan will enter the surge zone under the target load, the pre-feed control suggestion is automatically sent to the DCS, and the visual interface uses red contour lines to mark the device operation boundary, assisting the operator to adjust the operation strategy.

[0016] The beneficial effects of the present application are: The multi-source data acquisition module collects real-time temperature, pressure, vibration and current and other multi-dimensional signals of the boiler, steam turbine, generator and auxiliary system, the noise filtering and abnormal value correction function of the edge computing node is combined, so that the interference components in the sensor data are effectively eliminated, the reliability of the original data is improved, the digital twin modeling unit constructs a dynamic simulation model based on the physical model and historical data, which can accurately simulate the normal operation state of the equipment, and provides an accurate benchmark for anomaly detection, the mixed analysis engine integrates the LSTM neural network and the Bayesian inference algorithm, captures the fault characteristics under complex working conditions through deep learning, and optimizes the fusion weight of multi-source data by using probability reasoning, thereby significantly reducing the false alarm rate, the visual warning interface dynamically displays the health state and fault probability of the equipment, and adopts a hierarchical alarm mechanism, so that the operation personnel can quickly identify the real fault and avoid false alarms caused by data fusion deviation, the present application solves the problem of high false alarm rate caused by insufficient data fusion of the traditional thermal power plant monitoring system through multi-source data collaborative analysis, dynamic model optimization and intelligent algorithm fusion, improves the accuracy and reliability of fault diagnosis, and ensures the safe and economic operation of the power plant. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The system framework schematic diagram of the present application is shown. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0019] Please refer to Figure 1 The present application provides an embodiment: a thermal power plant intelligent early warning and fault diagnosis system, comprising: A multi-source data acquisition module acquires real-time temperature, pressure, vibration and current signals of the boiler, steam turbine, generator and auxiliary system; An edge computing node performs noise filtering and abnormal value correction on the collected data; A digital twin modeling unit constructs a dynamic simulation model of the equipment based on a physical model and historical data; A mixed analysis engine includes a rapid fault screening module and a multivariate time series correlation analysis module, integrates an LSTM neural network and a Bayesian inference algorithm, and is used for early anomaly detection and fault root location; A visual warning interface dynamically displays the health state and fault probability of the equipment, and generates a hierarchical alarm signal.

[0020] Further, the system first collects key parameters such as temperature, pressure, vibration and current of the boiler, steam turbine, generator and auxiliary system of the thermal power plant in real time through the multi-source data acquisition module, forms a comprehensive equipment operation state data set, and transmits the collected raw data to the edge computing node for real-time noise filtering and abnormal value correction processing, thereby effectively eliminating the data distortion caused by sensor drift and electromagnetic interference. The preprocessed high-quality data is synchronously input into the digital twin modeling unit, which constructs a dynamic simulation model based on the equipment physical mechanism model and historical operation data, accurately captures the subtle deviation of equipment performance by comparing the virtual simulation data with the actual measurement value in real time, and the mixed analysis engine comprehensively uses the time series prediction ability of the LSTM neural network and the probability diagnosis advantage of the Bayesian inference to perform deep feature extraction and correlation analysis on multi-source heterogeneous data, which can not only identify the early fault trend of slow development (such as bearing wear), but also quickly locate the root cause of sudden faults (such as electrical short circuit), and finally dynamically display the quantitative evaluation results of the equipment health state through the visual early warning interface, and automatically generate graded alarm signals according to the fault severity, so that the operation personnel can intuitively distinguish between slight abnormalities, faults requiring intervention and emergency dangerous states.

[0021] The present application realizes the following contents through the whole-process collaborative optimization of "data acquisition-edge processing-digital simulation-intelligent diagnosis-graded early warning": first, significantly reduces the false alarm rate caused by traditional single threshold alarm, second, shortens the fault identification time from several hours to minutes, and third, through the combination of digital twin and mixed algorithm, the early fault detection rate is improved to more than 95%.

[0022] The edge computing node uses a sliding window adaptive Kalman filtering algorithm to adjust the confidence weight of the sensor data. First, the initial covariance matrix is set according to the historical error rate of the sensor, the signal-to-noise ratio of the current data segment is calculated in real time through the sliding window mechanism, and when it is detected that the sensor data continuously deviates from the predicted value, its weight is automatically reduced. At the same time, a temperature drift compensation module is introduced to eliminate the influence of environmental factors on the pressure sensor, and finally the optimized standardized data stream is output to the digital twin modeling unit.

