Unit multi-physical field coupling monitoring system based on relief-fcc algorithm

Abstract

The present application relates to a unit multi-physical field coupling monitoring system based on a Relief-FCC algorithm, and belongs to the technical field of electric digital signal processing, aiming to solve the problems of ignoring multi-field coupling effect, high fault omission rate and extensive control strategy of existing pumped storage unit monitoring. The system collects unit electromagnetic field, temperature field and mechanical vibration field operation data through a sensor network, introduces an inter-field coupling coefficient FCC in the Relief feature weight to construct a Relief-FCC algorithm, and selects dominant coupling features; the inter-field coupling coefficient is used to complete multi-field coupling consistency cross-validation, distinguish unit body and sensor abnormalities, and complete adaptive correction; based on the dominant coupling features, fault identification, health assessment and failure root cause analysis are realized, and dynamic optimization control of the unit is simultaneously realized. The present application significantly improves the fault prediction accuracy of pumped storage units and ensures efficient and safe operation of the units.

Description

Technical Field

[0001] This invention belongs to the field of electrical digital signal processing technology, specifically relating to a multi-physics coupling monitoring system for generator units based on the Relief-FCC algorithm. Background Technology

[0002] Pumped storage power station generators are core energy conversion devices in power systems for peak shaving, frequency and voltage regulation, and emergency backup. Under ultra-high head and fixed speed operating conditions, their safe and stable operation faces the following technical challenges: First, electromagnetic field distortion. High head input power can easily lead to excessive stator winding current density and rotor magnetic pole magnetic flux saturation, causing magnetic field harmonic distortion and accelerating insulation aging. Second, temperature field imbalance. When the unit operates at high power density, the local temperature rise gradient of stator bars, rotor damping windings, and other parts increases sharply, and traditional air-cooled or water-cooled systems are difficult to effectively suppress local hot spots. Third, mechanical vibration coupling. The interaction between electromagnetic force waves and rotor dynamic eccentricity can easily cause abnormal vibration of the stator core and end windings.

[0003] Existing operational monitoring systems for pumped-storage generator motors mostly employ threshold alarm mechanisms based on a single physical field (such as temperature or vibration), failing to adequately consider the coupling effects between multiple physical fields, resulting in two core shortcomings:

[0004] Firstly, the failure to capture fault precursor information related to electromagnetic-thermal coupling leads to a high rate of false alarms, making it impossible to achieve early fault warning. Secondly, the unit control strategy is relatively crude, often suppressing abnormal temperature rise by reducing the load, resulting in a loss of unit operating efficiency and making it difficult to balance the safety and economy of unit operation. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems by providing a multi-physics coupling monitoring system for pumped storage units based on the Relief-FCC algorithm. This system collects multi-physics operating data covering the unit's electromagnetic field, temperature field, and mechanical vibration field. It combines the Relief algorithm with the Field-to-Field Coupling Coefficient (FCC) to construct the Relief-FCC algorithm for screening dominant coupling features. Cross-validation of multi-physics coupling consistency is performed using the FCC to distinguish between unit-related anomalies and sensor-related anomalies, and adaptive correction is applied to sensor outputs exhibiting drift. A fault classification model is constructed based on the dominant coupling features to achieve fault type identification, unit health status assessment, and root cause analysis of multi-field coupling failure paths, improving the accuracy of pumped storage unit fault prediction and avoiding missed fault reports. Through a coupling feature screening-intelligent fault diagnosis-dynamic control technology chain, precise unit status perception and optimized control are achieved.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a multi-physics coupling monitoring system for generator units based on the Relief-FCC algorithm, comprising the following modules: Multiphysics data acquisition module: Collects multiphysics operation data through a sensor network covering the unit's electromagnetic field, temperature field and mechanical vibration field, and constructs a multiphysics feature set after preprocessing the collected data; Coupling Feature Filtering Module: Used to construct an inter-field feature coupling model, combining physical equation-driven and data-driven modeling to quantify the inter-field coupling strength, introducing the inter-field coupling coefficient FCC into the RELIEF feature weights, dynamically adjusting the coupling strength factor to complete feature weight optimization, and outputting the dominant coupling feature after feature dimensionality reduction. Sensor health assessment and correction module: Based on the pre-learned sensor health baseline, it completes the detection of sensor output anomalies, performs cross-validation of multi-physics coupling consistency through inter-field coupling coefficient, distinguishes between unit-related anomalies and sensor-related anomalies, performs adaptive correction of sensor outputs with drift, and performs reliability marking and early warning for faulty sensors. Fault diagnosis module: Based on the dominant coupling characteristics, a fault classification model is constructed to complete fault type identification, unit health status assessment and root cause analysis of multi-field coupled failure paths; Dynamic control and visualization module: Based on the unit health status output by the fault diagnosis module, AI algorithm is used to solve the optimal operation model of the unit with efficiency, temperature rise and vibration amplitude as optimization objectives, to obtain the optimal combination of operating parameters of the unit, and to perform real-time dynamic optimization control of the unit; the multi-physics field thermogram and the unit health index curve of the unit are visualized and displayed in real time.

