Water turbine bearing temperature and lubricating oil state monitoring and fault diagnosis method and system

Abstract

The invention belongs to the technical field of hydroelectric generating set monitoring, and relates to a water turbine bearing temperature and lubricating oil state monitoring and fault diagnosis method and system. According to the method, triple-redundancy temperature measurement values of a water turbine bearing are adopted, multiple parameters are fused, redundancy design is adopted, sensor failure diagnosis is executed, a temperature-vibration-oil pressure combined feature space is constructed according to the triple-redundancy temperature measurement values and online oil quality parameters, a fault propagation map is established, and the fault propagation map is analyzed. And finally, based on the multi-level diagnosis model and in combination with the fault propagation atlas, composite fault mode recognition is realized, a fault classification result is obtained, and the composite fault mode recognition is realized based on the multi-level diagnosis model, so that the problems of low reliability, single monitoring dimension and insufficient fault recognition capability of a sensor in the prior art are effectively solved. And the equipment reliability and the operation safety are obviously improved.

Description

Technical Field

[0001] The invention belongs to the technical field of hydropower unit monitoring, and relates to a method and system for monitoring the temperature and lubricating oil status of a turbine bearing and for diagnosing its faults. Background Art

[0002] During the operation of the turbine, the turbine bearing temperature and lubricating oil status are one of the key factors to ensure the normal operation of the turbine. Therefore, the power station operation and management will monitor the turbine bearing temperature and lubricating oil status in real time to ensure equipment reliability and safety.

[0003] In the existing technology, the monitoring of turbine bearing temperature and lubricating oil status mainly has the following problems: Bearing temperature monitoring mainly relies on temperature sensors to collect temperature data. However, during long-term operation, the sensor may cause signal drift or failure due to vibration, corrosion, or electromagnetic interference, which in turn leads to failure of bearing temperature monitoring. Regarding the monitoring of bearing lubricant status, the main focus is on the oil temperature or oil pressure, while ignoring key indicators such as oil viscosity, water content, and particle contamination, making it difficult to fully assess the lubrication status; The monitoring of each parameter is generally isolated and lacks correlation analysis, making it difficult to identify complex failure modes. Summary of the Invention

[0004] The present invention aims to provide a method and system for monitoring and diagnosing turbine bearing temperature and lubricating oil status, addressing the technical issues of isolated parameter monitoring, lack of correlation analysis, and difficulty identifying complex fault modes. In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for monitoring and diagnosing the temperature and lubricating oil status of a turbine bearing, comprising the following steps: Obtain triple-redundant temperature measurements of turbine bearings; Obtaining online oil quality parameters of the lubricating oil, wherein the online oil quality parameters include temperature, pressure, viscosity, water content, and metal particle concentration parameters; Construct a temperature-vibration-oil pressure joint feature space based on triple-redundant temperature measurements and online oil quality parameters, and establish a fault propagation map; Based on the multi-level diagnosis model and the fault propagation map, composite fault pattern recognition is realized to obtain the fault classification result.

[0005] In a second aspect, the present invention provides a turbine bearing temperature and lubricating oil status monitoring and fault diagnosis system, comprising: Temperature measurement value acquisition module: used to obtain triple-redundant temperature measurement values ​​of turbine bearings; Online oil quality parameter acquisition module: used to obtain online oil quality parameters of lubricating oil, including temperature, pressure, viscosity, water content and metal particle concentration parameters; Fault propagation map acquisition module: used to construct the temperature-vibration-oil pressure joint feature space based on triple-redundant temperature measurement values ​​and online oil quality parameters, and establish a fault propagation map; Fault classification result acquisition module: used to realize composite fault pattern recognition based on multi-level diagnosis model combined with fault propagation map to obtain fault classification results.

