METHOD AND SYSTEM FOR PREDICTING THE STRESS STATE OF PILE FOUNDATIONS OF A HIGH-PILE QUARTER BASED ON A TRANSFORMER

Abstract

The present invention provides a method and a system for predicting the stress state of pile foundations of a transformer-based high-pile quay. The method comprises: acquiring initial stress data sequences, which are acquired in real time by a sensor network installed on the pile foundations of the high-pile quay; performing data cleaning, alignment, smoothing, noise reduction, and standardization of the initial stress data sequences to obtain stress sequence data; using a long-short-term memory (LSTM) network based on the stress sequence data to obtain stress trend vectors; and using a graphical neural network based on the stress trend vectors to obtain spatial feature maps.Generating a comprehensive spatiotemporal feature set through multimodal feature fusion based on spatial feature maps and stress trend vectors; and creating a predictive model based on the transformer and combining it with the comprehensive spatiotemporal feature set to obtain predictive results for the stress state of the pile foundations of the high-pile quay. The present invention significantly improves the scientific rigor and reliability of decisions for the safe operation of pile foundations of the high-pile quay by combining spatiotemporal feature fusion and dynamic prediction.

Description

2. Interrupted changes occur as they approach failure. This continuous temporal change is inextricably linked to the spatial stress distributions at different locations and forms a complex spatiotemporal relationship. Since this spatiotemporal coupling cannot be precisely captured, existing technologies are often unable to accurately determine the location and extent of damage. For example, 5 during the operation of a quay, a fine crack can form in a pile foundation due to long-term tidal forces at the base. However, since the development process of this local stress concentration is not captured in time, the crack gradually propagates and eventually leads to the failure of the entire pile foundation, which entails significant safety risks for the quay. 10 Therefore, the question arises as to how the stress changes in the pile foundations can be used to detect and detect the resulting ruptured changes in the pile foundations. spatial correlation features are extracted, the stress state of the pile foundations can be accurately predicted, as well as the origin and severity of damage can be precisely predicted.a key factor for the safe operation of high pile quays.15 CONTENT OF THE PRESENT INVENTION To solve the problems of the prior art, the present invention provides a method and a system for predicting the stress state of pile foundations of a high pile quay based on a transformer. This invention aims to extract 20 spatiotemporal correlation features from the stress changes of the pile foundations, to accurately predict the stress state of the pile foundations, and to precisely predict the location and severity of damage to the pile foundations. To achieve the aforementioned objectives, the present invention offers the following 25 solutions. A method for predicting the stress state of pile foundations of a high pile quay based on a transformer, and the method comprises: capturing initial stress data sequences, which are captured in real time by a sensor network installed on the pile foundation of the high pile quay; performing data cleansing, alignment, smoothing,Noise suppression and standardization of the initial stress data sequences to obtain stress sequence data; using a long-short-term memory (LSTM) network based on the stress sequence data to obtain stress trend vectors; 35 BE2025 / 7079 3 using a graphical neural network based on the stress trend vectors to obtain spatial feature maps; generating a comprehensive spatiotemporal feature set through multimodal feature fusion based on the spatial feature maps and the stress trend vectors; and 5 creating a prediction model based on the transformer and combining it with the comprehensive spatiotemporal feature set to obtain prediction results of the stress state of the pile foundations of the high pile quay. Preferably, a method for performing data cleansing, alignment, smoothing, noise reduction, and standardization of the initial voltage data sequences to obtain data from voltage sequences,The following: removing outliers and filling in missing data in the initial voltage data sequences using dynamic threshold detection and linear interpolation methods to obtain cleaned data; performing high-precision time alignment of the cleaned data using PrecisionTimeProtocol (PTP); performing smoothing and noise reduction of the aligned data using Kalman filters and wavelet thresholding methods; and performing standardization of the smoothed and noise-reduced data using Z-score standardization or normalization methods in combination with a dynamic adjustment strategy to obtain the data from voltage sequences. Preferably, the LSTM network comprises: an input layer, a hidden layer, and an output layer, wherein the hidden layer contains