[0007] Preferably, the sensor network that is communicatively connected to the multiphysics data acquisition module includes a Hall sensor deployed at the stator end, a flux probe installed at the rotor magnetic pole, a fiber optic grating sensor embedded in the stator bar and rotor damping winding, an infrared thermal imager installed at the thrust bearing, a triaxial accelerometer installed at the stator core end, and a laser displacement sensor deployed in the rotor shaft system. Hall sensors are used to collect the magnetic field strength and harmonic distortion rate of the unit; magnetic flux probes are used to monitor the distribution of magnetic pole eddy current loss; fiber optic grating sensors are used to acquire the temperature rise gradient in real time; infrared thermal imagers are used to capture the temperature field distribution of bearing oil film; triaxial accelerometers are used to extract the vibration spectrum energy entropy after collecting the unit's vibration signal; and laser displacement sensors are used to monitor the dynamic eccentricity of the rotor.

[0008] Furthermore, in the multiphysics data acquisition module, the preprocessing of the acquired data includes: performing wavelet packet decomposition on the vibration signal to extract the frequency band energy entropy features; performing harmonic analysis on the magnetic field data to calculate the amplitude ratio of each harmonic; and normalization processing to eliminate dimensional differences.

[0009] Preferably, in the coupling feature screening module, the combination of physical equation-driven and data-driven modeling to quantify the inter-field coupling strength includes: quantifying the interaction strength between physical quantities based on the physical equations of the electromagnetic-thermal coupling relationship of the unit; and generating the inter-field coupling coefficient FCC by calculating the statistical correlation of the characteristics of the electromagnetic field, temperature field, and mechanical vibration field of the unit through the mutual information method.

[0010] Preferably, in the sensor health assessment and correction module, the inter-field feature coupling model constructed based on the coupling feature screening module learns the expected correlation pattern between different sensors in the health state, continuously monitors the real-time output of each sensor, compares it with the learned baseline, and uses the machine learning-based anomaly detection algorithm SVM to identify abnormal fluctuations, trend changes or sudden jumps in the output of a single sensor. Through coupling consistency verification, it is determined whether the sensor has drift or malfunction rather than a fault in the unit itself. Based on the degree of drift of the sensor output data and the coupling consistency verification results, an adaptive correction factor is dynamically generated. This correction factor is applied in real time to the sensor's original output data to compensate for its drift.

[0011] The correction factor is updated using adaptive Kalman filtering or adaptive least squares method, enabling it to be dynamically adjusted according to the actual drift of the sensor.

[0012] Preferably, in the fault diagnosis module, the LightGBM algorithm is used, with the dominant coupling features output by the coupling feature screening module as input, to identify and classify unit faults. The identified fault types include rotor inter-turn short circuit, stator insulation breakdown, and thrust bearing wear.

