Deep foundation pit space-time deformation prediction method and system
The spatiotemporal deformation prediction method for deep foundation pits, which combines large language models and physical constraints, solves the problems of multimodal information fusion and cross-project generalization in deep foundation pit construction. It achieves high-precision, interpretable dynamic risk warning and intelligent decision support, thereby improving engineering safety and construction efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies struggle to achieve real-time and accurate spatiotemporal deformation prediction in deep foundation pit construction, especially in complex urban environments. They also fail to effectively integrate multimodal engineering information and cross-project generalization, resulting in delayed risk warnings and a lack of physical consistency.
A large language model-driven approach is adopted to generate geological semantic embedding vectors by semantically parsing geological exploration reports. Combined with multi-source monitoring data, cross-engineering domain invariant features are generated using domain adversarial neural networks, embedded in an adaptive graph convolutional recurrent network, and a physical constraint loss function is introduced to achieve dynamic risk warning and intelligent decision support.
It significantly improves the prediction accuracy and physical consistency of deep foundation pit construction, provides efficient risk warning and intelligent decision support, enhances engineering safety and construction efficiency, and has strong generalization ability.
Smart Images
Figure CN121809457A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deformation prediction technology, specifically to a method and system for predicting spatiotemporal deformation of deep foundation pits, and more particularly to a method and system for predicting spatiotemporal deformation of deep foundation pits based on physical constraints and a large language model. Background Technology
[0002] In deep foundation pit construction, deformation prediction and risk control are core aspects of ensuring project safety. Accurately predicting the spatiotemporal deformation evolution trends of the retaining structure and surrounding surface is crucial not only for the stability of the foundation pit itself but also for the safety of adjacent buildings, underground pipelines, and transportation facilities. It is a key prerequisite for achieving dynamic risk warning and intelligent decision support throughout the entire process. Existing technologies mainly rely on three types of methods, but all have significant limitations and cannot meet the comprehensive requirements of modern urban underground engineering for real-time performance, physical consistency, cross-scenario generalization capabilities, and multi-source information fusion. 1. Risk assessment methods based on empirical rules: Traditional construction risk assessment relies heavily on engineers' experience or analogies from historical cases. Such methods lack fine-grained perception of on-site geological conditions (such as soil layer thickness, type, mechanical parameters, etc.) and cannot efficiently perceive the spatiotemporal evolution trend of deformation implied in historical deformation data, resulting in delayed risk warnings, strong subjectivity, and difficulty in adapting to the complex and ever-changing urban underground environment.
[0003] 2. Static Numerical Simulation-Based Analysis Methods: While numerical models, such as the Finite Element Method (FEM), can reflect the mechanical response mechanism of soil to some extent, they generally suffer from low computational efficiency and high sensitivity to material parameters. These methods are typically based on static or quasi-static assumptions, making it difficult to effectively integrate spatiotemporal monitoring data acquired in real-time during construction. More importantly, their predictive performance heavily relies on the accurate input of specific engineering parameters. In the absence of sufficient measured data or when parameter uncertainties are significant, the predicted results often deviate significantly from the actual deformation.
[0004] 3. Purely Data-Driven Monitoring Data Analysis Methods: In recent years, deep learning models (such as LSTM and GRU) have been used to process monitoring time-series data. However, these methods are essentially "black box" models, ignoring the fundamental physical laws of geotechnical engineering and easily producing prediction results that violate soil mechanics principles. Furthermore, they typically utilize only numerical monitoring data, failing to effectively leverage the rich geological semantic information contained in unstructured text data, thus limiting the model's generalization ability and weakening its perception of key physical conditions. In addition, due to significant differences in geological conditions among different foundation pit projects, existing models generally suffer from weak cross-project generalization ability, resulting in a sharp decline in performance when deployed in new projects.
