Machine learning based deep self-diagnosis method and device for wind turbine

Abstract

The embodiment of the application provides a kind of based on machine learning's wind turbine deep self-diagnosis method and device, method includes: by dividing wind turbine overall structure into structural components and relevant space, and from overall structure, multiple physical field sensor data is collected, the structural components of wind turbine are discretized into multiple finite elements, and relevant space is defined as the node connected finite element, corresponding wind turbine network topology structure is constructed, corresponding physical field control equation is established for finite element based on network topology structure, corresponding global coupling equation group is determined, and numerical method is used to solve global coupling equation group, corresponding physical response data is determined, real-time multiple physical field sensor data and real-time physical response data are input into the set hybrid learning model to carry out unsupervised and supervised hybrid learning, to determine the life cycle prediction result of wind turbine, the accuracy and efficiency of the application can improve wind turbine fault diagnosis and prediction.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and device for deep self-diagnosis of wind turbine generators based on machine learning. Background Technology

[0002] Wind turbines, as a crucial component of renewable energy, are typically deployed in harsh natural environments such as mountaintops, deserts, or oceans, often operating unattended or with minimal maintenance. This environment leads to frequent damage to wind turbine components, severely impacting operational efficiency and lifespan. To effectively extend the lifespan of wind turbines and reduce maintenance costs, it is essential to detect early-stage faults as early as possible, predicting and revealing their causes, extent, location, and development trends. This provides on-site maintenance personnel with a scientific basis for online adjustments or shutdown maintenance.

[0003] Currently, wind turbine generators, as complex mechanical systems, face challenges in fault diagnosis and remaining life prediction, including scarce labeled data, unclear mechanisms of unknown faults, and a lack of research on the internal relationships between multiple components or faults. Traditional fault diagnosis methods based on threshold alarms or single physical models are insufficient to cope with the complexity of the wind turbine operating environment and the diversity of fault modes.

[0004] Therefore, there is an urgent need for a deep self-diagnosis method for wind turbines based on machine learning, which can improve the accuracy and efficiency of fault diagnosis and prediction for wind turbines. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides a method and device for deep self-diagnosis of wind turbine generators based on machine learning, which can improve the accuracy and efficiency of fault diagnosis and prediction of wind turbine generators.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] Firstly, this application provides a deep self-diagnosis method for wind turbine generators based on machine learning, including:

[0008] The overall structure of the wind turbine is divided into structural components and related spaces. Multiphysics sensor data is collected from the overall structure. The structural components include blades, hubs, nacelles and towers. The related spaces are used to characterize the connection or relative relationship between the structural components.

[0009] The structural components are discretized into multiple finite elements, and the relevant space is defined as nodes connecting the finite elements. The corresponding wind turbine network topology is determined. Based on the network topology, the corresponding physical field control equations are established for the finite elements. The corresponding global coupling equation set is determined, and the global coupling equation set is solved using numerical methods to determine the corresponding physical response data.

[0010] A hybrid machine learning model is constructed, comprising an unsupervised learning module and a supervised learning module. Historical data of wind turbines is acquired, and the unsupervised and supervised learning modules are initialized and trained based on the historical data. Real-time multiphysics sensor data and real-time physical response data are then input into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbines. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results, while the supervised learning module is used to predict the life cycle of the wind turbines based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0011] Furthermore, the step of inputting real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine includes:

[0012] Based on the preset variational autoencoder and the preset adversarial autoencoder, feature dimensionality reduction and noise reduction are performed on real-time multiphysics sensor data and real-time physical response data to determine the corresponding feature representation of the same magnitude.

[0013] The same-order feature representation is input into the unsupervised learning module to determine the anomaly detection result and the operating condition clustering result output by the unsupervised learning module;

[0014] The same-scale feature representation is input into the supervised learning module to determine the fault diagnosis results and remaining life prediction results of the wind turbine output by the supervised learning module.

[0015] Further, the step of inputting the feature representation of the same magnitude into the unsupervised learning module and determining the anomaly detection result and operating condition clustering result output by the unsupervised learning module includes:

[0016] An unsupervised learning module is constructed based on a preset anomaly detection layer and a preset operating condition clustering layer, wherein the anomaly detection layer and the anomaly operating condition clustering layer are computed in parallel.

[0017] The same-scale feature representation is input into the anomaly detection layer. Based on the reconstruction error of the isolated forest, a first-class support vector machine, and the autoencoder, anomaly detection is performed on the input feature representation to determine the corresponding anomaly detection result, including anomaly score or binary alarm. The anomaly alarm output based on the anomaly detection result is used to trigger incremental updates of the supervised learning module.

[0018] The same-scale feature representation is input into the operating condition clustering layer. Based on the deep embedding clustering algorithm and the temporal clustering algorithm, the input feature representation is divided into wind turbine operating state clusters to determine the corresponding operating condition clustering results. The operating condition clustering results are used as input to the supervised learning module for model training and prediction.

[0019] Further, the step of inputting the same-order feature representation into the supervised learning module to determine the fault diagnosis results and remaining life prediction results of the wind turbine output by the supervised learning module includes:

[0020] The same-level feature representation is input into the feature fusion layer, and the same-level feature representation is extracted according to the preset one-dimensional convolutional neural network and wavelet packet transform to determine the corresponding high-dimensional feature vector.

[0021] The high-dimensional feature vector and the clustering results of the operating conditions are input into the spatiotemporal feature learning layer. The spatial relationship and temporal series dependency between wind turbine structural nodes are modeled according to the graph neural network and the long short-term memory network to determine the corresponding spatiotemporal features. The graph neural network is constructed based on the network topology of the wind turbine.

[0022] The spatiotemporal features are input into the degradation trajectory modeling layer, and the spatiotemporal features are visualized and the trajectory is mined according to the nonlinear dimensionality reduction algorithm to determine the corresponding equipment degradation path;

[0023] The degradation path of the equipment is input into the fault diagnosis and life prediction layer. Fault classification is performed based on the attention mechanism and multilayer perceptron algorithm, and the remaining life is predicted using Transformer to determine the fault diagnosis result and remaining life prediction result of the corresponding wind turbine.

[0024] Further, the step of acquiring historical data of the wind turbine generator and initializing the training of the unsupervised learning module and the supervised learning module based on the historical data includes:

[0025] Based on historical data of wind turbines in a healthy state, a reconstruction model based on an autoencoder is trained to establish a baseline for the normal state and determine the corresponding unsupervised learning module after training.

[0026] Based on historical data containing labeled fault types, a supervised learning model is trained, and the corresponding supervised learning module is determined.

[0027] Further, the step of discretizing the structural components into multiple finite elements and defining the relevant space as nodes connecting the finite elements to determine the corresponding wind turbine network topology includes:

[0028] The continuous structures within the four structural components—blade, hub, nacelle, and tower—are each divided into multiple discrete finite element units.

[0029] The related space is defined as the boundary node connecting the finite element elements between different structural parts, wherein the related space includes the blade-hub connection space, the hub-nacelle connection space, the nacelle-tower connection space, and the tower-blade relative space;

[0030] Based on the actual physical connection and energy transfer path of the wind turbine, the connection relationship between the boundary node and the finite element unit of the adjacent structural part is established, and the corresponding wind turbine network topology is determined.

[0031] Further, the step of establishing corresponding physical field control equations for the finite element based on the network topology, determining the corresponding global coupling equation set, and solving the global coupling equation set using numerical methods to determine the corresponding physical response data includes:

[0032] Based on the network topology, in the local coordinate system, the control equations for the fluid field, structural force field and spatial position field are established for each finite element, and the control equations of all finite elements are combined to determine the corresponding global coupling equation set.

[0033] Based on the sensor data acquisition volume of each finite element, the global coupled equations are solved using numerical methods to determine the physical response data of the corresponding wind turbine unit across the entire field.

[0034] Secondly, this application provides a machine learning-based deep self-diagnosis device for wind turbine generators, comprising:

[0035] The wind turbine structure division module is used to divide the overall structure of the wind turbine into structural components and related spaces, and to collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0036] The wind turbine coupling equation determination module is used to discretize the structural components into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, solve the global coupling equation set using numerical methods, and determine the corresponding physical response data.

[0037] A wind turbine lifecycle self-diagnosis module is used to construct a hybrid machine learning model. The hybrid machine learning model includes an unsupervised learning module and a supervised learning module. It acquires historical data of the wind turbine, initializes and trains the unsupervised learning module and the supervised learning module based on the historical data, and then inputs real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results, and the supervised learning module is used to predict the lifecycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0038] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the machine learning-based deep self-diagnosis method for wind turbine generators.

[0039] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the machine learning-based deep self-diagnosis method for wind turbine generators.

[0040] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the machine learning-based deep self-diagnosis method for wind turbine generators.

[0041] As can be seen from the above technical solution, this application provides a deep self-diagnosis method and device for wind turbines based on machine learning. This method divides the overall structure of the wind turbine into structural components and related spaces, collects multi-physics sensor data from the overall structure, discretizes the structural components of the wind turbine into multiple finite elements, and defines the related spaces as nodes connecting these finite elements, constructing a corresponding wind turbine network topology. Based on the network topology, corresponding physical field control equations are established for each finite element, determining the corresponding global coupling equation set. Numerical methods are then used to solve the global coupling equation set to determine the corresponding physical response data. Real-time multi-physics sensor data and real-time physical response data are input into a hybrid learning model for unsupervised and supervised hybrid learning to determine the lifecycle prediction results of the wind turbine. This improves the accuracy and efficiency of wind turbine fault diagnosis and prediction. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is one of the flowcharts illustrating the machine learning-based deep self-diagnosis method for wind turbines in this application embodiment;

[0044] Figure 2 This is a structural diagram of the machine learning-based deep self-diagnosis device for wind turbines in an embodiment of this application;

[0045] Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0046] Figure label:

[0047] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0050] Considering that wind turbine generators are complex mechanical systems, their fault diagnosis and remaining life prediction face challenges such as scarce data annotations, unclear mechanisms of unknown faults, and a lack of research on the internal connections between multiple components or faults. This application provides a deep self-diagnosis method and device for wind turbine generators based on machine learning. By dividing the overall structure of the wind turbine generator into structural components and related spaces, and collecting multi-physics sensor data from the overall structure, the structural components of the wind turbine generator are discretized into multiple finite elements. The related spaces are defined as nodes connecting these finite elements, constructing a corresponding wind turbine generator network topology. Based on the network topology, corresponding physical field control equations are established for the finite elements, determining the corresponding global coupling equations. Numerical methods are then used to solve the global coupling equations to determine the corresponding physical response data. Real-time multi-physics sensor data and real-time physical response data are input into a hybrid learning model for unsupervised and supervised learning to determine the life cycle prediction results of the wind turbine generator. This improves the accuracy and efficiency of wind turbine generator fault diagnosis and prediction.