[0013] Preferably, in the fault diagnosis module, a Bayesian network model is constructed to complete the multi-field coupled chain failure path analysis, specifically including: The Bayesian network model adopts a directed acyclic graph structure, where the root node represents the root cause variables of unit faults, including magnetic field harmonic distortion, rotor dynamic eccentricity, stator current density exceeding limits, and cooling system anomalies; the intermediate nodes represent multi-physics coupling characteristic anomaly state variables, including rotor magnetic pole magnetic flux density saturation, vibration spectrum energy entropy anomalies, stator bar temperature rise rate exceeding limits, and bearing oil film temperature anomalies; the directed edges between nodes are constructed based on the physical coupling relationships determined by the inter-field characteristic coupling model, and the edge weights correspond to the inter-field coupling strength; the leaf nodes represent the final fault phenomenon variables of the unit, including stator insulation breakdown, rotor inter-turn short circuit, and thrust bearing wear; Based on the historical fault dataset and rated operating data of the unit, a conditional probability table for each node is constructed using the Bayesian parameter estimation method. The method for multi-field coupled chain failure path analysis is as follows: when a unit fault phenomenon or abnormal health index is detected, the fault state of the leaf node is used as evidence input into the Bayesian network model. The joint tree inference algorithm is used to perform reverse causal inference to calculate the posterior probability of each root node and intermediate node. The nodes with the highest probability are selected according to the posterior probability from high to low, which is the causal transmission link of the node with the highest probability. This is the multi-field coupled chain failure path, and the root cause of the fault is located.

[0014] Secondly, the present invention provides a unit health monitoring method based on the Relief-FCC algorithm, comprising the following steps: S1. Multiphysics Data Acquisition: Unit operating data is acquired through electromagnetic field, temperature field, and mechanical vibration field sensors. S2. Data preprocessing: After wavelet packet decomposition, harmonic analysis and normalization, a multiphysics feature set is constructed. S3. Coupling Feature Screening: Calculate the inter-field coupling coefficient FCC, superimpose the FCC contribution term into the traditional RELIEF weights, dynamically adjust the coupling strength factor to complete weight optimization and feature dimensionality reduction, and output the dominant coupling feature. S4. Sensor Health Assessment and Calibration: Based on the health baseline, sensor anomalies are detected, and multi-physics coupling consistency verification is performed through FCC. The abnormalities of the unit itself and the sensor are distinguished. Drift sensors are adaptively calibrated, and faulty sensors are marked and warned. S5. Fault Diagnosis: Based on the dominant coupling characteristics, typical faults are identified through the LightGBM model, the health index HI is calculated and an early warning is issued, the remaining life of components is predicted by combining the Weibull distribution, and the root cause is located by using Bayesian networks to infer multi-field coupled chain failure paths. S6. Dynamic Control and Visualization: Optimize operating parameters based on diagnostic results and display multi-physics field status and health index in real time.

[0015] Compared with the prior art, the beneficial effects of the present invention include: 1) This invention utilizes the Relief-FCC algorithm, integrating physical mechanisms and data-driven methods, to accurately extract the multi-field coupling characteristics of the electromagnetic field, temperature field, and mechanical vibration field of the generator motor. This solves the problem of insufficient analysis of complex interactive effects of multiple physical fields by traditional monitoring systems. By combining intelligent diagnostic models and dynamic control strategies, a closed-loop system of "perception-analysis-decision-execution" is constructed to achieve early warning and adaptive optimization control of faults such as stator insulation breakdown and rotor inter-turn short circuit. This significantly improves the operation and maintenance efficiency and operational safety of ultra-high head pumped storage units, providing an innovative technical solution for the safe and efficient operation of pumped storage power stations.

