Dynamic mode decomposition based dynamic analysis method for wind turbine drivetrain rigid-flexible coupled multi-body system

Abstract

The application belongs to the technical field of multi-body dynamics analysis, and discloses a wind driven generator transmission chain rigid-flexible coupling multi-body dynamics analysis method based on dynamic mode decomposition, which firstly embeds multi-degree-of-freedom time series data into a high-dimensional feature space through an encoder neural network; extracts dominant modes by using proper orthogonal decomposition (POD), and constructs a low-dimensional feature space; adopts parameterized dynamic mode decomposition and radial basis function regression to establish the mapping relationship between system parameters and Koopman operators, and realizes accurate prediction of dynamic characteristics under variable working conditions; and finally reconstructs physical responses through a decoder, and optimizes model parameters in combination with an error driving mechanism. The application effectively solves the problems of low calculation efficiency, poor nonlinear adaptability and difficult multi-parameter coupling prediction of traditional methods, significantly improves the efficiency and accuracy of transmission chain dynamics analysis, and provides reliable technical support for wind turbine condition monitoring and life prediction.

Description

Technical Field

[0001] This invention belongs to the field of multibody dynamics analysis technology, and particularly relates to a method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition. Background Technology

[0002] As a core component of complex electromechanical coupling systems, the precise analysis and prediction of the rigid-flexible coupling multibody dynamics of wind turbine drive trains is of great engineering value for improving the overall reliability of the machine, reducing component fatigue damage, and optimizing control strategies. In recent years, it has received widespread attention in the field of new energy equipment design and operation and maintenance.

[0003] Current research on the analysis of transmission chain dynamics largely relies on traditional finite element models or multibody dynamics simulation methods. These methods typically depend on high-resolution modeling and full-order numerical solutions, requiring the construction of large-scale nonlinear differential equation systems that include components such as gearboxes, main bearings, and couplings. This leads to an exponential increase in computational complexity with the number of degrees of freedom. Especially when considering time-varying wind loads, flexible deformation of the transmission chain, and multiphysics coupling effects, conventional numerical solvers face bottlenecks such as repetitive assembly of mass / stiffness matrices and high-frequency time-step integration, making it difficult to meet the requirements of real-time analysis and parameter optimization.

[0004] To address the computational efficiency challenge, existing research often employs model order reduction (MOR) techniques, achieving dimensionality compression by truncating low-energy modes. However, traditional POD-Galerkin-like methods rely on global projection bases for modal truncation. When faced with variations in multiple transmission chain parameters (such as variable gear meshing stiffness and dynamic damping ratio), the reconstruction error accumulates nonlinearly, failing to effectively maintain the predictability of dynamic characteristics. Furthermore, parameterized MOR models based on offline training typically use linear interpolation or low-order polynomial regression to construct parameter-modal relationships, making it difficult to capture the strongly nonlinear mapping between complex control systems and structural dynamics.

[0005] In recent years, data-driven dynamics analysis methods based on the Koopman operator have shown potential to solve the aforementioned problems. By embedding nonlinear systems into an infinite-dimensional linear space, they can achieve global linearization of nonlinear dynamics in a high-dimensional feature space. However, existing Koopman methods mostly rely on batch processing algorithms such as Dynamic Mode Decomposition (DMD), requiring independent computation of approximate operators for each parameter. This leads to multiple SVD decompositions and matrix inversion operations, resulting in a surge in computational costs in multi-parameter coupled analysis scenarios of transmission chains. Furthermore, the traditional DMD framework does not fully leverage the feature extraction capabilities of deep learning, making it difficult to effectively handle the complex motion mode aliasing phenomenon in flexible multibody systems, thus limiting its practical application in industrial equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a multibody dynamics analysis method for rigid-flexible coupling of wind turbine drive train based on dynamic mode decomposition, so as to solve the technical problems of multiple degrees of freedom of the model, low time-domain solution efficiency, and difficulty in extracting multi-parameter coupling features in the rigid-flexible coupling dynamics analysis of wind turbine drive train.