[0051] To improve the accuracy and efficiency of fault diagnosis and prediction for wind turbine generators, this application provides an embodiment of a deep self-diagnosis method for wind turbine generators based on machine learning. See [link to embodiment]. Figure 1 The machine learning-based deep self-diagnosis method for wind turbines specifically includes the following:

[0052] Step S101: Divide the overall structure of the wind turbine into structural components and related spaces, and collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0053] Optionally, in this embodiment, this step is part of the core algorithm framework and involves the wind turbine structure theory innovation system.

[0054] Specifically, the wind turbine is divided into four parts and four related spaces according to its structure.

[0055] The four parts and their detailed components are:

[0056] 1. Leaves

[0057] ① Leaf tip

[0058] ② Middle part of the leaf

[0059] ③ Leaf base

[0060] 2. Wheel hub

[0061] ① Wheel hub A-side

[0062] ② Wheel hub B side

[0063] ③ Wheel hub C-side

[0064] ④ Fairing

[0065] 3. Cabin

[0066] ① Cabin shell

[0067] ② Main spindle

[0068] ③ Gearbox

[0069] ④ Generator

[0070] 4. Tower

[0071] ① Top of the tower

[0072] ② Each layer in the middle of the tower

[0073] ③ Lower part of the tower

[0074] ④ Bottom of the tower

[0075] The four related spaces are:

[0076] 1. Blade-hub connection space

[0077] 2. Hub-engine bay connection space

[0078] 3. Nacelle-Tower Connection Space

[0079] 4. Relative space between tower and blades

[0080] Optionally, in this embodiment, after the above structural division, we deploy sensors at various key locations in each component and space of the wind turbine to collect multi-physics sensor data to reflect the true state of the wind turbine.

[0081] Specifically, the following 16 types of data will be obtained for subsequent model building and calculations:

[0082] 1. Ambient altitude (air pressure), temperature, wind speed data, spindle speed data, gearbox speed data at each stage, generator speed data, and power generation data.

[0083] Data acquisition methods: The fan comes with a thermometer, anemometer, speed encoder, power meter, SCADA system, etc.

[0084] Belonging to the component / space: natural environment and power generation operation status.

[0085] 2. Vibration data, stress load data, and temperature data within the blade.

[0086] Data acquisition methods: ultra-low frequency triaxial vibration sensor, stress sensor, load sensor, and temperature sensor.

[0087] Parts or related spaces: leaf (leaf tip - leaf middle - leaf root).

[0088] 3. Blade root flange clearance data and blade root bolt tightening force data.

[0089] Data acquisition methods: flange gap sensor, preload sensor.

[0090] Related components or spaces: blade root, blade-hub connection space.

[0091] 4. Vibration data, speed data, and current data of the in-hub pitch motor.

[0092] Data acquisition methods: vibration sensors, speed encoders, SCADA systems, etc.

[0093] Related components or spaces: hub (A, B, C surfaces), blade orientation, blade-hub connection space.

[0094] 5. Vibration data of the pitch bearing inside the wheel hub.

[0095] Data acquisition method: vibration sensor.

[0096] Related components or spaces: hub (A, B, C surfaces), blade orientation, blade-hub connection space.

[0097] 6. Hub flange clearance data.

[0098] Data acquisition method: flange gap sensor.

[0099] Related components or spaces: wheel hub, wheel hub-engine compartment connection space.

[0100] 7. Data on shaking inside the cabin.

[0101] Data acquisition method: sway sensor, etc.

[0102] Related components or spaces: nacelle shell, hub-to-nacelle connection space, nacelle-to-tower connection space.

[0103] 8. Spindle axial movement data and main bearing vibration data.

[0104] Data acquisition methods: displacement motion sensor, acceleration vibration sensor.

[0105] Part or related space: spindle.

[0106] 9. Gearbox axial movement data and vibration data.

[0107] Data acquisition methods: displacement motion sensor, acceleration vibration sensor.

[0108] Part or related space: Gearbox.

[0109] 10. Gearbox oil quality data.

[0110] Data acquisition method: oil quality sensor.

[0111] Part or related space: Gearbox.

[0112] 11. Generator oscillation data, vibration data, and electrical data.

[0113] Data acquisition methods: displacement sensors, acceleration and vibration sensors, SCADA systems, etc.

[0114] Related component or space: generator.

[0115] 12. Vibration data, speed data, and current data of the yaw motor at the bottom of the engine room.

[0116] Data acquisition methods: vibration sensors, speed encoders, SCADA systems, etc.

[0117] Related component or space: Nacelle-tower connection space.

[0118] 13. Blade tip to tower clearance data.

[0119] Data acquisition method: Clearance radar.

[0120] Related component or space: Tower-blade relative space.

[0121] 14. Tower sway posture data.

[0122] Data acquisition method: composite tilt sensor, etc.

[0123] Components or related spaces: Tower (top-middle-bottom-base).

[0124] 15. Tightening force data of flange bolts at each level of the tower.

[0125] Data acquisition method: preload sensor.

[0126] Related components or spaces: Tower (top-middle-lower part).

[0127] 16. Tower bottom anchor bolt tightening data.

[0128] Data acquisition method: preload sensor.

[0129] Part or related space: Bottom of the tower.

[0130] Through the above steps, we are able to obtain real-time information on the wind turbine's physical structure, key stress points, and operating status through various deployed sensors.

[0131] Step S102: Discretize the structural component into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, and solve the global coupling equation set using numerical methods to determine the corresponding physical response data.

[0132] Optionally, in this embodiment, this step constructs a wind turbine network topology and further utilizes the geometric and physical knowledge of wind turbines to establish a solution framework, which is used to break down data silos and convert the aforementioned scattered sensor data into an overall wind turbine status, accurately reflecting the real-time operating status of the wind turbine.

[0133] After the deployment of the sensor in step S101, each part and related space has its own data acquisition device to collect relevant data. Next, based on the key nodes of the wind turbine operation, we discretize the overall structure of the wind turbine and divide the continuous structure in each part into discrete small units.

[0134] In this embodiment, more than 100 discrete small units are set on the basis of the overall structure of the wind turbine. Each small unit is a key node in the operation of the wind turbine and is used to accurately reflect the status of each component of the wind turbine. Each discrete small unit is an independent finite element, and the relevant space is used as the node between finite elements to form a network.

[0135] Within each finite element, based on different data acquisition parameters, the following parameters are used: fluid field (temperature, airflow, noise, oil, etc.), structural force field (vibration, stress, load, preload, etc.), and spatial position field (tilt, displacement, deformation, clearance, gap), all based on the local coordinate system of the finite element. Each mathematical function and spatial boundary is established, and a bidirectional failure propagation link is established in four key connection spaces.

[0136] The equations of all elements are combined into a large system of equations, taking into account boundary conditions and loads, ultimately forming a global linear system of equations. The system of equations is solved using numerical methods (such as Gaussian elimination and iterative methods) to obtain the unknowns at the nodes, and then the response of the entire system (such as stress, strain, temperature distribution, etc.) is derived, i.e., physical response data.

[0137] The following are mathematical function expressions for different physical fields within a finite element, organized by category:

[0138] 1. Mathematical functions of fluid fields

[0139] 1.1 Temperature Field (Heat Conduction Equation)

[0140]

[0141] Temperature field

[0142] Thermal conductivity; :density; Specific heat capacity

[0143] Heat source item

[0144] Application example: Gearbox overheat warning. Based on oil temperature sensor data ( ) and gear friction heat source ( Coupling can predict the risk of local overheating (e.g., in the embodiment, the overheating warning for gears and bearings is 240 hours in advance).

[0145] 1.2 Wind Field (Simplified Navier-Stokes Equations)

[0146]

[0147] v Velocity field

[0148] :pressure; Kinematic viscosity; Volume force

[0149] Application example: Blade aerodynamic load analysis. Combined with anemometer data ( The blade stress field (Formula 2.2) is used to predict flutter risk (the accuracy is improved by 41% in the example).

[0150] 1.3 Noise Field (Wave Equation)

[0151]

[0152] Sound pressure field; Speed ​​of sound

[0153] 1.4 Oil Flow (Continuity Equation)

[0154]

[0155] Oil density

[0156] 2. Mathematical functions of structural force fields

[0157] 2.1 Vibration Field (Harmonic Response Equation)

[0158]

[0159] Displacement field

[0160] :quality; Damping coefficient; Stiffness

[0161] Application example: Vibration fault diagnosis. Through vibration data ( Reverse incentive force It can identify faults such as pitting or friction damage.

[0162] 2.2 Stress Field (Hooke's Law)

[0163]

[0164] Stress tensor; : Strain tensor; Elastic tensor

[0165] 2.3 Distributed load (uniformly distributed load)

[0166]

[0167] : Foundation load; Spatial distribution coefficient

[0168] 2.4 Preload (bolt preload model)

[0169]

[0170] Bolt stiffness; Pre-stretch amount

[0171] Application example: Health monitoring of tower bolts. Combined with bolt stiffness coefficient ( ) and variables ( Dynamically assess connection reliability (such as tower anchor bolt monitoring).

[0172] 3. Mathematical functions of spatial position field

[0173] 3.1 Inclination Angle (Geometric Relationship)

[0174]

[0175] : Inclined function

[0176] 3.2 Displacement (Time-varying function)

[0177]

[0178] Amplitude; :frequency; Phase

[0179] 3.3 Deformation field (strain-displacement relationship)

[0180]

[0181] Displacement field

[0182] 3.4 Clearance / Gap (Geometric Constraints)

[0183]

[0184] : Position function of two surfaces; Minimum safe distance; Maximum safe distance

[0185] Application example: Blade tip-to-tower clearance monitoring. This is achieved through clearance radar data ( , Real-time calculation of clearance distance to avoid tower sweeping accidents (such as tower-blade relative space monitoring).

[0186] Formula Explanation

[0187] ① Variable range: All formulas are based on the local coordinate system of the finite element. .

[0188] ② Linear and nonlinear: Dynamic fields such as vibration and airflow need to consider nonlinear terms, such as convection terms.

[0189] ③ Boundary conditions: The equations are closed by using boundary conditions.

[0190] ④ To facilitate model calculations, the formulas have been appropriately simplified and numerically discretized without affecting the accuracy of the results.

