Wave-induced seabed response prediction method around pile foundation based on deep learning time series model

Abstract

The present application belongs to the technical field of ocean engineering, ocean geotechnical engineering and artificial intelligence, and particularly relates to a wave-induced pile foundation surrounding seabed response prediction method based on a deep learning time series model, which comprises generating multi-point time series physical field data of multi-directional wave and seabed dynamic response through CFD-FEM coupling simulation, and processing the data through random spatial sampling; using multi-order difference to enhance the nonlinear characteristics of waves, realizing time-space dynamic feature extraction through TimeDistributed-LSTM; constructing a four-dimensional space-time tensor based on relative coordinates and Euclidean distance, and completing local-global feature aggregation through local multilayer perceptron and maximum pooling; and finally predicting the seabed pore water pressure response by a multi-output decoding network in parallel. The present application is free from fixed grid and measurement point dependence, has strong spatial adaptability and generalization ability, and can realize real-time and high-precision prediction of the dynamic response of the seabed surrounding the pile foundation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of marine engineering, marine geotechnical engineering, and artificial intelligence, and specifically relates to a method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model. Background Technology

[0002] With the rapid growth of global offshore wind power installed capacity, monopile foundations, as the most commonly used support structure, are widely used in shallow and near-shore wind farms. In the real marine environment, wave propagation not only exhibits random variations in frequency and amplitude but also displays significant multidirectional distribution characteristics (i.e., multidirectional irregular waves). Waves around monopile foundations cause spatiotemporally varying hydrodynamic pressure fluctuations, which are further transmitted to the seabed, triggering dynamic responses such as pore water pressure oscillations, accumulation, and potential liquefaction. This can lead to rapid local settlement of the seabed, pile tilting, or even instability and failure.

[0003] For predicting seabed response to wave-structure-seabed interaction (WSSI), existing technical solutions can be mainly divided into the following three categories: (i) Traditional fluid-structure interaction numerical simulation methods employ a high-resolution numerical model that couples computational fluid dynamics (CFD) and finite element method (FEM). In the fluid domain, three-dimensional incompressible RANS equations combined with the VOF method are used to capture the free surface and calculate wave dynamics and the complex flow field structure around the pile foundation. In the solid domain of the seabed, a finite element framework based on Biot consolidation theory is used for solution. The two are coupled at the interface to transmit dynamic water pressure and response data in real time. However, this method has extremely high computational cost and cannot meet the needs of multi-scenario, ultra-real-time safety early warning during the operation and maintenance period of engineering projects.

[0004] (ii) Existing data-driven or machine learning models extract historical time-series data from a limited number of measurement points and input them into conventional machine learning models for training to establish a mapping relationship between wave parameters and seabed pore water pressure. However, this method generally relies heavily on fixed computational grids or sensor topologies at specific spatial locations, and cannot adapt to the random changes of on-site monitoring points. Moreover, conventional time-series prediction models often process time and spatial information separately, lacking the ability to accurately capture multi-directional wave spatial interference attenuation mechanisms and extremely steep nonlinear wave peaks. Furthermore, when processing massive discrete spatial measurement points, they are prone to memory and computing power overflow.

[0005] (III) Point cloud / graph neural network (GNN) based on fixed topology separates and processes temporal and spatial information. It relies on fixed monitoring points or artificially predefined adjacency matrices (fixed topology) to aggregate spatial features. However, this method is difficult to characterize the energy transfer path and phase interference relationship that change dynamically with time during multi-directional wave propagation. It has poor adaptability to changes in the spatial distribution of measurement points and has high computational complexity under large-scale discrete measurement point conditions. At the same time, it is prone to feature over-smoothing, which weakens the ability to express local extreme value responses.

[0006] Therefore, breaking away from the dependence of existing models on fixed topology and developing an accurate, fast, and spatially adaptive spatiotemporal prediction method to achieve full-domain high-fidelity early warning of the dynamic response of the seabed around a monopile foundation under complex multi-directional wave conditions is of great theoretical and engineering significance for the safety assessment, optimization design, and full life-cycle management of offshore wind power foundations. Summary of the Invention

[0007] This invention overcomes the above-mentioned defects and provides a method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model. It solves the problems of the existing technology, such as excessive reliance on fixed grid topology, weak perception of sudden changes in random wave waveforms, insufficient spatiotemporal separation, high computational cost, and low prediction accuracy for complex sea conditions.