[0007] To address the aforementioned technical problems, the specific technical solution of the wind turbine drive train rigid-flexible coupling multibody dynamics analysis method based on dynamic mode decomposition of the present invention is as follows:

[0008] A multibody dynamics analysis method for rigid-flexible coupling of wind turbine drive train based on dynamic mode decomposition includes the following steps:

[0009] S1: Generate and extract raw multi-degree-of-freedom time series data based on measured data or existing high-fidelity dynamic models;

[0010] S2: Define two neural network structures, including an encoder and a decoder, and initialize the neural network parameters;

[0011] S3: The time series data is mapped to a high-dimensional feature space through the encoder neural network to form a high-dimensional system state dataset;

[0012] S4: Perform intrinsic orthogonal decomposition on the high-dimensional system state dataset to extract the dominant modes and generate a low-dimensional feature space;

[0013] S5: Pre-train the regression model on low-dimensional feature data under different parameter settings and optimize the parameters of the Koopman operator predictor.

[0014] S6: For different control parameters of the dynamic system, use the Koopman operator of the output response of the trained regressor to perform time-step iterative prediction of the current low-dimensional feature state.

[0015] S7: Perform an inverse transformation of intrinsic orthogonal decomposition on the converged low-dimensional feature prediction results to reconstruct the system state data in the high-dimensional feature space.

[0016] S8: The high-dimensional feature space system state data is back-mapped to the physical space through a decoder neural network to obtain the dynamic response;

[0017] S9: Determine the convergence of the prediction result based on the preset error threshold. If it does not converge, update the network model parameters and repeat S3 to S8.

[0018] S10: Based on the multi-degree-of-freedom time-series data output by S8, calculate the dynamic characteristics of key components of the transmission chain under different system parameters.

[0019] Furthermore, step S2 includes the following steps:

[0020] S2.1: Define the encoder-decoder network architecture. Based on the spatiotemporal coupling characteristics of state variables in the multi-degree-of-freedom physical space of the transmission chain system, construct a deep fully connected neural network to map the system motion state data in the physical space to a high-dimensional feature space.

[0021] S2.2: Network parameter initialization and regularization strategy. The Xavier initialization method is used to assign initial weights to the network. In the network training loss function, [the following is applied]: L2 The regularization constraint term and the neural network bias term are initialized to a zero vector.

[0022] Furthermore, step S4 includes the following steps:

[0023] S4.1: Construct a time-domain snapshot matrix by arranging the system states in the high-bit feature space output from the encoder into a snapshot matrix according to the time series.

[0024] S4.2: Calculate the covariance matrix and perform singular value decomposition;

[0025] S4.3: Determine the modal cutoff number and calculate the cumulative energy percentage to determine the cutoff order;

[0026] S4.4: Generate a low-dimensional feature space projection to project the high-dimensional feature data into the dominant mode basis vector space.

[0027] Furthermore, step S5 includes the following steps:

[0028] S5.1: Construction of parameterized dynamic modal dataset. For each system parameter configuration, the system state time series data corresponding to the system in the feature space is obtained through step S4, a shift data matrix is ​​constructed, and finally a parameter-modal dataset is formed.

[0029] S5.2: Dynamic mode decomposition solves the approximate Koopman operator, and executes the DMD core algorithm for the system state time series sequence matrix corresponding to each system parameter configuration;

[0030] S5.3: Parametric Koopman operator regression modeling, using radial basis function regression model to construct the mapping relationship between system parameter configuration and Koopman operator.

[0031] Furthermore, S5.2 specifically includes:

[0032] S5.2.1: Perform singular value decomposition on the shifted data matrix;

[0033] S5.2.2: Construct the Koopman operator matrix.

[0034] Furthermore, S5.3 specifically includes:

[0035] S5.3.1: Define the RBF kernel function;

[0036] S5.3.2: Calculate the RBF regression model;

[0037] S5.3.3: The objective function is set to Frobenius norm loss.