[0191] Application scenario example:

[0192] Monitoring of blade root bolt preload loosening (a common progressive failure):

[0193] The "blade root unit" collects bolt data through a preload sensor (Formula 2.4).

[0194] The "blade-hub connection space" serves as a transmission node, and deformation is monitored by a flange gap sensor (Formula 3.4).

[0195] The vibration sensor inside the "hub" (Formula 2.1) synchronously detects abnormal vibrations.

[0196] The data from the above three aspects and the transmission link influencing factors are substituted into the global equation system for solution. The final calculation results not only comprehensively judge the degree of failure degradation and the prediction of remaining life from multiple dimensions, but also form a multi-source verification.

[0197] Step S103: Construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. Acquire historical data of the wind turbine. After initializing and training the unsupervised learning module and the supervised learning module based on the historical data, input real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results. The supervised learning module is used to predict the life cycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0198] Optionally, in this embodiment, the purpose of this step is to build an intelligent engine that accurately reflects the fault diagnosis results and remaining life prediction results of all components of the wind turbine based on the original sensor data and the calculated physical field response data.

[0199] The core intelligence of this step stems from its deeply integrated machine learning engine, which works in conjunction with a multiphysics coupled solver to form a dual-drive paradigm of "physics-driven" and "data-driven." It not only includes supervised learning models but also innovatively introduces an unsupervised learning module, forming a hybrid paradigm of "supervised-unsupervised" collaborative learning. It not only processes the macroscopic field data output by the solver but also directly integrates raw sensor data to achieve keen detection of early faults, accurate classification of fault modes, and high-precision prediction of remaining useful life (RUL). It primarily addresses challenges common in real-world industrial scenarios, such as scarce data annotations, unclear mechanisms of unknown faults, and a lack of research on the internal connections between multiple components or multiple faults, further enhancing the system's generalization and adaptability.

[0200] Specifically, a structured hybrid machine learning model is constructed, comprising parallel unsupervised learning modules and supervised learning modules. Before the model is deployed in real-time, it undergoes systematic initial training using historical data from wind turbines, endowing both modules with initial expertise.

[0201] Specifically, the pre-training of the unsupervised learning module relies on massive amounts of historical data from wind turbines under normal and various unknown conditions. This data requires no manual fault labeling. The core objective of the training is to enable the model to learn the operating data patterns of the turbines under "healthy" or "normal" conditions. Among these:

[0202] For anomaly detection, models such as deep autoencoders are employed. By training the model to reconstruct normal input data with high fidelity, the model will produce larger reconstruction errors for anomaly data that deviates significantly from the normal pattern. By statistically analyzing the reconstruction error distribution of historical normal data, a dynamic threshold can be set. When the reconstruction error of real-time data exceeds this threshold, an anomaly alarm is triggered. This alarm is then used by the system to trigger incremental updates in downstream supervised learning modules.

[0203] For the operating condition clustering function, algorithms such as deep embedding clustering are employed. This process automatically analyzes historical data (such as combinations of power, wind speed, and ambient temperature), automatically divides different operating states (such as "rated power generation," "start-stop process," "turbulent operation," and "standby") in the feature space, forming several clusters, and assigning an operating condition label to each cluster. The operating condition labels provide data input for subsequent supervised learning modules.

[0204] Unsupervised modules endow models with intuitive abilities, enabling them to discover potential, unexpected anomaly patterns in data without relying on predefined fault knowledge, and to automatically identify the current operating context of the unit. This lays the foundation for solving two major industry challenges: "unknown fault discovery" and "scarcity of data annotation."

[0205] Specifically, the pre-training of supervised learning modules relies on historical data that has been labeled by experts or historical maintenance records. For example, a certain segment of vibration data is explicitly labeled as "bearing inner ring failure," and a certain time point is recorded as "200 hours before gearbox failure." The training process allows the model to learn the complex mapping relationship between input data (such as vibration characteristics and temperature sequences) and output labels (failure type, remaining life).

[0206] The supervised module endows the model with professional knowledge capabilities, enabling it to draw on historical experience to accurately classify known types of failures (e.g., distinguish between bearing failure and gear failure), and to quantitatively predict how long the equipment can still operate safely (remaining life) based on current degradation signs.

[0207] Optionally, in this embodiment, the model can enter the real-time application stage after completing the initialization training.

[0208] Specifically, the unsupervised module provides real-time sensing. Real-time data streams enter the unsupervised learning module. This module performs two tasks: first, anomaly detection, calculating the "anomaly score" of the current data or determining whether it exceeds normal boundaries; second, operational condition identification, determining which predefined operational condition cluster the current data most belongs to (e.g., "high-power generation" or "low-wind-speed idling"). This unsupervised module provides the system with real-time operational condition background and preliminary warnings, answering the two key questions: "In what state is the wind turbine currently operating?" and "Does the current data appear normal?"

[0209] Specifically, the supervised learning module performs fusion decision-making. When performing final fault diagnosis and life prediction, the supervised learning module receives diversified and fused input information. It not only receives raw or pre-processed sensor data and physical response data, but more importantly, it also uses the anomaly detection results and operating condition clustering results output in real time by the unsupervised learning module as key auxiliary input features.

[0210] For example, when a model is judging a vibration signal, if it is known at the same time that "it is currently in a high-power condition" and "the anomaly detection score is high", it may be more inclined to judge it as a serious load-related fault rather than random noise.

[0211] When predicting remaining lifespan, different operating conditions (such as strong winds at sea versus light winds on land) have drastically different effects on equipment degradation. Incorporating operating condition information allows the prediction model to move beyond a "one-size-fits-all" approach and instead make accurate predictions that adapt to different conditions.

[0212] This data fusion mechanism greatly enhances the context awareness and decision-making reliability of supervised models. It ensures that the final diagnosis and prediction results are no longer based on isolated data analysis, but rather on a comprehensive judgment made within a complete operational context and health status.

[0213] This example demonstrates how this embodiment creatively decomposes the overall wind turbine structure into structural units and dynamic controls, creates a discretized model, solves the problem of cross-physical field coupling, introduces a real-time joint solver for the fluid field, structural force field, and spatial position field to achieve breakthroughs in multiple couplings and break down data silos, and finally establishes a supervised-unsupervised hybrid learning model to achieve self-evolutionary diagnosis of the wind turbine life cycle based on the knowledge graph of wind turbine fault factors.

[0214] As described above, the machine learning-based deep self-diagnosis method for wind turbines provided in this application can divide the overall structure of a wind turbine into structural components and related spaces, collect multi-physics sensor data from the overall structure, discretize the structural components of the wind turbine into multiple finite units, define the related spaces as nodes connecting the finite units, construct a corresponding wind turbine network topology, establish corresponding physical field control equations for the finite units based on the network topology, determine the corresponding global coupling equation set, solve the global coupling equation set using numerical methods, determine the corresponding physical response data, and input real-time multi-physics sensor data and real-time physical response data into a set hybrid learning model for unsupervised and supervised hybrid learning to determine the life cycle prediction results of the wind turbine. This can improve the accuracy and efficiency of wind turbine fault diagnosis and prediction.

[0215] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0216] Step S201: Perform feature dimensionality reduction and denoising on real-time multiphysics sensor data and real-time physical response data based on preset variational autoencoders and preset adversarial autoencoders, and determine the corresponding feature representations of the same magnitude;

[0217] Step S202: Input the same-scale feature representation into the unsupervised learning module, and determine the anomaly detection result and operating condition clustering result output by the unsupervised learning module;

[0218] Step S203: Input the same-scale feature representation into the supervised learning module to determine the fault diagnosis results and remaining life prediction results of the wind turbine output by the supervised learning module.

[0219] Optionally, in this embodiment, the system simultaneously inputs the collected high-dimensional, multi-source real-time sensor data (such as vibration waveforms and current signals) and physical field response data (such as stress distribution and temperature field) into a preset variational autoencoder (VAE) and a preset adversarial autoencoder (AAE). The VAE maps the input data to a probabilistic latent space through its encoder and then reconstructs it through the decoder. This process forces the model to learn robust and essential feature distributions in the data, thereby filtering out noise and discarding unimportant information. At the same time, the AAE ensures that the latent space features output by the encoder conform to a specific prior distribution (such as a Gaussian distribution) through adversarial training of its discriminator and generator (encoder-decoder). This further standardizes the feature representation and improves its regularity and separability. The two autoencoders work together to finally output a "feature representation of the same magnitude" with uniform dimensions, significant features, and removed redundant noise.

[0220] This step transforms multimodal data from different components, with varying dimensions and sampling frequencies, into a unified low-dimensional, dense vector with equal feature sensitivity. This lays the foundation for subsequent model fusion and computation. Furthermore, by removing noise and extracting robust features, it establishes a data foundation for the entire system with high generalization ability and high accuracy.

[0221] Optionally, in this embodiment, this step performs unsupervised preliminary state assessment and pattern discovery.

[0222] The module inputs feature representations of the same magnitude in parallel into two core parts of the unsupervised learning module. On one hand, the module utilizes anomaly detection algorithms such as isolated forests and support vector machines to learn the boundaries or distribution patterns of data under normal operating conditions. When the feature representation of real-time data deviates from this pattern or the reconstruction error exceeds a threshold, an anomaly alarm is triggered, enabling the initial capture of unknown faults and early, subtle anomalies. On the other hand, the module simultaneously employs algorithms such as deep embedding clustering to perform unlabeled clustering analysis on the feature representations, automatically dividing the data into different clusters. Each cluster corresponds to a specific operating condition (such as rated power generation, start-up / shutdown, turbulent operation, etc.) and outputs an operating condition label.

[0223] Optionally, in this embodiment, this step is the decision-making stage for accurate diagnosis and quantitative prediction.

[0224] Based on features of the same magnitude, the supervised learning module uses complex models (such as CNN, GNN, LSTM, Transformer, etc.) trained on labeled historical data to perform tasks.

[0225] For fault diagnosis, the model analyzes the patterns in the feature representation, accurately classifies them into predefined fault types (such as bearing inner ring faults, gear tooth breakage, etc.), and outputs specific fault classification results.

[0226] For remaining lifetime prediction, the model uses regression analysis to learn the complex mapping relationship between feature representation and equipment performance degradation trajectory, predicting the operating time required for the equipment health index to drop to the failure threshold, which is the remaining lifetime prediction result.

[0227] Through step S203, this embodiment realizes the transformation from data to decision, providing a quantitative scientific basis for maintenance decisions.