[0008] To achieve the above objectives, the present invention provides a method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model, comprising the following steps: S1. By simulating the response of the coupled wave fluid domain and the seabed solid domain using CFD-FEM, multi-point time-series physical field data with spatial coordinates are generated for the multi-directional wave surface elevation and seabed pore water pressure response. S2. Randomly sample the multi-point time-series physical field data to obtain an unstructured point cloud dataset. The dataset includes the three-dimensional absolute coordinates of wave measurement points and the corresponding wave height sequence data, the three-dimensional absolute coordinates of target prediction points and the corresponding seabed pore water pressure time-series data. S3 constructs a time-series prediction model of the seabed response around wave-induced pile foundations. The model performs the following operations: S31 performs multi-order difference feature enhancement on the wave measurement point height sequence data and performs global standardization preprocessing to obtain the wave spatiotemporal input tensor. S32 performs high-dimensional spatial mapping on the three-dimensional absolute coordinates of wave measurement points and concatenates them with the wave spatiotemporal input tensor to construct the TimeDistributed-LSTM module to independently extract waveform change feature vectors from multiple wave measurement points. S33 calculates the relative coordinates and Euclidean distance between the target prediction point and the wave measurement point, and concatenates them with the waveform change feature vector, the three-dimensional absolute coordinates of the wave measurement point, and the three-dimensional absolute coordinates of the target prediction point to form a four-dimensional feature tensor with multi-dimensional spatiotemporal physical characteristics. S34 performs local multilayer perceptron and global max pooling on the four-dimensional feature tensor along the wave measurement point dimension to extract global wave energy evolution features; S35 stitches the global wave energy evolution characteristics with the three-dimensional absolute coordinates of the target prediction point, and outputs the dynamic response prediction values ​​of multiple points on the seabed in parallel by the multi-output decoding network to complete the model training and verification. S4. Input the elevation time series data of the wave measuring points deployed on site and the coordinates of the target prediction points into the trained model to directly obtain the dynamic response curves of the seabed at multiple points around the pile foundation.

[0009] Further, in step S2, the unstructured point cloud dataset is obtained as follows: S21 Extracts the boundary extent of the wave fluid domain mesh in three-dimensional space. , , And generate within the range The three-dimensional absolute coordinates of a random wave measuring point that follows a uniform distribution; S22 calculates the spatial correlation weights between the three-dimensional absolute coordinates and the original mesh nodes based on the radial basis spatial interpolation algorithm; S23 maps the time-series data of the original grid nodes to the three-dimensional absolute coordinates of random wave measurement points based on the spatial correlation weights, thereby obtaining the time-series sequence of each wave measurement point coordinate point over the entire time step, and thus obtaining... The three-dimensional absolute coordinates of randomly distributed wave measurement points and their corresponding wave elevation time series are used as input data. S24 generates a grid in the seabed solid phase domain in the same manner as steps S21-S23. The output data consists of the three-dimensional absolute coordinates of randomly distributed seabed target points and their corresponding pore water pressure time-series labels.

[0010] Furthermore, in step S31, the multi-order difference feature enhancement is specifically implemented as follows: Let the first i The length of the high-order sequence data for each wave measurement point is [length missing]. T , is represented as: ,in, Indicates the measuring point i At time step t The wave surface elevation value, a. Decomposing the first-order difference sequence: ; b. Decompose the second-order difference sequence: ; c. Pad the difference sequence length to the original length T by extending interpolation; d. After concatenating the original sequence, the first-order difference sequence, and the second-order difference sequence along the channel dimension and standardizing them, we obtain the sequence containing... A wave spatiotemporal input tensor with 3 wave measurement points, a time step of T, and 3 feature channels.

[0011] Further, in step S32, the waveform change feature vector is extracted as follows: S321 maps the three-dimensional absolute coordinates of the wave measurement points into high-dimensional spatial feature vectors through a three-layer fully connected neural network. S322 performs a copy and broadcast operation along the time axis on the high-dimensional spatial feature vector and concatenates it with the wave spatiotemporal input tensor to form a four-dimensional wave spatiotemporal input tensor. The S323 TimeDistributed-LSTM module performs a slice traversal on the four-dimensional wave spatiotemporal input tensor, decomposing it into... Each independent temporal feature slice provides an independent encoding path for each wave measurement point; The S324 LSTM processes individual feature slices sequentially according to time steps, gradually processing wave dynamics signals to accurately capture the dynamic characteristics of each wave measurement point evolving over time. For each wave measurement point, the S325 uses the LSTM gating mechanism at the end of the sequence processing to filter and retain information, compressing the spatiotemporal sequence information of length T into a fixed-length compressed vector. The S326 TimeDistributed-LSTM module will... The compressed vectors generated independently at each wave measurement point are reassembled along the original wave measurement point dimensions to output the waveform change feature vector of multiple wave measurement points. .

[0012] Furthermore, all the independent encoding paths share the same set of network weight parameters.

[0013] Furthermore, the four-dimensional feature tensor is represented as: , in, Waveform change feature vector Temporal embedding vectors are obtained through dimensional expansion and duplication; This is the absolute coordinate vector obtained by copying the wave measurement point along the dimension of the target prediction point; Relative coordinates; The distance is Euclidean. This represents a vector concatenation operation.