[0038] Furthermore, step S6 includes the following steps:

[0039] S6.1: Online parameter input and Koopman operator prediction. For real-time input system parameters, the RBF regression model trained in S5.3 is called to output the corresponding Koopman operator estimate.

[0040] S6.2: Dynamic iterative prediction in feature space, based on initial low-dimensional features, performs T-step time step calculation.

[0041] Furthermore, step S7 includes the following steps:

[0042] S7.1: Read the basis vectors and means of the dominant modes, and load the mode matrix and system feature space mean vector stored in S4;

[0043] S7.2: Perform linear inverse projection reconstruction to reconstruct the high-dimensional feature space based on the low-dimensional system state feature sequence.

[0044] Furthermore, step S9 includes the following steps:

[0045] S9.1: Construction of multi-objective loss function, design of composite loss function to simultaneously optimize network encoding-decoding accuracy, dynamic linearity and state prediction performance;

[0046] S9.2: Loss convergence threshold judgment, set the convergence judgment threshold and the maximum number of iterations, and calculate the composite loss residual ratio;

[0047] S9.3: Backpropagation optimization uses the Adam optimizer of the neural network to update the network parameters. Optimization is performed after the network training loss is calculated. First, the gradient is calculated, and then the momentum adaptive update strategy is executed.

[0048] The multibody dynamics analysis method for rigid-flexible coupling of wind turbine drive train based on dynamic mode decomposition of the present invention has the following advantages:

[0049] 1. Significantly improves computational efficiency:

[0050] By combining a deep encoder-decoder network with POD (Programmable Optimization) reduction technology, high-dimensional nonlinear dynamic systems are mapped to a low-dimensional feature space for linear evolution prediction, significantly reducing computational complexity. Compared to traditional finite element or multibody dynamics full-order simulations, this method significantly reduces computation time and resource consumption while maintaining accuracy, making it suitable for rapid parametric analysis and real-time state prediction.

[0051] 2. Enhanced nonlinear dynamic feature extraction capabilities:

[0052] By introducing deep neural networks for nonlinear embedding of the feature space, the limitations of traditional linear mode decomposition methods in flexible body motion reconstruction are effectively overcome. This allows for more accurate capture of the complex dynamic behavior of transmission chain systems under varying operating conditions, such as nonlinear responses caused by changes in gear meshing stiffness and time-varying damping.

[0053] 3. Improve the adaptability of multi-parameter coupling:

[0054] By combining the parametric Koopman operator with the RBF regression model, a nonlinear mapping relationship between system parameters and dynamic characteristics is constructed, which enhances the model's adaptability to different operating parameters (such as wind speed and load) and avoids the mode mismatch problem that occurs when parameters change in traditional reduced-order models.

[0055] 4. Achieve efficient end-to-end prediction and reconstruction:

[0056] By combining the encoder-decoder structure with the forward and inverse POD transformation, a two-way high-precision mapping from physical space to feature space and back to physical space is achieved. It can quickly generate full-cycle dynamic responses under various parameter combinations after a single training, supporting the rapid design and optimization of transmission chain systems.

[0057] 5. Possesses promising prospects for engineering applications:

[0058] This method provides strong technical support for real-time status monitoring, fatigue life prediction, fault diagnosis and variable operating condition control of wind turbine drive trains. It is especially suitable for multibody dynamics analysis and intelligent operation and maintenance scenarios of large wind turbine units, and has good practicality and promotion value.

[0059] In summary, this invention not only achieves innovation in theoretical methods, but also demonstrates significant advantages in engineering practice, namely high efficiency, accuracy, and robustness, providing a new solution for the dynamic modeling and analysis of complex electromechanical systems. Attached Figure Description

[0060] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0061] Figure 2aA schematic diagram of multibody dynamics modeling for a rigid-flexible coupled transmission chain in an implementation case.

