Fan foundation prediction method and system based on adaptive noise reduction and spatio-temporal graph fusion

Abstract

The application discloses a fan foundation prediction method and system combining adaptive noise reduction and space-time graph fusion, and belongs to the field of wind power generation engineering structure health monitoring and deep learning prediction. In view of the problems of insufficient data quality, incomplete space-time feature fusion and lack of dynamic space relationship modeling in the prior art, the application adopts a three-stage adaptive noise reduction method to process fan foundation multi-pile monitoring data, which sequentially includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering and edge reservation post-processing; an input feature sequence is constructed based on an anti-information leakage strategy; a space-time graph fusion network is constructed, which includes a dynamic space graph attention module, a multi-scale receptive field module and a global time encoder module, and a time fusion mechanism and a space-time fusion mechanism are designed; and a mixed loss function is used for training and outputting a prediction value. The application can perform high-precision and advanced prediction on various monitoring indexes such as fan foundation settlement change and pile top axial force.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and deep learning prediction for wind power generation projects, and in particular to a method and system for predicting wind turbine foundations using adaptive noise reduction and spatiotemporal graph fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a critical load-bearing structure of the entire wind turbine generator, the service health of the foundation directly affects operational safety and power generation efficiency. In adverse site conditions such as soft soil foundations, collapsible loess, and frozen soil, key indicators such as foundation settlement and pile top axial force are affected by the coupling of geological conditions, cyclic loads, and environmental factors, potentially leading to tower tilting, blade-to-tower collisions, transmission system failures, and even collapse. Therefore, real-time monitoring and proactive prediction of the service status of wind turbine foundation components in adverse sites are of significant engineering importance and research value for ensuring the safe operation of wind turbine foundations.

[0004] In recent years, deep learning methods have been introduced into the field of structural health monitoring. For example, the settlement prediction method based on CNN-LSTM adopts a pure sequence architecture, completely ignoring the spatial topological relationships between sensor nodes, and its noise reduction relies solely on simple statistics, making it insufficiently adaptable to complex noise environments. Another anomaly recognition method based on spatiotemporal graph convolutional networks introduces a learnable adjacency matrix, but this matrix remains fixed after training and cannot reflect the dynamic evolution of spatial coupling relationships between piles under changing operating conditions. Furthermore, this method focuses on anomaly detection rather than temporal prediction, lacking systematic data denoising and multi-scale spatiotemporal feature fusion.

[0005] In summary, existing technologies suffer from three main shortcomings: First, data quality is insufficient. Monitoring data is severely affected by high-frequency sensor noise and environmental interference. Existing denoising methods typically employ isolated, single strategies, failing to systematically integrate global frequency domain filtering with local time domain adaptive filtering. Second, spatiotemporal feature fusion is incomplete. Existing models either only model time dependence or only spatial topology, failing to organically integrate spatial modeling, local time feature extraction, and global time dependency capture. Third, dynamic spatial relationship modeling is lacking. Methods based on static graphs implicitly assume topological invariance between pile locations, failing to reflect the dynamic evolution of spatial coupling relationships under changing operating conditions. These deficiencies make it difficult for existing methods to achieve high-precision, advanced prediction of wind turbine foundation health status. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as insufficient data quality, incomplete spatiotemporal feature fusion, and lack of dynamic spatial relationship modeling, this invention provides a wind turbine foundation prediction method and system based on adaptive denoising and spatiotemporal graph fusion. It improves data quality through three-stage adaptive denoising, recovering important temporal patterns lost by single denoising strategies; constructs input feature sequences through an information leakage prevention strategy to ensure the generalization of predictions; captures time-varying spatial dependencies through a dynamic spatial graph attention module in the spatiotemporal graph fusion network; and extracts local and global temporal features through a multi-scale receptive field module and a global time encoder module, respectively. These features are then organically integrated through a time and spatiotemporal fusion mechanism, achieving high-precision advanced prediction of monitoring indicators such as settlement and axial force.

