Floating fan structure damage detection method and system based on transfer learning

Abstract

The invention relates to a wind power monitoring technology, in particular to a floating fan structure damage detection method and system based on transfer learning, and the method comprises the steps: constructing a motion scaling model of a floating fan, designing a blade vibration data collection unit, collecting and processing structural damage monitoring data, and obtaining a scaling model test data set; constructing a source domain structure damage fault identification model based on deep learning, and training the identification model according to the scale model test data set; establishing a floating fan full-coupling simulation model, and constructing a damage simulation method; performing multi-working-condition analogue simulation on the floating fan to generate damage simulation data; and fusing actually-measured structural damage fault data and damage simulation data to obtain a target domain data set, carrying out transfer learning on an identification model to obtain a high-precision identification model, and carrying out real-time monitoring and fault early warning on a fan state. According to the method, the limitation caused by data scarcity is overcome, and floating fan structure damage identification modeling of effective knowledge migration from the model to a real machine is realized.

Description

Technical Field

[0001] This invention relates to wind power monitoring technology, and in particular to a method and system for detecting structural damage in floating wind turbines based on transfer learning. Background Technology

[0002] As a core component of the new energy industry, wind power is increasingly driven by the need for lower costs and broader development opportunities. Against this backdrop, deep-sea wind power development based on floating wind turbine technology has become an important direction. In recent years, the successful deployment of several demonstration turbines has preliminarily verified the feasibility of floating wind turbine technology.

[0003] However, the commercialization of deep-sea floating wind power still faces severe challenges. The high construction cost of floating wind turbines and their offshore deployment characteristics result in significant safety risks and high vessel leasing costs associated with traditional on-site personnel-based operation and maintenance. These factors contribute to persistently high overall operation and maintenance costs, severely hindering the achievement of grid parity. Therefore, developing unmanned intelligent sensing and fault early warning technologies has become a key path to reducing total lifecycle costs.

[0004] Currently, floating wind turbines have not yet entered the stage of large-scale application. The number of operating units is small and their operating time is short, resulting in a severe lack of real monitoring data that can reflect various faults in the entire system. This data shortage makes it difficult to fully train data-driven deep learning models, limiting their generalization ability and deployment in actual engineering. For faults in mechanical subsystems such as transmission chains (e.g., bearings, gears), high-fidelity simulations can be performed at the component level by constructing bench test systems. However, for damage such as delamination, debonding, and erosion of large composite material structures in wind turbines (e.g., blades), their dynamic response is strongly coupled with the overall structure, making it difficult to accurately reproduce their global dynamic effects on a full-scale wind turbine through scaled-down model tests.

[0005] Traditional wind turbine structural damage monitoring relies on long-term operational data from a large number of deployed wind turbines. Even with issues like high noise and low data quality, usable monitoring tools can still be obtained. However, floating wind turbines are more expensive, require a longer deployment period for large-scale deployment, and lack sufficient structural damage or fault monitoring data in the short term after deployment. Furthermore, the distance of floating wind turbines from the coast makes data calibration difficult, inevitably resulting in lower data quality and making it difficult to meet the requirements of efficient deep learning fault monitoring models. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a floating wind turbine structural damage detection method and system based on transfer learning, which can overcome the limitations caused by data scarcity and realize the floating wind turbine structural damage identification and modeling through effective knowledge transfer from model to actual machine.

[0007] On one hand, embodiments of the present invention provide a method for detecting structural damage in floating wind turbines based on transfer learning, comprising the following steps:

[0008] S01. Construct a scaled-down motion model of the floating wind turbine, including an environmental simulation device and a physical model test device;

[0009] S02. Based on the kinematic scale model of the floating wind turbine, a blade vibration data acquisition unit is designed to build a floating wind turbine model test system to monitor the structural damage of the wind turbine.

[0010] S03. Using the kinematic scale model and blade vibration data acquisition unit, conduct model tests on the normal model and the damaged model respectively, and collect structural damage monitoring data of the model.

[0011] S04. Process the collected structural damage monitoring data to obtain the processed scaled-down model test dataset.

[0012] S05. Construct a source domain structural damage fault identification model based on deep learning, and automatically optimize and train the structural damage fault identification model based on the scaled-down model test dataset.

[0013] S11. Establish a fully coupled simulation model of the floating wind turbine using simulation software;

[0014] S12. Construct a simulation method for typical structural damage of floating wind turbines;

[0015] S13. Perform multi-condition simulation on the floating wind turbine and generate damage simulation data;

[0016] S14. The structural damage and fault data obtained from actual wind turbine monitoring and the damage simulation data are fused to obtain the target domain dataset.

[0017] S21. Using a source domain structural damage fault identification model based on deep learning as a pre-trained model, transfer learning is performed on the pre-trained model according to the target domain dataset to obtain a high-precision identification model.

[0018] S22. Deploy the high-precision identification model in the condition monitoring system for real-world engineering projects to monitor the status of wind turbines in real time and provide early warning of faults.