[0228] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0229] Step S301: Construct an unsupervised learning module based on the preset anomaly detection layer and the preset operating condition clustering layer, wherein the anomaly detection layer and the anomaly operating condition clustering layer are computed in parallel;

[0230] Step S302: Input the same-scale feature representation into the anomaly detection layer, perform anomaly detection on the input feature representation based on the isolation forest, a first-class support vector machine and the autoencoder reconstruction error, and determine the corresponding anomaly detection result, including anomaly score or binary alarm. The anomaly alarm output based on the anomaly detection result is used to trigger the incremental update of the supervised learning module.

[0231] Step S303: Input the same-scale feature representation into the operating condition clustering layer, and perform wind turbine operating state clustering on the input feature representation according to the deep embedding clustering algorithm and the temporal clustering algorithm to determine the corresponding operating condition clustering result. The operating condition clustering result is used to input the supervised learning module for model training and prediction.

[0232] Optionally, in this embodiment, this step aims to establish an unsupervised learning framework for parallel processing and output auxiliary data for the supervised learning layer through the unsupervised learning layer.

[0233] The unsupervised learning framework consists of two core layers: a pre-defined anomaly detection layer and a pre-defined operating condition clustering layer, which are computationally parallel. The main responsibility of the anomaly detection layer is to identify outliers and abnormal patterns in the data; its design aims to capture potential faults or abnormal events that deviate from normal operating conditions. The operating condition clustering layer focuses on analyzing the operating data of wind turbines under different external environments and internal loads, automatically classifying massive amounts of unlabeled historical operating data into meaningful operating state categories with similar characteristics.

[0234] The construction of this module provides the foundational data understanding capabilities for the entire diagnostic system. It does not rely on predefined fault labels or human experience to classify operating conditions, allowing the model to autonomously discover the inherent structure and patterns of the data itself. This provides pre-screened and structured input for subsequent supervised learning, forming the cornerstone for the system's intelligence and adaptability.

[0235] Optionally, in this embodiment, there is an anomaly detection layer.

[0236] This layer employs a hybrid strategy, integrating three methods—isolation forest, a type of support vector machine, and autoencoder reconstruction error—to perform multi-faceted and complementary anomaly detection analysis on the input feature representation. The isolation forest algorithm excels at quickly screening isolated points in high-dimensional feature spaces, demonstrating high efficiency in detecting sparse anomalies. The support vector machine constructs a hyperplane describing the boundary of the "normal" state, effectively identifying anomalous samples outside this boundary. The autoencoder learns by compressing and reconstructing the data itself, using the magnitude of the reconstruction error to determine whether the input data matches the learned normal pattern; a larger error indicates a higher probability of an anomaly. By fusing the outputs of these three methods, a comprehensive anomaly detection result is determined. This result can be an anomaly score or a explicit binary alert. The anomaly score reflects the severity or confidence level of the anomaly, providing trend information for continuous monitoring; while the binary alert directly triggers an early warning mechanism.

[0237] These output alerts are used by the system to trigger incremental updates in downstream supervised learning modules. When the system continuously detects a new, unidentified abnormal pattern, these samples can be automatically marked as high-value samples, guiding expert intervention or directly entering the model's online learning queue, thereby driving the diagnostic system to continuously evolve and acquire the ability to discover and adapt to "unknown unknowns" faults.

[0238] Optionally, in this embodiment, a runtime clustering layer is used.

[0239] This layer utilizes deep embedding clustering and temporal clustering algorithms to perform refined, unsupervised classification of wind turbine operating states. Deep embedding clustering first maps high-dimensional features to a lower-dimensional latent space that is more conducive to clustering through a neural network, and then performs clustering optimization in this space, thereby discovering natural clusters hidden in complex nonlinear data structures. Temporal clustering algorithms are specifically designed for time-series data and can identify operating conditions with similar patterns over time, such as startup, shutdown, rated operation, and turbulent fluctuations.

[0240] Through the combined work of these two algorithms, this step can automatically classify the data of all operating moments in the history of the wind turbine into several operating condition categories with clear physical meaning, and output the corresponding "operating condition clustering results".

[0241] The clustering results of operating conditions provide crucial contextual information for subsequent supervised learning modules. During model training, the system can train targeted diagnostic and predictive models based on different operating condition categories, avoiding the mixing of data from different operating states and significantly improving model accuracy and generalization ability. In the prediction phase, the system first determines the operating condition category of the current data and then calls the corresponding dedicated model for inference, ensuring that diagnostic and predictive results are derived within the correct operating context, significantly improving the scientific rigor and accuracy of decision-making.

[0242] Through step S303, this embodiment successfully constructed an unsupervised learning layer. The construction of this layer creates input conditions for the subsequent supervised learning output of wind turbine life cycle prediction results. This includes the abnormal alarms output by the unsupervised learning layer being used to trigger incremental updates of the supervised learning module, and the operating condition labels output by the unsupervised learning layer being used as data "labels" for the supervised learning module to select the corresponding specialized operating condition model and improve prediction accuracy.

[0243] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0244] Step S401: Input the same-level feature representation into the feature fusion layer, and extract features from the same-level feature representation according to the preset one-dimensional convolutional neural network and wavelet packet transform to determine the corresponding high-dimensional feature vector;

[0245] Step S402: Input the high-dimensional feature vector and the clustering result of the operating conditions into the spatiotemporal feature learning layer, and model the spatial relationship and time series dependency between wind turbine structural nodes according to the graph neural network and the long short-term memory network to determine the corresponding spatiotemporal features. The graph neural network is constructed based on the network topology of the wind turbine.

[0246] Step S403: Input the spatiotemporal features into the degradation trajectory modeling layer, visualize and mine the spatiotemporal features using a nonlinear dimensionality reduction algorithm, and determine the corresponding equipment degradation path;

[0247] Step S404: Input the equipment degradation path into the fault diagnosis and life prediction layer, classify the faults according to the attention mechanism and multilayer perceptron algorithm, and use Transformer to predict the remaining life, and determine the fault diagnosis result and remaining life prediction result of the corresponding wind turbine.

[0248] Optionally, in this embodiment, this step is the core process of the supervised learning layer.

[0249] Specifically, the feature fusion and extraction layer. Before unsupervised data is input into the supervised learning layer, the supervised learning layer performs feature extraction. Preferably, the feature extraction step of the supervised learning layer can be processed in parallel with the two unsupervised steps of the unsupervised learning layer, further increasing data processing efficiency.

[0250] Feature extraction is performed by combining a one-dimensional convolutional neural network (1D-CNN) with wavelet packet transform (WPD). Wavelet packet transform first performs a deeper time-frequency analysis of the time-series signal (such as vibration or current), enabling a more refined extraction of the energy distribution across different frequency bands compared to traditional wavelet transform. This allows for the extraction of sensitive features that characterize subtle changes in equipment status, such as energy entropy and scale entropy. Subsequently, these time-frequency features refined by WPD, along with the original or pre-processed time-series signal, are input into the 1D-CNN. The 1D-CNN uses its convolutional kernels to slide along the time dimension, automatically learning local patterns, periodicities, and subtle fault signs (such as specific impact characteristic frequencies and electrical harmonics) in the signal. Its advantage lies in its ability to capture the translational invariance of the signal along the time axis. Finally, this layer outputs a high-dimensional feature vector that integrates the physical meaning of the time-frequency domain with data-driven deep patterns.

[0251] This step can effectively extract critical fault-sensitive information from potentially noisy data, enhancing the ability to detect early, subtle fault characteristics.

[0252] Specifically, the spatiotemporal feature joint learning layer.

[0253] This process aims to simulate the propagation of faults within the physical structure of a wind turbine and their evolution over time. It receives high-dimensional feature vectors from step S401 and the output from the unsupervised clustering layer. The discretized network topology of the wind turbine is abstracted into a graph structure. Each node carries a high-dimensional feature vector that changes over time. A graph neural network (GNN) is responsible for learning the spatial topological relationships between nodes. It updates its own node representation by aggregating information from neighboring nodes, thereby simulating cross-component fault propagation effects such as how gearbox faults affect the generator through the main shaft, or how blade loads are transmitted to the pitch bearing and hub clearance. Simultaneously, a long short-term memory (LSTM) network is attached to each node, responsible for learning the long-term dependencies of that node's features over time, capturing the dynamic trends of performance degradation. Operating condition clustering results are used to provide context for this process, ensuring that the model understands spatiotemporal features within the correct operating condition context.

[0254] This step breaks through the limitations of traditional models that only focus on a single component or ignore the relationships between components. By combining spatial structure dependence with time series evolution, it not only reflects "what the current state of a certain component is", but also reveals "how this state is affected by other components" and "how it itself evolves over time", greatly enriching the descriptive dimensions of health status.

[0255] Specifically, degradation trajectory mining and visualization.

[0256] This step aims to intuitively reveal the complete evolution path of a device from health to failure from the high-dimensional and complex spatiotemporal characteristics.

[0257] The high-dimensional spatiotemporal features output from step S402 are received, and a nonlinear dimensionality reduction algorithm (such as UMAP or t-SNE) is used to faithfully represent the complex manifold structure in the high-dimensional data in two-dimensional or three-dimensional space. By reducing the spatiotemporal features at each time point throughout the device's lifecycle and projecting them into a lower-dimensional space, a series of scattered points will naturally form. Connecting and analyzing these scattered points in chronological order (i.e., trajectory mining) can clearly outline a "degradation trajectory." This trajectory can show the phased transitions in the device's health status (such as the healthy zone, the slow degradation zone, the accelerated failure zone), and the final failure cluster area.

[0258] The common degradation path patterns discovered in this step provide data support for understanding the failure mechanism of the equipment.

[0259] Specifically, fault diagnosis and accurate lifespan prediction.

[0260] This step is the final decision-making layer, receiving the equipment degradation path (or its high-dimensional spatiotemporal feature representation) mined by S403. This step performs two core tasks in parallel: fault classification and remaining lifetime (RUL) prediction.

[0261] For fault classification, the model employs a combination of an attention mechanism and a multilayer perceptron (MLP). The attention mechanism automatically assesses the importance weights of different nodes and time points in the spatiotemporal features for the current fault judgment, "focusing" on the most critical evidence, which greatly enhances the model's interpretability. The weighted features are then input into the MLP to complete accurate fault type classification (such as "pitting corrosion in the bearing inner ring" or "broken gear tooth").

[0262] For RUL prediction, the Transformer model is used. With its powerful self-attention mechanism, the Transformer can capture long-range global dependencies in the degradation sequence, thereby more accurately predicting the future decline trajectory of the Health Indicator (HI) until it reaches the preset failure threshold, and finally outputs the RUL value in units of time.

[0263] Through step S404, this embodiment successfully achieves accurate fault identification and location, and can predict the remaining service life of the equipment with high precision.