[0007] On the one hand, an adaptive noise reduction and spatiotemporal graph fusion method for wind turbine foundation prediction is provided, including: Acquire and preprocess monitoring data from multiple pile locations of the wind turbine foundation; An adaptive noise reduction method is used to perform three-stage noise reduction on the preprocessed monitoring data. The three-stage noise reduction process includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing. An input feature sequence is constructed based on an information leakage prevention strategy. This input feature sequence is then fed into a spatiotemporal graph fusion network to obtain a prediction result. The processing flow of the spatiotemporal graph fusion network includes: extracting spatial features through a dynamic spatial graph attention module; extracting local temporal features through a multi-scale receptive field module; extracting global temporal features through a global temporal encoder module; fusing the local and global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fusing the spatial and temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and mapping the spatiotemporal fusion features to the predicted value of the target pile location through a prediction head.

[0008] Furthermore, the acquisition of multi-pile monitoring data for the wind turbine foundation includes: acquiring monitoring data for N piles arranged in a circular pattern on the wind turbine foundation, wherein the monitoring data includes settlement change values ​​or pile top axial force values, and is collected at fixed time intervals.

[0009] Furthermore, the Fourier frequency domain filtering includes: estimating the noise level using a median absolute deviation estimator, calculating the signal-to-noise ratio, adaptively selecting the cutoff frequency ratio based on the signal-to-noise ratio, constructing a soft cutoff filter to achieve a smooth transition from the passband to the stopband, while applying attenuation to frequency components with power below a preset percentile threshold, and obtaining the filtered signal through inverse Fourier transform. The Bayesian adaptive Kalman filter includes: establishing a one-dimensional state-space model, updating the process noise covariance and observation noise covariance online through local window statistics, and obtaining a smooth signal through Kalman prediction and recursive update. The edge-preserving post-processing includes: detecting edge points using the 60th percentile of the differential signal as a threshold, preserving the frequency domain filtered signal at the edge with a 95% weight, fusing the Kalman smoothed signal with a 5% weight, and using the Kalman smoothed signal at non-edge locations.

[0010] Furthermore, the information leakage prevention strategy specifically involves: excluding the target pile location from the input features, using the remaining N-1 pile locations as context nodes, constructing an input feature sequence using the historical denoising data of the context nodes, and predicting the state value of the target pile location at the next time step; the spatiotemporal graph fusion network constructs a spatial topology graph using the N-1 context pile locations as nodes.

[0011] Furthermore, the extraction of spatial features through the dynamic spatial graph attention module specifically involves: for each time step, extracting the denoised input features of each node at the current time and encoding them into a high-dimensional space; dynamically generating an adjacency matrix through a multi-head self-attention mechanism; after symmetry and self-loop processing, extracting node embeddings through a multi-layer multi-head graph attention network; sending the time step sequence into a bidirectional temporal aggregation network for temporal aggregation; obtaining spatial features through linear projection and flattening; and retaining the original denoised input through a learnable residual shortcut.

[0012] Furthermore, the extraction of local temporal features through the multi-scale receptive field module specifically involves: the multi-scale receptive field module employs a causal dilated temporal convolutional network, stacking multiple residual blocks with an exponentially increasing inflation rate. Each residual block contains dilated convolution, batch normalization, activation functions, and Dropout, and outputs local temporal features that are linearly projected to a uniform dimension.

[0013] Furthermore, the global time encoder module adopts a self-attention network with an encoder-decoder structure, which linearly maps the noise reduction input and adds position encoding. The encoder consists of multiple layers of self-attention and feedforward networks, and the decoder applies a causal mask to prevent future information leakage and outputs global time features. The temporal fusion mechanism is specifically as follows: the local temporal features and the global temporal features are mutually enhanced through bidirectional cross attention, and the temporal fusion features are obtained after adaptive fusion with learnable Sigmoid gating. The spatiotemporal fusion mechanism is as follows: the spatial features and the temporal fusion features are input into three parallel attention branches, and temporal self-attention, spatial self-attention and spatiotemporal cross-attention are calculated respectively. The spatiotemporal fusion features are obtained by adaptive weighted fusion through a two-level network of dynamic gating and global gating.