[0019] On the other hand, embodiments of the present invention also provide a floating wind turbine structural damage detection system based on transfer learning, implemented using the above-mentioned floating wind turbine structural damage detection method, wherein the floating wind turbine structural damage detection system includes:

[0020] A scaled-down testing apparatus is constructed, comprising a scaled-down motion model of a floating wind turbine and a blade vibration data acquisition unit. The scaled-down motion model includes an environmental simulation device and a physical model testing device. A floating wind turbine model testing system is built using the scaled-down testing apparatus to monitor structural damage to the wind turbine. Model tests are conducted on both normal and damaged models using the scaled-down motion model and the blade vibration data acquisition unit, and structural damage monitoring data of the models are collected. The collected structural damage monitoring data is processed to obtain a processed scaled-down model test dataset.

[0021] The pre-trained model building module constructs a source domain structural damage fault identification model based on deep learning. It automatically optimizes and trains the structural damage fault identification model based on a scaled-down model test dataset to obtain the source domain structural damage identification model.

[0022] The software simulation device establishes a fully coupled simulation model of the floating wind turbine through simulation software, and constructs a simulation method for typical structural damage of the floating wind turbine; it performs multi-condition simulation on the floating wind turbine to generate damage simulation data; and it fuses the structural damage and fault data obtained from actual wind turbine monitoring with the damage simulation data to obtain a target domain dataset.

[0023] The transfer module uses a deep learning-based source domain structural damage fault identification model as a pre-trained model and performs transfer learning on the pre-trained model based on the target domain dataset to obtain a high-precision identification model.

[0024] The deployment module deploys the high-precision identification model in a status monitoring system for real-world engineering projects, enabling real-time monitoring and fault warning of wind turbine status.

[0025] As can be seen from the above technical solutions, this invention constructs a floating wind turbine equipment and environmental simulation system, designs a physical model testing device and a structural damage simulation method, generates a large amount of structural damage model test data through model tests, and establishes a source domain deep learning structural damage monitoring model based on model tests. Through a real-scale simulation model, a structural damage simulation method is developed, forming a simulation-based floating wind turbine structural damage data generation method; based on actual conditions, limited data on actual floating wind turbine structural damage is integrated to form a hybrid dataset. Through transfer learning, real-scale floating wind turbine structural damage deep learning modeling is achieved.

[0026] Compared with the prior art, the technical effects achieved by the present invention include:

[0027] 1. This invention designs a physical model testing device and a structural damage simulation method. The physical model testing device efficiently generates a large amount of high-quality physical model test data. Through learning from the physical model tests, the model gains a comprehensive understanding of damage characteristics. The structural damage simulation method generates simulation data, which is then fused with real-scale floating wind turbine structural damage monitoring data. Based on the fused data, near-domain transfer learning is performed, endowing the model with the ability to identify real-scale floating wind turbine structural damage, thus improving monitoring and identification capabilities and engineering application value. Compared with existing deep learning-enhanced wind turbine structural damage or fault identification technologies, this invention significantly improves the feasibility of efficient monitoring under conditions of scarce floating wind turbine data and limited research.

[0028] 2. This invention comprehensively optimizes data preprocessing methods, deep learning model algorithms, and model hyperparameters by extending hyperparameter optimization technology. It achieves automated selection of the optimal model algorithm, architecture, and transfer learning strategy, avoiding the problems of low efficiency and limited model performance improvement faced by traditional manual tuning methods, and improving the model's recognition accuracy and generalization ability. Attached Figure Description

[0029] Figure 1 This is a flowchart of the floating wind turbine structural damage detection method based on transfer learning in an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of the kinematic scale model of the floating wind turbine in an embodiment of the present invention;

[0031] Figure 3 This is a framework diagram of the fully coupled simulation model of the floating wind turbine in this embodiment of the invention;

[0032] In the diagram: 1-Fan matrix; 2-Rectifier grid; 3-Fan blade; 4-Nacelle; 5-Six-degree-of-freedom motion platform; 6-Platform control box; 7-Load; 8-Load controller; 9-Monitoring sensor; 10-Signal acquisition unit; 11-Control center. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0034] Example

[0035] This embodiment provides a method for detecting structural damage in floating wind turbines based on transfer learning. (See also...) Figure 1 This includes the following steps:

[0036] S01. Construct a scaled-down motion model of the floating wind turbine, including an environmental simulation device and a physical model test device.

[0037] This embodiment first establishes a scaled-down model simulating a floating wind turbine and its motion under wind and wave action, mainly including a physical model testing device and an environmental simulation device. For example... Figure 2 As shown, the environmental simulation device is mainly used to simulate wind fields and waves, including wind field generation components and wave simulation components.

[0038] The wind field generation component includes a fan matrix 1 and a rectifier grid 2 located at the air outlet of the fan matrix 1. The fan matrix 1 is mounted on a frame cabinet and fixed to the ground via the cabinet. It contains nine independently operating fans arranged in three rows and three columns, but is not limited to this single configuration. Each fan in the fan matrix 1 is a bladed axial flow fan, connected to the corresponding position inside the frame cabinet using bolts and nuts. The frame cabinet can be a metal frame cabinet. The core function of the rectifier grid 2 is to improve and homogenize the output wind field quality of the fan matrix, generating a stable and somewhat controllable wind field.