[0014] Furthermore, the extraction of the global wave energy evolution features is specifically as follows: S341 inputs the four-dimensional feature tensor into a layer-by-layer expanding local multilayer perceptron to mine the nonlinear features of directional differences, energy transfer patterns, and propagation paths between a single wave measurement point and the target seabed prediction point, thus obtaining local features: , in, Indicates local features, Represents a four-dimensional feature tensor. This represents a local multilayer perceptron. Indicates the target prediction point; S342 performs max pooling on the local features along the wave measurement point dimension to obtain global wave energy evolution features: , in, This represents the global wave energy evolution characteristics. This indicates max pooling. Indicates the wave measurement point.

[0015] Furthermore, in step S35, the decoding network consists of multiple fully connected layers, performs nonlinear mapping on the spliced ​​features in the dimension of the target prediction points, and reduces the dimension to restore the output space of the seabed pore water pressure time series label, and outputs the dynamic response prediction values ​​of all target prediction points in parallel.

[0016] Furthermore, the distribution range of the three-dimensional absolute coordinates of the random wave measuring points is constrained to... , , Within this range, it simultaneously encompasses the incident wave region upstream of the monopile structure, the diffraction and reflection region around the structure, and the wave attenuation region downstream.

[0017] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: 1. This invention introduces an unstructured spatiotemporal point cloud representation mechanism and relative geometric coordinate encoding. By calculating the relative coordinates and Euclidean distance between the wave input point and the target prediction point, the physical energy attenuation of the wave is perceived, which is no longer limited by the absolute position of the sensor. This completely breaks the dependence on fixed grids and has strong spatial generalization and field deployment adaptive capabilities. 2. This invention constructs a TimeDistributed-LSTM network. This architecture forces all discrete wave measurement points to completely share the same set of network weight parameters when performing independent feature encoding. This not only ensures from a physical mechanism that the evolutionary features of different spatial nodes do not experience non-physical crosstalk, but also simplifies the massive spatial dimension calculation, completely solves the problem of dimension and memory explosion, and achieves efficient parallel processing of massive point clouds. 3. This invention utilizes a local multilayer perceptron to extract the influence law of a single wave propagation path point by point. Then, by utilizing the arrangement invariance of the maximum pooling layer, it autonomously selects the core wave crest propagation path that contributes the most to the response of the current seabed target point from the disordered massive wave point cloud, thereby realizing local-global feature aggregation and capturing the strong spatiotemporal coupling and diffraction mechanism of multi-directional waves. 4. In the preprocessing stage, this invention explicitly extracts the first-order difference representing the transient deformation rate and the second-order difference representing the waveform curvature and local acceleration from the original wave height sequence. By introducing these physical higher-order derivatives that reflect extreme nonlinear kinematics as new feature channels into the network, the model is forced to pay attention to the abrupt changes in the wave, which significantly improves the transient perception accuracy of extreme nonlinear waves. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall prediction method of the present invention; Figure 2 This invention provides a time-series prediction model framework for the wave-induced seabed response around the pile foundation. Figure 3 This is a schematic diagram of the boundary condition settings for the CFD model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the CFD-FEM model coupling calculation in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing different input point cloud quantities in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the training iteration loss performance of the model with different input point cloud quantities in an embodiment of the present invention; Figure 7 This is an embodiment of the present invention. A comparison of the numerical simulation results of the seabed pore pressure field and the prediction results of the model of the present invention is shown in the figure, where (a) is a schematic diagram of the numerical simulation results of the seabed pore pressure field and (b) is a schematic diagram of the prediction results of the seabed pore pressure field by the model of the present invention. Figure 8 This is an embodiment of the present invention. A comparison of the numerical simulation results of the seabed shear stress field around the pile foundation and the prediction results of the model of the present invention is shown in the figure. (a) is a schematic diagram of the numerical simulation results of the seabed shear stress field, and (b) is a schematic diagram of the prediction results of the seabed shear stress field by the model of the present invention. Figure 9 This is a comparison chart showing the changes in prediction error indices under different multidirectional wave concentration S conditions. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model provided by this invention is as follows: I. Multiphysics Coupled Simulation and Dataset Construction This invention employs a coupled model of fluid dynamics (CFD) and finite element method (FEM) to simulate the interaction between multidirectional waves and monopile foundations and the seabed under different sea conditions, deriving free surface time-series data in the fluid domain and seabed pore water pressure / stress response data in the solid domain. Specifically: (1) Set input parameters: Input parameters include basic marine dynamic environmental parameters (including significant wave height, wave spectrum such as JONSWAP spectrum, directional concentration parameter s, etc.), as well as the three-dimensional geometric dimensions and physical and mechanical parameters of the monopile foundation and seabed soil (including soil density, porosity, shear modulus, Poisson's ratio, permeability coefficient, etc.). (2) CFD-FEM three-dimensional multiphysics coupling solution: a. Fluid Domain Wave Dynamics Simulation (CFD Stage): Establishing a three-dimensional multiphase fluid computational domain ( Figure 3 At the inlet boundary, segmented motion boundary wave generation technology is used to inject the wave energy of multi-directional irregular waves or multi-directional focused waves into the fluid domain according to different propagation directions; the three-dimensional incompressible RANS equations combined with the VOF method are used to capture the transient changes of the free liquid surface and calculate the propagation, diffraction and local dynamic water pressure evolution of waves around the monopile foundation. b. Solid phase domain seabed response simulation (FEM stage): Establish the seabed-monopile finite element solid phase calculation domain, and use the finite element framework based on Biot consolidation theory to solve the dynamic response of the porous media seabed; c. Fluid-Solid Interface Coupled Loading: The fluid domain (CFD) and solid domain (FEM) interact through a physical interface. Within a set computation time step, the local seafloor hydrodynamic pressure calculated by the CFD model is used as a dynamic load, which is transmitted in real time and applied to the non-permeable foundation boundary and mud surface of the FEM model based on interpolation, for fluid-solid coupled joint solution. Figure 4 ).