[0014] Furthermore, the method also includes training the spatiotemporal graph fusion network using a hybrid loss function, wherein the hybrid loss function is a weighted sum of mean squared error and mean absolute error; when the number of predicted output nodes is greater than 1, the hybrid loss function also includes a spatial trend consistency loss, which is used to constrain the consistency between the predicted difference and the actual difference between adjacent nodes.

[0015] On the other hand, a wind turbine foundation prediction system based on adaptive noise reduction and spatiotemporal graph fusion is also provided, including: The data acquisition module is configured to acquire and preprocess monitoring data from multiple pile locations of the wind turbine foundation. The adaptive noise reduction module is configured to perform three-stage noise reduction processing on the preprocessed monitoring data using an adaptive noise reduction method. The three-stage noise reduction processing includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing in sequence. The sample construction module is configured to: construct input feature sequences based on information leakage prevention strategies; The prediction module is configured to: input the input feature sequence into a spatiotemporal graph fusion network to obtain a prediction result; wherein, the spatiotemporal graph fusion network is configured to perform the following steps: extract spatial features through a dynamic spatial graph attention module, extract local temporal features through a multi-scale receptive field module, and extract global temporal features through a global temporal encoder module; fuse the local temporal features and the global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fuse the spatial features and the temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and map the spatiotemporal fusion features to the predicted value of the target pile position through a prediction head.

[0016] In another aspect, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the method described in the first aspect.

[0017] The above technical solution has the following advantages or beneficial effects: (1) Adaptive three-stage noise reduction can suppress broadband noise while preserving the details of sedimentation abrupt changes and improving the input signal-to-noise ratio.

[0018] (2) The dynamic spatial graph attention module generates an adjacency matrix by time step, automatically capturing the coupling relationship between piles that changes over time, thus avoiding human prior bias.

[0019] (3) The multi-scale receptive field module and the global time encoder model local short-range fluctuations and long-range trends respectively. The two are fused through bidirectional cross-attention to balance sensitivity and stability.

[0020] (4) The spatiotemporal fusion mechanism of three-way parallel attention + two-level gating can adaptively allocate the weights of spatial, temporal and spatiotemporal cross information under different working conditions. On the test set of pile monitoring point 1, the complete model R²=0.9640 and the ablation variant has the lowest R²=0.9341, which verifies the effectiveness of multi-branch collaboration. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is an overall flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the wind turbine foundation pile location arrangement in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the spatiotemporal graph fusion network in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the time fusion mechanism in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the spatiotemporal fusion mechanism in Embodiment 1 of the present invention; Figure 6 This is a comparison chart of the predicted and measured settlement values ​​of monitoring point 1 at the pile location in Embodiment 1 of the present invention. Figure 7 This is a performance comparison chart of the ablation experimental model in Embodiment 1 of the present invention, wherein, Figure 7 (a) in the figure is a comparison chart of the root mean square error of each module combination. Figure 7 (b) in the figure is a comparison of the determination coefficients of each module combination. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Those skilled in the art should understand that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 This embodiment provides a wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion. Figure 1 This is an overall flowchart of the method according to Embodiment 1 of the present invention. The method includes the following steps: S1: Acquire monitoring data of multiple pile locations of the wind turbine foundation and perform preprocessing; S2: An adaptive noise reduction method is used to perform three-stage noise reduction on the preprocessed monitoring data. The three-stage noise reduction process includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing. S3: Constructing an input feature sequence based on an information leakage prevention strategy; S4: Input the input feature sequence into the spatiotemporal graph fusion network to obtain the prediction result; the processing flow of the spatiotemporal graph fusion network includes: extracting spatial features through the dynamic spatial graph attention module, extracting local temporal features through the multi-scale receptive field module, and extracting global temporal features through the global temporal encoder module; fusing local and global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fusing spatial and temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and mapping the spatiotemporal fusion features to the predicted value of the target pile position through the prediction head.

[0026] Taking the settlement monitoring data of the wind turbine foundation piles of a certain wind power base as an example, in step S1, the monitoring data of multiple piles of the wind turbine foundation are obtained and preprocessed.