[0023] Further, the edge computing node first dynamically initializes the covariance matrix according to the historical error rate of each sensor to establish an initial data confidence evaluation benchmark, and in the real-time running stage, the signal-to-noise ratio of the current data segment is calculated through a sliding time window, when the systematic deviation of the pressure sensor caused by temperature drift is detected, the compensation module is automatically activated to correct the original signal, and at the same time, the algorithm continuously monitors the deviation degree of each channel data and the predicted value, if the output of a certain vibration sensor exceeds the 3 sigma range for 3 consecutive periods, the weight coefficient is gradually reduced from the default value 1.0 to below 0.3, and the decision weight of other reliable sensors is correspondingly improved, and the multi-source data optimized dynamically is finally output to the upper system in a standardized format.

[0024] The application improves the signal-to-noise ratio of the key parameters (such as the shaft vibration of the steam turbine) by more than 40% through adaptive weight adjustment, and controls the long-term stability error of the temperature-sensitive sensor (such as the boiler wall temperature thermocouple) to the order of ±0.5%. Compared with the traditional fixed threshold filtering method, the algorithm improves the false data filtering efficiency of the edge node by 55%, and provides a high-fidelity data basis for subsequent digital twin modeling.

[0025] The digital twin modeling unit generates a device performance degradation curve by comparing the simulation data with the actual running data in real time, and predicts the remaining useful life. First, a multi-field coupled simulation model of high-temperature creep of the boiler pipe wall is established, and the actually collected metal temperature and flue gas composition are used as boundary conditions. The damage accumulation of the pipe microstructure is calculated daily, and when the deviation between the simulated stress value and the actual vibration frequency spectrum feature exceeds 5%, the material aging warning is triggered. The remaining useful life is extrapolated by combining the Arrhenius equation, and the result is displayed in hours on the visual warning interface.

[0026] Further, the digital twin modeling unit constructs a creep model of the microstructure of the boiler pipe wall material based on multi-physical field coupling simulation technology, and receives the boundary condition data such as metal temperature and flue gas composition from the edge computing node in real time. High-temperature stress field calculation is performed every 5 seconds, and when the deviation between the simulated stress distribution of the pipe wall and the actual vibration frequency spectrum feature exceeds the threshold of 5%, the material aging warning mechanism is automatically triggered. The system synchronously integrates the creep dynamics process described by the Arrhenius equation, and combines the real-time updated temperature load spectrum to calculate the damage accumulation, and finally generates a remaining useful life (RUL) prediction curve in hours, which is displayed in a color gradient form on the visual interface (green > 1000h, yellow 500-1000h, red < 500h). The life prediction accuracy of the high-temperature pressure-bearing components of the boiler is improved from ±30% of the traditional statistical method to ±12%.

[0027] The mixed analysis engine includes a two-stage diagnosis structure: the first stage is a rapid fault screening module based on frequency domain analysis of vibration signals, which uses wavelet packet decomposition to extract energy entropy features of the turbine bearing in the 1-3 kHz frequency band, and matches with a pre-set fault template library of gear wear and shaft misalignment; the second stage is a multivariate time series correlation analysis module that constructs a causal graph model of the temperature-flow-insulation resistance of the generator cooling system. When a decrease in cooling water flow is detected before a temperature rise in the winding, the "heat exchanger blockage" fault mode is marked. The two-stage results are fused by Bayesian network to output the comprehensive diagnosis conclusion.

[0028] Further, the first-stage rapid fault screening module performs 128-order wavelet packet decomposition on the turbine bearing vibration signal, extracts the energy entropy features of 8 sub-bands in the 1-3 kHz frequency band, and then matches with a pre-trained random forest classifier and a typical fault template library (including 12 types of fault features such as gear wear and shaft misalignment) for real-time matching. The preliminary diagnosis can be completed within 50ms; the second-stage multivariate time series correlation analysis module uses Granger causality test method to construct a dynamic Bayesian network of temperature-flow-insulation resistance of the generator cooling system. When a decrease in cooling water flow (ΔQ>15%) is detected before a temperature rise in the winding (Δt>2min), the "heat exchanger blockage" fault mode is automatically marked, and the joint probability of each potential fault is calculated. The two-stage diagnosis results are finally fused by D-S evidence theory to output a fault conclusion with a comprehensive confidence of more than 85%.