[0016] 2) This invention provides a sensor health assessment and correction method based on inter-field coupling coefficients. It completes the initial detection of sensor output anomalies through a pre-learned sensor health baseline and conducts cross-validation of multi-physics coupling consistency by combining the inter-field coupling coefficients. This method can effectively distinguish between unit-related anomalies and sensor-related anomalies, achieve adaptive correction for sensor outputs with drift, and complete reliability marking and early warning for faulty sensors. This method can effectively avoid the distortion of monitoring data caused by sensor anomalies, ensure the accuracy and reliability of multi-physics operation data acquisition, and provide a high-quality data foundation for subsequent coupling feature analysis, fault diagnosis, and control decision-making.

[0017] 3) The fault diagnosis method based on dominant coupling features of this invention combines the dominant coupling features selected by the Relief-FCC algorithm with the LightGBM fault classification model and the health status quantitative assessment system. The dominant coupling features selected by the Relief-FCC algorithm retain multi-field coupling correlation information that is strongly related to unit faults, while eliminating redundant and irrelevant features. This can simultaneously reduce the input dimension and computational complexity of the fault classification model, and improve the model's sensitivity and classification accuracy for faults caused by multi-field coupling. The two support each other functionally, and the combined technical effect is better than the sum of the effects of single feature selection or single full feature fault diagnosis. It can realize the accurate identification of typical unit faults and the quantitative assessment of health status.

[0018] 4) This invention is a root cause analysis method for generating units based on multi-field coupled chain failure path analysis. It constructs a Bayesian network model based on the physical coupling relationship determined by the inter-field characteristic coupling model, and uses the monitored fault state or abnormal health index as evidence to perform reverse causal reasoning. It can accurately screen out the multi-field coupled chain failure path with the highest probability and complete the fault root cause location. This method breaks through the limitation of existing technologies that can only identify fault phenomena and cannot trace the multi-field coupled failure transmission link, and can provide a clear technical basis for the accurate handling of unit faults and the formulation of operation and maintenance strategies.

[0019] 5) This invention proposes a real-time dynamic optimization control method based on dominant coupling characteristics and unit health status to solve for the optimal combination of unit operating parameters. It deeply integrates the unit fault diagnosis results and real-time health status obtained based on the dominant characteristics of multi-field coupling with the multi-objective optimization operation model. The real-time health status of the unit and the root cause analysis results of the fault provide accurate optimization boundaries and constraints for the multi-objective optimization model. The multi-objective optimization model can simultaneously take into account multiple coupled optimization objectives such as unit operating efficiency, temperature rise, and vibration amplitude. The two are mutually supportive in function, and the combined technical effect is better than the sum of the effects of a single fault warning or a single fixed threshold control. It avoids the loss of unit operating efficiency caused by coarse control and can suppress the further development of multi-field coupling anomalies through precise control, thereby achieving a simultaneous improvement in unit operating safety and economy. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Figure 1 This is a schematic diagram of a unit multiphysics coupling monitoring system based on the Relief-FCC algorithm according to an embodiment of the present invention. Detailed Implementation

[0022] Example 1 like Figure 1 As shown, the unit multiphysics coupling monitoring system based on the Relief-FCC algorithm includes the following modules: Multi-physics data acquisition module: Collects multi-physics operation data through a sensor network covering the unit's electromagnetic field, temperature field, and mechanical vibration field, and constructs a multi-physics feature set after preprocessing the collected data.

[0023] In this embodiment, the sensor network includes a Hall sensor deployed at the stator end, a flux probe installed at the rotor magnetic pole, a fiber optic grating sensor embedded in the stator bar and rotor damping winding, an infrared thermal imager installed at the thrust bearing, a triaxial accelerometer installed at the stator core end, and a laser displacement sensor deployed in the rotor shaft system.