[0027] Specifically: First, acquire monitoring data from multiple pile locations on the wind turbine foundation. This data includes settlement variation values ​​(in mm) or pile top axial force values ​​(in kN). For example... Figure 2 As shown, in this embodiment, the wind turbine foundation adopts a circular layout, with a total of 8 pile monitoring points (numbered 1 to 8). Settlement change values ​​are collected at 15-minute intervals using a hydrostatic leveling system (HLS), with a resolution of 0.1 mm and a range of 0.2 to 3000 mm. The monitoring period is approximately 32 days, resulting in approximately 3072 observation data points. The spatial coordinates of each pile monitoring point are evenly distributed clockwise on a unit circle, starting from true north. The settlement change data of each pile monitoring point are statistically analyzed in Table 1.

[0028] Table 1 shows the settlement change data at each pile location monitoring point.

[0029]

[0030] The collected data was then preprocessed, including: automatically identifying file encoding formats, sorting by time and resetting the index, using forward and backward padding to handle missing values, and using the 3σ principle to remove outliers.

[0031] Thus, the cleaned multivariate time series was obtained, with a length of 3072 and 8 pile monitoring points. The feature dimension of each pile monitoring point is F=1, meaning that each pile monitoring point corresponds to a settlement observation value (unit: mm) at each time step.

[0032] In step S2, to remove sensor noise and environmental interference while preserving key trends and abrupt change characteristics, this embodiment employs an adaptive denoising method to perform three-stage denoising on the preprocessed 3072 settlement observations. The specific denoising process is illustrated using the settlement data of pile monitoring point 1 as an example. Using the settlement observation sequence of pile monitoring point 1 for 24 consecutive hours (96 15-minute time steps) in Table 1 as input, after Fourier filtering, the dominant frequency band is preserved, and low-power high-frequency components are effectively attenuated; Kalman smoothing further reduces signal variance and eliminates noise components in short-time jumps; edge detection triggers protection at settlement abrupt changes, and the original denoised input signal and the Kalman smoothed signal are fused at several detected abrupt change points with a weight of 0.95 / 0.05, thereby suppressing broadband noise while completely preserving the settlement jump characteristics. The first stage is Fourier frequency domain filtering. First, the noise level is estimated and the signal-to-noise ratio (SNR) is calculated using a median absolute deviation (MAD) estimator. The cutoff frequency ratio is adaptively selected based on the SNR: 0.75 for SNR > 10, 0.65 for SNR <= 10, and 0.55 for SNR < 5. A soft cutoff filter is constructed, with exponential attenuation to achieve a smooth transition from the passband to the stopband and suppress Gibbs ringing. Frequency components with power below the 70th percentile are further suppressed by an attenuation factor of 0.5, preserving the dominant frequency components. The filtered signal is obtained through inverse FFT.

[0033] The second stage involves Bayesian adaptive Kalman filtering. For the settlement time series of monitoring point 1, a one-dimensional linear Gaussian state-space model is established step-by-step: the state equation is x_t = x_(t-1) + w_t, and the observation equation is y_t = x_t + v_t, where w_t~N(0,Q) and v_t~N(0,R) represent the process noise and observation noise, respectively. The process noise covariance Q and the observation noise covariance R are adaptively estimated online using local window statistics. The Kalman filter executes according to the standard prediction-update recursion, outputting a smooth signal.

[0034] The third stage is edge-preserving post-processing. The first-order absolute difference d_i = |s_i - s_(i-1)| (with boundary d_0 = 0) is calculated for the Kalman-smoothed signal s as the differential signal. Using the 60th percentile τ = Q60(d) of the differential signal as a threshold, when d_i > τ, the i-th time step is determined to be an edge point. At non-edge locations, the Kalman-smoothed signal is directly applied; at edge points, a weighted fusion is performed using 0.95 × the original denoised input signal + 0.05 × the Kalman-smoothed signal, thus suppressing noise while preserving the details of abrupt settlement changes.