[0039] The wave simulation component includes a six-degree-of-freedom motion platform 5, a platform control box 6, and a control center 11 connected in sequence. It mainly simulates the motion effect of waves on the floating wind turbine through the motion of the six-degree-of-freedom motion platform 5. The control center 11 controls the motion mode, amplitude, speed and other parameters of the six-degree-of-freedom motion platform 5 through the platform control box 6 to simulate the six-degree-of-freedom motion of the wind turbine caused by different waves at sea.

[0040] The physical model test setup mainly consists of a scaled-down wind turbine model comprising wind turbine blades (3), a nacelle (4), and a six-degree-of-freedom motion platform (5). The wind turbine blades are connected to the nacelle, which is connected to the six-degree-of-freedom motion platform via a tower. A near-rigid connection between the tower and the platform is achieved through clamps that match the six-degree-of-freedom motion platform. The wind turbine blades can rotate normally to generate electricity under the action of the inflow wind generated by the fan matrix, and the speed of the wind turbine blades can be controlled by the control center to simulate different speeds of an actual wind turbine.

[0041] S02. Based on the kinematic scale model of the floating wind turbine, a blade vibration data acquisition unit is designed to build a floating wind turbine model test system to monitor wind turbine structural damage.

[0042] The blade vibration data acquisition unit includes a load 7, a load controller 8, a monitoring sensor 9, a signal acquisition unit 10, and... 11.

[0043] Load 7 is connected between load controller 8 and the nacelle of the physical model test device. Load controller 8 is connected to control center 11. The load current is controlled by control center 11. Load controller 8 executes control commands from control center to regulate load current and power, and adjust wind turbine generator torque and speed. All components are connected by ordinary wires. This embodiment only shows one possible form of load design; the current can also be controlled by battery storage and generator controller.

[0044] For structural damage monitoring, monitoring sensor 9 is installed on the wind turbine tower to detect characteristic signals of the wind turbine, such as acceleration, strain signals, and acoustic signals. Specifically, it can be a structural vibration sensor, strain gauge, or acoustic sensor, etc., that matches the type of damage being monitored. Monitoring sensor 9 is connected to signal acquisition unit 10 via an anti-interference data cable to reduce environmental interference to the signal. Signal acquisition unit 10 is connected to control center 11 via a data transmission cable, and control center 11 acquires and stores the signals from monitoring sensor through signal acquisition unit 10.

[0045] The floating wind turbine model test system works as follows for acquiring blade vibration data: the motion of the six-degree-of-freedom motion platform 5 simulates the effect of the ocean wave environment on the floating wind turbine. Simultaneously, an external fan matrix 1 simulates the sea breeze environment, which, after passing through a rectifier net 2, acts on the wind turbine blades 3 of the model. Therefore, the six-degree-of-freedom motion platform and the fan matrix together constitute the external load (motion) input of the floating wind turbine. When the wind turbine operates under the external load, the wind turbine blades absorb wind energy, thereby driving the generator to generate electricity. The electrical energy is consumed through the load 7, and the load power is controlled by the load controller 8 and transmitted and stored to the control center. Simultaneously, monitoring sensors collect acceleration data, strain signals, and acoustic signals from the tower, which are transmitted to the control center 11 through the signal acquisition unit 10, and then to the central computer. The motion data of the six-degree-of-freedom motion platform simulates the motion data of the floating wind turbine platform and is transmitted to the central computer through the control center. The strain signal reflects the degree of deformation of the wind turbine material, and the acoustic signal includes the frequency and amplitude of the sound waves.

[0046] In addition, this embodiment can simulate different types of damage to floating wind turbines, specifically focusing on typical damage simulation, including: (1) blade or tower damage, which can be simulated by artificially drilling or cutting specific areas of the blade or tower; (2) bearing damage, which can be simulated by artificially corroding or grinding the bearing balls or inner and outer rings; (3) critical bolt damage, which can be simulated by cutting or loosening.

[0047] S03. Under the preset inflow wind field and platform motion conditions, the normal model and the damaged model are tested using the motion scaled-down model of the floating wind turbine and the blade vibration data acquisition unit, and the structural damage monitoring data of the model is collected.

[0048] Using a scaled-down model of a floating wind turbine and a blade vibration data acquisition unit, model tests were conducted on a normal model and a damaged model under preset inflow wind field and platform motion conditions. The structural monitoring data of the model, such as acceleration, strain signals and acoustic signals, were acquired through the information acquisition unit, and the structural response of the model under wave load and wind force was recorded.

[0049] By adjusting the wind speed and wind speed distribution of the fan matrix, different wind speeds and turbulence in the actual environment are simulated to achieve the simulation of different wind field environments; by simulating different motion forms of the platform, the platform motion under different wave inputs is simulated, increasing the diversity of environmental elements and generated data, which is conducive to deep learning.

[0050] S04. Process the collected structural damage monitoring data to obtain the processed scaled-down model test dataset.

[0051] First, the data collected was removed from the first and last two minutes of the model test to eliminate transient responses during the start-up and shutdown of the wind turbine model.