[0264] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0265] Step S501: Based on historical data of the wind turbine in a healthy state, train a reconstruction model based on an autoencoder to establish a baseline for the normal state and determine the corresponding unsupervised learning module after training.

[0266] Step S502: Based on historical data containing labeled fault types, train a supervised learning model and determine the corresponding trained supervised learning module.

[0267] Optionally, in this embodiment, the core of the unsupervised learning module is to enable the machine to learn and memorize the normal operating mode of the wind turbine in a "fully healthy" state. To this end, it is necessary to collect a large amount of historical data generated during the turbine's trouble-free operation, covering readings from various sensors such as vibration, temperature, and pressure. The model used for training is typically a deep autoencoder, a special type of neural network. During training, healthy, normal data is input into the autoencoder, forcing the model to learn how to first compress the data to a low-dimensional "bottleneck" layer (capturing the most core features of the data), and then reconstruct the original data as accurately as possible from this compressed representation. Through repeated iterations, the model gradually masters the inherent structure, variation patterns, and noise levels of healthy data. After training, this autoencoder, capable of accurately reconstructing normal data, becomes the benchmark for measuring whether the state is "abnormal."

[0268] Optionally, in this embodiment, the goal of the supervised learning module is to establish an expert system capable of accurate diagnosis and prediction. The training data consists of clearly labeled historical data; for example, a vibration signal segment might be labeled as "bearing outer race failure," and the end point of a time series might be labeled as "gearbox failure." The training process involves the supervised learning model (such as a deep convolutional neural network or graph neural network) learning the complex mapping relationship between the complex input data and these specific fault labels or remaining life values. By adjusting millions of internal parameters, the model continuously narrows the gap between its predictions and the true labels, ultimately learning to extract patterns highly correlated with specific faults or degradation stages from the data features.

[0269] Through step S502, this embodiment successfully trained both the unsupervised learning module and the supervised learning module, laying the foundation for subsequent prediction of wind turbine lifecycles.

[0270] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0271] Step S601: Divide the continuous structure within the four structural parts—blade, hub, nacelle, and tower—into multiple discrete finite element units.

[0272] Step S602: Define the related space as the boundary node connecting the finite element units between different structural parts, wherein the related space includes the blade-hub connection space, the hub-nacelle connection space, the nacelle-tower connection space, and the tower-blade relative space;

[0273] Step S603: Based on the actual physical connection and energy transfer path of the wind turbine, establish the connection relationship between the boundary node and the finite element units of the adjacent structural parts, and determine the corresponding wind turbine network topology.

[0274] Optionally, in this embodiment, the structural parts are first discretized, which is the basis for physical field numerical modeling.

[0275] The large, complex, continuous structure of a wind turbine (such as blades and towers) is meshed, transforming it into a set of numerous small, regularly shaped elements (such as tetrahedrons and hexahedrons). Each discretized finite element is assigned precise geometric properties (such as size and shape) and physical properties (such as material density, elastic modulus, and thermal conductivity). Its core function is to transform the partial differential equations describing the continuous medium, which cannot be directly solved, into a system of algebraic equations solvable on each small element. This discretization process makes it possible to simulate the complex physical behavior of the entire wind turbine, laying the model foundation for subsequent high-precision calculations.

[0276] Optionally, in this embodiment, the next step is to implement the key innovation of cross-component coupling analysis.

[0277] Four key connection or interaction regions (such as the blade-hub connection space) are explicitly defined as boundary nodes in the network, moving away from traditional "blank areas" or simple boundaries. These nodes are given specific physical meanings and data processing functions as logical entities, carrying data on interactions between adjacent components, such as forces, torques, heat flow, and displacements. This step overcomes the limitations of traditional component-level modeling, organically linking physically independent but functionally tightly coupled components at the data level through these nodes, creating conditions for accurately simulating the propagation paths of faults or loads between structures.

[0278] Optionally, in this embodiment, the topology graph structure is constructed to transform the physical model into a computational model that can be recognized and processed by data-driven algorithms.

[0279] Based on the actual mechanical connections and physical interactions of the wind turbine (such as force transmission and vibration propagation), a clear connection relationship (edge) is established between the finite element elements (vertices) defined in step S301 and the boundary nodes (vertices) defined in step S302. Ultimately, a complete graph data structure that can reflect the actual structure of the wind turbine is formed.

[0280] This step has two effects: First, at the physical simulation level, the graph structure defines the mathematical relationships between the variables in the global coupled equation set; second, at the machine learning diagnostic level, the graph structure is directly used as the input of the graph neural network (GNN), enabling the AI ​​model to understand the "propagation relationship" of faults in the wind turbine mechanical structure in the same way it understands social networks, thereby achieving accurate diagnosis and prediction of cross-component coupled faults.

[0281] Through step S603, this embodiment successfully achieved the transformation from physical entity to digital network model.

[0282] In one embodiment of the machine learning-based deep self-diagnosis method for wind turbines in this application, the method may further include the following:

[0283] Step S701: Based on the network topology, in the local coordinate system, establish the control equations for the fluid field, structural force field and spatial position field for each finite element, and combine the control equations of all finite elements to determine the corresponding global coupling equation set.

[0284] Step S702: Based on the sensor data acquisition volume of each finite element, the global coupled equations are solved using numerical methods to determine the physical response data of the corresponding wind turbine unit in the entire field.

[0285] Optionally, in this embodiment, the core of this process lies in transforming the continuous physical entity of the wind turbine into a discretized mathematical model that can be numerically solved by a computer.

[0286] First, based on the wind turbine network topology, and according to the physical characteristics of the specific component (such as blades or gearboxes) or dynamic space (such as connection space) to which the finite element belongs, a unique physical field governing equation is assigned to each discretized finite element. For example, an element located within a gearbox is assigned fluid field equations involving heat conduction and oil flow, as well as structural force field equations describing gear meshing vibrations, while an element located within the tower structure mainly involves structural force field and spatial position field equations. After determining the equation type, physical and material parameters characterizing the inherent properties of the element, such as density, elastic modulus, and thermal conductivity, are substituted as constants into the equations, thus completing the mathematical definition of a single finite element in a physical sense.

[0287] Subsequently, an independent mathematical solution environment is established for each finite element with assigned equations and parameters. By establishing a local coordinate system on each element and defining spatial dimensions and introducing time variables, the originally static equations are transformed into four-dimensional functions that dynamically describe the changes of physical quantities (such as temperature, displacement, and stress) within the element as a function of spatial location and time. This step precisely anchors the physical phenomena from a macroscopic, holistic description to each tiny discrete element, achieving a precise characterization of complex physical fields at minute details, and laying the foundation for subsequently revealing how local faults evolve.

[0288] Finally, all discrete, independent unit equations are systematically integrated into a global, interconnected, coupled system. The link between these systems is the node connections within the network topology. At common nodes of adjacent units, physical continuity conditions (such as temperature continuity and displacement compatibility) and force balance conditions are enforced. These conditions, as mandatory boundary constraints, are applied to the entire equation set. In this way, the physical state of a single unit is no longer isolated; its changes are transmitted to adjacent units through nodes, thus simulating the propagation path of faults or loads within the overall wind turbine structure. Ultimately, all the boundary-constrained unit control equations are mathematically combined and assembled into a unified, globally coupled equation set with all node unknowns as the solution objective.

[0289] Next, after obtaining the coupled equations, the global coupled equations are solved.

[0290] We first use the multiphysics sensor data corresponding to each finite element as the boundary conditions, initial conditions, and external excitation loads of the corresponding control equations, and substitute them into the global coupling equation set, where:

[0291] Boundary conditions refer to the forced application of observed data from the system boundaries to the model. For example, environmental wind speed and temperature data can be set as boundary conditions for the fluid and temperature field equations, while the fixed constraints at the bottom of the tower can be set as displacement boundary conditions for the structural field. This ensures that the solution domain of the model remains consistent with the boundaries of the real physical world.

[0292] Initial conditions refer to the starting states provided for the governing equations describing transient processes. For example, the initial temperatures and stress distributions of each component measured at system startup can be used as initial values ​​for the heat conduction equations and structural force field equations, allowing the dynamic evolution of the model to begin from a realistic state.

[0293] External excitation loads refer to the physical quantities that directly drive the system response and are substituted into the source or force terms of the equations. For example, the power generation data is applied to the system as the generator load, the measured wind pressure on the blade surface is applied to the structural dynamics equations as an external load, and the frictional heat generated by the gearbox is substituted into the temperature field equations as a heat source term.

[0294] After this step, the theoretical mathematical model will be instantiated into a high-fidelity digital twin for the current specific operating conditions.

[0295] Next, after parameterizing the equation system, a numerical iterative algorithm is used to solve this large, coupled global system of equations. Preferably, the numerical iterative algorithm can be the Newton-Raphson method or the conjugate gradient method in the finite element method. The physical response data obtained from the solution across the entire field can be used to calculate physical quantities that can be obtained even in locations where no sensors are installed. For example, by solving the equations, the stress distribution at key points inside the blade, the contact stress on the gearbox bearing raceway, and the temperature field contour map of the entire nacelle can be obtained.

[0296] Through step S702, this embodiment successfully reconstructs and extrapolates the physical state of the space field from discrete sensor data points, providing a solid data foundation for subsequent lifecycle prediction.

[0297] To improve the accuracy and efficiency of fault diagnosis and prediction for wind turbine generators, this application provides an embodiment of a machine learning-based deep self-diagnosis device for wind turbine generators, which implements all or part of the aforementioned machine learning-based deep self-diagnosis method. See [link to embodiment]. Figure 2 The machine learning-based deep self-diagnosis device for wind turbines specifically includes the following components:

[0298] The wind turbine structure division module 10 is used to divide the overall structure of the wind turbine into structural components and related spaces, and to collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0299] The wind turbine coupling equation determination module 20 is used to discretize the structural components into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, solve the global coupling equation set using numerical methods, and determine the corresponding physical response data.

[0300] The wind turbine lifecycle self-diagnosis module 30 is used to construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. It acquires historical data of the wind turbine, initializes and trains the unsupervised and supervised learning modules based on the historical data, and then inputs real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module outputs anomaly detection results and operating condition clustering results, while the supervised learning module predicts the lifecycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0301] As described above, the machine learning-based deep self-diagnosis device for wind turbines provided in this application can divide the overall structure of a wind turbine into structural components and related spaces, collect multi-physics sensor data from the overall structure, discretize the structural components of the wind turbine into multiple finite units, define the related spaces as nodes connecting the finite units, construct a corresponding wind turbine network topology, establish corresponding physical field control equations for the finite units based on the network topology, determine the corresponding global coupling equation set, solve the global coupling equation set using numerical methods, determine the corresponding physical response data, and input real-time multi-physics sensor data and real-time physical response data into a set hybrid learning model for unsupervised and supervised hybrid learning to determine the life cycle prediction results of the wind turbine. This can improve the accuracy and efficiency of wind turbine fault diagnosis and prediction.