[0029] Further, a three-layer knowledge network including equipment components, fault modes and maintenance measures is constructed based on Neo4j graph database, and the nodes are connected by "cause" "need to handle" "prevention measures" relationship edges. When the mixed analysis engine detects abnormal vibration signals, the system automatically starts three-hop association query, first locates the vibration source equipment (such as "turbine bearing B"), then retrieves 12 types of fault modes that may be caused by the equipment, and finally matches maintenance records with a similarity of more than 80% in the historical case library. The query results are arranged in descending order of disposal success rate, and 3 preventive maintenance work orders that have not been executed in the current system are automatically labeled, such as "lubricating oil filter cartridge overdue".

[0030] The system dynamically updates the LSTM neural network weights using newly occurring fault data through an online incremental learning mechanism.

[0031] Further, the online incremental learning mechanism is described as follows: An elastic weight consolidation (EWC) module is added after the LSTM neural network hidden layer. When a new fault is confirmed as an unknown type, the importance matrix of the original synapse is frozen, and only the last fully connected layer parameter is allowed to update. The training data uses a sliding time window to retain the last 30 days of samples. The learning rate is dynamically adjusted according to the delay time of fault confirmation. Samples with a delay of more than 8 hours are assigned a weight of 0.1 times to ensure that the model does not suffer catastrophic forgetting when adapting to new faults.

[0032] The trigger logic of the hierarchical alarm signal includes: the first-level alarm corresponds to a single parameter exceeding the static threshold but not reaching the protection value, and the system automatically records the deviation duration; the second-level alarm requires two or more associated parameters to be abnormal, triggering a voice prompt and unlocking the relevant operation screen; the third-level alarm detects protection refusal risk, directly sends a hard-wired trip instruction to the ETS system, and starts the accident recall function to record the full plant data 5 minutes before the trip.

[0033] Further, when the boiler drum water level fluctuation is monitored to exceed ±50mm but not reach ±75mm protection value, the first-level alarm is triggered, the system automatically records the deviation duration and starts the self-adjusting program to try to restore the water level. If the associated abnormality of condenser vacuum drop accompanied by circulating water outlet temperature rise is detected at the same time, it is upgraded to a second-level alarm, automatically unlocking the relevant operation screen and issuing a voice prompt "condenser abnormal, please check the circulating water system". When the steam turbine overspeed signal (>3300rpm) lasts for 200ms without triggering the AST solenoid valve action, the third-level alarm is immediately started, sending a trip instruction to the ETS system through hard-wiring, and synchronously activating the accident recall function to record the full plant 3000+ measurement point data 5 minutes before the trip.

[0034] The fault diagnosis output JSON format contains four nested fields: the fault location field is accurate to the device ID and component number, the probability field shows the confidence of Bayesian network calculation, the influence range field lists the possible affected associated systems, and the maintenance suggestion field provides the spare parts inventory status and estimated man-hours, and is signed by SHA-256 algorithm to ensure that the data cannot be tampered with during transmission to the MIS system.

[0035] Further, when the mixed analysis engine completes fault diagnosis, a structured JSON data packet is automatically generated, wherein the "fault location" field adopts double-level positioning of "equipment number / part code", such as "TUR-204 / low-pressure cylinder 3rd stage moving blade", the "probability" field displays 85.7% confidence through Bayesian network calculation, the "impact range" field lists possible related systems and expected deterioration speed, such as "running for 24 hours will cause temperature of adjacent bearing to exceed the standard", the "maintenance suggestion" field intelligently recommends an optimal disposal scheme in association with a spare parts inventory database, such as "use existing spare part #B-5029 in the warehouse, and 4 man-hours are expected", and the data packet is signed by SHA-256 algorithm and transmitted to the power plant MIS system through OPC UA protocol, so that any tampering in the transmission process will cause hash verification to fail.

[0036] The hardware configuration of the edge computing node uses an explosion-proof industrial computer, which is deployed in a cabinet adjacent to the DCS system of the power plant, the shell satisfies the IP54 protection level and is internally provided with a positive pressure ventilation system, and the automatic restart of a dead machine is realized through a hardware watchdog circuit.

[0037] The digital twin modeling unit supports linkage with the SCADA system, receives real-time load instructions provided by the SCADA, and deduces the change of the equipment state in the simulation environment 60 seconds in advance, automatically sends a feedforward control suggestion to the DCS when it is predicted that the induced draft fan will enter the surge zone under the target load, and the visual interface marks the equipment operation boundary with red contour lines to assist the operator to adjust the operation strategy.