[0005] While some studies have attempted to alleviate these problems by introducing Physical Information Neural Networks (PINNs) or domain adaptation techniques, no method has yet been found that can systematically integrate multimodal engineering information (numerical monitoring data + unstructured text) with embedded dynamic physical constraints and achieve a unified spatiotemporal deformation framework with strong generalization capabilities. In particular, there is a lack of a technical solution that can organically combine the semantic understanding capabilities of Large Language Models (LLMs), the generalization capabilities of domain adversarial learning, and physical constraint mechanisms to support dynamic risk prediction and intelligent decision-making throughout the entire deep foundation pit construction process.
[0006] To overcome the aforementioned shortcomings, this invention proposes a method and system for predicting and controlling the spatiotemporal deformation of deep foundation pits driven by physical information and a large language model. This method constructs a unified framework integrating large language model semantic encoding, domain adversarial feature interaction, physical constraint spatiotemporal prediction, and a three-stage collaborative training strategy. This achieves high-precision, highly generalizable, and interpretable dynamic risk warning, providing core technical support for the intelligent safety management of deep foundation pit projects. Summary of the Invention
[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting the spatiotemporal deformation of deep foundation pits.
[0008] A method for predicting the spatiotemporal deformation of deep foundation pits according to the present invention includes: Step S1: Obtain the geological survey report text and multi-source monitoring time series data for the deep foundation pit project; Step S2: Use a large language model to perform semantic parsing on the geological exploration report text and generate geological semantic embedding vectors; Step S3: Input the geological semantic embedding vector and multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features; Step S4: Input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and embed the physical constraint loss function of the formation loss theory during the training process of the model to jointly optimize the prediction accuracy loss and physical consistency loss. Step S5: Based on the trained model, perform spatiotemporal joint prediction of wall deformation and ground settlement, and output dynamically updated formation loss rate parameters.
[0009] Preferably, the large language model includes Qwen3-Embedding-4B, which extracts high-dimensional embedding vectors by semantically parsing the stratigraphic descriptions, soil distributions, and rock and soil parameters in the geological exploration report, and uses the embeddings corresponding to the EOS words at the end of the report as the comprehensive semantic representation of the entire geological exploration report.
[0010] Preferably, the domain adversarial neural network includes a feature extractor and a domain discriminator, which are optimized collaboratively through an adversarial training mechanism. The feature extractor minimizes the sum of prediction accuracy loss and adversarial loss, while the domain discriminator maximizes domain classification accuracy.
[0011] Preferably, the adaptive graph convolutional recurrent network includes a node adaptive parameter learning module and a data adaptive graph generation module; the adaptive parameter learning module generates exclusive convolution weights for each monitoring node, and the data adaptive graph generation module dynamically constructs a spatial relationship graph between nodes.
[0012] Preferably, the physical constraint loss function is based on the formation loss theory, and calculates the predicted formation loss rate by numerically integrating the predicted wall deformation with the surface settlement curve. m pred and compared with the measured value m act By comparing the results, a physical consistency loss is constructed, and it is then weighted and fused with the prediction accuracy loss to form a hybrid training objective.
[0013] Preferably, a three-stage training strategy is adopted. In the first stage, the domain adversarial neural network and the adaptive graph convolutional recurrent network are trained together to achieve domain-invariant feature learning. In the second stage, the feature extractor is frozen and the adaptive graph convolutional recurrent network is optimized only with accuracy loss. In the third stage, physical constraint loss is introduced to jointly optimize prediction accuracy and physical consistency.
[0014] Preferably, the multi-source monitoring time series data includes wall horizontal displacement and ground settlement, which are unified into fixed dimensions after engineering-independent preprocessing. The original monitoring points are then curve-fitted and resampled at equal intervals along the normalized depth / horizontal axis to generate a virtual monitoring sequence with consistent dimensions.
[0015] A deep foundation pit spatiotemporal deformation prediction system provided by the present invention includes: The data acquisition module is used to acquire the geological survey report text and multi-source monitoring time series data for deep foundation pit projects; The large language model encoding module uses a large language model to perform semantic parsing on the geological exploration report text and generate geological semantic embedding vectors. The domain adversarial training module is used to input the geological semantic embedding vector and the multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features. The physical information-enhanced spatiotemporal deformation prediction module is used to input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and to embed a physical constraint loss function based on the formation loss theory during the model training process to jointly optimize the prediction accuracy loss and physical consistency loss; based on the trained model, spatiotemporal joint prediction of wall deformation and surface settlement is performed, and dynamically updated formation loss rate parameters are output.
