A wind turbine fault prediction method and system based on digital twinning
By constructing a high-fidelity multiphysics mechanism simulation model and real-time data calibration, and combining the method of virtual and physical feature fusion, the problems of low accuracy and lag in wind turbine fault prediction are solved. It achieves high-precision prediction with individual self-adaptation and sensitivity to early faults, and supports closed-loop self-evolutionary learning of digital twins.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing wind turbine fault prediction methods suffer from low accuracy, delayed updates, and insensitivity to early faults. Current digital twin technology has failed to achieve dynamic and deep integration of physical mechanisms and data-driven models, resulting in prediction models that cannot achieve adaptive and high-precision early fault warnings.
A high-fidelity multiphysics mechanism simulation model is constructed, multi-source monitoring data is collected in real time for dynamic online calibration, and a calibrated real-time mechanism model is generated. Virtual sensing features and physical sensing features are extracted, and their physical meanings are aligned and deeply fused to form a high-dimensional hybrid feature set. This set is then input into a hybrid intelligent prediction model for fault prediction, and the model is updated through reverse simulation and residual analysis.
It realizes individual adaptive fault prediction of wind turbine units, improves early fault sensitivity, dynamically optimizes the prediction model, supports virtual and real interactive verification, and realizes the upgrade from predictive maintenance to proactive maintenance.
Smart Images
Figure CN122106836A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power equipment condition monitoring and fault diagnosis technology, specifically to a wind turbine fault prediction method and system based on digital twins. Background Technology
[0002] Wind turbines operate in complex and variable environments for extended periods, making critical components (such as gearboxes, generators, blades, and main bearings) prone to failure, leading to unplanned downtime and significant economic losses. Existing fault prediction methods mainly fall into two categories: one is mechanistic analysis based on physical models, which offers high accuracy but suffers from complex modeling and difficulty adapting to individual differences and degradation; the other is statistical or machine learning methods based on historical data, which rely on large amounts of fault data and are insensitive to early, slowly changing faults, exhibiting weak generalization ability. Current applications of digital twin technology in the wind power field are mostly limited to visualization or offline simulation, failing to achieve dynamic deep integration, real-time interaction, and bidirectional optimization of physical mechanisms and data-driven models. This results in lagging prediction model updates and a decline in prediction accuracy as equipment degrades, making it impossible to achieve truly adaptive, high-precision early fault warnings. Summary of the Invention
[0003] Purpose of the invention: To provide a method and system for predicting wind turbine faults based on digital twins, so as to solve the above-mentioned problems existing in the prior art.
[0004] Technical solution: A method for predicting wind turbine faults based on digital twins, comprising the following steps:
[0005] Construct an initial digital twin of the target wind turbine, the initial digital twin including a high-fidelity multiphysics mechanism simulation model based on the turbine design parameters;
[0006] Real-time acquisition of multi-source monitoring data during the operation of the target wind turbine;
[0007] Based on the multi-source monitoring data, the boundary conditions and key physical parameters of the high-fidelity multiphysics mechanism simulation model are dynamically calibrated online to make the simulation output consistent with the current operating state of the actual unit, and to generate a calibrated real-time mechanism model.
[0008] Virtual sensing features characterizing the health status of components are extracted from the calibrated real-time mechanism model; simultaneously, physical sensing features are extracted from the multi-source monitoring data.
[0009] The virtual sensing features and the physical sensing features are physically aligned and deeply fused to form a high-dimensional hybrid feature set;
[0010] The high-dimensional hybrid feature set is input into the hybrid intelligent prediction model, which integrates data-driven algorithms and physical mechanism constraints.
[0011] The hybrid intelligent prediction model outputs prediction results of fault type, fault probability, and remaining useful life.
[0012] The abnormal patterns in the prediction results are fed back to the real-time mechanism model for reverse simulation and residual analysis; if the residual exceeds the threshold, an update instruction for the structure of the real-time mechanism model or the fault mode library is triggered.