[0021] (3) Generation of high-dimensional multi-field spatiotemporal datasets: After completing the simulation over the set time period, two types of physical field data are simultaneously exported from the calculation results: Output the three-dimensional spatial coordinate set (x, y, z) of the discrete mesh nodes of the entire wave surface from the wave fluid domain, and the historical time series of the free surface wave height corresponding to each node. ; Output the three-dimensional spatial coordinate set (x, y, z) of specified grid nodes within the seabed solid domain, along with the corresponding pore water pressure time series data for each node. and shear stress time series data; The multi-point temporal physical field data with precise spatial coordinates exported above together constitute the basic dataset for wave-pillar-seabed coupling in subsequent neural network training, verification and testing.

[0022] This invention provides a method for obtaining rigorous basic training data with a solid physical mechanism through CFD-FEM three-dimensional multiphysics coupling, and removes the dependence of deep learning models on fixed network topology.

[0023] II. Random Spatial Sampling of Unstructured Point Clouds Random space sampling is performed on the computational domain and computational results in step one. Figure 5 Randomly sampled from the wave fluid domain grid. Several points are used as input wave point clouds, and at the same time, points are randomly selected from the seabed grid. Each point is used as the target prediction point cloud; to ensure optimal spatial generalization and prediction accuracy, the wave input point cloud spatially needs to simultaneously cover the incident wave region upstream of the monopile structure, the diffraction / reflection region around the structure, and the wave attenuation region downstream. The specific process is as follows: (1) For the dataset generated in step one, extract the boundary range of the original wave fluid domain mesh in three-dimensional space. , , Generate within this continuous coordinate space A set of random three-dimensional coordinates follows a uniform distribution, the distribution of which is explicitly constrained to simultaneously cover the incident wave region upstream of the monopile structure, the diffraction / reflection region around the structure, and the wave attenuation region downstream; similarly, the same scheme is used to randomly generate coordinates within the seabed region. Coordinates of the predicted target points; (2) For the generated random free coordinate set, the spatial correlation weight between the random free coordinate and the original grid node is calculated based on the radial basis spatial interpolation algorithm; the time series data of the original grid node is mapped to the generated random coordinate point based on the weight matrix, so as to accurately calculate the physical quantity evolution sequence of each random coordinate point at the full time step; (3) After completing the above operations, an unstructured point cloud dataset completely detached from the original finite element mesh topology is obtained, including: The three-dimensional absolute coordinates of a randomly distributed wave measurement point and its corresponding wave elevation time series; The three-dimensional absolute coordinates of a randomly distributed seabed target point, and its corresponding pore water pressure time-series label.

[0024] The extracted point cloud data will be used as independent and discrete spatial physical nodes and fed into the subsequent network for spatiotemporal feature encoding.

[0025] The random recombination sampling mechanism used in this embodiment of the invention cuts off the model's dependence on a specific computational grid or fixed sensor topology, giving the model the ability to adapt to sensor layouts in any spatial location.

[0026] III. Constructing a time-series prediction model for the seabed response around wave-induced pile foundations, such as... Figure 2 As shown, the model performs the following operations: 1. Multi-level feature enhancement and normalization preprocessing (1) Targeting The historical time series of wave surface elevations at each wave monitoring point expands the original free surface wave height time series of each monitoring point into three physical components: the original elevation signal, a first-order difference sequence (characterizing the rate of change of water surface elevation and the degree of waveform deformation), and a second-order difference sequence (characterizing waveform curvature), as detailed below: Let the first i The original wave height program sequence length for each measuring point is: T , represented as a vector ,in Indicates the measuring point i At time step t The wave surface elevation value, a. The first-order difference sequence is: ; b. The second-order difference sequence is: ; c. Pad the length of the difference sequence to the original length T by using an extended interpolation method, while maintaining the same time dimension length as the original sequence. d. Concatenate the original sequence, the first-order difference sequence, and the second-order difference sequence along the channel dimension to expand the single-point feature from one dimension to three dimensions; (2) The above-mentioned concatenated input features and The pore water pressure time-series labels of each seabed target point are preprocessed using a globally standardized Z-score to preserve the physical distribution differences of extreme waves. The Z-score standardization formula is as follows: , in, x These are the original data points. i It is the first in the specified column i One point, μ It is the mean of the data column. σ It is the standard deviation of the data column. z These are standardized data points.