[0016] Preferably, the large language model includes Qwen3-Embedding-4B, which extracts high-dimensional embedding vectors by semantically parsing the stratigraphic descriptions, soil distributions, and rock and soil parameters in the geological exploration report, and uses the embeddings corresponding to the EOS words at the end of the report as the comprehensive semantic representation of the entire geological exploration report.
[0017] Preferably, the domain adversarial neural network includes a feature extractor and a domain discriminator, which are optimized collaboratively through an adversarial training mechanism. The feature extractor minimizes the sum of prediction accuracy loss and adversarial loss, while the domain discriminator maximizes domain classification accuracy.
[0018] Preferably, the adaptive graph convolutional recurrent network includes a node adaptive parameter learning module and a data adaptive graph generation module; the adaptive parameter learning module generates exclusive convolution weights for each monitoring node, and the data adaptive graph generation module dynamically constructs a spatial relationship graph between nodes.
[0019] Preferably, the physical constraint loss function is based on the formation loss theory, and calculates the predicted formation loss rate by numerically integrating the predicted wall deformation with the surface settlement curve. m pred and compared with the measured value m act By comparing the results, a physical consistency loss is constructed, and it is then weighted and fused with the prediction accuracy loss to form a hybrid training objective.
[0020] Preferably, a three-stage training strategy is adopted. In the first stage, the domain adversarial neural network and the adaptive graph convolutional recurrent network are trained together to achieve domain-invariant feature learning. In the second stage, the feature extractor is frozen and the adaptive graph convolutional recurrent network is optimized only with accuracy loss. In the third stage, physical constraint loss is introduced to jointly optimize prediction accuracy and physical consistency.
[0021] Preferably, the multi-source monitoring time series data includes wall horizontal displacement and ground settlement, which are unified into fixed dimensions after engineering-independent preprocessing. The original monitoring points are then curve-fitted and resampled at equal intervals along the normalized depth / horizontal axis to generate a virtual monitoring sequence with consistent dimensions.
[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention organically integrates large language models, domain adversarial learning, and physical information neural networks to construct a multimodal, strong generalization, and high-reliability deformation prediction framework for deep foundation pit engineering. By semantically parsing geological text, generating cross-engineering domain invariant features, and embedding geotechnical physical mechanisms, it significantly improves the model's prediction accuracy, physical consistency, and cross-project transferability in complex urban environments.
[0023] 2. The method provided by this invention not only overcomes the shortcomings of subjective experience judgment, static lag in traditional numerical simulation, and poor physical interpretability of pure data-driven models, but also provides dynamic risk warning and intelligent decision support for deep foundation pit construction that combines numerical accuracy and physical reliability. It significantly improves engineering safety, construction efficiency and intelligent management level, and has outstanding engineering practical value and promotion prospects. Attached Figure Description
[0024] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the implementation of the spatiotemporal deformation prediction method for deep foundation pits based on physical constraints and a large language model driven by the present invention.
[0025] Figure 2 This is a schematic diagram of the structure of the system of the present invention and the relationship between its modules.
[0026] Figure 3 This is a schematic diagram of the large language model encoding module of the present invention.
[0027] Figure 4 This is a schematic diagram of the adversarial training module in the field of this invention.
[0028] Figure 5 This is a schematic diagram of the spatiotemporal deformation prediction module with enhanced physical information of the present invention.
[0029] Figure 6 This is a schematic diagram of the spatiotemporal deformation prediction results of deep foundation pits according to the present invention. Detailed Implementation
[0030] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0031] like Figure 1 As shown, a method for predicting the spatiotemporal deformation of deep foundation pits includes: Step S1: Obtain the geological survey report text and multi-source monitoring time series data for the deep foundation pit project; Multi-source monitoring time series data includes wall horizontal displacement and ground settlement. After engineering-independent preprocessing, the data is unified into a fixed dimension: curve fitting is performed on the original monitoring points and resampling is performed at equal intervals along the normalized depth / horizontal axis to generate a virtual monitoring sequence with consistent dimensions to support unified modeling across projects.