[0006] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes triple-redundant temperature measurements of turbine bearings, measured simultaneously by three temperature sensors. Subsequently, data is fused using median selection, weighted averaging, or voting mechanisms to reduce the risk of single-point failures and improve the reliability and accuracy of temperature monitoring. This avoids signal failures often associated with traditional single sensors due to vibration, corrosion, or electromagnetic interference. Next, online lubricating oil quality parameters are acquired, including temperature, pressure, viscosity, water content, and metal particle concentration. Dynamic data acquisition through online detection technology avoids the lag associated with laboratory testing. This addresses the problem of traditional methods that monitor only oil temperature and pressure while ignoring key indicators such as viscosity and contamination. A temperature-vibration-oil pressure joint feature space is constructed based on the triple-redundant temperature measurements and online oil quality parameters, and a fault propagation map is created. This breaks down parameter silos and establishes cross-domain feature correlations, addressing the inability of traditional isolated monitoring to capture complex fault modes. Complex fault mode identification is achieved based on a multi-level diagnostic model combined with the fault propagation map, resulting in fault classification. Model parameters are corrected in real time using the fault propagation map, improving adaptability and diagnostic accuracy. This eliminates the inability of a single diagnostic model to handle complex faults.

[0007] The present invention integrates multiple parameters, adopts redundant design, performs sensor failure diagnosis, constructs a temperature-vibration-oil pressure joint feature space, establishes a fault propagation map, and finally realizes composite fault pattern recognition based on a multi-level diagnostic model, thereby effectively solving the problems of low sensor reliability, single monitoring dimension, and insufficient fault identification capability in the existing technology, and significantly improving equipment reliability and operational safety.

[0008] The present invention adopts the Kalman filter algorithm to perform dynamic drift compensation for temperature measurement values. It integrates the Kalman filter algorithm with vibration characteristic analysis, online parameter calibration, physical modeling and other technologies to form a significantly innovative temperature drift compensation scheme, thereby improving the accuracy of temperature measurement values.

[0009] By constructing a fault knowledge graph and a fault propagation graph, the present invention can effectively solve the problems of isolated parameter monitoring and compound fault identification, and realize the transformation from data to knowledge.

[0010] The present invention combines the random forest model, deep residual network and digital twin model to realize composite fault pattern recognition, forming a multi-level diagnostic model with fast diagnostic response and high accuracy.

[0011] The digital twin model of the present invention introduces fault influencing factors and fault excitation forces, and the unknown additional forces caused by the fault The unknown additional force caused by the fault is controlled by the fault influence factor. The initial fault is detected by a small change in the fault influence factor value, and the specific fault degree (such as wear depth) corresponding to the fault influence factor value is quantified. At the same time, the remaining service life can also be predicted based on the trend of the fault influence factor. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart of an embodiment of the present invention; Figure 2 is a flow chart of a method according to an embodiment of the present invention; Figure 3 It is a module diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0014] It should be noted that the terms "first," "second," and the like in the description of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0015] The present invention is described in further detail below with reference to the accompanying drawings: See also Figure 2 The present invention discloses a method for monitoring the temperature and lubricating oil status of a turbine bearing and diagnosing its faults, comprising the following steps: Step S1: Acquire the three redundant temperature measurement values ​​of the turbine bearings, as follows: Step S1.1: Synchronously obtain triple-redundant temperature measurement values ​​from the fiber Bragg grating, infrared thermal imaging, and PT100 sensor.

[0016] Step S1.2: Use the Kalman filter algorithm to dynamically compensate for the drift of the triple-redundant temperature measurement values, as follows: Step S1.2.1: Establish a state space model of the temperature sensor, wherein the state space model includes a state equation and an observation equation; The state equation is as follows:

[0017] The observation equation is as follows:

[0018] in, is the state vector at the current moment, ， is the true temperature, is the drift amount, is the bearing vibration accelerometer measurement value, is the state transition matrix, is the control input matrix, ~ (0, ), ~N(0, ) are process noise and observation noise, represents the covariance matrix of the observation noise, To represent the normal distribution, used to describe process noise and observation noise The probability distribution form of The mean is 0, and the covariance matrix is ​​the process noise covariance matrix The normal distribution of The mean is 0 and the covariance matrix is The normal distribution of Update the observation matrix for online calibration, is the observed value; Step S1.2.2: Dynamically adjust the process noise covariance matrix based on vibration data , when the vibration amplitude exceeds the threshold, the process noise covariance matrix The adjustment formula is as follows:

[0019] in, is the baseline drift noise intensity, is the effective value of vibration, is the threshold; Step S1.2.3: Perform online calibration to update the observation matrix : Under preset stable conditions:

[0020] When a sensor abnormality is detected: =[0

[0021] Step S1.2.4: Use the variational Bayesian method to estimate the noise statistical characteristics in real time, as follows: A variational distribution is established, and the variational distribution formula is:

[0022] in, is the variational distribution; Update the noise parameters by iteratively optimizing the KL divergence; Step S1.2.5: Obtain a compensated temperature value based on the result of converting the first element of the state vector, the Kalman gain, the sensor observation value, and the state prediction value into the observation space. The calculation formula of the compensated temperature value is as follows:

[0023] in, is the temperature value after compensation, is the Kalman gain, For the moment For the moment The state prediction value, is the observation matrix, is the result of converting the state prediction value to the observation space, Kalman gain, a weight coefficient, is the state vector The first element of , for example: .