[0013] In a further embodiment, the dynamic online calibration specifically includes: using some measurable state variables in the multi-source monitoring data as constraint targets, and employing an adaptive filtering algorithm or a real-time parameter identification algorithm to reversely adjust the friction coefficient, stiffness damping coefficient, efficiency parameter, or environmental load coefficient in the mechanism simulation model, so as to minimize the error between the simulated state variables and the measured state variables.
[0014] In a further embodiment, the virtual sensing features include at least one of the following calculated from the calibrated mechanistic model: gearbox tooth surface contact stress distribution, main bearing raceway subsurface stress, cumulative fatigue damage at the blade root, and generator air gap magnetic flux density distortion rate.
[0015] In a further embodiment, the physical meaning alignment and deep fusion specifically involves: synchronizing, normalizing, and performing correlation analysis on virtual features and physical features representing the same physical phenomenon or the state of the same component in the time-frequency domain, and then performing feature-level or decision-level fusion based on the analysis results.
[0016] In a further embodiment, the physical mechanism constraint of the hybrid intelligent prediction model is embodied in the following way: the component failure physical equation or degradation trajectory derived from the real-time mechanism model is used as a regularization term of the neural network loss function, or as a boundary constraint of the feasible region of the prediction result.
[0017] In a further embodiment, after the trigger update instruction is given, the method further includes: injecting a simulated fault corresponding to the predicted fault into the digital twin, running the updated mechanism model, observing the changes in virtual sensing characteristics, and comparing and verifying them with subsequent actual monitoring data to complete a closed-loop self-evolutionary learning of the digital twin.
[0018] A system for implementing the digital twin-based wind turbine fault prediction method described above includes:
[0019] The digital twin construction and simulation module is used to build and maintain the high-fidelity multiphysics mechanism simulation model.
[0020] The multi-source data real-time sensing and acquisition module is used to acquire the operating data of the physical unit;
[0021] The mechanism model dynamic calibration module is used to realize online parameter adjustment of the mechanism model;
[0022] The virtual-real feature fusion module is used to generate the high-dimensional hybrid feature set;
[0023] The hybrid intelligent prediction and decision-making module, which embeds the hybrid intelligent prediction model, is used to output prediction results;
[0024] The two-way interaction and evolution module is used to enable the prediction results to provide feedback to the digital twin and drive its updates;
[0025] The digital twin collaborative computing platform is used to support data interaction, model scheduling, and closed-loop control flow of all the above modules.
[0026] In a further embodiment, the mechanism model dynamic calibration module and the hybrid intelligent prediction and decision module are deployed in parallel on the edge computing gateway to achieve real-time calibration and rapid early warning of faults under high-frequency vibration data; the digital twin collaborative computing platform is deployed in the cloud to perform large-scale simulation, deep model training and global optimization.
[0027] In a further embodiment, the bidirectional interaction and evolution module also includes a fault mode knowledge graph for storing historical fault cases, simulation inference rules and residual-fault mapping relationships; when a model update instruction is triggered, the knowledge graph assists in generating candidate model correction schemes.
[0028] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a digital twin-based wind turbine fault prediction method.
[0029] Beneficial Effects: This invention relates to a method and system for predicting wind turbine faults based on digital twins, which has the following beneficial effects:
[0030] 1. Dynamic fusion and bidirectional optimization: It breaks through the limitation of the mechanism model and data model being "independent" in traditional methods, and realizes online, bidirectional calibration and collaborative optimization of the two, enabling the digital twin to have self-evolution capabilities.
[0031] 2. High sensitivity to early faults: By fusing virtual features with clear physical meaning generated by the mechanism model with real sensing features, the weak signals of early faults are amplified, improving the predictive foresight.
[0032] 3. Strong individual adaptability: The digital twin continuously learns the operating data of a specific unit and constantly adjusts the model parameters in a personalized manner to accurately match the actual health status and degradation trajectory of the unit.
[0033] 4. Virtual-Real Interaction Verification: Supports fault injection and simulation in the digital space to verify the effectiveness of the prediction algorithm, optimize operation and maintenance decisions, and upgrade from predictive maintenance to proactive maintenance. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the composition framework of the wind turbine fault prediction system based on digital twins described in this invention.
[0035] Figure 2 This is a flowchart illustrating the wind turbine fault prediction method based on digital twins as described in this invention. Detailed Implementation
[0036] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.