[0062] Figure 2b A schematic diagram of the finite element mesh along the principal axis of the flexible body;

[0063] Figure 3a This is a comparison diagram of the absolute values ​​of the nodal strains of the main axis of the flexible body predicted by the method proposed in this invention and PDMD.

[0064] Figure 3b This diagram illustrates the relationship between the prediction error and prediction time of the method proposed in this invention and PDMD. Detailed Implementation

[0065] To better understand the purpose, structure, and function of this invention, the following detailed description, in conjunction with the accompanying drawings, provides a more comprehensive understanding of the rigid-flexible coupling multibody dynamics analysis method for wind turbine drive train based on dynamic mode decomposition.

[0066] This invention proposes a novel multibody dynamics analysis method for rigid-flexible coupled wind turbine drivetrains based on deep Koopman feature learning. It achieves nonlinear embedding of multi-degree-of-freedom time-series data into a high-dimensional feature space by constructing a deep coding network. Dominant modes are extracted using intrinsic orthogonal decomposition (POD) to establish a low-dimensional quasi-steady-state feature space. A parameterized Koopman operator predictor is established, and a pre-trained regression model is used to achieve online mapping and association between parameters under different operating conditions and dynamic operators. Reconstruction and back-projection algorithms complete high-precision full-order restoration of the low-dimensional feature data. A bidirectional nonlinear mapping channel between the feature space and the physical space is established in conjunction with a deep decoding network. This method effectively solves the challenge of efficient prediction of cross-scale dynamic evolution in rigid-flexible coupled systems through end-to-end feature evolution modeling.

[0067] like Figure 1 As shown, the method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition of the present invention includes the following steps:

[0068] S1: Generate and extract raw multi-degree-of-freedom time series data based on measured data or existing high-fidelity dynamic models.

[0069] S2: Define two neural network structures, including an encoder and a decoder, and initialize the neural network parameters.

[0070] S2.1: Define the encoder-decoder network architecture based on the spatiotemporal coupling characteristics of state variables (displacement, velocity, and acceleration fields) in the multi-degree-of-freedom physical space of the transmission chain system. (a joint distribution of degrees of freedom) to construct a deep fully connected neural network, mapping the system motion state data in physical space to a high-dimensional feature space. The encoder network structure is defined as follows:

[0071]

[0072] in, Represents the encoder network function. It represents the system state vector at any time in physical space, which can be summarized as the system degrees of freedom in multibody dynamics; Let be the trainable parameters of the encoder neural network, denoted as . ; Indicates the number of encoder network layers. These represent the encoder network's first and second generations, respectively. The network weight parameters and biases of each layer; Indicates the first The activation function of the layer, L Indicates the number of network layers.

[0073] The decoder network is symmetrical to the encoder and can be represented as follows:

[0074]

[0075] in, Indicates the decoder network function, This represents the system state vector at any time within the high-order feature space; Let be the trainable parameters of the decoder neural network, denoted as . , Indicates the number of layers in the decoder network. These represent the decoder network's... The network weight parameters and biases of the layers.

[0076] S2.2: Network parameter initialization and regularization strategy. The Xavier initialization method is used to assign initial weights to the network, and its mathematical expression is:

[0077]

[0078] in, Indicates a uniform random distribution. and The first The input and output dimensions of a layered network. In the network training loss function, the following is applied... L2 Regularization constraint terms The expression is:

[0079]

[0080] in, This is a hyperparameter, and in the experiment it takes the value of [value missing]. ; It is the Frobenius norm.

[0081] Neural network bias terms and Initialize to a zero vector.