[0052] Subsequently, the collected structural damage monitoring data, such as acceleration data, were filtered using a low-pass filter to remove high-frequency noise. The cutoff frequency of the low-pass filter was selected considering the frequency characteristics of the damage type. The selection criterion for the cutoff frequency is to ensure the predicted characteristic frequency of coverage structure damage or failure. That is, satisfying .

[0053] Secondly, based on the frequency characteristics of the model's structural response, the filtered data is segmented in the time domain to reduce the dimensionality of individual data points and increase the data volume. The principle of signal segmentation is that a single data point must contain at least one cycle of structural damage characteristic signals. To ensure sufficient information content, the longest platform motion cycle can also be used as the basis; for example, if the platform motion cycle is 10 seconds, then the length of a single data point after segmentation must be at least 10 seconds.

[0054] Finally, considering the characteristics of the fault signal, the data after segmentation is enhanced by taking into account the periodicity of the signal caused by the rotation of the wind turbine blades. Specifically, considering the differences in the characteristics of blade damage with blade rotation (i.e., with time), time-frequency transformation techniques such as wavelet transform or Hilbert-Huang transform are used to enhance the key features related to wind turbine blade damage in the signal, thereby improving the signal-to-noise ratio of the data. The scaled-down model test dataset obtained after processing is used for subsequent training and debugging of the deep learning model.

[0055] S05. Construct a source domain structural damage fault identification model based on deep learning, and automatically optimize and train the constructed structural damage fault identification model based on the scaled-down model test dataset.

[0056] With deep learning algorithm set and the corresponding hyperparameters To optimize the variables, where These represent different deep learning frameworks, such as ANN, CNN, RNN, etc. This represents different sets of model hyperparameters, used to test the classification and recognition accuracy of the model. Using the scaled-down model test dataset as an indicator, the constructed structural damage fault identification model is automatically optimized and trained.

[0057] The above steps S1-S5 use the constructed scaled-down motion model of the floating wind turbine to conduct simulation tests, obtain the corresponding source domain measured data and source domain experimental structure damage identification model, which can be transferred to the damage prediction of the real model in subsequent steps.

[0058] S11. Establish a fully coupled simulation model of the floating wind turbine using simulation software.

[0059] This embodiment considers the scarcity of actual floating wind turbine structural damage and fault data, and uses simulation software to perform fully coupled simulation of the floating wind turbine to generate damage simulation data for transfer learning. Simultaneously, based on actual conditions, the damage simulation data is fused with structural damage and fault data obtained from actual wind turbine monitoring to construct a target domain dataset.

[0060] To improve the accuracy of floating wind turbine dynamics simulation and enhance its similarity to actual floating wind turbine fault data, the framework of the fully coupled simulation model established in this embodiment is as follows: Figure 3As shown, the core of the fully coupled simulation model of the floating wind turbine lies in the unidirectional and bidirectional dynamic coupling between its subsystems. Under inflow wind input, aerodynamic loads act bidirectionally with the floating wind turbine blades and tower structure, driving the turbine blades and transmission chain; bidirectional electromechanical coupling vibration exists between the transmission chain and the tower and nacelle; platform motion and tower motion constitute bidirectional coupling, while platform motion also influences the aerodynamic performance of the turbine blades, forming a bidirectional interaction between tower / nacelle dynamics and aerodynamics. Under inflow wave / current action, hydrodynamics couples bidirectionally with platform motion, while the mooring system couples bidirectionally with the platform, controlling platform positioning and recovery. This multi-coupling mechanism is key to accurately predicting the dynamic response and structural damage characteristics of the floating wind turbine.

[0061] Based on the simulation and modeling requirements for wind turbine structural damage or failure, Figure 3 The partial or complete coupling characteristics of mid-floating wind turbines require accurate selection of platforms and simulation methods to reproduce the true impact of structural damage or failure on dynamic characteristics.

[0062] S12. Construct a simulation method for typical structural damage of floating wind turbines.

[0063] based on Figure 3 The dynamic coupling characteristics of mid-floating wind turbines necessitate the selection of simulation methods for each module or subsystem to ensure the accuracy of the coupled model. Based on verifying the simulation accuracy, it is necessary to select appropriate simulation methods for target components or subsystems undergoing structural damage monitoring, according to the actual structural damage characteristics, and establish simulation methods for typical structural damage of the wind turbine. Specifically, this includes:

[0064] (1) The dynamics of the blade are simulated by using the geometrically accurate beam theory, and the blade is considered as a beam structure with initial bending and initial twist angle.

[0065] set up The position vector represents any point along the length of the reference axis of the initial beam structure. Therefore, the governing equations of the geometrically precise beam theory are as follows:

[0066] ;

[0067] ;

[0068] in For linear motion momentum, Angular momentum; For time; For cross-sectional force, For torque; This is the displacement vector in the inertial coordinate system; the "~" in the parameter symbol indicates the antisymmetric tensor of the corresponding parameter; For the distributed force applied to the beam structure, This refers to the distributed moment applied to the beam structure.