[0032] Step S2: Semantic parsing of the geological exploration report text is performed using a large language model to generate geological semantic embedding vectors; the large language model is Qwen3-Embedding-4B. It extracts high-dimensional embedding vectors by semantically parsing the stratigraphic descriptions, soil type distributions, and rock and soil parameters in the geological exploration report, using the embeddings corresponding to the EOS (End-of-Sentence) tokens at the end of the report as the comprehensive semantic representation of the entire geological exploration report.
[0033] Step S3: Input the geological semantic embedding vector and multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features; Domain Adversarial Neural Networks (DANNs) consist of a feature extractor and a domain discriminator, which are optimized collaboratively through an adversarial training mechanism: the feature extractor minimizes the sum of prediction accuracy loss and adversarial loss, while the domain discriminator maximizes domain classification accuracy. This mechanism generates domain-invariant features (DIFs) that have strong generalization performance for different foundation pit projects, thereby improving the model's generalization ability in unseen engineering scenarios.
[0034] Step S4: Input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and embed the physical constraint loss function of the formation loss theory during the training process of the model to jointly optimize the prediction accuracy loss and physical consistency loss. The Adaptive Graph Convolutional Recurrent Network (AGCRN) includes a Node Adaptive Parameter Learning (NAPL) module and a Data Adaptive Graph Generation (DAGG) module: the NAPL module generates exclusive convolution weights for each monitoring node, and the DAGG module dynamically constructs a spatial relationship graph between nodes, thereby jointly modeling the spatiotemporal heterogeneity and spatial dependence of wall and surface deformation.
[0035] Step S5: Based on the trained model, perform spatiotemporal joint prediction of wall deformation and ground settlement, and output dynamically updated formation loss rate parameters.
[0036] The physical constraint loss function is based on the formation loss theory. It calculates the predicted formation loss rate by numerically integrating the predicted wall deformation with the surface settlement curve. m pred and compared with the measured value m act Compare and construct physical consistency loss L phyand compare it with the prediction accuracy loss. L acc Weighted fusion to form a hybrid training objective L 2. This ensures both prediction accuracy and physical consistency.
[0037] (1) (2) (3) In the formula, and Representing the first i The true and predicted values of formation loss rate for each sample; and Representing the first i The true and predicted values of the deformation of a sample retaining wall; n The total number of samples; yes L phy The weight of is set to 0.3 in this study.
[0038] The method of this invention adopts a three-stage training strategy: the first stage is to co-train DANN and AGCRN to achieve domain-invariant feature learning; the second stage is to freeze the feature extractor and optimize AGCRN only with accuracy loss; the third stage introduces physical constraint loss to jointly optimize prediction accuracy and physical consistency, ensuring training stability and physical interpretability.
[0039] In one embodiment, the method is deployed on an intelligent monitoring platform, and the prediction results are visualized in the form of a spatiotemporal prediction map, which displays the deformation trend of the wall and the surface and the evolution of the formation loss rate in real time, providing engineers with interpretable and traceable decision-making basis.
[0040] Example 1 A method for predicting the spatiotemporal deformation of deep foundation pits based on physical constraints and a large language model includes: Step S1: Obtain the geological survey report text and multi-source monitoring time series data of the target deep foundation pit project. The monitoring data includes the spatiotemporal sequence of retaining wall deformation and surface settlement. The geological survey report needs to describe the geological conditions of the site in detail, including soil layer type, thickness, groundwater level, etc. Multi-source monitoring time-series data includes the spatiotemporal sequences of horizontal displacement of the retaining wall and ground settlement. These data are collected in real time by high-precision automated sensors and transmitted to the central processing system via an industrial-grade wireless network to ensure the integrity, timeliness, and representativeness of the project.