[0082] S3: The time series data is mapped to a high-dimensional feature space through the encoder neural network to form a high-dimensional system state dataset;

[0083] S4: Perform intrinsic orthogonal decomposition (POD) on the high-dimensional system state dataset to extract the dominant modes and generate a low-dimensional feature space;

[0084] S4.1: Construct a temporal snapshot matrix by arranging the system states in the high-bit feature space output from the encoder into a snapshot matrix according to the time series. ,in The dimension of the system state variables in the high-dimensional feature space. The column vectors of the time sampling point matrix correspond to the feature space states:

[0085]

[0086] S4.2: Calculate the covariance matrix and perform singular value decomposition. The covariance matrix can be represented as:

[0087]

[0088] in and It is an orthogonal matrix. Given a singular value diagonal matrix arranged in descending order, to achieve order reduction, we need to extract... The dominant mode matrix of order , can be represented as:

[0089]

[0090] S4.3: Determine the modal cutoff number, assuming an energy threshold. Calculate the cumulative energy percentage to determine the cutoff order. :

[0091]

[0092] in For the first A singular value, accumulated through a loop. ,until .

[0093] S4.4: Generate a low-dimensional feature space projection, transforming high-dimensional feature data... Projected onto the dominant mode basis vector space, it is represented as: The expression is as follows:

[0094]

[0095] S5: Pre-train the regression model on low-dimensional feature data under different parameter settings and optimize the parameters of the Koopman operator predictor.

[0096] S5.1: Construction of parameterized dynamic modal datasets, for each system parameter configuration Step S4 obtains the system state time series data corresponding to the system in the feature space. Construct a shifted data matrix and The expression is as follows:

[0097]

[0098] Final constituent parameters - modal dataset ,in The total number of parameter samples.

[0099] S5.2: Dynamic mode decomposition to solve for the approximate Koopman operator, for each system parameter configuration. The corresponding system state time series matrix executes the DMD core algorithm, specifically including:

[0100] S5.2.1: For the shifted data matrix Perform Singular Value Decomposition (SVD):

[0101]

[0102] S5.2.2: Constructing the Koopman operator matrix:

[0103] .

[0104] S5.3: Parametric Koopman operator regression modeling, using radial basis function (RBF) regression model to construct system parameter configuration. To the Koopman operator The mapping relationships specifically include:

[0105] S5.3.1: Define the RBF kernel function:

[0106]

[0107] in The center point of the uniform distribution of parameters in space. For bandwidth function, The number of basis functions.

[0108] S5.3.2: The regression model expression is:

[0109]

[0110] in For coefficient tensors, This is the bias matrix.

[0111] S5.3.3: The objective function is set to Frobenius norm loss:

[0112]

[0113] in This is the regularization coefficient.

[0114] S6: For different control parameters of the dynamic system, use the Koopman operator of the output response of the trained regressor to perform time-step iterative prediction of the current low-dimensional feature state.

[0115] S6.1: Online parameter input and Koopman operator prediction for real-time input system parameters Call the RBF regression model trained in S5.3 and output the corresponding Koopman operator estimate. ;

[0116] S6.2: Dynamic iterative prediction in feature space, based on initial low-dimensional features ,implement Step time step calculation:

[0117]

[0118] S7: Perform the inverse POD transformation on the converged low-dimensional feature prediction results to reconstruct them into high-dimensional feature space system state data;

[0119] S7.1: Read the basis vectors and mean of the dominant modes, and load the mode matrix stored in S4. and the mean vector of the system feature space ,

[0120] S7.2: Perform linear inverse projection reconstruction based on low-dimensional system state feature sequences. Reconstructing the high-dimensional feature space:

[0121]

[0122] S8: The high-dimensional feature space system state data is back-mapped to the physical space through a decoder neural network to obtain the dynamic response;

[0123] S9: Determine the convergence of the prediction result based on the preset error threshold. If it does not converge, update the network model parameters and repeat S3 to S8.

[0124] S9.1: Construction of multi-objective loss function. A composite loss function is designed to simultaneously optimize network encoding-decoding accuracy, dynamic linearity, and state prediction performance. The expression of the loss function is as follows:

[0125]

[0126] in , and These are the reconstruction loss, linear dynamic loss, and prediction loss, respectively, and their expressions are as follows:

[0127]

[0128] in , and These are the weighting coefficients. For regularization strength, and These represent the computation processes of the encoder network and the decoder network, respectively. This represents the mean square error.