[0038] Further, the workflow of the application is described as follows: The multi-source data acquisition module acquires real-time temperature, pressure, vibration and current and other key parameters of the boiler, steam turbine, generator and auxiliary system of the thermal power plant, forming a comprehensive equipment operation state data set to provide real-time monitoring basis for the system.

[0039] The edge computing node performs real-time noise filtering and outlier correction processing on the collected raw data, dynamically adjusts the confidence weight of each sensor data by using a sliding window adaptive Kalman filtering algorithm, eliminates data distortion caused by sensor drift and electromagnetic interference, and outputs high-quality standardized data.

[0040] The digital twin modeling unit constructs a dynamic simulation model based on the equipment physical model and historical data, compares the virtual simulation data with the actual measured values in real time, accurately captures the subtle deviation of the equipment performance through multi-physics field coupling calculation, and generates a performance degradation curve and a remaining useful life prediction.

[0041] The mixed analysis engine comprehensively uses an LSTM neural network and a Bayesian inference algorithm to perform deep feature extraction and correlation analysis on the preprocessed multi-source heterogeneous data, realizes rapid fault screening and accurate fault positioning through a two-level diagnosis structure, and outputs a diagnosis result with a confidence level.

[0042] The fault knowledge graph system automatically retrieves and matches historical fault cases and maintenance schemes based on a Neo4j graph database, provides optimal disposal suggestions through three-layer knowledge network correlation queries, and prompts unexecuted preventive maintenance items to assist decision-making.

[0043] The visual early warning interface dynamically displays the equipment health state evaluation results, automatically generates graded alarm signals (level one to level three) according to the fault probability and severity output by the mixed analysis engine according to a preset logic, and provides differentiated early warning prompts.

[0044] The standardized output module encapsulates the diagnosis result as a structured JSON data packet, including fields such as fault positioning, probability evaluation, impact range, and maintenance suggestions, transmits the data packet to the power plant MIS system after digital signature, and ensures data integrity and traceability.

[0045] Further, the present application provides an embodiment: in a 600MW supercritical unit boiler system, the present application is implemented, 48 temperature sensors and 32 vibration sensors are installed on each heating surface of the boiler to collect data in real time, the edge computing node uses a 0.5 second sliding window to prioritize processing of high temperature area sensors, the digital twin model accurately predicts the abnormal rising trend of the tube wall temperature of the screen superheater, the mixed analysis engine identifies the risk of spalling of the oxide skin 3 days in advance, triggers a secondary alarm and guides the operator to adjust the combustion air distribution, thereby avoiding unplanned shutdown and directly recovering economic losses of about 800,000 yuan.

[0046] Further, the present application provides an embodiment: in the monitoring of the generator cooling system, when the system finds that the stator cooling water flow has dropped to the alarm threshold, it identifies the precursor characteristics of ion exchanger resin leakage through multivariate time series analysis, the digital twin model simulates and predicts that an insulation decline accident will occur after 12 hours, triggers a tertiary alarm and automatically starts the standby cooling system, avoids the burning of the stator bar worth 20 million yuan, and at the same time, the knowledge graph recommends the optimal resin replacement scheme, shortening the maintenance period by 3 days.

[0047] Through the above steps, the multi-dimensional signals of the equipment are collected in real time by the multi-source data acquisition module, the precise simulation model is constructed by the digital twin modeling unit after filtering and denoising by the edge computing node, the mixed analysis engine fuses LSTM and Bayesian algorithms, deeply mines fault features and optimizes data weights, the system significantly improves the diagnosis accuracy and timeliness, ensures the safe and economic operation of the power plant, and solves the problem of high false alarm rate caused by insufficient multi-source data fusion of existing systems.

Claims

1. A power plant intelligent early warning and fault diagnosis system, characterized in that, Comprise: Multi-source data acquisition module, real-time acquisition of temperature, pressure, vibration and current signals of boiler, steam turbine, generator and auxiliary system; Edge computing node, noise filtering and outlier correction are performed on the collected data; Digital twin modeling unit, based on physical model and historical data, a dynamic simulation model of the device is constructed; Mixed analysis engine, including a fast fault screening module and a multivariate time series correlation analysis module, integrating LSTM neural network and Bayesian inference algorithm, for early abnormal detection and fault root location; Visual warning interface, dynamically displaying the health status and fault probability of the device, and generating a hierarchical alarm signal.