[0035] Optionally, the specific values ​​given in this embodiment (such as signal-to-noise ratio thresholds of 10 and 5, the 70th percentile of the low-power frequency component, the 60th percentile of the differential edge detection, and the 95% weight at the edge, etc.) are preferred parameters for the data in this example and are not a strict limitation on the scope of protection of this invention. In practical applications, those skilled in the art can adjust the above parameters through conventional experiments or adaptive methods based on the characteristics of the wind turbine foundation monitoring data, such as the sampling frequency, noise level, and signal mutation degree. For example, when the monitoring data noise is high, the cutoff frequency ratio threshold of the Fourier filter can be appropriately increased or the low-power percentile can be decreased; when the settlement change is relatively gentle, the differential percentile of the edge detection can be decreased to improve the sensitivity to weak mutations; the retention weight of the frequency domain filtered signal in edge processing can be selected between 80% and 100% to balance the needs of noise reduction and edge preservation.

[0036] After the above three stages, noise-reduced data that both removes noise and retains key change characteristics is obtained.

[0037] In step S3, in order to prevent the prediction model from directly or indirectly using the future information of the target pile location itself during training or prediction, this embodiment adopts a strict anti-information leakage strategy to construct the input feature sequence.

[0038] Specifically, pile monitoring point 1 is selected as the prediction target. Historical data of pile monitoring point 1 is completely excluded from the input features. The remaining seven pile monitoring points (pile monitoring points 2 to 8) are used as feature piles. Historical denoised data from these piles are used to construct an input feature sequence to predict the state value of the target pile at the next time step. In the subsequent step S4, the spatiotemporal graph fusion network uses the seven feature piles (i.e., pile monitoring points 2 to 8) as context nodes to construct a spatial topology graph. This aims to capture the dynamic spatial relationships between context piles, thereby indirectly inferring the changing trend of the target pile.

[0039] In this embodiment, the input feature sequence length is 10 time steps (corresponding to 150 minutes), and the prediction step length is 1 step (15 minutes). The data is divided in chronological order, with the first 80% of the samples used as the training set, the middle 10% as the validation set, and the last 10% as the test set. No random shuffling is performed to prevent future information leakage. Robust standardization is then applied to both the input features and the prediction target: the denoised settlement sequence x for each pile monitoring point is transformed using x'=(x-median(x)) / IQR(x), where median(·) is the median and IQR(·)=Q75(x)-Q25(x) is the interquartile range. This method is insensitive to outliers and effectively suppresses the influence of extreme noise on the dimensions.

[0040] In step S4, a spatiotemporal graph fusion network is constructed and the prediction results are obtained. For example... Figure 3 As shown, the spatiotemporal graph fusion network includes the following key modules.

[0041] The Dynamic Spatial Graph Attention (DSGA) module is used to capture the spatial dependencies between piles that change over time. First, the denoised input features are fed into the spatial modeling module. For each time step, the denoised input features of each node at the current time are extracted and encoded into a high-dimensional space. An adjacency matrix is ​​dynamically generated at each time step through multi-head self-attention, symmetricized, and self-loops are added. Optionally, learnable static priors based on physical layout are fused. Based on the dynamic adjacency matrix, node embeddings are extracted through multi-layer multi-head graph attention convolution. The time step sequence is fed into a multi-layer bidirectional temporal aggregation network to aggregate the temporal context. The forward and backward hidden states are concatenated and linearly projected to flatten the node dimensions and obtain spatial features. Furthermore, a learnable residual shortcut is used to preserve the original input information to accelerate network training.

[0042] The Multi-Scale Receptive Field (MSRF) module is used to extract local multi-scale temporal features. MSRF employs a causal dilated temporal convolutional network, stacking L temporal convolutional residual blocks with an exponentially increasing dilation rate and a kernel size of k. Each residual block contains dilated convolutions, batch normalization, an activation function, and Dropout. After the input features are processed by the multi-layer causal dilated convolutional residual network, the output of the MSRF module is linearly projected to a unified dimension to obtain the local temporal features.