[0041] Step S2: The geological survey report is subjected to deep semantic parsing using a Large Language Model (LLM) to transform it into a high-dimensional geological semantic embedding vector rich in stratigraphic structure, soil parameters and engineering background information. This embodiment uses the pre-trained large language model Qwen3-Embedding-4B to perform deep semantic parsing on geological exploration reports, transforming natural language descriptions into high-dimensional geological semantic embedding vectors. These vectors implicitly encode prior knowledge such as soil layer sequences, rock and soil types, and engineering background, effectively compensating for the shortcomings of pure numerical models in utilizing unstructured geological information.
[0042] Step S3: Input the geological semantic embedding vector and the monitoring time series data into the Domain-Adversarial Neural Network (DANN). Through the adversarial training mechanism between the feature extractor and the domain discriminator, generate domain-invariant features (DIFs) that are insensitive to the engineering domain.
[0043] Domain Adversarial Neural Networks (DANNs) use an adversarial training mechanism to enable the feature extractor to generate domain-invariant features (DIFs) that are insensitive to the source of the engineering project. This significantly reduces the data distribution shift caused by differences in geological conditions between different foundation pits, thereby greatly improving the model's cross-project generalization ability in scenarios where no construction site has been seen.
[0044] Step S4: Input the DIFs into the decoder built based on the Adaptive Graph Convolutional Recurrent Network (AGCRN), and introduce a physical constraint loss function based on the formation loss theory during model training to construct a loss function based on prediction accuracy (…). L acc Formula 1) and physical consistency loss ( L phy The composite optimization function composed of Formula 2) L 2, Formula 3), to achieve synergistic optimization of data-driven and physical mechanisms.
[0045] Specifically: (1) In the first stage of training (1-20 epochs), the DANN and AGCRN models are trained together to improve the generalization performance of DIFs and make them suitable for downstream prediction tasks. (2) In the second phase of training (21-30 epochs), with accuracy loss (L accThe AGCRN model is further trained with ) as the primary focus, thereby avoiding the premature introduction of physical loss (L) phy This could lead to gradient explosion.
[0046] (3) In the third stage of training, a composite optimization function is used ( L 2) Further train the model to ensure that the prediction results conform to the basic laws of soil and rock while maintaining high prediction accuracy.
[0047] The Adaptive Graph Convolutional Recurrent Network (AGCRN) decoder dynamically models the spatiotemporal dependencies between monitoring points, adaptively learning spatial adjacency relationships without requiring a pre-defined graph structure. Simultaneously, a physically constrained loss function based on formation loss theory is embedded in the training process, ensuring that the prediction results strictly adhere to the fundamental principles of soil mechanics, and outputs dynamically updated formation loss rate parameters. m This enhances the physical reliability and engineering interpretability of the predictions.
[0048] Step S5: Based on the trained model, perform high-precision, highly physically interpretable spatiotemporal joint prediction of retaining wall deformation and ground settlement, and simultaneously output dynamically updated ground loss rate parameters. m This provides decision support with both numerical accuracy and physical reliability for risk warning, safety assessment, and dynamic optimization of support schemes during deep foundation pit construction. Specifically, it provides real-time feedback to engineering management personnel on foundation pit deformation trends and ground loss rates. m The evolution of the situation can help take timely and targeted prevention and control measures, such as strengthening support, adjusting the excavation sequence, or implementing grouting, so as to effectively control construction risks and ensure project safety.
[0049] The prediction results can be visualized in real time, allowing construction personnel to view the spatial distribution and temporal evolution trends of wall deformation and ground settlement. In addition, the system simultaneously provides physical interpretative early warning indicators (such as ground loss rate). m This helps decision-makers dynamically adjust the excavation sequence, the timing of support application, or local reinforcement schemes, thereby achieving proactive risk prevention and control.