[0302] To further illustrate this solution, this application also provides a specific application example of implementing the machine learning-based deep self-diagnosis method for wind turbines using the aforementioned machine learning-based deep self-diagnosis device, which specifically includes the following:

[0303] 1. Machine Learning Methodology

[0304] For different tasks and data types in the whole life cycle prediction of wind turbines, this invention adopts a hierarchical hybrid machine learning model architecture, including unsupervised learning and supervised learning. The two architectures are parallel and hybrid, and the specific process is as follows:

[0305] First, unsupervised pre-training is performed using variational autoencoders (VAEs) and adversarial autoencoders (AAEs) to reduce the dimensionality and noise of the original sensor data, extract robust feature representations, and enhance the generalization ability of subsequent models. Simultaneously, dimensionality reduction of different types of data from different parts to the same order of magnitude of sensitivity facilitates their combination during subsequent modeling and fusion.

[0306] Next, the preprocessed data will be input into the unsupervised learning layer:

[0307] A) Unsupervised learning layer, where A1 and A2 also work in parallel and in combination.

[0308] A1. Anomaly Detection and Unknown Fault Discovery Layer

[0309] Methods: Combining Isolation Forest, One-Class SVM, and autoencoder-based reconstruction error detection.

[0310] Description: This layer combines isolation forest, one-class support vector machine (SVM) and autoencoder reconstruction error. Since most of the data on site is normal data, while the number and types of fault data samples are scarce, it is convenient to label normal data. By learning the boundary of the "normal" state, it is possible to discover abnormal data and unknown faults (outliers) without relying on fault labels.

[0311] Implementation: Train a One-Class SVM or deep autoencoder on device health status data to learn the boundaries or reconstruction patterns of "normal" states. During online runtime, calculate the reconstruction error of new data or the distance to the decision boundary in real time. Trigger an anomaly alert when the error or distance exceeds an adaptive threshold. Isolation forests are used to quickly screen outliers in high-dimensional feature spaces, and are particularly effective for scenarios with high-dimensional data and few outliers.

[0312] enter:

[0313] Content: Equipment health status data or feature-enhanced data, including sensor readings or feature vectors under normal conditions. For example, vibration data and temperature data.

[0314] Data type: Multidimensional feature vectors or raw time series data. Input can be a numerical array (such as feature vectors) or a raw signal (such as vibration waveform), and data types include floating-point numbers (continuous values) and integers (discrete sampling points).

[0315] Output:

[0316] Content: Anomaly alerts or outlier detection results, such as reconstruction error score and distance to the decision boundary. When the error exceeds the adaptive threshold, a binary alert or confidence score is triggered.

[0317] Data type: scalar or binary label (e.g., 0 / 1 for normal / abnormal), or probability value (e.g., anomalous score). Output is numeric and used to trigger alerts or subsequent processing.

[0318] Benefits: Early anomaly detection can be achieved without fault labels, and it has the ability to initially discover novel fault modes (Unknown-Unknowns) that have not appeared in the training set, which greatly complements the limitation of supervised models that can only identify known faults.

[0319] A2. Unsupervised Clustering and Operational Mode Partitioning

[0320] Methods: Deep Embedded Clustering (DEC) and temporal clustering (such as K-Shape).

[0321] Description: This layer uses deep embedding clustering (DEC) and temporal clustering (such as K-Shape) to automatically identify different operating conditions of the wind turbine (such as rated power generation, start-up and shutdown). No manual labeling is required; only simple joint pre-training of categories is needed to provide a fine data foundation for fault diagnosis. In the first chapter of this patent, it is mentioned that a total of 16 types of data were collected. It can be seen that the data in categories (2) to (16) are the operating or status data of various internal parts of the wind turbine, while most of the data in category (1) are the environmental data of the external environment of the wind turbine (such as seasonal temperature and humidity, air density, atmospheric pressure, instantaneous wind speed and direction). These data need to be clustered to determine the operating mode of the wind turbine at each moment. These are closely related to the state, health and life prediction of the wind turbine. The original dimension of these data is low, and the data itself does not need to be dimensionality reduced. However, the wind turbine normal, abnormal and unknown fault data discovered in step A1 need to be based on the different operating modes divided in this layer.

[0322] Implementation: Unsupervised clustering analysis (DEC) is performed using the features obtained from the feature extraction layer. DEC first acquires features using a pre-trained autoencoder, then optimizes the clustering objective, and simultaneously fine-tunes the feature extractor to make the feature space more conducive to clustering. The number of clusters K in Deep Embedded Clustering (DEC) is automatically determined based on the Bayesian Information Criterion (BIC). For time-series data (such as power curves and wind speed sequences), time-series clustering is performed using algorithms such as K-Shape to automatically identify different operating conditions of wind turbines (such as rated power generation, start-up and shutdown, idling, and turbulent operation).

[0323] enter:

[0324] Content: Features after processing by the feature extraction layer, or original time-series data (such as power curves, wind speed sequences).

[0325] Data type: Feature vector (numerical array) or time series data (e.g., power value series, unit: kW). Input is multidimensional data and includes spatiotemporal features (e.g., node features changing over time).

[0326] Output:

[0327] Content: Clustering labels or operational status divisions, such as category identifiers for different operating conditions (e.g., cluster centers) or time-series segmentation results.

[0328] Data type: Category labels (integers, such as 0, 1, 2 representing different states) or cluster assignment vectors. The output is discrete data, used for state identification and sub-state modeling.

[0329] Benefits: It automatically discovers hidden natural clustering structures in the data, identifies different operating states and potential health degradation stages without manual annotation, and provides a more refined data foundation for subsequent state-specific modeling, fault diagnosis and RUL prediction.

[0330] B. Supervised learning layer

[0331] B1. Feature Extraction and Fusion Layer (Bottom Layer Processing)

[0332] Method: A combination of one-dimensional convolutional neural network (1D-CNN) and wavelet packet transform (WPD).

[0333] Description: For high-frequency time-series data (such as vibration and current signals), time-frequency domain analysis is first performed using Wavelet Processing (WPD) to extract features such as energy entropy and scale entropy. These features, along with the preprocessed original signal, are then input into a 1D-CNN. The CNN's convolutional kernels automatically learn the local dependencies and translation invariance features in the signal, significantly improving the detection capability for weak fault signs such as vibration characteristic frequencies and electrical harmonics. (For example: wavelet packet decomposition uses the db4 wavelet basis with 5 decomposition layers; the 1D-CNN uses 4 convolutional layers with kernel sizes of 64, 32, 16, and 8 respectively; the GNN uses a Graph Attention Network (GAT) with 4 heads.)

[0334] enter:

[0335] Content: High-dimensional raw sensor data, such as time-series data like vibration waveforms and current signals, or time-frequency characteristics (such as energy entropy and scale entropy) preprocessed by WPD. This data comes directly from the multiphysics coupled solver or the original sensor acquisition system.

[0336] Data type: One-dimensional time series data, such as vibration acceleration signals (unit: m / s²) and current values ​​(unit: A). The data is usually continuous numerical and has a high sampling frequency (e.g., kHz level).

[0337] Output:

[0338] Content: The fused high-dimensional deep feature vector contains local dependencies and physical features (such as failure frequency).

[0339] Data type: High-dimensional feature vector (floating-point array), the dimension of which is determined by the CNN architecture (such as multi-layer convolution output). The output is numerical and serves as the input to the intermediate fusion layer.

[0340] Finally, a high-dimensional deep feature vector that incorporates physical meaning is generated and used as input to the upper-level model.

[0341] Benefits: This step effectively removes noise and extracts robust features, especially when labeled data is insufficient, it can significantly improve the generalization ability and training efficiency of subsequent supervised models (such as CNNs).

[0342] B2. Spatiotemporal Feature Learning Layer (Mid-Layer Fusion)

[0343] Methods: Graph Neural Network (GNN) and Long Short-Term Memory Network (LSTM).

[0344] Description: This layer uses a graph neural network (GNN) and a long short-term memory network (LSTM) to learn the spatial topological relationships and temporal dependencies in the wind turbine structure and simulate the fault propagation path.

[0345] Implementation: This invention treats the discretized wind turbine structure (4 main parts + 4 main spaces) as a graph structure, where each finite element element or node is a vertex in the graph, and physical connections or energy transfer paths are edges. Each vertex carries data on the time-varying deep features extracted from the lower layer. A GNN is responsible for learning the spatial topological relationships between nodes, simulating the propagation paths of fault factors in the structure (e.g., how gearbox failures affect generator and nacelle sway). An LSTM is responsible for learning the long-term dependencies of each node's own features over time, capturing performance degradation trends.

[0346] enter:

[0347] Contents: 1. Graph structure data, where nodes represent discretized units of the wind turbine (such as blades and gearboxes), and edges represent physical connections. Each node carries data on the time-varying depth characteristics of the B1 output; 2. Anomaly probability values ​​output by the A1 anomaly detection and unknown fault discovery layer; 3. Operating condition category results output by the A2 unsupervised clustering and operating mode segmentation layer.

[0348] Data type: Graph data (node ​​feature matrix and adjacency matrix), where node features are time-series sequences (floating-point arrays) with time steps potentially in the range of seconds or minutes. Data types include graph structure information (such as node IDs) and feature values.

[0349] Output:

[0350] Content: The hidden state of each node at each time step encodes the spatiotemporal health information of itself and its neighbors.

[0351] Data type: Hidden state vector (floating-point array), the dimension of which is determined by the parameters of GNN and LSTM (e.g., 4 attention heads). The output is numerical and used for top-level decision-making.

[0352] B3. Unsupervised Degenerate Trajectory Modeling Layer

[0353] Methods: Nonlinear dimensionality reduction (UMAP) and trajectory mining.

[0354] Description: This layer uses nonlinear dimensionality reduction (such as UMAP) and trajectory mining algorithms to visualize high-dimensional spatiotemporal features, analyze the degradation path of equipment from health to failure, use fault analysis as the criterion for failure degradation, and finally assist in health assessment and modeling.