[0037] This invention proposes a method and system for predicting wind turbine faults based on digital twins, aiming to address the problems in existing technologies where digital twin applications are limited to visualization or offline simulation, lack dynamic deep fusion of physical mechanisms and data-driven models, lag in prediction model updates, and low sensitivity to early faults. By constructing a high-fidelity multiphysics mechanism model, real-time calibration, fusion of virtual and real features, and a closed-loop self-evolution mechanism, this invention achieves individualized and adaptive fault prediction for wind turbines.
[0038] The wind turbine fault prediction method based on digital twins involved in this invention specifically includes the following steps:
[0039] Construct an initial digital twin of the target wind turbine, the initial digital twin including a high-fidelity multiphysics mechanism simulation model based on the turbine design parameters;
[0040] Real-time acquisition of multi-source monitoring data during the operation of the target wind turbine;
[0041] Based on the multi-source monitoring data, the boundary conditions and key physical parameters of the high-fidelity multiphysics mechanism simulation model are dynamically calibrated online to make the simulation output consistent with the current operating state of the actual unit, and to generate a calibrated real-time mechanism model.
[0042] Virtual sensing features characterizing the health status of components are extracted from the calibrated real-time mechanism model; simultaneously, physical sensing features are extracted from the multi-source monitoring data.
[0043] The virtual sensing features and the physical sensing features are physically aligned and deeply fused to form a high-dimensional hybrid feature set;
[0044] The high-dimensional hybrid feature set is input into the hybrid intelligent prediction model, which integrates data-driven algorithms and physical mechanism constraints.
[0045] The hybrid intelligent prediction model outputs prediction results of fault type, fault probability, and remaining useful life.
[0046] The abnormal patterns in the prediction results are fed back to the real-time mechanism model for reverse simulation and residual analysis; if the residual exceeds the threshold, an update instruction for the structure of the real-time mechanism model or the fault mode library is triggered.
[0047] The dynamic online calibration specifically includes: using some measurable state variables in the multi-source monitoring data as constraint targets, and employing an adaptive filtering algorithm or a real-time parameter identification algorithm to reversely adjust the friction coefficient, stiffness damping coefficient, efficiency parameter, or environmental load coefficient in the mechanism simulation model, so as to minimize the error between the simulated state variables and the measured state variables.
[0048] The virtual sensing features include at least one of the following calculated from the calibrated mechanistic model: gearbox tooth surface contact stress distribution, main bearing raceway subsurface stress, cumulative fatigue damage at the blade root, and generator air gap magnetic flux density distortion rate.
[0049] The physical meaning alignment and deep fusion specifically refers to: synchronizing, normalizing, and performing correlation analysis on virtual features and physical features representing the same physical phenomenon or the state of the same component in the time and frequency domain, and then performing feature-level or decision-level fusion based on the analysis results.
[0050] The physical mechanism constraint of the hybrid intelligent prediction model is manifested as follows: the component failure physical equation or degradation trajectory derived from the real-time mechanism model is used as the regularization term of the neural network loss function, or as the boundary constraint of the feasible region of the prediction result.
[0051] After the update command is triggered, the process further includes: injecting a simulated fault corresponding to the predicted fault into the digital twin, running the updated mechanism model, observing the changes in virtual sensing characteristics, and comparing and verifying them with subsequent actual monitoring data to complete a closed-loop self-evolutionary learning of the digital twin.
[0052] A system for implementing the digital twin-based wind turbine fault prediction method described above includes:
[0053] The digital twin construction and simulation module is used to build and maintain the high-fidelity multiphysics mechanism simulation model.
[0054] The multi-source data real-time sensing and acquisition module is used to acquire the operating data of the physical unit;
[0055] The mechanism model dynamic calibration module is used to realize online parameter adjustment of the mechanism model;
[0056] The virtual-real feature fusion module is used to generate the high-dimensional hybrid feature set;
[0057] The hybrid intelligent prediction and decision-making module, which embeds the hybrid intelligent prediction model, is used to output prediction results;
[0058] The two-way interaction and evolution module is used to enable the prediction results to provide feedback to the digital twin and drive its updates;
[0059] The digital twin collaborative computing platform is used to support data interaction, model scheduling, and closed-loop control flow of all the above modules.