multiple LSTM units and multiple fully networked layers; dropout layers are inserted between LSTM units.so that the LSTM network discards neurons according to a predefined probability, while a LeakyReLU activation function is used between fully networked layers: ; where x represents input data. 30 Preferably comprises a method for generating a prediction model based on the transformer and combining it with the comprehensive spatiotemporal feature set to obtain the prediction results of the stress state of the pile foundations of the high pile quay, as follows: BE2025 / 7079 4 Generating a prediction model by inputting the comprehensive spatiotemporal feature set into an improved hierarchical transformer architecture to obtain an accurate output for a predicted stress state of the pile foundations of the high pile quay; A loss function of the improved hierarchical transformer architecture for creating a prediction model consists of four components: Reconstruction Loss Lrecon, Kullback-Leibler Divergence Lkl,Cross-entropy loss Lce and deviation regression error loss Lreg. ; ;10 Total loss function L: ; where Ne represents the number of features, Ne is a preliminary prediction result, Ni is a true value, Ni denotes a posterior distribution, N(0,1) is a standard normal distribution, Ne is a distribution of true values, and Nq is a distribution of multimodal prediction results, and the coefficients of the individual loss components γάβα,,, can be fitted. The present invention further provides a system for predicting the stress state of pile foundations of a high pile quay based on a transformer, which is used to implement the above-mentioned method. The system comprises: a data acquisition module, a preprocessing module, an extraction module of the temporal features, an analysis module of the spatial features, a fusion module of the features, and a prediction module; the data acquisition module serves to acquire the initial stress data sequences.which is detected in real time by a sensor-25 network installed in the pile foundation of the high pile quay; BE2025 / 7079 5 the preprocessing module is used to perform data cleaning, alignment, smoothing, noise reduction and standardization of the initial voltage data sequences in order to obtain the data of voltage sequences; the temporal feature extraction module is used to employ a Long-Short-Term Memory (LSTM) network based on the data of voltage sequences to obtain the 5 trend vectors of the voltages; the spatial feature analysis module is used to employ a graphical neural network based on the trend vectors of the voltages,to obtain the spatial feature maps; the feature fusion module serves to generate a comprehensive spatiotemporal feature set through multimodal feature fusion based on the spatial feature maps and the stress trend vectors; and the prediction module serves to create a prediction model based on the transformer and combine it with the comprehensive spatiotemporal feature set to obtain the predicted stress state results of the pile foundation of the high pile quay. Preferably, the preprocessing module comprises a data cleaning unit, an alignment unit, a smoothing and noise reduction unit, and a standardization unit; the data cleaning unit serves to remove outliers and fill in missing data in the initial stress data sequences using dynamic threshold detection and linear interpolation methods.to obtain the cleaned data; the alignment unit is used to perform high-precision temporal alignment of the cleaned data using PTP; 25 the smoothing and noise reduction unit is used to perform smoothing and noise reduction of the aligned data using Kalman filters and wavelet thresholding methods; and the standardization unit is used to perform standardization of the smoothed and noise-reduced data using Z-score standardization or 30 normalization methods in combination with the dynamic adjustment strategy to obtain the data of voltage sequences. Preferably, the LSTM network comprises: the input layer, the hidden layer, and the output layer, wherein the hidden layer contains the multiple LSTM units and the multiple fully networked layers; the dropout layers are inserted between the LSTM units so that the BE2025 / 7079 6 LSTM network discards the neurons according to the predefined probability,while the LeakyReLU activation function is used between the fully interconnected layers: ; where x represents the input data. 5 Preferably, the method for creating a prediction model based on the Transformer and combining it with the comprehensive spatiotemporal feature set to obtain the prediction results of the stress state of the pile foundation of the high pile quay comprises the following: Creating a prediction model by inputting the comprehensive spatiotemporal feature set into the improved hierarchical Transformer architecture to obtain the precise output for the predicted stress state of the pile foundation of the high pile quay; The loss function of the improved hierarchical transformer architecture for constructing the prediction model consists of four components: Reconstruction loss Lrecon, Kull-15 back-Leibler divergence