[0024] Step S1.3: Identify the mechanical damage state of the sensor through vibration spectrum analysis combined with the compensated triple-redundant temperature measurement values ​​to obtain the sensor failure diagnosis result.

[0025] Step S2: Obtaining online oil quality parameters of the lubricating oil, which include temperature, pressure, viscosity, water content, and metal particle concentration parameters, as follows: Step S2.1: Measuring the water content of the oil using a micro dielectric constant sensor; Step S2.2: Calculating the real-time viscosity value based on the pressure difference change of the microfluidic detection channel; Step S2.3: detecting the concentration of metal particles using laser-induced breakdown spectroscopy; Step S2.4: dynamically calculating the oil health index based on temperature, pressure, water content, viscosity, and metal particle concentration; The oil quality health index calculation formula is as follows:

[0026] Where, is the oil quality health index, is the temperature of the oil, is the lubricating oil pressure, is the viscosity of the oil, is the water content in the oil, and is the weight coefficient.

[0027] Step S3: constructing a temperature-vibration-oil pressure joint feature space based on the triple-redundant temperature measurement values ​​and the online oil quality parameters, and establishing a fault propagation map based on the temperature-vibration-oil pressure joint feature space; In an embodiment of the present invention, the following steps are further included: obtaining a probability distribution of a fault propagation path according to the fault propagation map, specifically as follows: Step S3.1: Use the attention mechanism LSTM network to extract the temporal correlation features of temperature, vibration, and oil pressure parameters; Step S3.2: Matching the timing correlation features with the preset fault mode in the fault knowledge graph to obtain a matching result; Step S3.3: Based on the matching results, the probability distribution of the fault propagation path is obtained through the graph neural network.

[0028] In an embodiment of the present invention, the method for constructing the fault knowledge graph is as follows: Step S3.2.1: Fault mode ontology modeling; used to define the knowledge structure of the fault domain, including fault types (such as bearing wear, gear fracture), attributes (such as severity, frequency of occurrence) and basic relationships (such as "cause" and "accompany").

[0029] Step S3.2.2: Extracting data features based on monitoring data, wherein the monitoring data includes temperature signals, vibration spectra, oil testing, and historical maintenance records; It should be noted that the features need to be aligned with the fault mode (e.g., high-frequency components of the vibration spectrum → bearing failure) to provide input data for relationship mining; Step S3.2.3: Perform association mining based on the fault mode ontology modeling and data features to obtain mining results; The association relationship mining includes prior relationships based on physical models and data-driven association discovery, wherein the prior relationships based on physical models include thermodynamic associations and mechanical transmission chains; Step S3.2.4: Graph storage and representation learning based on mining results; storing entities and relationships, the purpose of which is to transform the structured knowledge of the previous step into a computable form, providing a basis for dynamic updates (S3.2.5) Step S3.2.5: Dynamically update the graph storage and representation learning results as follows: If the dynamic update trigger condition is met: sensor data distribution drift, the update strategy is executed: node attribute recalibration; If the dynamic update trigger condition is met: a new fault case is stored in the database, the update strategy is executed: incremental graph expansion.

[0030] In the embodiment of the present invention, the establishment of the fault propagation map is specifically as follows: The temperature gradient change rate, vibration harmonic components, and oil particle concentration are extracted as key nodes; Establish a probability transfer matrix between key nodes; The fault propagation weights are updated through a graph convolutional network combined with a probabilistic transfer matrix.

[0031] Step S4: Based on the multi-level diagnosis model and the fault propagation map, composite fault pattern recognition is implemented to obtain the fault classification results, which are as follows: Step S4.1: Use the random forest model to perform anomaly detection on the time domain features and complete the primary judgment of the abnormal state; The time domain features include the mean, variance and kurtosis of temperature and vibration signals; Step S4.2: When an anomaly is detected, the fault type is classified through the deep residual network and the fault classification result is output; Step S4.3: Obtain the confidence level of the fault classification result. If the confidence level is less than the threshold, perform fault evolution deduction through the digital twin model. Step S4.4: Dynamically update model parameters at each level.