[0027] (3) Calculate the contents The wave spatiotemporal input tensor with 3 measurement points, a time step of T, and 3 feature channels (original, first-order, and second-order), as well as the time series tensor of the target point pore pressure, including the standardized tensor.

[0028] The embodiments of the present invention scale based on mean and standard deviation, which can preserve the true distribution difference between normal waves and extreme abrupt waves to the greatest extent while unifying the dimensions. This not only enhances the physical perception ability of neural networks of the high steepness and abrupt change characteristics of extreme waves, but also eliminates the dimensional differences of multi-field data.

[0029] 2. Spatial coordinate mapping and spatiotemporal dynamic feature extraction based on TimeDistributed-LSTM (1) Input the three-dimensional absolute coordinates of the wave measurement point into a three-layer fully connected neural network and map them to generate a high-dimensional spatial feature vector; (2) The high-dimensional spatial feature vector is copied and broadcast along the time axis, so that it is concatenated with the wave spatiotemporal input tensor in the time dimension to form a unified four-dimensional wave spatiotemporal input tensor. This tensor contains the time series features of the wave (such as the original wave height and multi-order difference sequence) and also integrates the corresponding spatial geometric structure information (absolute coordinate mapping vector). Its tensor dimension structure is usually defined as four dimensions: (batch size, number of wave measurement points) (Time step T, fusion feature dimension); (3) TimeDistributed-LSTM Independent Time Coding: The above four-dimensional wave spatiotemporal input tensor is input into TimeDistributed-LSTM. TimeDistributed provides a logically independent encoding path for each wave sensor and forces parameter sharing. While avoiding memory explosion, it independently extracts dynamic wave features for each wave point and finally outputs waveform change feature vectors for multiple wave measurement points, as follows: a. Number of wave measurement points distributed by TimeDistribution Spatial dimension: The input high-dimensional four-dimensional tensor X is sliced ​​and traversed, decomposing the high-dimensional tensor into... An independent two-dimensional temporal feature slice This provides a logically independent LSTM encoding path for each wave measurement point, preventing mutual interference between features of different spatial nodes in the early stages of encoding. Here, T is the time step. Indicates the number of input feature channels; b. In each independent encoding path, LSTM processes individual feature slices sequentially at time step T. By progressively reading wave dynamics signals, the system accurately captures the dynamic characteristics of local acceleration, wavefront slope, and nonlinear wave group structure at a specific measurement point, which evolve over time. It then processes all... During the process of measuring each point, the network forces all LSTM encoding paths to share the same set of network weight parameters, which ensures that parallel feature extraction of massive discrete point clouds can be completed with low memory consumption.

[0030] c. For each measurement point, at the end of the sequence processing, information is filtered and retained through the gating mechanism inside the LSTM, compressing the long spatiotemporal sequence information of length T into a fixed-length compressed vector. ,in, The number of hidden layer nodes in the LSTM; d. The TimeDistributed-LSTM module will handle all... Compression vectors generated independently at each measurement point The original measurement points are reassembled along their dimensions and packaged into a dimensionality-reduced waveform change feature vector. It successfully extracted dynamic features and directly output them to the next layer for the fusion calculation of relative geometric features.

[0031] The TimeDistributed-LSTM module is designed to extract spatiotemporal dynamic features independently, in parallel, and efficiently from multiple discretely distributed wave monitoring points.

[0032] This invention embodiment is based on spatial coordinate mapping and the TimeDistributed-LSTM module to initially fuse the spatial distribution structure information and temporal evolution characteristics of waves, and extract spatiotemporal dynamic features for each independent wave measurement point.

[0033] 3. Relative geometric feature encoding This invention employs the construction of relative geometric features to characterize the attenuation and directional differences of wave energy along a specific spatial path. Specifically: (1) Calculation of relative geometric features: Calculate the relative coordinates and Euclidean distance between the target prediction point and the wave monitoring point: , , in, Relative coordinates Represents the three-dimensional absolute coordinates of the wave measurement point. Represents the three-dimensional absolute coordinates of the target prediction point. Euclidean distance; (2) Construction of four-dimensional feature tensor: After dimensional expansion and broadcasting mechanism, the temporal embedding vector, absolute coordinates of wave measurement points, absolute coordinates of target points, relative coordinates and Euclidean distance output in step 4 are spliced ​​point by point to construct a four-dimensional feature tensor containing multi-dimensional spatiotemporal physical features. , is represented as: , in, For the temporal embedding tensor, it is composed of the waveform change feature vector after dimensionality reduction. Obtained through dimension expansion and copying operations (tiles), ensuring dimension alignment during model computation. The feature dimension of the temporal embedding vector; This is the absolute coordinate vector obtained by copying the wave measurement point along the dimension of the target point; Relative coordinates; Euclidean distance; the first three dimensions ( ) represent the number of samples in the model training, the number of predicted target location points, and the number of wave measurement points, respectively; This represents a vector concatenation operation.