[0069] Based on the small deformation assumption, the relationships between momentum and velocity, and between strain and cross-sectional force are as follows:

[0070] ;

[0071] ;

[0072] in Indicates the mass of a unit segment. Indicates linear velocity. Indicates angular velocity; Indicates the cross-sectional stiffness. Represents one-dimensional strain. It represents one-dimensional curvature.

[0073] (2) Establish a damage model of blade material damage on cross-sectional stiffness and unit segment mass matrix to simulate blade damage; the established damage model can be solved by one of the following two methods.

[0074] Simplified method: Assume that the stiffness and mass loss caused by blade material damage is achieved through an approximate scaling criterion. This is done by roughly assessing the location of blade damage after the damage (…). ) and before injury ( The cross-sectional area ratio of ) By assuming homogeneity of the segments, the stiffness after damage is obtained. ) and quality ( The matrix is ​​as follows:

[0075] ;

[0076] ;

[0077] in, For shear stiffness in the swing direction, Shear stiffness in the oscillation direction For axial stiffness, For oscillation stiffness, To enhance swing stiffness, For torsional stiffness; , , and These are the unit segment mass of the blade, the moment of inertia in the flapping direction, the moment of inertia in the flaring direction, and the moment of inertia in the torsional direction, respectively. The ratio of the spanwise length of the damaged area along the blade to the length of the segment in which it is located; simultaneously, the mass scaling factor. It can be represented as:

[0078] .

[0079] Precise Method: Based on classical laminated plate theory, the location of damage to the composite blade material (e.g., in the spanwise and circumferential directions) and the severity of the damage (including the area and depth of the damage) are first identified. Then, the damaged segments of the blade are remodeled, and the damaged areas are simulated by deleting materials of corresponding thickness and circumferential length. Next, the mass and stiffness matrices of the damaged segments and other segments of the blade are solved. Finally, the stiffness and mass matrices of the reference blade model are replaced.

[0080] (3) To simulate the tower damage of the wind turbine, the geometrically accurate beam theory is used to simulate the tower dynamics. Similar to the geometrically accurate beam damage simulation of the blade, the mass loss and stiffness degradation of the structure at the corresponding height of the tower are simulated by modifying the mass matrix and stiffness matrix at the corresponding positions, so as to realize the prediction of the dynamic response of the tower structure under different damage levels.

[0081] (4) To simulate mooring cable damage or breakage, a lumped mass model method is used to simulate the mooring cable. Different degrees of damage are simulated by changing the stiffness of the lumped mass spring element, while ignoring mass loss. For stiffness loss, a residual stiffness coefficient is proposed. Define the mooring cable stiffness after damage for:

[0082] ;

[0083] in, Let be the tensile stiffness of the i-th mooring cable segment. The elastic modulus of the mooring cable material. This represents the cross-sectional area of ​​a segment of the mooring cable material. Based on this, the tension between two adjacent nodes of the mooring cable is... for:

[0084] ;

[0085] in This is the critical length at which elastic force begins to be generated between the two nodes. Let be the distance between the two nodes at any given time.

[0086] A mooring cable breakage can also be simulated by removing the entire mooring cable unit.

[0087] S13. Perform multi-condition simulation on the floating wind turbine to generate damage simulation data.

[0088] For the target floating wind turbine, a series of engineering-representative wind and wave combination simulation conditions are defined. The wind and wave data can be selected from the actual environmental measurement data of the target wind field to fit the wind and wave model parameters. The wind and wave combination simulation conditions meet the entire operating range from cut-in wind speed to cut-out wind speed, as well as extreme survival conditions. Preferably, the wind and wave combination simulation conditions are defined according to international standards (such as IEC 61400-3). Subsequently, based on the damage simulation method constructed in step S12, multi-condition simulations under structural damage or failure are carried out to generate and extract typical dynamic response data.

[0089] Different damage types reflect different data types of damage. In this embodiment, the following typical key signals for simulation monitoring are initially selected for the above damage types:

[0090] (1) Blade structure damage or breakage. Monitoring of relatively mild blade damage is more difficult. Acceleration near the blade tip can be used. Considering the blade imbalance caused by blade damage or mass loss, as well as the difficulty and cost of signal monitoring, acceleration at the nacelle position can be used.

[0091] (2) The difficulty of tower damage monitoring is comparable to that of blades, and acceleration monitoring can be carried out in the nacelle or at the top of the tower.

[0092] (3) The most direct way to detect a mooring cable breakage is to monitor the tension of the mooring cable. However, considering the monitoring and maintenance costs, the acceleration of the engine room or platform can also be monitored.

[0093] By using fully coupled simulation of floating wind turbines, different design load conditions are set, and the key monitoring signals of normal and damaged wind turbines are calculated to generate damage simulation data.

[0094] S14. The structural damage and fault data obtained from actual wind turbine monitoring and the damage simulation data are fused together.

[0095] The damage simulation data is analyzed and integrated with multiple design load conditions. Following the processing method in step S04, the initial 200 seconds of data in the damage simulation data are removed to eliminate transient response. At the same time, the damage simulation data is subjected to low-pass filtering and signal segmentation to reduce the dimensionality of the input data and increase the amount of data.