[0060] The dynamic calibration module of the mechanism model and the hybrid intelligent prediction and decision module are deployed in parallel on the edge computing gateway to realize real-time calibration and early warning of faults under high-frequency vibration data; the digital twin collaborative computing platform is deployed in the cloud to perform large-scale simulation, deep model training and global optimization.
[0061] The bidirectional interaction and evolution module also includes a fault mode knowledge graph, which stores historical fault cases, simulation inference rules and residual-fault mapping relationships; when a model update command is triggered, the knowledge graph assists in generating candidate model correction schemes.
[0062] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a digital twin-based wind turbine fault prediction method.
[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0064] Example 1: A fault prediction method for wind turbines based on digital twins
[0065] This embodiment provides a method for predicting wind turbine faults based on digital twins. Its core lies in constructing a digital twin that can interact with and dynamically evolve with the physical wind turbine in real time, and achieving high-precision and forward-looking fault prediction through deep fusion of mechanisms and data. Figure 1 The overall architecture of this method is shown.
[0066] Step S101: For the target wind turbine (e.g., a certain model of 2MW doubly-fed asynchronous wind turbine), construct a high-fidelity multiphysics mechanism simulation model based on its design drawings, material properties, aerodynamic characteristics, transmission chain parameters, etc. This model should cover multiple physical domains such as aerodynamics, structure, transmission, and electrical systems, and be able to simulate the dynamic response of the turbine under different boundary conditions such as wind speed, turbulence, and temperature in the simulation environment. The initial model parameters are set based on the design values.
[0067] Step S102: Collect operational data in real time through a sensor network installed on the physical unit, including but not limited to:
[0068] SCADA data includes wind speed, power, engine speed, blade pitch angle, generator temperature, oil temperature, etc., with a sampling frequency typically ranging from 1 to 10 Hz.
[0069] Vibration data: Vibration signals of key components such as gearbox, main bearing, and generator in multiple directions are collected using accelerometers. The sampling frequency needs to reach several kHz to tens of kHz to capture high-frequency fault characteristics.
[0070] Stress / strain data: Dynamic stress in key components such as blade roots and tower sections is measured using fiber optic grating sensors or strain gauges.
[0071] Oil analysis data: Indicators such as abrasive concentration, viscosity, and moisture content provided by online oil sensors.
[0072] Acoustic emission / ultrasonic data: used to monitor crack initiation and propagation. This multi-source heterogeneous data is aggregated and preprocessed (such as noise reduction and timestamp alignment) through an industrial IoT gateway before being transmitted to a digital twin collaborative computing platform.
[0073] Step S103: Due to manufacturing tolerances, assembly differences, material property dispersion, and performance degradation caused by long-term operation, the simulation output of the initial mechanism model will inevitably deviate from the actual operating state of the physical unit. Therefore, online calibration using real-time monitoring data is required.
[0074] The specific calibration process employs a hybrid algorithm combining adaptive ensemble Kalman filtering (AEnKF) and parameter identification.
[0075] 1. Joint estimation of state and parameters: This involves estimating key adjustable physical parameters of the mechanistic model (such as gearbox meshing stiffness and damping coefficient c). g The transmission efficiency η, blade aerodynamic coefficient correction factor Ccorr, etc., are expanded into state variables, which together with the original dynamic state variables of the model (such as transmission chain torsional displacement and velocity) form an augmented state vector. .
[0076] 2. Define the observation vector: Select some easily measurable physical quantities that can reflect the overall state of the system as the observation vector y, such as generator speed. Gearbox input shaft torque Cabin acceleration wait.
[0077] 3. AEnKF Iterative Updates:
[0078] Prediction Step: Based on the prior distribution set of the augmented state vectors, forward simulation is performed using a mechanistic model to obtain the state prediction set. .