[0355] Implementation: The high-dimensional spatiotemporal features (GNN / LSTM output) are reduced to 2D or 3D using unsupervised UMAP, with parameters set to n_neighbors=15 and min_dist=0.1. The degradation trajectory of the device throughout its entire lifecycle is visualized and analyzed. In the low-dimensional space, a trajectory mining algorithm automatically identifies common patterns in degradation paths, stage inflection points, and final failure clusters.

[0356] enter:

[0357] Content: Spatiotemporal feature data, from the B2 layer (GNN and LSTM outputs), such as node hidden state sequences, which encode the health information of the device throughout its entire lifecycle.

[0358] Data type: High-dimensional spatiotemporal feature vector (numerical array), which may contain time step and node dimension (as shown in the node features in the structure). The data type is a floating-point number, representing feature strength or relation weight.

[0359] Output:

[0360] Content: Dimensionally reduced trajectory coordinates (such as 2D or 3D point sets) or degenerate path patterns, such as sequences of trajectory points in low-dimensional space, for visualizing health evolution.

[0361] Data type: Low-dimensional coordinates (floating-point arrays, such as 2D or 3D points) or trajectory descriptors (such as path vectors). Output is numerical, used for qualitative assessment and health metric construction.

[0362] B4. Fault Diagnosis and Life Prediction Layer (Top-Level Decision Making)

[0363] Description: This layer uses an attention mechanism and a multilayer perceptron (MLP) for fault classification and a Transformer for remaining lifetime prediction to make the final decision. The sequence is as follows: fault diagnosis is performed first, followed by remaining lifetime prediction based on fault severity and impact on the overall system lifespan.

[0364] enter:

[0365] Content: Spatiotemporal feature sequences, from the output of layer B3 (such as node hidden state sequences), which may be combined with attention weights to highlight key nodes.

[0366] Data type: Sequence data (numerical array), representing features changing over time, potentially containing multidimensional time series. Floating-point data type, used for regression or classification.

[0367] Output:

[0368] Contents: Fault type classification results (e.g., bearing fault codes) and remaining life prediction (RUL, in hours).

[0369] Data type: Categorical output is a category label (integer, such as fault type ID), and regression output is a continuous value (such as RUL hours). The output can be discrete or continuous numerical, used for maintenance decisions.

[0370] Among them, the fault type classification results (classification task):

[0371] Method: Attention mechanism + Multilayer Perceptron (MLP).

[0372] Implementation: The spatiotemporal features learned by GNN and L are input into the attention layer. The model automatically learns and highlights the sensor nodes and time points most relevant to the current fault, significantly improving the model's interpretability and ability to distinguish complex faults. Finally, MLP is used to achieve accurate classification of fault types (e.g., bearing inner ring fault, gear tooth breakage, bolt loosening, etc.).

[0373] Among them, the remaining lifetime prediction (regression task):

[0374] Method: Transformer encoder or autoregressive model.

[0375] Implementation: Spatiotemporal features are input into the Transformer as a sequence. Its powerful global dependency modeling capabilities can accurately predict the future trajectory of the performance degradation curve. We define RUL as the time it takes for the device health index (HI) to drop to the failure threshold. The health index is constructed by weighting key indicators output by the multiphysics solver (such as maximum stress and temperature gradient) and features extracted by machine learning.

[0376] 2. Model Training Process

[0377] Model training is a phased, continuous learning process:

[0378] 2.1 Unsupervised-Supervised Collaborative Learning Mechanism

[0379] Semi-supervised learning: It uses unsupervised modules (such as clustering and anomaly detection) to process massive amounts of unlabeled data, filter out high-confidence "pseudo-label" samples, or automatically label different running states, thereby expanding the training set of supervised models and significantly reducing the reliance on manual labeling.

[0380] Active learning: Novelty samples discovered by the unsupervised anomaly detection module can automatically trigger manual review requests or be prioritized for entry into the online learning queue, guiding experts to focus on the most valuable samples and efficiently constructing high-quality labeled datasets.

[0381] Model self-updating: When the unsupervised module continuously detects a new pattern (such as a new cluster or anomaly pattern) and it is confirmed by experts or verified by maintenance, the system can automatically treat it as a new fault category or health status, update the classifier structure or output layer of the supervised model, and realize the autonomous evolution of diagnostic knowledge.

[0382] 2.2 Offline Pre-training Stage

[0383] Data: Training is conducted using historical SCADA data, manual maintenance records, and a fault case library generated from multiphysics simulation.

[0384] Loss function:

[0385] Diagnostic model: Focal Loss, which effectively solves the problem of extremely imbalanced class of fault samples.

[0386] Prediction model: combining mean squared error (MSE) and coefficient of determination (R²).

[0387] Process: First, train the CNN feature extractor, GNN and LSTM respectively. Then, train the top classifier and regressor while freezing the bottom parameters. Finally, perform global fine-tuning.

[0388] 2.3 Online Self-Learning and Optimization Phase (Core Innovation)

[0389] Methods: Online learning and Bayesian optimization framework.

[0390] Implementation:

[0391] After the system is deployed, the model continuously receives real-time data and makes predictions.

[0392] When the predicted faults are confirmed to be consistent with actual maintenance records, or when sensor data indicates that the model predictions are biased, these new, validated samples will automatically be placed into a priority queue.

[0393] The system periodically (e.g., weekly) initiates an incremental learning process to fine-tune the model with new samples. At the same time, it automatically adjusts hyperparameters (such as learning rate and number of network layers) through Bayesian optimization to prevent catastrophic forgetting and ensure continuous improvement in model performance.

[0394] This process achieves the "parameter update" and "equation update" (referring to model update) described in the instruction manual.

[0395] 3. Final Result

[0396] The real-time data stream first passes through an unsupervised anomaly detection module (based on pre-trained autoencoder reconstruction error). If the reconstruction error exceeds a dynamic threshold, an alarm is triggered and the sample is temporarily stored. For normal data, deep features are further extracted and input into an unsupervised clustering module to automatically identify the current operating condition. Finally, based on the identified operating condition, the corresponding supervised-trained fault diagnosis and RUL prediction model (such as GNN-LSTM) is invoked for precise analysis. All suspected unknown pattern samples discovered by the unsupervised module are periodically submitted to an expert system or online learning process for review and learning.

[0397] By introducing the aforementioned machine learning engine, this invention achieves a qualitative leap compared to traditional methods:

[0398] ① Improved early fault detection rate: For faults such as early pitting of bearings and micro-pitting of gears, the average warning window has been extended from the traditional <72 hours to more than 240 hours, and the false alarm rate has been reduced to below 2%.

[0399] ② Remaining lifetime prediction accuracy: The RUL prediction error has been reduced from more than ±15% in the traditional model to within ±5% (for example, in the example, the predicted lifetime is 863 hours, the actual lifetime is 842 hours, and the error is only 2.5%).

[0400] ③ Generalization performance: The hybrid model architecture demonstrates strong adaptability to different types of wind turbines (direct drive / doubly fed) and different operating environments (offshore / land). It can be applied without redesigning the model, requiring only minor model structure modifications and transfer learning, thus solving the industry's generalization problem.

[0401] ④ Interpretability: Attention mechanisms and GNNs provide clear fault attribution analysis, indicating "where" and "what" faults occurred, as well as the severity of the faults, providing valuable decision support for maintenance personnel, rather than a "black box" alarm.

[0402] ⑤ Handling unknown faults: The system has the initial ability to detect and discover new faults that have not been seen during the training phase, reducing the risk of missed detections.

[0403] ⑥ Reduce dependence on annotations: Make full use of the large amount of easily accessible unannotated historical data from wind farms to reduce the dependence on expensive expert annotations in model development and maintenance.

[0404] ⑦ Enhance generalization: Unsupervised learning helps the model learn the more fundamental distribution and structure of the data, improving its adaptability to different wind turbine models and different operating environments.

[0405] ⑧ Enhance interpretability: Clustering and visualization results help engineers understand the distribution of wind turbine operating states and degradation mechanisms.

[0406] From a hardware perspective, in order to improve the accuracy and efficiency of fault diagnosis and prediction for wind turbine generators, this application provides an embodiment of an electronic device for implementing all or part of the aforementioned machine learning-based deep self-diagnosis method for wind turbine generators. The electronic device specifically includes the following components:

[0407] The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the machine learning-based deep self-diagnosis method for wind turbines and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the machine learning-based deep self-diagnosis method for wind turbines in the previous embodiments, and the content of the embodiments of the machine learning-based deep self-diagnosis method for wind turbines is incorporated herein, and repeated parts will not be described again.

[0408] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0409] In practical applications, the machine learning-based deep self-diagnosis method for wind turbines can be partially executed on the electronic device side as described above, or all operations can be completed on the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed on the client device, the client device may further include a processor.

[0410] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0411] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0412] In one embodiment, the machine learning-based deep self-diagnostic method for wind turbines can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0413] Step S101: Divide the overall structure of the wind turbine into structural components and related spaces, and collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0414] Step S102: Discretize the structural component into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, and solve the global coupling equation set using numerical methods to determine the corresponding physical response data.

[0415] Step S103: Construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. Acquire historical data of the wind turbine. After initializing and training the unsupervised learning module and the supervised learning module based on the historical data, input real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results. The supervised learning module is used to predict the life cycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0416] As described above, the electronic device provided in this application divides the overall structure of a wind turbine into structural components and related spaces, collects multi-physics sensor data from the overall structure, discretizes the structural components of the wind turbine into multiple finite units, defines the related spaces as nodes connecting the finite units, constructs a corresponding wind turbine network topology, establishes corresponding physical field control equations for the finite units based on the network topology, determines the corresponding global coupling equation set, solves the global coupling equation set using numerical methods, determines the corresponding physical response data, and inputs real-time multi-physics sensor data and real-time physical response data into a set hybrid learning model for unsupervised and supervised hybrid learning to determine the life cycle prediction results of the wind turbine. This can improve the accuracy and efficiency of wind turbine fault diagnosis and prediction.

[0417] In another implementation, the machine learning-based deep self-diagnosis method for wind turbines can be configured separately from the central processing unit 9100. For example, the machine learning-based deep self-diagnosis method for wind turbines can be configured as a chip connected to the central processing unit 9100, and the function of the machine learning-based deep self-diagnosis method for wind turbines can be realized through the control of the central processing unit.

[0418] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0419] like Figure 3As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0420] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0421] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0422] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0423] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0424] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0425] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Furthermore, the audio processor 9130 is also coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored sound via the speaker 9131.

[0426] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the machine learning-based deep self-diagnosis method for wind turbines, where the execution subject is a server or client, as described in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the machine learning-based deep self-diagnosis method for wind turbines, where the execution subject is a server or client, as described in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0427] Step S101: Divide the overall structure of the wind turbine into structural components and related spaces, and collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0428] Step S102: Discretize the structural component into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, and solve the global coupling equation set using numerical methods to determine the corresponding physical response data.