[0043] The Global Temporal Encoder (GTE) module is used to capture global long-range temporal dependencies. This embodiment employs a self-attention network with an encoder-decoder structure. First, the denoised input is linearly mapped and multiplied by a scaling factor, followed by the addition of a sinusoidal positional encoding. The encoder consists of multiple layers of self-attention and feedforward networks stacked together. The decoder input is constructed by right-shifting the encoder input, using a learnable starting vector as the first position, and applying a causal mask to ensure that the decoder can only notice information at positions less than or equal to t-1 at time t. The decoder output contains global temporal features for each time position. Subsequently, global average pooling in the prediction head aggregates the full temporal dimensions into a single feature vector, which is then regressed through a multilayer perceptron to obtain the predicted value for the next time step of the target pile position. This encoder-decoder structure retains the ability to extend to multi-step prediction.

[0044] In response to the spatial features, local temporal features, and global temporal features obtained above, this invention proposes a multi-level feature fusion mechanism, which includes two stages: temporal fusion and spatiotemporal fusion.

[0045] Specifically, such as Figure 4As shown, the first stage is temporal fusion: the local temporal features (output of the MSRF module) and the global temporal features (output of the GTE module) are fed into the bidirectional cross-attention fusion module, and then adaptively weighted and fused through the Sigmoid gated network. After layer normalization, the temporal fusion features are obtained.

[0046] like Figure 5 As shown, the second stage is spatiotemporal fusion: the aforementioned temporal fusion features and spatial features are input into a three-way parallel attention module. These three attention branches respectively calculate temporal self-attention, spatial self-attention, and spatiotemporal cross-attention. The first-level dynamic gating network adaptively balances the three branches at each time step, and performs weighted fusion by outputting the time-step weight vectors of the three branches through Softmax. The second-level global gating introduces a Sigmoid gating to adaptively balance the stitched features of linear projection and the dynamically weighted features, and obtains the final fused features after layer normalization.

[0047] The prediction head will compress the time dimension of the final fused features through global average pooling, and then output the settlement prediction value of the target pile position (pile monitoring point 1) for the next time step through two layers of multilayer perceptrons.

[0048] In the model training and evaluation phase, this embodiment is a single-node prediction scenario (the number of predicted output nodes is equal to 1). The loss function uses a weighted combination of mean squared error (MSE) and mean absolute error (MAE). If the number of predicted output nodes is greater than 1, the loss function also needs to introduce spatial trend consistency loss, which maintains the spatial trend relationship by constraining the consistency between the predicted difference and the true difference between adjacent nodes. When the number of predicted output nodes is equal to 1, the trend consistency term is zero.

[0049] The model uses the Adam optimizer for parameter updates, and combines learning rate decay and early stopping mechanisms to prevent overfitting and accelerate convergence. Gradients are clipped during training to prevent gradient explosion.

[0050] Finally, experimental results on the settlement dataset show that the root mean square error (RMSE) of the method of this invention is 0.6051 mm, and the coefficient of determination (R²) is 0.9640. Ablation experiments further confirm the effectiveness of each module: the complete model reduces the RMSE by 8.1% compared to the combination without the Global Temporal Encoder (GTE) module, and reduces the RMSE by 26.0% compared to the combination without the Multi-Scale Receptive Field (MSRF) module.

[0051] Furthermore, to further verify the generalization ability of the present invention, the same model architecture was applied to the axial force monitoring data at the top of the same wind turbine foundation. Axial force data was collected at the same 15-minute intervals using pile top stress sensors, with units in kN. The data preprocessing and noise reduction procedures were consistent with those for settlement data. On the axial force dataset, the method of the present invention achieved an R² of 0.9818, verifying the good applicability of the method to different monitoring indicators.

[0052] Specifically, this embodiment selects 8 pile monitoring points (numbered 1 to 8, with pile monitoring point 1 being the target pile) of a certain unit in an onshore wind farm. Data is collected continuously for approximately one month at 15-minute sampling intervals, totaling 3072 time steps. The dataset is divided into a training set (80%), a validation set (10%), and a test set (10%) in chronological order. The model is trained using the Adam optimizer with an initial learning rate of 1e-3, a batch size of 64, 15 early stopping rounds, an input sequence length of 10 time steps, and a prediction length of 1.