[0034] 4. Local-Global Feature Aggregation This invention extracts the overall coupling influence of wave groups on a single target point through local-global feature aggregation, as detailed below: (1) The four-dimensional feature tensor is input into a layer-by-layer extended local multilayer perceptron (Local MLP) to extract the nonlinear influence between a single wave point and the target point layer by layer to obtain the local feature matrix. This includes three Local MLPs. Through the layer-by-layer extended neural network structure (64 neurons → 256 neurons → 512 neurons), the relationship between a single wave point and the target point is fully explored. The directional differences, energy transfer modes, and nonlinear characteristics of the propagation paths between the components are used to obtain the local feature matrix: , in, For local feature matrices, Represents a four-dimensional feature tensor; (2) Max Pooling is performed on the local feature matrix in the wave sensor dimension to aggregate the contribution of all wave points to the target point, and the wave crest propagation path features with the highest energy are selected and extracted to obtain the target point. Global characteristics of the response of all wave sensors: , in, The global feature matrix, after dimensionality reduction via max pooling, becomes a three-dimensional tensor with the shape (batch size, number of target prediction points, global feature dimension). Indicates the wave measurement point.

[0035] After relative geometric feature encoding and local-global feature aggregation, the embodiments of the present invention obtain a global feature matrix. Each feature vector in this tensor integrates the spatiotemporal coupling influence of the entire random multidirectional wave field on a specific seabed prediction point. It is the core link to realize spatial generalization and arbitrary location perception. It is input to the final decoder to output the pore pressure value.

[0036] 5. Multi-output decoding network prediction response and model evaluation This invention rebinds deep features incorporating global wave influences with the absolute spatial location of target prediction points. Through a decoding network, it achieves synchronous and parallel prediction of the dynamic responses (such as pore water pressure) of any multiple target points within the seabed. Simultaneously, it defines a loss function and evaluation metrics to complete end-to-end network weight optimization and generalization capability verification. Specifically: (1) Global features absolute coordinates of the corresponding target prediction point Cascaded concatenation is performed along the feature channel dimension, and the concatenated feature tensor is input into a multi-output decoding network composed of multiple fully connected layers. The nonlinear mapping is performed in parallel along the dimension to reduce its dimension and restore it to the output space of the target physical quantity, thus obtaining the final predicted value of all target points; (2) During the model training phase, mean squared error (MSE) is used to quantify the error between the predicted value and the true label (i.e., the seabed response time series label). The batch-level global loss calculation formula is as follows: , in, and They represent the first b In the batch, the first jThe predicted and actual values ​​of each target point are compared. Based on the calculated loss, an adaptive moment estimation optimizer (such as the Adam optimizer) is used to calculate the gradient through backpropagation and iteratively update all weight parameters in the network until the model converges.

[0037] The dataset was divided into a training set (80%) and a test set (20%). Forward propagation inference was performed on the test set, and the prediction results were de-standardized to restore them to the true physical dimensions. The coefficient of determination R was then used. 2 The model's prediction accuracy and spatial distribution reproduction were fully verified by four statistical indicators: mean absolute error (MAE), root mean square error (RMSE), and prediction bias (PBIAS). (3) After training and verification, a spatiotemporal neural network with fixed parameters is obtained, and the reliability of the model is verified based on the above evaluation indicators.

[0038] IV. In the actual engineering application stage, when the coordinate range of the prediction area does not exceed the coordinate range of the training model, the model can output the dynamic response curve of the seabed at any point around the monopile foundation with high fidelity by simply using the elevation time series data measured by wave sensors at any location on the input site and the coordinates of the target prediction point.

[0039] Figure 6 The graph shows the model training iteration loss performance curves for different numbers of input point clouds provided in this embodiment of the invention. The horizontal axis represents the number of iterations (Epochs), and the vertical axis represents the loss value (Loss). It can be seen that after using the TimeDistributed parameter sharing mechanism, even if the number of input points increases to 100 (i.e., high-batch discrete measurement points), the model can still converge quickly without memory overflow. Furthermore, the more input points, the lower the validation set loss and the more stable the model. When the number of wave input measurement points increases from 25 to 100, the model not only does not experience memory overflow, but also achieves a faster convergence speed with an extremely limited number of parameters. In N... input Under the optimal sampling condition of 100, the training loss and validation loss curves smoothly decrease to the minimum and perfectly overlap in a very short time (about 30 epochs), which proves the excellent stability and strong generalization performance of the present invention with extremely low computing power consumption.