[0032] In the embodiment of the present invention, the method for constructing the digital twin model is as follows: Establish the finite element dynamic equation of the bearing-lubricant system:

[0033] in, is the mass matrix, is the damping matrix, is the stiffness matrix, is the displacement vector, represents the velocity vector, is the first-order derivative of the displacement, represents the acceleration vector, The second derivative of displacement, External excitation force, failure coefficient , Fault incentives.

[0034]

[0035] in, is the impact amplitude, is the fault characteristic period, is the attenuation coefficient, is an exponential function, is the moment, which indicates the moment of the entire fault excitation force change process. For the integer parameter used for summation, set the value, is the Dirac function.

[0036] In the embodiment of the present invention, the failure coefficient is solved by inverse problem optimization, specifically as follows: Construct the objective function: β

[0037] Compute the gradient using the adjoint method:

[0038] in, is the accompanying variable, is the objective function, is the partial derivative, is the system equation, Based on parameters The model simulation output value, is the observed data obtained from actual measurement, β is the minimization objective function.

[0039] Based on the above method, the present invention also discloses a turbine bearing temperature and lubricating oil state monitoring and fault diagnosis system, see Figure 3 ,include: Temperature measurement value acquisition module: used to obtain triple-redundant temperature measurement values ​​of turbine bearings; Online oil quality parameter acquisition module: used to obtain online oil quality parameters of lubricating oil, including temperature, pressure, viscosity, water content and metal particle concentration parameters; Fault propagation map acquisition module: used to construct a temperature-vibration-oil pressure joint feature space based on triple-redundant temperature measurement values ​​and online oil quality parameters, and to establish a fault propagation map based on the temperature-vibration-oil pressure joint feature space; Fault classification result acquisition module: used to realize composite fault pattern recognition based on multi-level diagnosis model combined with fault propagation map to obtain fault classification results.

[0040] Example 2: like Figure 1 As shown, this embodiment provides a method for real-time online monitoring and fault diagnosis of turbine bearing temperature and lubricating oil status. The design principle of this method is as follows: multiple parameters are integrated and redundant design is adopted to perform sensor failure diagnosis, then a temperature-vibration-oil pressure joint feature space is constructed, a fault propagation map is established, and finally, composite fault pattern recognition is realized based on a multi-level diagnostic model.

[0041] In this embodiment, the specific implementation steps of the method for real-time online monitoring and fault diagnosis of turbine bearing temperature and lubricating oil status are as follows: Step S1: Collect turbine bearing temperature data According to the redundant design, turbine bearing temperature data is collected and sensor failure diagnosis is performed; the layout of the turbine bearing sensors is as follows: including temperature sensors: fiber Bragg grating, infrared thermal imaging and PT100 sensors are set, and a vibration acceleration sensor is installed coaxially with the temperature sensor. The vibration acceleration sensor is introduced to assist in diagnosing sensor failure modes. The sensor failure diagnosis includes: physical layer: analyzing the vibration spectrum characteristics of the sensor installation location; data layer: detecting abnormal data through error reconstruction of variational autoencoders; model layer: comparing the difference in the measurement values ​​of the three redundant sensors.

[0042] The specific implementation of this step is as follows: Step S1.1: Synchronously obtain triple redundant temperature measurement values ​​of fiber Bragg grating, infrared thermal imaging and PT100 sensor; Step S1.2: Use Kalman filter algorithm to dynamically compensate for the temperature measurement value drift; Step S1.3: Identify the mechanical damage status of the sensor through vibration spectrum analysis.