2. The intelligent early warning and fault diagnosis system for a thermal power plant according to claim 1, characterized in that: The edge computing node uses a sliding window adaptive Kalman filter algorithm to adjust the confidence weight of the sensor data. First, the initial covariance matrix is set according to the historical error rate of the sensor. The signal-to-noise ratio of the current data segment is calculated in real time through the sliding window mechanism. When it is detected that the data of a certain sensor continuously deviates from the predicted value, its weight is reduced. At the same time, a temperature drift compensation module is introduced to eliminate the influence of environmental factors on the pressure sensor. Finally, the optimized standardized data stream is output to the digital twin modeling unit.

3. The intelligent early warning and fault diagnosis system for a thermal power plant according to claim 1, characterized in that: The digital twin modeling unit generates a device performance degradation curve by comparing simulation data with actual operation data in real time, and predicts the remaining useful life. First, a multi-field coupled simulation model of high-temperature creep of boiler pipe wall is established, using the actually collected metal temperature and flue gas composition as boundary conditions. The damage accumulation of the pipe material microstructure is calculated daily. When the deviation between the simulated stress value and the measured vibration frequency spectrum characteristics exceeds 5%, a material aging warning is triggered. Combined with the Arrhenius equation, the remaining useful life is extrapolated, and the result is displayed in hours on the visual warning interface.

4. The intelligent early warning and fault diagnosis system for thermal power plants according to claim 1, characterized in that, The mixed analysis engine includes a two-level diagnosis structure: the first level is a fast fault screening module based on frequency domain analysis of vibration signals, which uses wavelet packet decomposition to extract the energy entropy features of the 1-3 kHz frequency band of the steam turbine bearing, and matches them with the pre-set fault template library of gear wear and shaft misalignment. The second level is a multivariate time series correlation analysis module that constructs a causal graph model of the temperature-flow-insulation resistance of the generator cooling system. When a decrease in cooling water flow is detected before the winding temperature rises, the "heat exchanger blockage" fault mode is marked. The results of the two levels are fused through a Bayesian network to output the comprehensive diagnosis conclusion.

5. The intelligent early warning and fault diagnosis system for thermal power plants according to claim 1, characterized in that, The fault knowledge graph is stored in a Neo4j graph database, containing three types of entity nodes: device components, fault modes and maintenance measures. The nodes are connected through "cause", "need to handle" and "prevention measures" relationship edges. When the mixed analysis engine detects an anomaly, the knowledge graph automatically expands a three-hop association query, returning the disposal records and corresponding OEM manual chapters of similar historical cases, while highlighting the preventive maintenance items not performed in the current system.

6. The intelligent early warning and fault diagnosis system for thermal power plants of claim 1, wherein: The system dynamically updates the LSTM neural network weights using newly occurring fault data through an online incremental learning mechanism.

7. The intelligent early warning and fault diagnosis system for thermal power plants of claim 1, wherein, The trigger logic of the hierarchical alarm signal includes: the first-level alarm corresponds to a single parameter exceeding the static threshold but not reaching the protection value, and the system automatically records the deviation duration; the second-level alarm requires two or more associated parameters to be abnormal, triggers a voice prompt and unlocks the relevant operation screen; the third-level alarm directly sends a hard-wired trip instruction to the ETS system when the protection refusal risk is detected, and starts the accident memory function to record the full plant data 5 minutes before the trip.

8. The intelligent early warning and fault diagnosis system for thermal power plants of claim 1, wherein, The system outputs fault diagnosis, which includes four nested fields in JSON format: the fault location field is accurate to the device ID and component number, the probability field shows the confidence calculated by the Bayesian network, the impact range field lists the possible related systems, and the maintenance suggestion field provides the spare parts inventory status and estimated working hours, and is signed by SHA-256 algorithm to ensure that the data cannot be tampered with during transmission to the MIS system.

9. The intelligent early warning and fault diagnosis system for thermal power plants of claim 1, wherein: The hardware configuration of the edge computing node uses an explosion-proof industrial computer, which is deployed in the adjacent cabinet of the power plant DCS system, the shell meets the IP54 protection level and is built-in with a positive pressure ventilation system, and the hardware watchdog circuit is used to realize automatic restart of the dead machine.

10. The intelligent early warning and fault diagnosis system for thermal power plants of claim 1, wherein: The digital twin modeling unit supports linkage with the SCADA system, receives real-time load instructions provided by the SCADA, and deduces the device state change in the simulation environment 60 seconds in advance, when it is predicted that the induced draft fan will enter the surge zone under the target load, automatically sends a feedforward control suggestion to the DCS, and the visualization interface uses red contour lines to mark the device operation boundary, assisting the operator to adjust the operation strategy.

Images