Lkl, Cross-entropy loss Lce, and Deviation regression error loss Lreg. ; ; Total loss function L: 20 ; where N represents the number of features, the preliminary prediction result is,The true value is denoted as the posterior distribution, N(0,1) as the standard normal distribution, p as the distribution of the true values, and q as the distribution of the multimodal prediction results, and the coefficients of the individual loss components γάβα can be adjusted. Compared to the prior art, the present invention offers the following advantages. The present invention discloses an innovative method for the safety monitoring and damage prediction of pile foundations of a high pile quay. Eskon-5 focuses on overcoming operational challenges through the fusion of multidimensional time series data and deep learning techniques, thereby enabling the analysis of temporal development patterns of stresses in pile foundations and the precise identification of potential damage. The present invention acquires stress data in real time from a sensor network installed in the pile foundations, performs data cleaning, alignment, smoothing, noise reduction, and standardization,to obtain stress data sequences. Subsequently, dynamic trends in the time dimension are captured using an LSTM network and, in combination with a graphical neural network, spatial distribution features are extracted to form a comprehensive spatiotemporal feature set. By integrating this with a transformer-based prediction model, the prediction of the stress state of the pile foundations of the high pile quay is realized, thereby obtaining predictive results for the location and severity of damage to pile foundations. The present invention, through the combination of spatiotemporal feature fusion and dynamic prediction, significantly improves the scientific rigor and reliability of decisions for the safe operation of pile foundations of the high pile quay and offers a Technical safeguards for the long-term stability of infrastructures. DESCRIPTION OF THE DRAWING To illustrate the technical solution of the present invention more clearly, 25 chen,The drawings used in the embodiments are briefly presented below, and it is evident that the drawings accompanying the following descriptions are only some embodiments of the present invention, and persons skilled in the art can obtain further drawings according to these drawings without any inventive effort being involved. 30 Fig. 1 is a schematic representation of a method for predicting the stress state of pile foundations of a raised pile quay based on a transformer of the present invention; and BE2025 / 7079 8 Fig. 2 is a schematic diagram of a system for predicting the stress state of pile foundations of a raised pile quay based on a transformer of the present invention. DETAILED DESCRIPTION 5 The technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the attached drawings in the embodiments of the present invention, and these are obviously,that the described embodiments represent only a part of the embodiments of the present invention and not all embodiments. Starting from the embodiments of the present invention, all other embodiments that can be achieved by a person skilled in the art without inventive work fall within the scope of protection of the present invention. In order to present the aforementioned objectives, features and advantages of the present invention more clearly and understandably, a further detailed description of the invention is given below with reference to the attached drawings and specific embodiments. Embodiment 1 As shown in Fig. 1, the present invention provides a method for predicting 20 the stress state of pile foundations of a high pile quay based on a transformer, comprising: acquiring initial stress data sequences, which are acquired in real time by a sensor network installed on the pile foundation of the high pile quay; performing data cleansing, alignment, smoothing,Noise suppression and standardization of the initial stress data sequences to obtain stress sequence data; using a long-short-term memory (LSTM) network based on the stress sequence data to obtain stress trend vectors; using a graphical neural network based on the stress trend vectors to obtain spatial feature maps; generating a comprehensive spatiotemporal feature set by multimodal feature fusion based on the spatial feature maps and the stress trend vectors; and creating a prediction model based on the transformer and combining it with the comprehensive spatiotemporal feature set to obtain prediction results of the stress state of the pile foundations of the high pile quay. The specific implementation process of this invention is as follows. S1, Acquisition of initial voltage data sequences,which is acquired in real time by a sensor network installed at the pile foundations of the high pile quay. Fiber Bragg grating (FBG) sensors or strain gauge sensor networks are installed at the pile foundations of the high pile quay. The nodes of the sensor network are arranged according to the spatial topology and cover stress-sensitive areas such as the pile head, the middle pile