[0024] Hall sensors are used to collect the magnetic field strength and harmonic distortion rate of the unit; magnetic flux probes are used to monitor the distribution of magnetic pole eddy current loss; fiber optic grating sensors are used to acquire the temperature rise gradient in real time; infrared thermal imagers are used to capture the temperature field distribution of bearing oil film; triaxial accelerometers are used to extract the vibration spectrum energy entropy after collecting the unit's vibration signal; and laser displacement sensors are used to monitor the dynamic eccentricity of the rotor.

[0025] In this embodiment, wavelet packet decomposition is performed on the vibration signal to obtain energy entropy features in 16 frequency bands; harmonic analysis (FFT) is performed on the magnetic field data to calculate the amplitude ratio of each harmonic; normalization is performed to eliminate dimensional differences and construct an 80-dimensional multi-physics field feature set.

[0026] Coupling Feature Filtering Module: Used to construct inter-field feature coupling models, combining physical equation-driven and data-driven modeling to quantify the inter-field coupling strength, introducing the inter-field coupling coefficient FCC into the RELIEF feature weights, dynamically adjusting the coupling strength factor to complete feature weight optimization, and outputting the dominant coupling features after feature dimensionality reduction.

[0027] In this embodiment, the current density is quantified based on the physical equations governing the electromagnetic-thermal coupling relationship of the unit. J With temperature rise ΔT The interaction strength between physical quantities; the inter-field coupling coefficient FCC is generated by statistically correlating the electromagnetic field, temperature field, and mechanical vibration field characteristics of the computer group using the mutual information method.

[0028] In the inter-field characteristic coupling model, the formula for calculating the inter-field coupling coefficient FCC is: ; In the formula, Represents the mutual information of features i and j. , Representing features respectively The variance.

[0029] The expression for feature weight optimization is: ; In the formula, Representation of features RELIEF feature weights, diff Let be the distance function. In the k-th sampling, the feature is... The samples belong to the nearest neighbor samples of the same category. In the k-th sampling, the feature is... The samples belong to the nearest neighbor samples of different categories. This is the coupling strength factor, with a value range of [0.3, 0.7]. Represents the features after optimization of inter-field coupling coefficients The weight, m This represents the total number of randomly sampled samples in the traditional RELIEF algorithm.

[0030] In this embodiment, 12-15 dominant features (such as the third harmonic amplitude, rotor eccentricity, and bar temperature rise rate) are selected from 80-dimensional multi-physics field feature data.

[0031] Sensor health assessment and correction module: Based on the pre-learned sensor health baseline, it completes the detection of sensor output anomalies, performs cross-validation of multi-physics coupling consistency through inter-field coupling coefficient, distinguishes between unit-related anomalies and sensor-related anomalies, performs adaptive correction of sensor outputs with drift, and marks and warns of faulty sensors based on reliability.

[0032] In this embodiment, the inter-field feature coupling model constructed based on the coupling feature screening module learns the expected correlation patterns between different sensors under healthy conditions, continuously monitors the real-time output of each sensor, compares it with the learned baseline, and uses the machine learning-based anomaly detection algorithm SVM to identify abnormal fluctuations, trend changes or sudden jumps in the output of a single sensor. Through coupling consistency verification, it is determined whether the sensor has drift or malfunction rather than a fault in the unit itself. Based on the degree of drift of the sensor output data and the coupling consistency verification results, an adaptive correction factor is dynamically generated. This correction factor is applied in real time to the sensor's original output data to compensate for its drift.

[0033] The correction factor is updated using adaptive Kalman filtering or adaptive least squares method, enabling it to be dynamically adjusted according to the actual drift of the sensor.

[0034] Fault diagnosis module: Based on the dominant coupling characteristics, a fault classification model is constructed to complete fault type identification, unit health status assessment and root cause analysis of multi-field coupled failure paths.

[0035] In this embodiment, the LightGBM algorithm is used, with the dominant coupling features output by the coupling feature screening module as input, to identify and classify unit faults. The identified fault types include rotor inter-turn short circuit, stator insulation breakdown, and thrust bearing wear.