[0129] S9.2: Loss convergence threshold determination, setting a convergence determination threshold. and the maximum number of iterations Calculate the composite loss residual ratio:

[0130]

[0131] when satisfy or At that time, it was assumed that the network training had converged, and during testing, settings were made... ,set up .

[0132] S9.3: Backpropagation optimization uses the Adam optimizer of the neural network to update the network parameters. Optimization is performed after the loss is calculated during network training. First, the gradient is calculated:

[0133]

[0134] Then execute the momentum adaptive update strategy:

[0135]

[0136] in , It is the exponentially decaying momentum term. The learning rate for network training is set to [value] during testing. , To avoid numerical overflow caused by the denominator being close to zero, a smoothing term is usually set. .

[0137] S10: Based on the multi-degree-of-freedom time-series data output by S8, calculate the dynamic characteristics of key components of the transmission chain under different system parameters, such as displacement, rotational speed, stress, and strain.

[0138] Implementation Case:

[0139] To verify the effectiveness of the method of this invention, this example uses a rigid-flexible coupled multibody dynamics model of a real in-service wind turbine three-stage planetary gearbox transmission chain system established using Simpack. The method proposed in this invention is then compared with the parametric dynamic mode decomposition method (DMD for parametric dynamic systems, PDMD) proposed by Andreuzzi et al. (Andreuzzi F, Demo N, Rozza G. A dynamic mode decomposition extension for the forecasting of parametric dynamical systems[J]. SIAMJournal on Applied Dynamical Systems, 2023, 22(3): 2432-2458.).

[0140] like Figure 2a Figure 2b As shown, in this test case, the operating condition for the multibody dynamics simulation is set to steady laminar wind, and the wind speed range is set to [range missing]. The strain data of 4000 nodes b on the flexible body principal axis under different working conditions within 50 seconds were extracted as the initial time series data. Wind speed was set as the input to the system and model. Nineteen sets of data were selected for model training under wind speed conditions ranging from 3 to 21 m / s, predicting the strain of the flexible body principal axis at a wind speed of 11.5 m / s. The network training used the Adam optimizer with an initial learning rate of 1×10⁻⁶. -3 .

[0141] The strain results of the flexible principal axis nodes of the multibody dynamics system solved using Simpack under steady-state laminar wind conditions at a wind speed of 11.5 m / s were used as the benchmark solution. The results obtained using the method of this invention were compared with the strain results of the flexible principal axis nodes solved using the method proposed by F Andreuzzi et al. The results are as follows: Figure 3a , Figure 3b As shown.

[0142] Under the same initial conditions and test cases, the method of this invention significantly improves the accuracy of solving for the displacement of the flexible body's principal axis under different working conditions compared to the parametric dynamic mode decomposition method proposed by F. Andreuzzi et al., demonstrating that the method of this invention can greatly improve the efficiency and accuracy of rigid-flexible coupled multibody dynamics analysis. This is mainly because this method introduces a neural network-based autoencoder, which can better capture the dynamic evolution characteristics of complex multibody dynamics systems in the original physical space. Simultaneously, the introduced parametric Koopman operator based on the RBF kernel function regression model can dynamically capture dynamic modes under different system state parameters, thereby achieving efficient and accurate analysis of the dynamic characteristics of rigid-flexible coupled multibody dynamics systems.