section, and the pile base. Data on axial stress, bending stress, and shear stress are acquired in real time with predefined sampling frequencies for important cross-sections of the pile foundations of the high pile quay. At the same time, the spatial positions of the sensor nodes are acquired to obtain initial stress data sequences. S2, Performing data cleansing, alignment, smoothing, noise reduction, and standardization of the initial voltage data sequences to obtain data from voltage sequences. For the initial voltage data sequences,The sensor network installed in the pile foundation of the high pile quay, which is acquired in real time, is first cleaned: Outliers caused by temporary sensor failures or environmental disturbances are identified and removed by dynamic threshold detection. At the same time, missing data segments are supplemented by linear interpolation or voltage field reconstruction algorithms based on physical models to ensure temporal continuity. Subsequently, the cleaned data undergoes high-precision temporal alignment. The timestamps of the individual nodes are synchronized using the PTP (Precision Time Protocol), and the unevenly sampled data are unified into a fixed timestamp sequence by spline interpolation, thus achieving data alignment. During the smoothing and noise reduction phase, a Kalman filter is set to -30.to smooth data fluctuations. In combination with the frequency range properties of the voltage signal, wavelet threshold noise reduction is applied to separate the effective voltage waveform components from high-frequency noise. Finally, a standardization strategy is dynamically adapted to the operating conditions of the quays: sliding window and Z-score standardization serves to eliminate dimensional differences in long-term monitoring data, while peak normalization is used for transient shock events such as ship docking to highlight the local voltage concentration effect. The processed voltage sequence data are output in the form of an [N×T]-dimensional tensor, where the spatial dimension N preserves the sensor topology relationships, while the temporal dimension T embeds combined time-frequency domain features. S3,Use a Long-Short-Term-Memory (LSTM) network based on the voltage sequence data,to obtain trend vectors of the voltages. Based on the preprocessed voltage sequence data, a Long Short-Term Metric-10 mory (LSTM) network is used to model the dynamic evolution trends of the voltage data. The LSTM network is fed with time-series voltage data from individual sensor nodes. Through multiple layers of LSTM units, it extracts dependency features across different time scales, encompassing both short-term fluctuations such as temporary shock loads and long-term evolution such as voltage relaxation due to foundation settlement. LSTM networks represent an improved variant of conventional recurrent neural networks (RNNs). By incorporating memory modules, they exhibit more complex activation mechanisms for nodes in the hidden layers.This enables effective learning of long-term dependencies. This solves the problems of gradient explosion and gradient disappearance in conventional RNNs. The LSTM network of the present invention comprises an input layer, a hidden layer, and an output layer. The hidden layer consists of several LSTM units and several fully interconnected layers. The neural network formed by the LSTM units can be viewed as an extremely complex nonlinear function. Through data training, its internal weights are adjusted to approximate the input data, thereby achieving the purpose of data recognition. To prevent overfitting in the LSTM network, a dropout layer is added between the LSTM cells in the present invention. This causes the LSTM network to discard neurons according to a preset probability.thereby avoiding excessive dependence on certain local features. At the same time, the LeakyReLU activation function is used between fully interconnected layers: ; where x represents input data. 35 BE2025 / 7079 11 S4, Use a graphical neural network based on the trend vectors of the stresses to obtain spatial feature maps. Based on the stress trend vectors extracted by LST, a graphical neural network (GNN) is used to model the spatial topological relationships of the pile foundations of a high pile quay in order to generate a spatial feature map with 5 physical meaning. Specifically, each sensor node serves as a node of the map,where the node features are initialized by the stress trend vectors output by LSTM. The edge weights are determined based on the physical connection relationships of the pile foundations or data-driven similarity metrics.10 Using the Message Passing Neural Network (MPNN) framework, neighborhood information is aggregated through multi-layered graph convolutions. For example, a graph attention mechanism (GAT) is employed to dynamically learn the weights between nodes and thus highlight key nodes in the stress transfer path. Additionally, edge features such as material properties and 15 angular orientations between pile foundations are introduced to enhance the physical interpretability of the spatial modeling. By