[0096] In addition, if there is structural damage and fault data obtained from actual wind turbine monitoring, the same processing method described above can be used to obtain actual monitoring data; the integration and fusion of damage simulation data and actual monitoring data can be carried out to improve the efficiency of transfer learning and the application value of actual engineering monitoring.

[0097] The fused data is processed to generate a training dataset, forming a target domain dataset that includes normal and damaged operating conditions under different environments, for subsequent transfer learning.

[0098] Since actual wind turbine structural damage data is scarce, this embodiment uses simulation software to obtain damage simulation data through the simulation experiments in steps S11-S14 above, and integrates it with the monitoring data of actual wind turbines to obtain the target domain dataset, so as to solve the problem of data scarcity making it difficult to train the model.

[0099] S21. Using a source domain structural damage fault identification model based on deep learning as a pre-trained model, transfer learning is performed on the pre-trained model based on the target domain dataset to obtain a high-precision identification model.

[0100] The structural damage fault identification model from step S05 is used as a pre-trained model, retaining all its network parameters and weights. This pre-trained model has been trained using source domain data from the floating wind turbine, learning the basic data characteristics of the wind turbine's health and damage status.

[0101] Based on the target domain dataset, several transfer learning strategies were initially selected. , Different transfer strategies, such as direct transfer, partial freezing, or overall network fine-tuning, are represented by the parameters of the corresponding transfer learning strategy, which are then selected as optimization parameters. Based on this, parameter optimization is performed. Considering the differences in problem complexity and data volume between the source and target domains, appropriate model hyperparameters are selected, such as the number of fully connected network layers and the learning rate. With the goal of achieving the highest damage recognition accuracy, the optimal transfer network is optimized to obtain the optimal transfer strategy and optimal network hyperparameters.

[0102] Based on the pre-trained model and the optimal network hyperparameters, the source domain structural damage and fault identification model is trained on the target domain dataset to obtain a high-precision identification model for structural damage or faults of actual wind turbines.

[0103] S22. Deploy the high-precision identification model in the condition monitoring system for real-world engineering projects to monitor the status of wind turbines in real time and provide early warning of faults.

[0104] The high-precision recognition model optimized through transfer learning in step S21 is encapsulated and deployed to the condition monitoring system of the floating wind turbine. This condition monitoring system can receive and process data collected by sensors in real time, automatically and online to assess the structural health status of the wind turbine and identify faults, and output diagnostic results immediately, thereby achieving engineering-level real-time condition monitoring and fault early warning.

[0105] In addition, to maintain and optimize the system's long-term performance, the system will encrypt and upload online operational data and corresponding diagnostic conclusions to the cloud for archiving. Once the accumulated operational data reaches a certain scale, incremental model learning can be initiated, using new data to iteratively optimize the online model, generating an improved version of the model, which will be updated after rigorous verification. This serves as an auxiliary technical means for the system to continuously maintain high accuracy.

[0106] Through the above steps, an efficient system for constructing, learning, and transferring modeling of floating wind turbine structural damage identification models has been established, along with the design of related supporting devices.

[0107] Based on the same inventive concept, this embodiment also provides a floating wind turbine structural damage detection system based on transfer learning, implemented using the above-described floating wind turbine structural damage detection method. The floating wind turbine structural damage detection system of this embodiment specifically includes:

[0108] A scaled-down testing apparatus is constructed, comprising a scaled-down motion model of a floating wind turbine and a blade vibration data acquisition unit. The scaled-down motion model includes an environmental simulation device and a physical model testing device. A floating wind turbine model testing system is built using the scaled-down testing apparatus to monitor structural damage to the wind turbine. Model tests are conducted on both normal and damaged models using the scaled-down motion model and the blade vibration data acquisition unit, and structural damage monitoring data of the models are collected. The collected structural damage monitoring data is processed to obtain a processed scaled-down model test dataset.

[0109] The pre-trained model building module constructs a source domain structural damage fault identification model based on deep learning, and automatically optimizes and trains the structural damage fault identification model based on a scaled-down model test dataset.

[0110] The software simulation device establishes a fully coupled simulation model of the floating wind turbine through simulation software, and constructs a simulation method for typical structural damage of the floating wind turbine; it performs multi-condition simulation on the floating wind turbine to generate damage simulation data; and it fuses the structural damage and fault data obtained from actual wind turbine monitoring with the damage simulation data to obtain a target domain dataset.

[0111] The transfer module uses a deep learning-based source domain structural damage fault identification model as a pre-trained model, and performs transfer learning on the pre-trained model based on the target domain dataset to obtain a high-precision target domain identification model.

[0112] The deployment module deploys the high-precision target domain recognition model in a status monitoring system for real-world engineering projects, enabling real-time monitoring and fault warning of wind turbine status.

[0113] In the above modules, the scaled-down experimental device is implemented through steps S01-S04, the pre-trained model construction module is implemented through step S05, the software simulation device is implemented through steps S11-S14, and the migration module and deployment module are implemented through steps S21 and S22, respectively. For the specific implementation process, please refer to the detailed description of each of the above steps.