[0079] Update step: Map the predicted state to the observation space to obtain the predicted observation set. Calculate the sample covariance of the predicted observations. and the cross-covariance with the augmented state Subsequently, actual observation data were used. Calculate Kalman gain And update the augmented state set:
[0080]
[0081] in This process involves observing random disturbances in the noise. Simultaneously, the dynamic state and model parameters are updated.
[0082] 4. Parameter Identification Assistance: For parameters that change slowly or have clear physical meaning, parameter identification algorithms based on minimizing model output error (such as Particle Swarm Optimization, PSO) can be used periodically for batch optimization. The results can be used to supplement or correct the prior values of parameters in AEnKF. Through the above process, the boundary conditions (such as actual wind profiles) and internal parameters of the mechanistic model can be continuously adjusted, so that its simulation output (such as torque of each shaft segment, gear meshing force, bearing load, etc.) remains highly consistent with the actual state of the physical unit under the current operating conditions, generating a calibrated real-time mechanistic model.
[0083] Step S104: Virtual Sensing Feature Extraction: From the calibrated real-time mechanism model, calculate and extract physical quantities that are difficult or costly to measure directly on the physical unit and are closely related to the health status of components as virtual features. For example:
[0084] Gearbox tooth surface contact stress distribution characteristics: Based on the calibrated gear dynamics model, the time-varying contact stress between meshing teeth is calculated. Extract its peak value mean and statistical characteristics reflecting the non-uniformity of stress distribution (such as skewness) kurtosis ).
[0085] Stress characteristics of the main bearing raceway subsurface: Based on Hertzian contact theory, the maximum shear stress of the subsurface in the contact area between the roller and the raceway is calculated. The stress and its depth are the main causes of bearing fatigue spalling.
[0086] Accumulated fatigue damage at the blade root: Based on the calibrated aeroelastic model and load spectrum, the fatigue damage degree D(t) at the critical section of the blade root is calculated in real time using Miner's linear cumulative damage rule.
[0087]
[0088] in The stress amplitude is The number of loops, This corresponds to the material fatigue life.
[0089] Generator air gap eccentricity flux density distortion rate: Based on electromagnetic field simulation, the ratio of the fundamental wave to the amplitude of each harmonic of the air gap flux density is calculated under rotor dynamic eccentricity fault, which serves as a sensitive indicator of early electrical faults.
[0090] Entity sensing feature extraction: Extracting traditional fault features from multi-source monitoring data using signal processing techniques, such as:
[0091] Time-domain characteristics of vibration signals (root mean square value, peak value, kurtosis, etc.).
[0092] Frequency domain characteristics (characteristic frequency amplitudes obtained through Fast Fourier Transform (FFT), such as gear meshing frequency and its sidebands, bearing fault characteristic frequencies).
[0093] Time-frequency domain characteristics (energy entropy, marginal spectrum, etc. obtained through wavelet packet transform or empirical mode decomposition EMD).
[0094] Step S105: To avoid feature redundancy or conflicts caused by simple splicing, physical alignment and deep fusion are required:
[0095] 1. Spatiotemporal synchronization: Using a unified timestamp, the time series of virtual features are strictly aligned with the corresponding physical features (such as sensor data of the same parts or similar physical locations).
[0096] 2. Normalization and Correlation Analysis: Normalize both virtual and entity features (e.g., Z-score standardization). Then, calculate the correlation coefficient and mutual information between virtual and entity features to analyze their inherent physical relationship. For example, the sideband amplitude of the meshing frequency in a gearbox vibration signal should have a strong correlation with the tooth surface contact stress fluctuation in the virtual feature.
[0097] 3. Feature-level fusion: For strongly correlated feature pairs, weighted fusion based on an attention mechanism can be used. A dynamic weight is assigned to each feature. The weight is determined by the correlation between the feature and the target fault mode and its signal-to-noise ratio. The fused features for:
[0098]
[0099] in, The temperature parameter controls the smoothness of the weight distribution.
[0100] 4. Decision-level fusion: For feature groups from different information sources and with strong independence, they can be input separately into sub-prediction models for preliminary judgment, and then their outputs (such as fault probabilities) can be fused. For example, using DS evidence theory, the outputs of the vibration feature sub-model and the virtual stress feature sub-model can be used as evidence to synthesize a comprehensive fault confidence allocation.