[0036] In this embodiment, the health status of the unit is assessed based on the dominant coupling feature computer group health index output by the coupling feature screening module. The formula for calculating the unit health index HI is as follows: ; ; In the formula, This represents the degradation degree of the i-th dominant coupling feature. n The number of dominant coupling features, This represents the real-time monitoring value of the i-th dominant coupling feature. This represents the baseline value of the i-th dominant coupling characteristic under the rated healthy operating conditions of the unit. This represents the fault warning threshold corresponding to the i-th dominant coupling feature.

[0037] In this embodiment, a Bayesian network model is constructed to complete the multi-field coupled chain failure path analysis, specifically including: The Bayesian network model adopts a directed acyclic graph structure, where the root node represents the root cause variables of unit faults, including magnetic field harmonic distortion, rotor dynamic eccentricity, stator current density exceeding limits, and cooling system anomalies; the intermediate nodes represent multi-physics coupling characteristic anomaly state variables, including rotor magnetic pole magnetic flux density saturation, vibration spectrum energy entropy anomalies, stator bar temperature rise rate exceeding limits, and bearing oil film temperature anomalies; the directed edges between nodes are constructed based on the physical coupling relationships determined by the inter-field characteristic coupling model, and the edge weights correspond to the inter-field coupling strength; the leaf nodes represent the final fault phenomenon variables of the unit, including stator insulation breakdown, rotor inter-turn short circuit, and thrust bearing wear; Based on the historical fault dataset and rated operating data of the unit, a conditional probability table for each node is constructed using the Bayesian parameter estimation method. The method for multi-field coupled chain failure path analysis is as follows: when a unit fault phenomenon or abnormal health index is detected, the fault state of the leaf node is used as evidence input into the Bayesian network model. The joint tree inference algorithm is used to perform reverse causal inference to calculate the posterior probability of each root node and intermediate node. The nodes with the highest probability are selected according to the posterior probability from high to low, which is the causal transmission link of the node with the highest probability. This is the multi-field coupled chain failure path, and the root cause of the fault is located.

[0038] Dynamic Control and Visualization Module: Based on the unit health status output by the fault diagnosis module, the NSGA-II algorithm is used to solve the optimal operation model of the unit with efficiency, temperature rise, and vibration amplitude as optimization objectives. The optimal combination of operating parameters such as excitation current (±5% adjustment), cooling water flow rate (0-100% opening), and rotor counterweight adjustment (0-200g) is obtained to perform real-time dynamic optimization control of the unit; the multi-physics field thermogram and unit health index curve of the unit are visualized and displayed in real time.

[0039] Example 2 This embodiment provides a unit health monitoring method based on Embodiment 1, including the following steps: S1. Multiphysics Data Acquisition: Unit operating data is acquired through electromagnetic field, temperature field, and mechanical vibration field sensors. S2. Data preprocessing: After wavelet packet decomposition, harmonic analysis and normalization, a multiphysics feature set is constructed. S3. Coupling Feature Screening: Calculate the inter-field coupling coefficient FCC, superimpose the FCC contribution term into the traditional RELIEF weights, dynamically adjust the coupling strength factor to complete weight optimization and feature dimensionality reduction, and output the dominant coupling feature. S4. Sensor Health Assessment and Calibration: Based on the health baseline, sensor anomalies are detected, and multi-physics coupling consistency verification is performed through FCC. The abnormalities of the unit itself and the sensor are distinguished. Drift sensors are adaptively calibrated, and faulty sensors are marked and warned. S5. Fault Diagnosis: Based on the dominant coupling characteristics, typical faults are identified through AI models, the computer group health index HI is generated and an early warning is issued, the remaining life of components is predicted by combining Weibull distribution, and Bayesian networks are used to infer multi-field coupled chain failure paths and locate the root cause. S6. Dynamic Control and Visualization: Based on real-time unit-dominant coupling characteristics and unit health index, AI algorithms are used to solve the unit's optimal operation model with efficiency, temperature rise, and vibration amplitude as optimization objectives, to obtain the optimal combination of unit operating parameters, and to perform real-time dynamic optimization control of the unit; the unit's multi-physics field thermogram and unit health index curve are visualized and displayed in real time.