[0114] This invention constructs a method for monitoring structural damage or faults of floating wind turbines using transfer learning as a basic strategy. It addresses the problem of scarce and low-quality structural damage data for floating wind turbines by using transfer learning. By establishing a physical model to simulate the main structural damage, it integrates the authenticity of physical experimental data with the richness of numerical simulation data to carry out model training, which significantly improves the reliability of the basic pre-trained model and reduces the need for real-scale structural damage and fault data of floating wind turbines.

[0115] Secondly, the optimal monitoring model and data preprocessing method were selected through extended hyperparameter optimization: taking hyperparameter optimization as the basic means, the range of hyperparameters was expanded, and additional hyperparameters such as data feature extraction algorithms and model architecture were integrated to construct an optimal model architecture selection strategy for simultaneous selection of data preprocessing algorithms and model algorithms.

[0116] In addition, a whole-machine dynamic model and environmental testing system for simulating structural damage or failure of floating wind turbines were constructed, realizing the selection of key monitoring signals for different failure types and efficient acquisition of monitoring signals; for the main structural damage of floating wind turbines, such as blade damage / fracture, tower damage, and mooring damage / fracture, a simulation method for the corresponding damage was constructed based on the establishment of the whole-machine dynamic characteristic requirements of floating wind turbines.

[0117] In summary, this invention not only solves the problems of traditional methods relying on manual labor and having poor adaptability, but also makes up for the lack of structural damage data for floating wind turbines. It provides a feasible fault diagnosis solution based on transfer learning for the development of offshore floating wind power, and has significant technological innovation and engineering application value.

[0118] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for detecting structural damage in floating wind turbines based on transfer learning, characterized in that, Includes the following steps: S01. Construct a scaled-down motion model of the floating wind turbine, including an environmental simulation device and a physical model test device; S02. Based on the kinematic scale model of the floating wind turbine, a blade vibration data acquisition unit is designed to build a floating wind turbine model test system to monitor the structural damage of the wind turbine. S03. Using the kinematic scale model and blade vibration data acquisition unit, conduct model tests on the normal model and the damaged model respectively, and collect structural damage monitoring data of the model. S04. Process the collected structural damage monitoring data to obtain the processed scaled-down model test dataset. S05. Construct a source domain structural damage fault identification model based on deep learning, and automatically optimize and train the structural damage fault identification model based on the scaled-down model test dataset. S11. Establish a fully coupled simulation model of the floating wind turbine using simulation software; S12. Construct a simulation method for typical structural damage of floating wind turbines; S13. Perform multi-condition simulation on the floating wind turbine and generate damage simulation data; S14. The structural damage and fault data obtained from actual wind turbine monitoring and the damage simulation data are fused to obtain the target domain dataset. S21. Using a source domain structural damage fault identification model based on deep learning as a pre-trained model, transfer learning is performed on the pre-trained model according to the target domain dataset to obtain a high-precision identification model. S22. Deploy the high-precision identification model in the condition monitoring system for real-world engineering projects to monitor the status of wind turbines in real time and provide early warning of faults.

2. The method for detecting structural damage of a floating wind turbine according to claim 1, characterized in that, The environmental simulation device is used to simulate wind fields and waves, including wind field generation components and wave simulation components; The wind field generation component includes a fan matrix and a rectifier mesh set at the air outlet of the fan matrix. The rectifier mesh is used to improve and homogenize the output wind field quality of the fan matrix, generating a stable and controllable wind field. The wave simulation component includes a six-degree-of-freedom motion platform, a platform control box, and a control center connected in sequence. The motion of the six-degree-of-freedom motion platform simulates the motion effect of waves on the floating wind turbine. The control center controls the motion mode, amplitude, and speed of the six-degree-of-freedom motion platform through the platform control box.

3. The method for detecting structural damage of a floating wind turbine according to claim 1, characterized in that, The physical model test device includes wind turbine blades, a nacelle, and a six-degree-of-freedom motion platform. The wind turbine blades are connected to the nacelle, and the nacelle is connected to the six-degree-of-freedom motion platform through a tower. The tower and the platform are nearly rigidly connected by a clamp that matches the six-degree-of-freedom motion platform.

4. The method for detecting structural damage of a floating wind turbine according to claim 1, characterized in that, The blade vibration data acquisition unit includes a load, a load controller, monitoring sensors, a signal acquisition unit, and... ; The load is connected between the load controller and the nacelle of the physical model test device. The load controller is connected to the control center. The load controller executes the control commands of the control center to regulate the load current and power, and adjust the torque and speed of the wind turbine generator. Monitoring sensors are installed on the wind turbine tower to detect characteristic signals of the wind turbine, including acceleration, strain signals and acoustic signals. The signal acquisition unit is connected to the control center via a data transmission line, and the control center acquires and stores the signals from the monitoring sensors through the signal acquisition unit.

5. The method for detecting structural damage of a floating wind turbine according to claim 1, characterized in that, Step S04 involves filtering the collected structural damage monitoring data using a low-pass filter; then, the filtered data is segmented in the time domain to reduce the dimensionality of individual data and increase the data volume; finally, considering the periodicity of the signal caused by the rotation of the wind turbine blades, the segmented data is enhanced to obtain a scaled-down model test dataset.