[0043] The specific implementation method of step S1.2 is as follows: Step S1.2.1: Establish a state-space model of the temperature sensor, including: State equation: , observation equation: ,in, is the state vector at the current moment, is the state vector at the previous moment, , is the true temperature, is the drift amount, is the bearing vibration accelerometer measurement value, is the state transition matrix, is the control input matrix, ~ (0, ), ~N(0, ) are process noise and observation noise, which obey Gaussian distribution. The covariance matrix representing the observation noise can be set by the sensor calibration data or empirical value; the state transfer matrix The determination method includes: establishing the temperature conduction partial differential equation: + ,in, is the thermal diffusivity of the material, is the vibration coupling coefficient, is the vibration heat generation coefficient, Temperature The Laplace operator describes the temperature variation in space, reflects the non-uniformity of temperature distribution and the diffusion trend of heat in space, is discretized into a state space form through the finite difference method, and the matrix parameters are identified using the least squares method using the measured data; Step S1.2.2: Dynamically adjust the process noise covariance matrix based on vibration data Q： When the vibration amplitude exceeds the threshold hour, , is the baseline drift noise intensity, rms is the effective value of vibration; threshold Dynamic settings are used, as follows: historical vibration data is collected to construct a Weibull distribution model, the 95% quantile of the cumulative distribution function is taken as the dynamic threshold, and the threshold parameters are updated every 24 hours; Step S1.2.3: Perform online calibration to update the observation matrix : Under preset stable conditions: =[1 (Normal observation mode) When a sensor abnormality is detected: =[0 (Drift monitoring mode) Step S1.2.4: Use variational Bayesian method to estimate noise statistics in real time: Establish variational distribution: , update the noise parameters by iteratively optimizing the KL divergence; the specific method of iteratively optimizing the KL divergence to update the noise parameters is as follows: Construct the ELBO objective function: , use stochastic gradient descent method to optimize variational parameters, and set the noise parameter update period to 10 seconds; Step S1.2.5: Output the compensated temperature value: , is the Kalman gain, which is calculated as follows: ,in, is the prior estimated covariance matrix, indicating that at time based on The uncertainty of the state estimate obtained from the moment information, is the observation noise covariance matrix, is the observation matrix The transposed matrix of Step S2: Online detection of lubricating oil temperature, pressure, viscosity, water content and metal particle concentration parameters The specific implementation of this step is as follows: Step S2.1: Measuring the water content of the oil using a micro dielectric constant sensor; Step S2.2: Calculate real-time viscosity based on the pressure differential across the microfluidic detection channel. This includes pumpless oil sample circulation via capillary effects, nanoparticle-enhanced Raman spectroscopy for oil oxidation detection, and self-cleaning backflushing every two hours. Step S2.3: Detecting the concentration of metal particles using laser-induced breakdown spectroscopy (LIBS). This technique is a mature technology and will not be described in detail. Step S2.4: Dynamically calculate the oil health index. The formula is as follows:

[0044] Where, is the temperature of the oil (infrared temperature measurement), is the lubricating oil pressure (detected by pressure sensor), is the viscosity of the oil, is the water content in the oil, and is the weight coefficient, which is updated every N minutes through principal component analysis. For example, N is 30 minutes. The principal component analysis method is an existing method, so it will not be described here. It can also be determined by expert experience to reflect the sensitivity and importance of each parameter under different working conditions.

[0045] Step S3: Constructing the temperature-vibration-oil pressure joint feature space and establishing the fault propagation map; The specific implementation of this step is as follows: Step S3.1: Use the attention mechanism LSTM network to extract the temporal correlation features of temperature, vibration, and oil pressure parameters. Using the attention mechanism to extract temporal correlation features is a mature technology and will not be described in detail here. The innovation of this step lies in the multi-source parameters. Step S3.2: Match the current feature vector with the preset fault mode in the fault knowledge graph. The fault knowledge graph is constructed as follows: Step S3.2.1: Fault mode ontology modeling (using OWL to define the fault ontology; using the Protégé tool to build a class hierarchy and define object properties (e.g., hasSymptom, leadsTo)). Step S3.2.2: Extract data features, including temperature signals (extraction method: sliding window statistics (mean, gradient)), vibration spectra (extraction method: wavelet packet decomposition energy entropy), oil testing (extraction method: microfluidic chip output (water content ppm)), and historical maintenance records (extraction method: NLP entity recognition (fault location, action)). Step S3.2.3: Association mining: This includes prior relationships based on physical models and data-driven association discovery (using the Apriori algorithm to mine frequent item sets (e.g., [temperature > 80°C] ∧ [water content > 150ppm] → Bearing corrosion), using a temporal causal discovery algorithm (such as the PC algorithm) to identify causal relationships), where the prior relationships based on physical models include thermodynamic associations (e.g., high temperature → decreased oil film viscosity → increased vibration) and mechanical transmission chains (e.g., bearing wear → shaft misalignment → impeller vibration); Step S3.2.4: Graph storage and representation learning (using Neo4j to store nodes and edges and GraphSAGE to generate node embeddings); S3.2.5: Dynamic update: (1) Trigger condition: sensor data distribution drift, update strategy: node attribute recalibration (technical implementation: online clustering); (2) Trigger condition: new fault case storage, update strategy: incremental graph expansion (technical implementation: Neo4j batch import + deduplication); Step S3.3: Calculate the probability distribution of the fault propagation path through the graph neural network.