Claims

1. A multi-physics coupled monitoring system for generator units based on the RELIEF-FCC algorithm, characterized in that, Includes the following modules: Multi-physics data acquisition module: Collects multi-physics operation data through a sensor network covering the unit's electromagnetic field, temperature field, and mechanical vibration field, and constructs a multi-physics feature set after preprocessing the collected data; Coupling Feature Filtering Module: Used to construct an inter-field feature coupling model, combining physical equation-driven and data-driven modeling to quantify the inter-field coupling strength, introducing the inter-field coupling coefficient FCC into the RELIEF feature weights, dynamically adjusting the coupling strength factor to complete feature weight optimization, and outputting the dominant coupling feature after feature dimensionality reduction. Sensor health assessment and correction module: Based on the pre-learned sensor health baseline, it completes the detection of sensor output anomalies, performs cross-validation of multi-physics coupling consistency through inter-field coupling coefficient, distinguishes between unit-related anomalies and sensor-related anomalies, performs adaptive correction of sensor outputs with drift, and performs reliability marking and early warning for faulty sensors. Fault diagnosis module: Based on the dominant coupling characteristics, a fault classification model is constructed to complete fault type identification, unit health status assessment and root cause analysis of multi-field coupled failure paths.

2. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, It also includes a dynamic control and visualization module: based on the unit health status output by the fault diagnosis module, an AI algorithm is used to solve the unit's optimal operation model with efficiency, temperature rise, and vibration amplitude as optimization objectives, to obtain the optimal combination of unit operating parameters, and to perform real-time dynamic optimization control of the unit; and to visualize and display the unit's multi-physics field thermogram and unit health index curve in real time.

3. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, The sensor network includes Hall sensors deployed at the stator end, flux probes installed at the rotor poles, fiber optic grating sensors embedded in the stator bars and rotor damping windings, infrared thermal imagers installed at the thrust bearing, triaxial accelerometers installed at the stator core end, and laser displacement sensors deployed in the rotor shaft system. Hall sensors are used to collect the magnetic field strength and harmonic distortion rate of the unit; magnetic flux probes are used to monitor the distribution of magnetic pole eddy current loss; fiber optic grating sensors are used to acquire the temperature rise gradient in real time; infrared thermal imagers are used to capture the temperature field distribution of bearing oil film; triaxial accelerometers are used to extract the vibration spectrum energy entropy after collecting the unit's vibration signal; and laser displacement sensors are used to monitor the dynamic eccentricity of the rotor.

4. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, In the multiphysics data acquisition module, the preprocessing of the acquired data includes: wavelet packet decomposition of the vibration signal to extract the frequency band energy entropy features; harmonic analysis of the magnetic field data to calculate the amplitude ratio of each harmonic; and normalization to eliminate dimensional differences.

5. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, In the coupling feature screening module, the combination of physical equation-driven and data-driven modeling to quantify the inter-field coupling strength includes: quantifying the interaction strength between physical quantities based on the physical equations of the electromagnetic-thermal coupling relationship of the unit; and generating the inter-field coupling coefficient FCC by calculating the statistical correlation of the characteristics of the electromagnetic field, temperature field, and mechanical vibration field of the unit through the mutual information method.

6. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, In the inter-field characteristic coupling model, the formula for calculating the inter-field coupling coefficient FCC is: ； In the formula, Represents the mutual information of features i and j. , Representing features respectively The variance.

7. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 6, characterized in that, In the coupled feature selection module, the expression for feature weight optimization is: ； In the formula, Representation of features RELIEF feature weights, diff Let be the distance function. This represents the nearest neighbor sample belonging to the same category as the target sample in the k-th sampling. This represents the nearest neighbor sample that belongs to a different category from the target sample in the k-th sampling. This is the coupling strength factor; Represents the features after optimization of inter-field coupling coefficients The weight, m This represents the total number of randomly sampled samples in the traditional RELIEF algorithm.

8. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, In the sensor health assessment and correction module, the inter-field feature coupling model built based on the coupling feature screening module learns the expected correlation patterns between different sensors under the health state, continuously monitors the real-time output of each sensor and compares it with the learned baseline, and uses the machine learning-based anomaly detection algorithm SVM to identify abnormal fluctuations, trend changes or sudden jumps in the output of a single sensor. Through coupling consistency verification, it is determined whether the sensor has drift or malfunction rather than a fault in the unit itself. Based on the degree of drift of the sensor output data and the coupling consistency verification results, an adaptive correction factor is dynamically generated. This correction factor is applied in real time to the sensor's original output data to compensate for its drift.

9. The unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm according to claim 1, characterized in that, In the fault diagnosis module, a Bayesian network model is constructed to complete the analysis of multi-field coupled chain failure paths, specifically including: The Bayesian network model adopts a directed acyclic graph structure, where the root node represents the root cause variables of unit faults, including magnetic field harmonic distortion, rotor dynamic eccentricity, stator current density exceeding limits, and cooling system anomalies; the intermediate nodes represent multi-physics coupling characteristic anomaly state variables, including rotor magnetic pole magnetic flux density saturation, vibration spectrum energy entropy anomalies, stator bar temperature rise rate exceeding limits, and bearing oil film temperature anomalies; the directed edges between nodes are constructed based on the physical coupling relationships determined by the inter-field characteristic coupling model, and the edge weights correspond to the inter-field coupling strength; the leaf nodes represent the final fault phenomenon variables of the unit, including stator insulation breakdown, rotor inter-turn short circuit, and thrust bearing wear; Based on the historical fault dataset and rated operating data of the unit, a conditional probability table for each node is constructed using the Bayesian parameter estimation method. The method for multi-field coupled chain failure path analysis is as follows: when a unit fault phenomenon or abnormal health index is detected, the fault state of the leaf node is used as evidence input into the Bayesian network model. The joint tree inference algorithm is used to perform reverse causal inference to calculate the posterior probability of each root node and intermediate node. The nodes with the highest probability are selected according to the posterior probability from high to low, which is the causal transmission link of the node with the highest probability. This is the multi-field coupled chain failure path, and the root cause of the fault is located.

10. The method of the unit multiphysics coupling monitoring system based on the RELIEF-FCC algorithm as described in any one of claims 1-9, characterized in that, Includes the following steps: 1) Multi-physics data acquisition and preprocessing: The unit operation data is acquired through electromagnetic field, temperature field and mechanical vibration field sensors, and then processed by wavelet packet decomposition, harmonic analysis and normalization to construct a multi-physics feature set; 2) Coupling feature selection: Calculate the inter-field coupling coefficient FCC, superimpose the FCC contribution term into the traditional RELIEF weights, dynamically adjust the coupling strength factor to complete weight optimization and feature dimensionality reduction, and output the dominant coupling feature; 3) Sensor health assessment and calibration: Based on the health baseline, sensor anomalies are detected, and multi-physics coupling consistency verification is performed through FCC. The abnormalities of the unit itself and the sensor are distinguished. Drift sensors are adaptively calibrated, and faulty sensors are marked and given early warning. 4) Fault diagnosis: Based on the dominant coupling characteristics, typical faults are identified through the LightGBM model, the health index HI is calculated and an early warning is issued, the remaining life of components is predicted by combining the Weibull distribution, and the root cause is located by using Bayesian network to infer multi-field coupled chain failure paths. 5) Dynamic control and visualization: Optimize operating parameters based on diagnostic results and display multi-physics field status and health index in real time.

Images