embedding Gated Temporal Convolutions (TCNs) between GNN layers, the time-spatial state of the node features is synchronously updated, thus preventing a loss of time series information due to spatial aggregation.20 S5,Generate a comprehensive spatiotemporal feature set through multimodal feature fusion based on the spatial feature maps and the stress trend vectors. Based on the spatial feature maps generated by GNN and the stress trend vectors output by LSTM25, a comprehensive spatiotemporal feature set for the pile foundations of the high pile quay is generated through multimodal feature fusion and correlation analysis. Specifically, this includes the following: graph node centrality analysis is used to quantify the criticality of each pile foundation in the stress network, to identify potential weak nodes, to detect local spatial patterns of stress anomaly propagation through subgraph clustering, and to obtain spatial correlation indicators. The Dynamic Time Warping (DTW) distance matrix of the stress trend vectors is calculated,to analyze the synchronicity or delay of stress development in different pile foundations. The common time-frequency domain features are extracted as temporal correlation indicators to characterize the non-stationary properties of the stress fluctuations. Based on the spatiotemporal cross-attention (ST cross-attention) mechanism, spatial correlation indicators and temporal correlation indicators are dynamically weighted to generate a comprehensive spatiotemporal feature set. This determines the interaction relationships between the spatial feature map and the stress trend vectors, which are used for the subsequent end-to-end prediction of the transformer. S6, Creating a prediction model based on the transformer and combining it with the comprehensive spatiotemporal feature set,To obtain prediction results of the stress state of the pile foundation of the high pile quay, the comprehensive time-spatial feature set is fed into a hierarchical transformer architecture to achieve a precise output for a prediction of the stress state of the pile foundation of the high pile quay. Specifically, this includes the following: The present invention creates a prediction model based on an improved hierarchical transformer architecture: within the window, ST cross-attention is performed, where the interaction weights are calculated based on the query vector (Q), the key vector (K), and the value vector (V): Attention(Q,K,V) = Softmax((QKT) / √d + Mmask)V. The feature dimension is , and the mask matrix Mmask is generated from the mechanical connection relationship of the pile foundation. The loss function of the overall model consists of four components: reconstruction loss Lrecon, Kullback-Leibler divergence Lkl,Cross-entropy loss Lce and deviation regression error loss Lreg. The reconstruction loss is used to calculate the L2 distance between the preliminary prediction result and the true value in order to provide an accurate initial prediction. The KL divergence is used to measure the similarity between the posterior distribution N and the standard normal distribution N(0,1) of endpoints, thereby learning the latent distribution of endpoints of multimodal stress. The cross-entropy loss measures the similarity between the distribution of true values ​​ep and the distribution of multimodal prediction results q in order to determine the optimal strategy for predicting multimodal stress states. The deviation regression loss is used to correct the initial prediction. BE2025 / 7079 13 prediction results, in order to obtain a more precise final prediction output. The specific formulas are as follows: ; ; Total loss function L:5 ; where N represents the number of features, and the coefficients of the individual loss components γάβα,,,can be adapted. Furthermore, reinforcement learning (RL) has achieved significant results in the area of ​​decision optimization. By integrating RL into Transformer models, a reinforcement learning-based prediction and optimization model for the stress state of pile foundations of high pile quays can be created. By introducing a reward function, the model can dynamically adapt its prediction strategy to maximize overall benefit. For example, a DeepQ Network (DQN-TF)-based prediction model could be developed that improves the model's prediction performance through continuous optimization of the policy network. The specific framework can be represented as follows: S stands for the current state and Q-network for the policy network.20 Furthermore, based on the predicted stress state results of the pile foundations of the high pile quay, a multi-scale damage identification algorithm can be used.to predict the location and severity of damage to the pile foundations of the high pile quay and thus achieve an accurate prediction of the location and severity of damage to the pile foundations. 