[0046] The fault propagation map is constructed as follows: first, the temperature gradient change rate, vibration harmonic components, and oil particle concentration are extracted as key nodes; second, a probability transfer matrix between nodes is established; finally, the fault propagation weights are updated through a graph convolutional network.

[0047] Through the above, a temperature-vibration-oil pressure joint monitoring matrix is ​​established to realize multi-physical field correlation analysis.

[0048] Step S4: Implementing composite fault pattern recognition based on a multi-level diagnostic model; The multi-layer diagnostic model includes a random forest model in the first layer, a deep residual network in the second layer, and a digital twin model in the third layer. The specific implementation method is as follows; Step S4.1: Perform anomaly detection on the time domain features through the random forest model to complete the primary judgment of the abnormal state; the response time is within 100ms. The anomaly detection method includes: (1) extracting the mean, variance and kurtosis of the temperature and vibration signals in the sliding window as time domain features; (2) using the feature importance weighting strategy to improve the random forest splitting criterion: , where the weight factor Determined by the information entropy of historical fault data, Representation characteristics Standard split importances (such as Gini gain or information gain), Indicates traversing all possible split features of the current node; Step S4.2: When an anomaly is detected, the fault type is classified through a deep residual network and the fault classification result is output. The response time is less than 1 second. The deep residual network contains: 3 cascaded residual blocks, each with 2 convolutional layers and 1 skip connection; attention mechanism module, calculates channel weights:

[0049] in, For the The global average pooling result of feature channels, For the The global average pooling result of feature channels, For the The weight matrix or weight vector corresponding to the feature channels, For the The specific structure of the residual block includes the input undergoing 1×5 convolution, batch normalization, and ReLU activation in sequence; the jump connection uses 1×1 convolution to match the dimension; the output is the sum of the convolution path and the jump path; Step S4.3: For low-confidence classification results, perform fault evolution deduction using a digital twin model; the response time is less than 5 seconds. The digital twin model is constructed as follows: Establish the finite element dynamics equation of the bearing-lubricant system:

[0050] The adaptive step-size Runge-Kutta method is used for numerical solution, where is the mass matrix, is the damping matrix, is the stiffness matrix, is the displacement vector, represents the velocity vector, is the first-order derivative of the displacement, represents the acceleration vector, The second derivative of displacement, External excitation force (known external excitation (such as water flow, electromagnetic force)), failure coefficient Based on the provided digital twin model, by adjusting the parameters , so that the model output As close as possible to the actual measured value , thereby optimizing the matching between the model and the real system, the digital twin model introduces the fault influencing factor and fault excitation , The unknown additional force caused by the fault is determined by Control, through Small changes in the value can detect early faults and quantify The value corresponds to the specific fault level (such as wear depth), and can also be based on Trend prediction of remaining useful life; Fault excitation force, as follows:

[0051] in, is the impact amplitude (force amplitude positively correlated with damage size (unit: N)), is the fault characteristic period, is the attenuation coefficient (system damping characteristic (unit: 1 / s)), Dirac Function, expressed in The instantaneous impact of the moment; diagnostic logic: when When it increases, bearing fault characteristic frequencies (such as BPFO and BPFI) will appear in the vibration spectrum; Failure coefficient The inverse problem is solved by optimization, as follows: Construct the objective function: β , and use the adjoint method to calculate the gradient: ,in, is the accompanying variable, is the objective function (usually an error function (such as the mean square error between simulation and measured data), the quantity to be minimized), is the partial derivative, is the system equation (i.e., the finite element dynamics equation), Based on parameters Model simulation output values ​​(such as simulation data, theoretical prediction values), The observed data is actually measured (such as the real value collected by the sensor); Step S4.4: Dynamically update the model parameters at each level. There are many methods for dynamically updating the model parameters at each level, including but not limited to: S4.41, dynamically updates model parameters at each level based on the online Bayesian method. Based on Bayesian inference, the model parameters are regarded as probability distributions, and parameter adjustment is achieved by online updating of the posterior distribution. Specifically, the first-layer model: Bayesian linear regression replaces random forest (computational efficiency needs to be weighed), and the second-layer model: Bayesian filtering is used for digital twin parameter updates.