[0040] Figure 7 For the embodiments of the present invention, the number of input points When comparing the numerical simulation results (a) and the model prediction results (b) of the present invention with the cloud map, it can be seen from the figure that the red high pressure zone and the blue negative pressure zone formed in the front, back and sides of the pile foundation are well matched in terms of location, range and shape. Especially in the high gradient pore pressure change area presented on the wave-facing side and wake of the pile foundation, the model can accurately capture its asymmetric distribution characteristics, proving that the present invention can not only predict the numerical distribution of pore water pressure, but also more accurately capture the local spatiotemporal interference effect of the seabed caused by multi-directional waves.

[0041] Figure 8 For the embodiments of the present invention, the number of input points When comparing the numerical simulation results (a) and the prediction results (b) of the deep learning model around the pile foundation, it can be seen from the cloud map that a significant high shear stress concentration area is formed in the vicinity of the pile foundation due to wave diffraction and local flow field acceleration. The model can accurately reproduce the location and amplitude characteristics of this area. At the same time, in the area far from the structure, the shear stress shows a smooth transition distribution. There is no significant deviation between the prediction results and the numerical solution. This further proves that the model also has extremely high fidelity in predicting other key mechanical responses (such as shear stress field) in the solid phase domain of the seabed, proving the reliability of the invention.

[0042] Figure 9 This is a comparison curve of the prediction error index changes under different multi-directional wave concentration S conditions according to an embodiment of the present invention. The horizontal axis represents the concentration parameter (s) of the multi-directional wave, and the vertical axis includes R. 2 The four evaluation metrics, MAE, RMSE, and PBIAS, demonstrate that the present invention maintains excellent prediction accuracy and robustness when facing complex sea conditions with varying degrees of dispersion (from unidirectional swells to highly discrete multidirectional wind waves).

[0043] This invention is not limited to the above-described embodiments. Without departing from the concept of this invention, the following substitutions and modifications can be made, all of which fall within the protection scope of this invention: (1) In step one, the acquisition of the basic multi-field dataset can be achieved not only by using CFD-FEM coupled numerical simulation to obtain training data, but also by using multi-channel wave meter and pore pressure gauge data collected from large-scale physical model experiments (such as wave flume experiments or centrifuge experiments), or by using in-situ observation data from multi-point wave buoys and seabed embedded sensors transmitted in real time from the on-site health monitoring system (SHM) of a real offshore wind farm. All of the above sources can provide multi-directional wave and seabed dynamic response labels with spatial coordinates and temporal evolution attributes to meet the model training requirements.

[0044] (2) In step 1, multi-order feature enhancement can be achieved not only by using first-order and second-order difference to extract physical features, but also by using frequency domain or time-frequency domain feature extraction algorithms. For example, continuous wavelet transform (CWT) or short-time Fourier transform (STFT) can be used to extract local high-frequency transient change features of wave signals; or a one-dimensional convolutional neural network (1D-CNN) with shared weights can be used as the front end to automatically extract the local rate of change and curvature features of wave surface elevation on the time axis through convolution kernels of different sizes. All of these methods can effectively extract nonlinear, asymmetric, and extremely steep features in the original wave elevation sequence, thereby improving the model's physical perception sensitivity to extreme sea conditions.

[0045] (3) In step 2, when the time-series feature independent encoding network module uses the TimeDistributed wrapper for independent time-series encoding, in addition to LSTM, it can also be replaced by GRU, Temporal Convolutional Network TCN or Transformer encoder, which can achieve independent extraction of time-series features and weight sharing of each measurement point, and achieve similar technical effects.

[0046] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model, characterized in that, Includes the following steps: S1. By simulating the response of the coupled wave fluid domain and the seabed solid domain using CFD-FEM, multi-point time-series physical field data with spatial coordinates are generated for the multi-directional wave surface elevation and seabed pore water pressure response. S2. Randomly sample the multi-point time-series physical field data to obtain an unstructured point cloud dataset. The dataset includes the three-dimensional absolute coordinates of wave measurement points and the corresponding wave height sequence data, the three-dimensional absolute coordinates of target prediction points and the corresponding seabed pore water pressure time-series data. S3 constructs a time-series prediction model of the seabed response around wave-induced pile foundations. The model performs the following operations: S31 performs multi-order difference feature enhancement on the wave measurement point height sequence data and performs global standardization preprocessing to obtain the wave spatiotemporal input tensor. S32 performs high-dimensional spatial mapping on the three-dimensional absolute coordinates of wave measurement points and concatenates them with the wave spatiotemporal input tensor to construct the TimeDistributed-LSTM module to independently extract waveform change feature vectors from multiple wave measurement points. S33 calculates the relative coordinates and Euclidean distance between the target prediction point and the wave measurement point, and concatenates them with the waveform change feature vector, the three-dimensional absolute coordinates of the wave measurement point, and the three-dimensional absolute coordinates of the target prediction point to form a four-dimensional feature tensor with multi-dimensional spatiotemporal physical characteristics. S34 performs local multilayer perceptron and global max pooling on the four-dimensional feature tensor along the wave measurement point dimension to extract global wave energy evolution features; S35 stitches the global wave energy evolution characteristics with the three-dimensional absolute coordinates of the target prediction point, and outputs the dynamic response prediction values ​​of multiple points on the seabed in parallel by the multi-output decoding network to complete the model training and verification. S4. Input the elevation time series data of the wave measuring points deployed on site and the coordinates of the target prediction points into the trained model to directly obtain the dynamic response curves of the seabed at multiple points around the pile foundation.

2. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 1, characterized in that, In step S2, the unstructured point cloud dataset is obtained as follows: S21 Extracts the boundary extent of the wave fluid domain mesh in three-dimensional space. , , And generate within the range The three-dimensional absolute coordinates of a random wave measuring point that follows a uniform distribution; S22 calculates the spatial correlation weights between the three-dimensional absolute coordinates and the original mesh nodes based on the radial basis spatial interpolation algorithm; S23 maps the time-series data of the original grid nodes to the three-dimensional absolute coordinates of random wave measurement points based on the spatial correlation weights, thereby obtaining the time-series sequence of each wave measurement point coordinate point over the entire time step, and thus obtaining... The three-dimensional absolute coordinates of randomly distributed wave measurement points and their corresponding wave elevation time series are used as input data. S24 generates a grid in the seabed solid phase domain in the same manner as steps S21-S23. The output data consists of the three-dimensional absolute coordinates of randomly distributed seabed target points and their corresponding pore water pressure time-series labels.

3. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 2, characterized in that, In step S31, the multi-order difference feature enhancement is specifically implemented as follows: Let the first i The length of the high-order sequence data for each wave measurement point is [length missing]. T , is represented as: ,in, Indicates the measuring point i At time step t The wave surface elevation value, a. Decomposing the first-order difference sequence: ; b. Decompose the second-order difference sequence: ; c. Pad the difference sequence length to the original length T by extending interpolation; d. After concatenating the original sequence, the first-order difference sequence, and the second-order difference sequence along the channel dimension and standardizing them, we obtain the sequence containing... A wave spatiotemporal input tensor with 3 wave measurement points, a time step of T, and 3 feature channels.

4. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 3, characterized in that, In step S32, the waveform change feature vector is extracted as follows: S321 maps the three-dimensional absolute coordinates of the wave measurement points into high-dimensional spatial feature vectors through a three-layer fully connected neural network. S322 performs a copy and broadcast operation along the time axis on the high-dimensional spatial feature vector and concatenates it with the wave spatiotemporal input tensor to form a four-dimensional wave spatiotemporal input tensor. The S323 TimeDistributed-LSTM module performs a slice traversal on the four-dimensional wave spatiotemporal input tensor, decomposing it into... Each independent temporal feature slice provides an independent encoding path for each wave measurement point; The S324 LSTM processes individual feature slices sequentially according to time steps, gradually processing wave dynamics signals to accurately capture the dynamic characteristics of each wave measurement point evolving over time. For each wave measurement point, the S325 uses the LSTM gating mechanism at the end of the sequence processing to filter and retain information, compressing the spatiotemporal sequence information of length T into a fixed-length compressed vector. The S326 TimeDistributed-LSTM module will... The compressed vectors generated independently at each wave measurement point are reassembled along the original wave measurement point dimensions to output the waveform change feature vector of multiple wave measurement points. .

5. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 4, characterized in that, All the independent encoding paths mentioned above share the same set of network weight parameters.

6. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 4, characterized in that, The four-dimensional feature tensor is represented as follows: ， in, Waveform change feature vector Temporal embedding vectors are obtained through dimensional expansion and duplication; This is the absolute coordinate vector obtained by copying the wave measurement point along the dimension of the target prediction point; Relative coordinates; The distance is Euclidean. This represents a vector concatenation operation.

7. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 6, characterized in that, The extraction of global wave energy evolution features is as follows: S341 inputs the four-dimensional feature tensor into a layer-by-layer expanding local multilayer perceptron to mine the nonlinear features of directional differences, energy transfer patterns, and propagation paths between a single wave measurement point and the target seabed prediction point, thus obtaining local features: ， in, Indicates local features, Represents a four-dimensional feature tensor. This represents a local multilayer perceptron. Indicates the target prediction point; S342 performs max pooling on the local features along the wave measurement point dimension to obtain global wave energy evolution features: ， in, This represents the global wave energy evolution characteristics. This indicates max pooling. Indicates the wave measurement point.

8. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 1, characterized in that, In step S35, the decoding network consists of multiple fully connected layers. It performs nonlinear mapping on the spliced ​​features in the dimension of the target prediction points, and reduces the dimension to restore the output space of the seabed pore water pressure time series label, and outputs the dynamic response prediction values ​​of all target prediction points in parallel.

9. The method for predicting the seabed response around wave-induced pile foundations based on a deep learning time series model according to claim 2, characterized in that, The distribution range of the three-dimensional absolute coordinates of the random wave measuring points is constrained to... , , Within this range, it simultaneously encompasses the incident wave region upstream of the monopile structure, the diffraction and reflection region around the structure, and the wave attenuation region downstream.

Images