[0143] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition, characterized in that, Includes the following steps: S1: Generate and extract raw multi-degree-of-freedom time series data based on measured data or existing high-fidelity dynamic models; S2: Define two neural network structures, including an encoder and a decoder, and initialize the neural network parameters; S3: The time series data is mapped to a high-dimensional feature space through the encoder neural network to form a high-dimensional system state dataset; S4: Perform intrinsic orthogonal decomposition on the high-dimensional system state dataset to extract the dominant modes and generate a low-dimensional feature space; S5: Pre-train the regression model on low-dimensional feature data under different parameter settings and optimize the parameters of the Koopman operator predictor. S5.1: Construction of parameterized dynamic modal dataset. For each system parameter configuration, the system state time series data corresponding to the system in the feature space is obtained through step S4, a shift data matrix is ​​constructed, and finally a parameter-modal dataset is formed. S5.2: Dynamic mode decomposition solves the approximate Koopman operator, and executes the DMD core algorithm for the system state time series sequence matrix corresponding to each system parameter configuration; S5.3: Parametric Koopman operator regression modeling, using radial basis function regression model to construct the mapping relationship between system parameter configuration and Koopman operator; S6: For different control parameters of the dynamic system, use the Koopman operator of the output response of the trained regressor to perform time-step iterative prediction of the current low-dimensional feature state. S7: Perform an inverse transformation of intrinsic orthogonal decomposition on the converged low-dimensional feature prediction results to reconstruct the system state data in the high-dimensional feature space. S8: The high-dimensional feature space system state data is back-mapped to the physical space through a decoder neural network to obtain the dynamic response; S9: Determine the convergence of the prediction result based on the preset error threshold. If it does not converge, update the network model parameters and repeat S3 to S8. S10: Based on the multi-degree-of-freedom time-series data output by S8, calculate the dynamic characteristics of key components of the transmission chain under different system parameters.

2. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, S2 includes the following steps: S2.1: Define the encoder-decoder network architecture. Based on the spatiotemporal coupling characteristics of state variables in the multi-degree-of-freedom physical space of the transmission chain system, construct a deep fully connected neural network to map the system motion state data in the physical space to a high-dimensional feature space. S2.2: Network parameter initialization and regularization strategy. The Xavier initialization method is used to assign initial weights to the network. In the network training loss function, [the following is applied]: L2 The regularization constraint term and the neural network bias term are initialized to a zero vector.

3. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, S4 includes the following steps: S4.1: Construct a time-domain snapshot matrix by arranging the system states in the high-bit feature space output from the encoder into a snapshot matrix according to the time series. S4.2: Calculate the covariance matrix and perform singular value decomposition; S4.3: Determine the modal cutoff number and calculate the cumulative energy percentage to determine the cutoff order; S4.4: Generate a low-dimensional feature space projection to project the high-dimensional feature data into the dominant mode basis vector space.

4. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, Specifically, S5.2 includes: S5.2.1: Perform singular value decomposition on the shifted data matrix; S5.2.2: Construct the Koopman operator matrix.

5. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, Specifically, S5.3 includes: S5.3.1: Define the RBF kernel function; S5.3.2: Calculate the RBF regression model; S5.3.3: The objective function is set to Frobenius norm loss.

6. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 5, characterized in that, S6 includes the following steps: S6.1: Online parameter input and Koopman operator prediction. For real-time input system parameters, the RBF regression model trained in S5.3 is called to output the corresponding Koopman operator estimate. S6.2: Dynamic iterative prediction in the feature space, based on initial low-dimensional features, performs... Step-time step calculation.

7. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, S7 includes the following steps: S7.1: Read the basis vectors and means of the dominant modes, and load the mode matrix and system feature space mean vector stored in S4; S7.2: Perform linear inverse projection reconstruction to reconstruct the high-dimensional feature space based on the low-dimensional system state feature sequence.

8. The method for rigid-flexible coupling multibody dynamics analysis of wind turbine drive train based on dynamic mode decomposition according to claim 1, characterized in that, S9 includes the following steps: S9.1: Construction of multi-objective loss function, design of composite loss function to simultaneously optimize network encoding-decoding accuracy, dynamic linearity and state prediction performance; S9.2: Loss convergence threshold judgment, set the convergence judgment threshold and the maximum number of iterations, and calculate the composite loss residual ratio; S9.3: Backpropagation optimization uses the Adam optimizer of the neural network to update the network parameters. Optimization is performed after the network training loss is calculated. First, the gradient is calculated, and then the momentum adaptive update strategy is executed.

Images