25 BE2025 / 7079 14 In summary, the present invention discloses a method for predicting the stress state of pile foundations of a high pile quay based on a transformer. It focuses on overcoming operational challenges through the fusion of multidimensional time series data and deep learning techniques, thereby enabling the analysis of spatiotemporal development patterns of stresses in pile foundations and the precise identification of potential damage. The present invention acquires in real time stress data from a sensor network installed in the pile foundations werks performs data cleansing, alignment, smoothing, noise reduction, and standardization.to obtain stress data sequences. Subsequently, dynamic trends in the time dimension are captured using an LSTM network and, in combination with a graphical neural network, spatial distribution features are extracted to form a comprehensive spatiotemporal feature set. By integrating with a transformer-based prediction model, the predicted stress state of the pile foundations of the high pile quay is realized, thereby obtaining prediction results for the location and severity of damage to pile foundations. The present invention, through the combination of spatiotemporal feature fusion and dynamic prediction, significantly improves the scientific rigor and reliability of decisions for the safe operation of pile foundations of the high pile quay and provides technical assurance for the long-term stability of infrastructures. 20 Embodiment 2 As shown in Fig. 2,The present invention, based on the same inventive concept, further provides a system for predicting the stress state of pile foundations of a high pile quay based on a transformer, which is used to implement the above-mentioned method. The system comprises: a acquisition module, a preprocessing module, an extraction module of the temporal features, an analysis module of the spatial features, a fusion module of the features, and a prediction module; the acquisition module serves to acquire the initial stress data sequences, which are acquired in real time by a sensor network installed at the pile foundation of the high pile quay; the preprocessing module serves to perform data cleansing, alignment, smoothing, noise reduction, and standardization of the initial stress data sequences in order to obtain the stress sequence data; BE2025 / 7079 15 the extraction module of the current features serves to utilize a Long-Short-Term-Memory (LSTM) network based on the data of voltage sequences,to obtain the trend vectors of the stresses; the spatial feature analysis module is used to employ a graphical neural network based on the trend vectors of the stresses to obtain the spatial feature maps; the feature fusion module is used to generate a comprehensive spatiotemporal feature set through multimodal feature fusion based on the spatial feature maps and the trend vectors of the stresses; and the prediction module is used to create a prediction model based on the transformer and combine it with the comprehensive spatiotemporal feature set to obtain the prediction results of the stress state of the pile foundation of the high pile quay. Furthermore, the preprocessing module includes a data cleaning unit, an alignment unit,a smoothing and noise reduction unit and a standardization unit; the data cleaning unit is used to remove outliers and fill in missing data in the initial voltage data sequences using dynamic threshold detection and linear interpolation methods to obtain the cleaned data; the alignment unit is used to perform high-precision temporal alignment of the cleaned data using PTP; the smoothing and noise reduction unit is used to perform smoothing and noise reduction of the aligned data using Kalman filters and wavelet thresholding methods; and the standardization unit is used to perform standardization of the smoothed and noise-reduced data using Z-score standardization or normalization methods in combination with the dynamic adjustment strategy to obtain the data of voltage sequences. Furthermore, the LSTM network includes: 30 the input layer, the hidden layer, and the output layer,where the hidden layer contains the multiple LSTM units and the multiple fully networked layers; the dropout layers are inserted between the LSTM units so that the LSTM network discards the neurons according to the predefined probability, while the LeakyReLU activation function is used between the fully networked layers: BE2025 / 7079 16 ; where x represents the input data. Furthermore, the procedure for creating a prediction model based on the transformer and combining it with the comprehensive spatiotemporal feature set to obtain the prediction results of the stress state of the pile foundations of the high pile quay 5 comprises the following: creating a prediction model by inputting the comprehensive spatiotemporal feature set into the improved hierarchical transformer architecture,to achieve the precise output for the predicted stress state of the pile foundation of the high pile quay; 10 the loss function of the improved hierarchical transformer architecture for constructing the prediction model consists of four components: reconstruction loss Lrecon, Kull-back Leibler divergence Lkl, cross-entropy loss Lce and deviation regression error loss Lreg. ; 15 ; Total loss function L: ; where N represents the number of features that the predicted,