[0101] Through the above steps, a high-dimensional hybrid feature set that reflects both the physical essence and is rich in data information is formed.
[0102] Step S106: Construct as follows Figure 2 The hybrid intelligent prediction model shown is based on the integration of data-driven models and physical mechanism constraints.
[0103] Model architecture: It uses deep neural networks (such as Long Short-Term Memory Network LSTM or Graph Convolutional Network GCN) as the backbone to learn the complex mapping relationship between high-dimensional mixed features and failure modes and remaining useful lifetime (RUL).
[0104] Introduction of physical mechanism constraints:
[0105] Loss function regularization: The physical equations of component degradation derived from the real-time mechanism model are used as prior knowledge and added to the loss function of the neural network. For example, for the fatigue degradation of bearings, the damage accumulation usually follows Paris's law. Define physical consistency loss. :
[0106]
[0107] Where D is the damage variable predicted by the network. It is the rate of change. The range of stress intensity factors is calculated from virtual features, where C and m are material constants. This is the regularization coefficient. The total loss function is... ,in This is the cross-entropy or mean squared error between the predicted value and the true label.
[0108] Output layer constraints: The range of variation of key physical quantities (such as vibration energy and temperature rise rate) of the component simulated by the mechanism model under different states such as health, minor failure and severe failure is used as the boundary constraint of the activation function of the neural network output layer to ensure that the predicted failure probability or RUL falls within the physically feasible range.
[0109] Model Training and Output: The hybrid model is trained using historical normal operation data, failure case data, and simulated failure data generated by digital twins. After training, the model receives a high-dimensional hybrid feature set generated in real time and outputs in parallel: failure type (classification result), failure probability (confidence level of belonging to each type of failure), and remaining useful life (RUL) (regression prediction value, such as the number of hours the gearbox can still operate normally).
[0110] Step S107: Feed back the abnormal prediction results (such as high failure probability, sudden drop in RUL) output by the hybrid intelligent prediction model to the real-time mechanism model and start the reverse analysis process:
[0111] Reverse simulation: In the digital twin, assuming the predicted fault type occurs (such as setting a tiny crack on a tooth of a gear), run the mechanism model and observe how virtual sensing characteristics (such as the distribution of contact stress on the tooth surface) change.
[0112] Residual analysis: Calculate the residual R between the virtual feature change pattern obtained from the reverse simulation and the entity feature change pattern extracted from the actual monitoring data.
[0113] Update decision: Set a residual threshold ϵ. If This indicates that the current mechanistic model or failure mode library fails to adequately explain the observed anomalies, indicating a model mismatch. An update command is then triggered.
[0114] Closed-loop self-evolutionary learning:
[0115] Model Structure / Parameter Update: Based on the residual patterns and combined with the failure mode knowledge graph (which stores historical failure cases, failure physical models, and residual-fault mapping rules), candidate model correction schemes are generated. For example, if the residuals exhibit abnormal high-frequency components, it may indicate the need to introduce a new gear local defect dynamics model. A Bayesian model update framework is used to evaluate and select candidate models, updating the structure or parameters of the mechanistic model.
[0116] Fault mode library update: The newly identified fault modes and their corresponding virtual and real feature evolution rules are added to the fault mode knowledge graph to enrich prior knowledge.
[0117] Verification and Iteration: In the updated digital twin, simulated faults are injected again for forward simulation. The changes in its virtual characteristics are compared with the actual monitoring data over a subsequent period to verify the effectiveness of the update. This completes one "prediction".
[0118] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.