[0053] Table 2 shows the RMSE and R² indices of the present invention and two ablation variants, namely, GTE removal and MSRF removal, on the test set at pile monitoring point 1.

[0054] Table 2. Comparison of ablation of test sets at monitoring point 1 of pile location.

[0055]

[0056] like Figure 6 As shown, the settlement prediction curve on the test set of pile monitoring point 1 in this invention is in high agreement with the measured curve, and there is no significant lag at key inflection points; as Figure 7 As shown, Figure 7 (a) in the figure is a comparison chart of the root mean square error of each module combination. Figure 7 (b) in the figure shows a comparison of the determination coefficients of each module combination. It can be seen that the RMSE increases significantly after removing the multi-scale receptive field module (MSRF) or the global temporal encoder module (GTE) separately, which confirms that local temporal modeling and global temporal modeling are not interchangeable and verifies the necessity of the synergy between the temporal dual branch and the DSGA spatial branch.

[0057] Example 2 This embodiment provides a wind turbine foundation prediction system based on adaptive noise reduction and spatiotemporal graph fusion, including: The data acquisition module is configured to acquire and preprocess monitoring data from multiple pile locations of the wind turbine foundation. The adaptive noise reduction module is configured to perform three-stage noise reduction on the preprocessed monitoring data using an adaptive noise reduction method. The three-stage noise reduction process includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing. The sample construction module is configured to: construct input feature sequences based on information leakage prevention strategies; The prediction module is configured to: input the input feature sequence into the spatiotemporal graph fusion network to obtain the prediction result; wherein, the spatiotemporal graph fusion network is configured to perform the following steps: extract spatial features through a dynamic spatial graph attention module, extract local temporal features through a multi-scale receptive field module, and extract global temporal features through a global temporal encoder module; fuse local and global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fuse spatial and temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and map the spatiotemporal fusion features to the predicted value of the target pile position through the prediction head.

[0058] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0059] Example 3 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion, characterized in that, include: Acquire and preprocess monitoring data from multiple pile locations of the wind turbine foundation; An adaptive noise reduction method is used to perform three-stage noise reduction on the preprocessed monitoring data. The three-stage noise reduction process includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing. An input feature sequence is constructed based on an information leakage prevention strategy. This input feature sequence is then fed into a spatiotemporal graph fusion network to obtain a prediction result. The processing flow of the spatiotemporal graph fusion network includes: extracting spatial features through a dynamic spatial graph attention module; extracting local temporal features through a multi-scale receptive field module; extracting global temporal features through a global temporal encoder module; fusing the local and global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fusing the spatial and temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and mapping the spatiotemporal fusion features to the predicted value of the target pile location through a prediction head.

2. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion according to claim 1, characterized in that, The acquisition of monitoring data for multiple pile positions of the wind turbine foundation includes: acquiring monitoring data for N pile positions arranged in a circular pattern on the wind turbine foundation, wherein the monitoring data includes settlement change values ​​or pile top axial force values, and is collected at fixed time intervals.

3. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion according to claim 1, characterized in that, The Fourier frequency domain filtering includes: estimating the noise level using a median absolute deviation estimator, calculating the signal-to-noise ratio (SNR), adaptively selecting the cutoff frequency ratio based on the SNR, constructing a soft cutoff filter to achieve a smooth transition from the passband to the stopband, and simultaneously applying attenuation to frequency components with power below a preset percentile threshold, and obtaining the filtered signal through inverse Fourier transform. The Bayesian adaptive Kalman filter includes: establishing a one-dimensional state-space model, updating the process noise covariance and observation noise covariance online through local window statistics, and obtaining a smooth signal through Kalman prediction and recursive update. The edge-preserving post-processing includes: detecting edge points using the 60th percentile of the differential signal as a threshold, preserving the frequency domain filtered signal at the edge with a 95% weight, fusing the Kalman smoothed signal with a 5% weight, and using the Kalman smoothed signal at non-edge locations.

4. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion according to claim 1, characterized in that, The information leakage prevention strategy specifically involves: excluding the target pile location from the input features, using the remaining N-1 pile locations as context nodes, constructing an input feature sequence using the historical denoising data of the context nodes, and predicting the state value of the target pile location at the next time step; the spatiotemporal graph fusion network constructs a spatial topology graph using the N-1 context pile locations as nodes.

5. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal map fusion according to claim 1, characterized in that, The extraction of spatial features through the dynamic spatial graph attention module specifically involves: for each time step, extracting the denoised input features of each node at the current time and encoding them into a high-dimensional space; dynamically generating an adjacency matrix through a multi-head self-attention mechanism; after symmetry and self-loop processing, extracting node embeddings through a multi-layer multi-head graph attention network; sending the time step sequence into a bidirectional temporal aggregation network for temporal aggregation; obtaining spatial features through linear projection and flattening; and preserving the original denoised input through a learnable residual shortcut.

6. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal map fusion according to claim 1, characterized in that, The extraction of local temporal features through the multi-scale receptive field module is specifically as follows: the multi-scale receptive field module adopts a causal dilated temporal convolutional network, stacks multiple residual blocks, and the inflation rate increases exponentially. Each residual block contains dilated convolution, batch normalization, activation function and Dropout, and outputs local temporal features that are linearly projected to a uniform dimension.

7. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion according to claim 1, characterized in that, The global time encoder module adopts a self-attention network with an encoder-decoder structure. It linearly maps the noise reduction input and adds position encoding. The encoder consists of multiple layers of self-attention and feedforward networks. The decoder applies a causal mask to prevent future information leakage and outputs global time features. The temporal fusion mechanism is specifically as follows: the local temporal features and the global temporal features are mutually enhanced through bidirectional cross attention, and the temporal fusion features are obtained after adaptive fusion with learnable Sigmoid gating. The spatiotemporal fusion mechanism is as follows: the spatial features and the temporal fusion features are input into three parallel attention branches, and temporal self-attention, spatial self-attention and spatiotemporal cross-attention are calculated respectively. The spatiotemporal fusion features are obtained by adaptive weighted fusion through a two-level network of dynamic gating and global gating.

8. The wind turbine foundation prediction method based on adaptive noise reduction and spatiotemporal graph fusion according to claim 1, characterized in that, The method further includes training the spatiotemporal graph fusion network using a hybrid loss function, wherein the hybrid loss function is a weighted sum of mean squared error and mean absolute error; when the number of predicted output nodes is greater than 1, the hybrid loss function also includes spatial trend consistency loss, which is used to constrain the consistency between the predicted difference and the actual difference between adjacent nodes.

9. A wind turbine foundation prediction system based on adaptive noise reduction and spatiotemporal graph fusion, characterized in that, include: The data acquisition module is configured to acquire and preprocess monitoring data from multiple pile locations of the wind turbine foundation. The adaptive noise reduction module is configured to perform three-stage noise reduction processing on the preprocessed monitoring data using an adaptive noise reduction method. The three-stage noise reduction processing includes Fourier frequency domain filtering, Bayesian adaptive Kalman filtering, and edge-preserving post-processing in sequence. The sample construction module is configured to: construct input feature sequences based on information leakage prevention strategies; The prediction module is configured to: input the input feature sequence into a spatiotemporal graph fusion network to obtain a prediction result; wherein, the spatiotemporal graph fusion network is configured to perform the following steps: extract spatial features through a dynamic spatial graph attention module, extract local temporal features through a multi-scale receptive field module, and extract global temporal features through a global temporal encoder module; fuse the local temporal features and the global temporal features through a temporal fusion mechanism to obtain temporal fusion features; fuse the spatial features and the temporal fusion features through a spatiotemporal fusion mechanism to obtain spatiotemporal fusion features; and map the spatiotemporal fusion features to the predicted value of the target pile position through a prediction head.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the wind turbine foundation prediction method of adaptive noise reduction and spatiotemporal graph fusion as described in any one of claims 1-8.

Images