6. The method for detecting structural damage of a floating wind turbine according to claim 1, characterized in that, Step S12 includes: (1) The dynamics of the blade are simulated using the geometrically accurate beam theory, and the blade is considered as a beam structure with initial bending and initial twist angle; (2) Establish a damage model of blade material damage on cross-sectional stiffness and unit segment mass matrix to simulate blade damage; (3) The dynamics of the tower are simulated by using the geometrically accurate beam theory. By modifying the mass matrix and stiffness matrix at the corresponding positions, the mass loss and stiffness degradation of the structure at the corresponding height of the tower are simulated respectively, so as to realize the prediction of the dynamic response of the tower structure under different damage levels. (4) To simulate mooring cable damage or breakage, the lumped mass model method is used to simulate the mooring cable, and the stiffness of the lumped mass spring unit is changed to simulate different degrees of damage.

7. The method for detecting structural damage of a floating wind turbine according to claim 6, characterized in that, In step (1), let... The position vector represents any point along the length of the reference axis of the initial beam structure. Therefore, the governing equations of the geometrically precise beam theory are: ； ； in For linear motion momentum, Angular momentum; For time; For cross-sectional force, For torque; This is the displacement vector in the inertial coordinate system; the "~" in the parameter symbol indicates the antisymmetric tensor of the corresponding parameter; For the distributed force applied to the beam structure, The distributed moment applied to the beam structure; The relationships between momentum and velocity, and between strain and cross-sectional force are as follows: ； ； in Indicates the mass of a unit segment. Indicates linear velocity. Indicates angular velocity; Indicates the cross-sectional stiffness. Represents one-dimensional strain. It represents one-dimensional curvature.

8. The method for detecting structural damage of a floating wind turbine according to claim 6, characterized in that, The damage model established in step (2) is solved using the following simplified or exact methods; Simplified method: Assuming that the stiffness and mass loss caused by blade material damage is achieved through an approximate scaling criterion; by evaluating the ratio of the cross-sectional area of ​​the blade at the damage location after damage to that before damage. By assuming homogeneity of the segments, the stiffness after damage is obtained. and quality The matrix is ​​as follows: ； ； in, For shear stiffness in the swing direction, Shear stiffness in the oscillation direction For axial stiffness, For oscillation stiffness, To enhance swing stiffness, For torsional stiffness; , , and These are the unit segment mass of the blade, the moment of inertia in the flapping direction, the moment of inertia in the flaring direction, and the moment of inertia in the torsional direction, respectively. The ratio of the spanwise length of the damaged area along the blade to the length of the segment in which it is located; simultaneously, the mass scaling factor. Represented as: ； Precise Method: Based on classical laminated plate theory, the location and severity of damage to the composite blade material are first identified; then, the damaged segment of the blade is remodeled, and the damaged area is simulated by deleting material with corresponding thickness and circumferential length; next, the mass and stiffness matrices of the damaged segment and other segments of the blade are solved; finally, the stiffness and mass matrices of the reference blade model are replaced.

9. The method for detecting structural damage of a floating wind turbine according to claim 6, characterized in that, In step (4), a residual stiffness loss coefficient is proposed to address the stiffness loss. Define the mooring cable stiffness after damage for: ； in, Let be the tensile stiffness of the i-th mooring cable segment. The elastic modulus of the mooring cable material. The cross-sectional area of ​​a segment of the mooring cable material; based on this, the tension between two adjacent nodes of the mooring cable. for: ； in This is the critical length at which elastic force begins to be generated between the two nodes. Let be the distance between the two nodes at any given time.

10. A floating wind turbine structural damage detection system based on transfer learning, implemented using any one of claims 1-9, characterized in that, The floating wind turbine structural damage detection system includes: A scaled-down testing apparatus is constructed, comprising a scaled-down motion model of a floating wind turbine and a blade vibration data acquisition unit. The scaled-down motion model includes an environmental simulation device and a physical model testing device. A floating wind turbine model testing system is built using the scaled-down testing apparatus to monitor structural damage to the wind turbine. Model tests are conducted on both normal and damaged models using the scaled-down motion model and the blade vibration data acquisition unit, and structural damage monitoring data of the models are collected. The collected structural damage monitoring data is processed to obtain a processed scaled-down model test dataset. The pre-trained model building module constructs a source domain structural damage fault identification model based on deep learning, and automatically optimizes and trains the structural damage fault identification model based on a scaled-down model test dataset. The software simulation device establishes a fully coupled simulation model of the floating wind turbine through simulation software, and constructs a simulation method for typical structural damage of the floating wind turbine; it performs multi-condition simulation on the floating wind turbine to generate damage simulation data; and it fuses the structural damage and fault data obtained from actual wind turbine monitoring with the damage simulation data to obtain a target domain dataset. The transfer module uses a deep learning-based source domain structural damage fault identification model as a pre-trained model and performs transfer learning on the pre-trained model based on the target domain dataset to obtain a high-precision identification model. The deployment module deploys the high-precision identification model in a status monitoring system for real-world engineering projects, enabling real-time monitoring and fault warning of wind turbine status.

Images