[0052] S4.42, based on continuous learning, dynamically updates the model parameters at each level, and prevents the model from forgetting historical knowledge through regularization, dynamic architecture or playback mechanism. Typical methods: EWC (elastic weight consolidation): protect important parameters; GEM (gradient context memory): constrain the gradient update direction; dynamic network expansion: add new nodes to adapt to new tasks; specifically, the second-layer model: use GEM to constrain the classifier update direction; the third-layer model: dynamically expand the digital twin model structure.

[0053] S4.43, online ensemble learning, dynamically maintains model sets and updates them through weighted voting or stacked generalization. Typical methods include: dynamic weighted ensemble: adjusting weights based on new data performance, incremental bagging: gradually adding new base learners; specifically, the first-layer model: incremental bagging updates the decision tree set; the second-layer model: integrating multiple residual network classifiers.

[0054] S4.44, online optimization algorithm, converts parameter update into an online optimization problem and uses gradient information to iteratively solve it. Typical methods include: Online Gradient Descent (OGD): single-sample gradient update, Adaptive Moment Estimation (AdaMod): improved Adam variant; specifically, the second-layer model: AdaMod optimizes the residual network; the third-layer model: online gradient descent adjusts the digital twin parameters.

[0055] In a further preferred solution, a hybrid update strategy may be adopted: The first-layer random forest model updates the decision tree set through the incremental bagging algorithm; The second-layer deep residual network uses elastic weight consolidation (EWC) to constrain parameter updates and the adaptive moment estimation (AdaMod) optimizer to adjust the learning rate; The third-layer digital twin model adjusts the digital twin parameters through online gradient descent to achieve updates.

[0056] Through the above method, the reliability of turbine bearing and lubricating oil status monitoring is high, the oil monitoring is comprehensive, and the ability to identify complex faults is improved.

[0057] The above content is only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for monitoring and diagnosing the temperature and lubricating oil status of a turbine bearing, characterized in that: The following steps are involved: Obtain triple-redundant temperature measurements of turbine bearings; Obtaining online oil quality parameters of the lubricating oil, wherein the online oil quality parameters include temperature, pressure, viscosity, water content, and metal particle concentration parameters; Construct a temperature-vibration-oil pressure joint feature space based on triple-redundant temperature measurements and online oil quality parameters, and establish a fault propagation map; Based on the multi-level diagnosis model and the fault propagation map, composite fault pattern recognition is realized to obtain the fault classification result.

2. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 1, characterized in that: The three redundant temperature measurement values ​​of the turbine bearing are obtained as follows: Synchronously obtain triple-redundant temperature measurements from fiber Bragg grating, infrared thermal imaging, and PT100 sensors; The Kalman filter algorithm is used to dynamically compensate for the drift of triple redundant temperature measurements; The mechanical damage status of the sensor is identified through vibration spectrum analysis combined with compensated triple-redundant temperature measurement values, and the sensor failure diagnosis results are obtained.

3. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 2, characterized in that: The Kalman filter algorithm is used to dynamically compensate for the drift of the triple redundant temperature measurement values, as follows: Establishing a state space model of the temperature sensor, wherein the state space model includes: a state equation and an observation equation; The state equation is as follows: The observation equation is as follows: in, is the state vector at the current moment, ， is the true temperature, is the drift amount, is the bearing vibration accelerometer measurement value, is the state transition matrix, is the control input matrix, ~ (0, ), ~N(0, ) are process noise and observation noise, represents the covariance matrix of the observation noise, To represent the normal distribution, used to describe process noise and observation noise The probability distribution form of The mean is 0, and the covariance matrix is ​​the process noise covariance matrix The normal distribution of The mean is 0 and the covariance matrix is The normal distribution of Update the observation matrix for online calibration, is the observed value; The compensated temperature value is obtained by converting the first element of the state vector, the Kalman gain, the observation value, and the state prediction value into the observation space. The calculation formula of the compensated temperature value is as follows: in, is the temperature value after compensation, is the Kalman gain, For the moment For the moment The state prediction value, is the observation matrix, is the result of converting the state prediction value to the observation space, Kalman gain, a weight coefficient, is the state vector The first element of .

4. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 1, characterized in that: The online oil quality parameters of the lubricating oil are obtained as follows: Measure the water content in oil through a micro dielectric constant sensor; Calculate real-time viscosity values ​​based on pressure difference changes in microfluidic detection channels; Laser-induced breakdown spectroscopy was used to detect the concentration of metal particles; The oil health index is dynamically calculated based on temperature, pressure, water content, viscosity and metal particle concentration.

5. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 1, characterized in that: The following steps are also included: The probability distribution of the fault propagation path is obtained according to the fault propagation map, as follows: The attention mechanism LSTM network is used to extract the temporal correlation features of temperature, vibration, and oil pressure parameters; Match the timing correlation features with the preset fault mode in the fault knowledge graph to obtain the matching results; According to the matching results, the probability distribution of the fault propagation path is obtained through the graph neural network.

6. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 5, characterized in that: The method for constructing the fault knowledge graph is as follows: Failure mode ontology modeling; Extracting data features based on monitoring data, wherein the monitoring data includes temperature signals, vibration spectra, oil testing, and historical maintenance records; Perform association mining based on fault mode ontology modeling and data features to obtain mining results; Perform graph storage and representation learning based on mining results; Dynamically update the graph storage and representation learning results as follows: If the dynamic update trigger condition is met: sensor data distribution drift, the update strategy is executed: node attribute recalibration; If the dynamic update trigger condition is met: a new fault case is stored in the database, the update strategy is executed: incremental graph expansion.

7. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 1, characterized in that: The establishment of the fault propagation map is as follows: The temperature gradient change rate, vibration harmonic components, and oil particle concentration are extracted as key nodes; Establish a probability transfer matrix between key nodes; The fault propagation weights are updated through a graph convolutional network combined with a probabilistic transfer matrix.

8. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 1, characterized in that: The multi-level diagnostic model is combined with the fault propagation map to realize the complex fault pattern recognition and obtain the fault classification results, which are as follows: Anomaly detection is performed on time domain features using a random forest model to complete a primary judgment of abnormal conditions; the time domain features include the mean, variance, and kurtosis of temperature and vibration signals; When an anomaly is detected, the fault type is classified through the deep residual network and the fault classification result is output; Obtain the confidence level of the fault classification result. If the confidence level is less than the threshold, perform fault evolution deduction through the digital twin model. Dynamically update model parameters at each level.

9. The method for monitoring and diagnosing the turbine bearing temperature and lubricating oil status according to claim 8, characterized in that: The method for constructing the digital twin model is as follows: Establish the finite element dynamic equation of the bearing-lubricant system: in, is the mass matrix, is the damping matrix, is the stiffness matrix, is the displacement vector, represents the velocity vector, is the first-order derivative of the displacement, represents the acceleration vector, The second derivative of displacement, External excitation force, failure coefficient , Fault incentives; in, is the impact amplitude, is the fault characteristic period, is the attenuation coefficient, is an exponential function, is the moment, which indicates the moment of the entire fault excitation force change process. For the integer parameter used for summation, set the value, is the Dirac function; The failure coefficient is solved by inverse problem optimization, as follows: Construct the objective function: β Compute the gradient using the adjoint method: in, is the accompanying variable, is the objective function, is the partial derivative, is the system equation, Based on parameters The model simulation output value, is the observed data obtained from actual measurement, β is the minimization objective function.

10. The turbine bearing temperature and lubricating oil status monitoring and fault diagnosis system is characterized by: include: Temperature measurement value acquisition module: used to obtain triple-redundant temperature measurement values ​​of turbine bearings; Online oil quality parameter acquisition module: used to obtain online oil quality parameters of lubricating oil, including temperature, pressure, viscosity, water content and metal particle concentration parameters; Fault propagation map acquisition module: used to construct the temperature-vibration-oil pressure joint feature space based on triple-redundant temperature measurement values ​​and online oil quality parameters, and establish a fault propagation map; Fault classification result acquisition module: used to realize composite fault pattern recognition based on multi-level diagnosis model combined with fault propagation map to obtain fault classification results.

Images