[0429] Step S103: Construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. Acquire historical data of the wind turbine. After initializing and training the unsupervised learning module and the supervised learning module based on the historical data, input real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results. The supervised learning module is used to predict the life cycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0430] As described above, the computer-readable storage medium provided in this application divides the overall structure of a wind turbine into structural components and related spaces, collects multi-physics sensor data from the overall structure, discretizes the structural components of the wind turbine into multiple finite units, defines the related spaces as nodes connecting the finite units, constructs a corresponding wind turbine network topology, establishes corresponding physical field control equations for the finite units based on the network topology, determines the corresponding global coupling equation set, solves the global coupling equation set using numerical methods, determines the corresponding physical response data, and inputs real-time multi-physics sensor data and real-time physical response data into a set hybrid learning model for unsupervised and supervised hybrid learning to determine the life cycle prediction results of the wind turbine. This can improve the accuracy and efficiency of wind turbine fault diagnosis and prediction.

[0431] Embodiments of this application also provide a computer program product capable of implementing all steps in the machine learning-based deep self-diagnosis method for wind turbines, where the execution subject is a server or client, as described in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the machine learning-based deep self-diagnosis method for wind turbines. For example, the computer program / instruction implements the following steps:

[0432] Step S101: Divide the overall structure of the wind turbine into structural components and related spaces, and collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components.

[0433] Step S102: Discretize the structural component into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, and solve the global coupling equation set using numerical methods to determine the corresponding physical response data.

[0434] Step S103: Construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. Acquire historical data of the wind turbine. After initializing and training the unsupervised learning module and the supervised learning module based on the historical data, input real-time multiphysics sensor data and real-time physical response data into the hybrid machine learning model to determine the corresponding fault diagnosis results and remaining life prediction results of the wind turbine. The unsupervised learning module is used to output anomaly detection results and operating condition clustering results. The supervised learning module is used to predict the life cycle of the wind turbine based on the results output by the unsupervised learning module, the original sensor data, and the physical response data.

[0435] As described above, the computer program product provided in this application divides the overall structure of a wind turbine into structural components and related spaces, collects multi-physics sensor data from the overall structure, discretizes the structural components of the wind turbine into multiple finite elements, defines the related spaces as nodes connecting the finite elements, constructs a corresponding wind turbine network topology, establishes corresponding physical field control equations for the finite elements based on the network topology, determines the corresponding global coupling equation set, solves the global coupling equation set using numerical methods, determines the corresponding physical response data, and inputs real-time multi-physics sensor data and real-time physical response data into a set hybrid learning model for unsupervised and supervised hybrid learning to determine the life cycle prediction results of the wind turbine. This can improve the accuracy and efficiency of wind turbine fault diagnosis and prediction.

[0436] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0437] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0438] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0439] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0440] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A machine learning based deep self-diagnosis method for wind turbine generators, characterized in that, The method includes: The overall structure of the wind turbine is divided into structural components and related spaces. Multiphysics sensor data is collected from the overall structure. The structural components include blades, hubs, nacelles and towers. The related spaces are used to characterize the connection or relative relationship between the structural components. The structural components are discretized into multiple finite elements, and the relevant space is defined as nodes connecting the finite elements. The corresponding wind turbine network topology is determined. Based on the network topology, the corresponding physical field control equations are established for the finite elements. The corresponding global coupling equation set is determined, and the global coupling equation set is solved using numerical methods to determine the corresponding physical response data. A hybrid machine learning model is constructed, comprising an unsupervised learning module and a supervised learning module. The unsupervised learning module outputs anomaly detection results and operating condition clustering results, while the supervised learning module predicts the lifecycle of the wind turbine based on the output of the unsupervised learning module, raw sensor data, and physical response data. Historical wind turbine data is acquired, and the unsupervised and supervised learning modules are initialized and trained using this historical data. Then, feature reduction and denoising are performed on the real-time multiphysics sensor data and real-time physical response data using a preset variational autoencoder and a preset adversarial autoencoder to determine corresponding feature representations of the same magnitude. Finally, an unsupervised learning module is constructed based on a preset anomaly detection layer and a preset operating condition clustering layer, wherein the anomaly detection layer and the operating condition clustering layer are computed in parallel. The same-scale feature representation is input to the anomaly detection layer. Based on the isolation forest, a first-class support vector machine, and the reconstruction error of the autoencoder, anomaly detection is performed on the input feature representation to determine the corresponding anomaly detection result, including anomaly score or binary alarm. The anomaly alarm output based on the anomaly detection result is used to trigger incremental updates of the supervised learning module. The same-scale feature representation is input to the operating condition clustering layer. Based on the deep embedding clustering algorithm and the temporal clustering algorithm, the input feature representation is clustered into wind turbine operating states to determine the corresponding operating condition clustering result. The operating condition clustering result is used to input the supervised learning module for model training and prediction. The same-scale feature representation is input to the supervised learning module to determine the fault diagnosis result and remaining life prediction result of the wind turbine unit output by the supervised learning module.

2. The machine learning based wind turbine generator deep self-diagnosis method of claim 1, wherein, The step of inputting the same-scale feature representation into the supervised learning module and determining the fault diagnosis results and remaining life prediction results of the wind turbine output by the supervised learning module includes: The same-level feature representation is input into the feature fusion layer, and the same-level feature representation is extracted according to the preset one-dimensional convolutional neural network and wavelet packet transform to determine the corresponding high-dimensional feature vector. The high-dimensional feature vector and the clustering results of the operating conditions are input into the spatiotemporal feature learning layer. The spatial relationship and temporal series dependency between wind turbine structural nodes are modeled according to the graph neural network and the long short-term memory network to determine the corresponding spatiotemporal features. The graph neural network is constructed based on the network topology of the wind turbine. The spatiotemporal features are input into the degradation trajectory modeling layer, and the spatiotemporal features are visualized and the trajectory is mined according to the nonlinear dimensionality reduction algorithm to determine the corresponding equipment degradation path; The degradation path of the equipment is input into the fault diagnosis and life prediction layer. Fault classification is performed based on the attention mechanism and multilayer perceptron algorithm, and the remaining life is predicted using Transformer to determine the fault diagnosis result and remaining life prediction result of the corresponding wind turbine.

3. The machine learning based deep self-diagnosis method of wind turbine generators as claimed in claim 1, wherein, The step of acquiring historical data of wind turbine generators and initializing the training of the unsupervised learning module and the supervised learning module based on the historical data includes: Based on historical data of wind turbines in a healthy state, a reconstruction model based on an autoencoder is trained to establish a baseline for the normal state and determine the corresponding unsupervised learning module after training. Based on historical data containing labeled fault types, a supervised learning model is trained, and the corresponding supervised learning module is determined.

4. The deep self-diagnosis method for wind turbine generators based on machine learning according to claim 1, characterized in that, The step of discretizing the structural components into multiple finite elements and defining the relevant space as nodes connecting the finite elements to determine the corresponding wind turbine network topology includes: The continuous structures within the four structural components—blade, hub, nacelle, and tower—are each divided into multiple discrete finite element units. The relevant space is defined as the boundary node connecting the finite element units between different structural components. The relevant space includes the blade-hub connection space, the hub-nacelle connection space, the nacelle-tower connection space, and the tower-blade relative space. Based on the actual physical connection and energy transfer path of the wind turbine, the connection relationship between the boundary node and the finite element unit of the adjacent structural part is established, and the corresponding wind turbine network topology is determined.

5. The deep self-diagnosis method for wind turbine generators based on machine learning according to claim 1, characterized in that, The process of establishing corresponding physical field control equations for the finite element based on the network topology, determining the corresponding global coupling equation set, and solving the global coupling equation set using numerical methods to determine the corresponding physical response data includes: Based on the network topology, in the local coordinate system, the control equations for the fluid field, structural force field and spatial position field are established for each finite element, and the control equations of all finite elements are combined to determine the corresponding global coupling equation set. Based on the sensor data acquisition volume of each finite element, the global coupled equations are solved using numerical methods to determine the physical response data of the corresponding wind turbine unit across the entire field.

6. A deep self-diagnostic device for wind turbine generators based on machine learning, characterized in that, The device includes: The wind turbine structure division module is used to divide the overall structure of the wind turbine into structural components and related spaces, and to collect multi-physics sensor data from the overall structure. The structural components include blades, hubs, nacelles and towers, and the related spaces are used to characterize the connection or relative relationship between the structural components. The wind turbine coupling equation determination module is used to discretize the structural components into multiple finite elements, define the relevant space as nodes connecting the finite elements, determine the corresponding wind turbine network topology, establish the corresponding physical field control equations for the finite elements based on the network topology, determine the corresponding global coupling equation set, solve the global coupling equation set using numerical methods, and determine the corresponding physical response data. A wind turbine lifecycle self-diagnosis module is used to construct a hybrid machine learning model, which includes an unsupervised learning module and a supervised learning module. The unsupervised learning module outputs anomaly detection results and operating condition clustering results, while the supervised learning module predicts the wind turbine lifecycle based on the output of the unsupervised learning module, raw sensor data, and physical response data. The module acquires historical wind turbine data, initializes and trains both the unsupervised and supervised learning modules based on this historical data, and then performs feature reduction and denoising on real-time multiphysics sensor data and real-time physical response data using preset variational autoencoders and preset adversarial autoencoders to determine corresponding feature representations of the same magnitude. Finally, an unsupervised learning module is constructed based on a preset anomaly detection layer and a preset operating condition clustering layer. The clustering layer is computed in parallel. The same-scale feature representation is input to the anomaly detection layer. Based on the isolation forest, a first-class support vector machine, and the reconstruction error of the autoencoder, anomaly detection is performed on the input feature representation, determining the corresponding anomaly detection result, including anomaly score or binary alarm. The anomaly alarm output based on the anomaly detection result is used to trigger incremental updates in the supervised learning module. The same-scale feature representation is input to the operating condition clustering layer. Based on deep embedding clustering and temporal clustering algorithms, the input feature representation is clustered according to wind turbine operating states, determining the corresponding operating condition clustering result. The operating condition clustering result is used to input into the supervised learning module for model training and prediction. The same-scale feature representation is input to the supervised learning module, determining the fault diagnosis result and remaining life prediction result of the wind turbine unit output by the supervised learning module.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep self-diagnosis method for wind turbines based on machine learning as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the machine learning-based deep self-diagnosis method for wind turbines as described in any one of claims 1 to 5.

Images