Claims
1. A method for predicting wind turbine faults based on digital twins, characterized by: Includes the following steps: Construct an initial digital twin of the target wind turbine, the initial digital twin including a high-fidelity multiphysics mechanism simulation model based on the turbine design parameters; Real-time acquisition of multi-source monitoring data during the operation of the target wind turbine; Based on the multi-source monitoring data, the boundary conditions and key physical parameters of the high-fidelity multiphysics mechanism simulation model are dynamically calibrated online to make the simulation output consistent with the current operating state of the actual unit, and to generate a calibrated real-time mechanism model. Virtual sensing features characterizing the health status of components are extracted from the calibrated real-time mechanism model; simultaneously, physical sensing features are extracted from the multi-source monitoring data. The virtual sensing features and the physical sensing features are physically aligned and deeply fused to form a high-dimensional hybrid feature set; The high-dimensional hybrid feature set is input into the hybrid intelligent prediction model, which integrates data-driven algorithms and physical mechanism constraints. The hybrid intelligent prediction model outputs prediction results of fault type, fault probability, and remaining useful life. The abnormal patterns in the prediction results are fed back to the real-time mechanism model for reverse simulation and residual analysis; if the residual exceeds the threshold, an update instruction for the structure of the real-time mechanism model or the fault mode library is triggered.
2. The wind turbine fault prediction method based on digital twin according to claim 1, characterized in that: The dynamic online calibration specifically includes: using some measurable state variables in the multi-source monitoring data as constraint targets, and employing an adaptive filtering algorithm or a real-time parameter identification algorithm to reversely adjust the friction coefficient, stiffness damping coefficient, efficiency parameter, or environmental load coefficient in the mechanism simulation model, so as to minimize the error between the simulated state variables and the measured state variables.
3. The wind turbine fault prediction method based on digital twin according to claim 1, characterized in that: The virtual sensing features include at least one of the following calculated from the calibrated mechanistic model: gearbox tooth surface contact stress distribution, main bearing raceway subsurface stress, cumulative fatigue damage at the blade root, and generator air gap magnetic flux density distortion rate.
4. The wind turbine fault prediction method based on digital twin according to claim 1, characterized in that: The physical meaning alignment and deep fusion specifically refers to: synchronizing, normalizing, and performing correlation analysis on virtual features and physical features representing the same physical phenomenon or the state of the same component in the time and frequency domain, and then performing feature-level or decision-level fusion based on the analysis results.
5. The wind turbine fault prediction method based on digital twin according to claim 1, characterized in that: The physical mechanism constraint of the hybrid intelligent prediction model is manifested as follows: the component failure physical equation or degradation trajectory derived from the real-time mechanism model is used as the regularization term of the neural network loss function, or as the boundary constraint of the feasible region of the prediction result.
6. The wind turbine fault prediction method based on digital twin according to claim 1, characterized in that: After the update command is triggered, the process further includes: injecting a simulated fault corresponding to the predicted fault into the digital twin, running the updated mechanism model, observing the changes in virtual sensing characteristics, and comparing and verifying them with subsequent actual monitoring data to complete a closed-loop self-evolutionary learning of the digital twin.
7. A system for implementing the wind turbine fault prediction method based on digital twins as described in any one of claims 1-6, characterized in that, include: The digital twin construction and simulation module is used to build and maintain the high-fidelity multiphysics mechanism simulation model. The multi-source data real-time sensing and acquisition module is used to acquire the operating data of the physical unit; The mechanism model dynamic calibration module is used to realize online parameter adjustment of the mechanism model; The virtual-real feature fusion module is used to generate the high-dimensional hybrid feature set; The hybrid intelligent prediction and decision-making module, which embeds the hybrid intelligent prediction model, is used to output prediction results; The two-way interaction and evolution module is used to enable the prediction results to provide feedback to the digital twin and drive its updates; The digital twin collaborative computing platform is used to support data interaction, model scheduling, and closed-loop control flow of all the above modules.
8. The system of the wind turbine fault prediction method based on digital twin according to claim 7, characterized in that: The dynamic calibration module of the mechanism model and the hybrid intelligent prediction and decision module are deployed in parallel on the edge computing gateway to realize real-time calibration and early warning of faults under high-frequency vibration data; the digital twin collaborative computing platform is deployed in the cloud to perform large-scale simulation, deep model training and global optimization.
9. The system for wind turbine fault prediction based on digital twins according to claim 7, characterized in that: The bidirectional interaction and evolution module also includes a fault mode knowledge graph, which is used to store historical fault cases, simulation inference rules and residual-fault mapping relationships; When a model update instruction is triggered, the knowledge graph assists in generating candidate model correction schemes.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 6.