[0050] Example 2 like Figure 2 As shown, a deep foundation pit spatiotemporal deformation prediction system based on physical constraints and a large language model includes: The data acquisition module is used to acquire the geological survey report text and multi-source monitoring time series data for deep foundation pit projects; The large language model encoding module uses a large language model to perform semantic parsing on the geological exploration report text and generate geological semantic embedding vectors. The domain adversarial training module is used to input the geological semantic embedding vector and the multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features. The physical information-enhanced spatiotemporal deformation prediction module is used to input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and to embed a physical constraint loss function based on the formation loss theory during the model training process to jointly optimize the prediction accuracy loss and physical consistency loss; based on the trained model, spatiotemporal joint prediction of wall deformation and surface settlement is performed, and dynamically updated formation loss rate parameters are output.
[0051] Each module of the system works closely together to achieve high-precision prediction of deep foundation pit deformation through data flow. In actual engineering, the data acquisition module collects geological survey reports and multi-source monitoring time-series data; the large language model encoding module performs semantic parsing on geological texts to generate geological semantic embedding vectors; the domain adversarial training module integrates the embedding vectors and time-series data, and extracts cross-engineering domain invariant features through adversarial training; the physically enhanced spatiotemporal deformation prediction module, based on these features and combined with physical constraints, achieves deformation prediction with strong physical interpretability and outputs key parameters such as soil loss ratio, providing data support for risk assessment.
[0052] like Figure 3 As shown, the large language model encoding module receives the original text of the geological exploration report, processes it through a word segmenter, and then inputs it into the Qwen3-embedding-4B large language model to generate a high-dimensional geological semantic embedding vector.
[0053] In practical engineering, geological survey reports are input into the system as unstructured text. A word segmenter divides the report into a sequence of terms, adding start markers (BOS) and end markers (EOS). This sequence is then fed into a pre-trained embedding model, Qwen3-embedding-4B, which outputs semantic embedding vectors for the corresponding dimensions. The embedding vectors of the EOS terms are selected as the global semantic information representation. These vectors implicitly encode information such as stratigraphic distribution and soil properties, providing a structured semantic foundation for subsequent cross-domain feature learning and physical enhancement prediction.
[0054] like Figure 4 As shown, the domain adversarial training module consists of a feature extractor, a discriminator, and a loss function calculation unit. It achieves the extraction of cross-domain invariant features (DIFs) through joint optimization.
[0055] During training, the raw data (source and target domains) are concatenated with the monitored data using semantic embedding vectors and then input into the feature extractor to generate domain-independent features (DIFs). These features are simultaneously fed into the discriminator to distinguish the data source domain. The feature extractor and discriminator optimize the mixture loss L1 and the discriminant loss L1, respectively. discThis enables adversarial training and ultimately outputs domain-invariant features with strong generalization capabilities, providing robust input for subsequent physics augmentation predictions.
[0056] like Figure 5 As shown, the physical information-enhanced spatiotemporal deformation prediction module adopts the Adaptive Graph Convolutional Recurrent Network (AGCRN) model. By fusing the Node Adaptive Parameter Learning (NAPL) and Data Adaptive Graph Generation (DAGG) mechanisms, it dynamically models the spatiotemporal dependencies between monitoring points, thereby achieving high-precision joint prediction of wall and ground surface deformation.
[0057] During the prediction process, the AGCRN model uses domain-invariant features as input and calculates node-specific weights and spatial adjacency matrices through an improved GCN module. This allows it to capture the heterogeneous responses at different locations in the foundation pit without requiring a pre-defined graph structure. Simultaneously, the model embeds a physical constraint loss function based on the theory of soil loss during the training phase. This ensures that the prediction results not only conform to data patterns but also satisfy the basic principles of soil mechanics. Ultimately, it outputs highly physically interpretable deformation prediction values and soil loss ratio parameters, providing reliable data support for construction safety assessment.
[0058] like Figure 6 As shown, this invention enables high-precision spatiotemporal prediction of retaining wall deformation and ground settlement during deep foundation pit construction. The prediction results are in good agreement with the measured data, verifying the effectiveness and engineering applicability of the model.
[0059] Based on multi-source monitoring data and geological semantic information, the system continuously outputs deformation prediction values at different construction stages, providing reliable data support for on-site safety assessment. The prediction results have good physical interpretability and spatiotemporal consistency, which can help engineers identify potential risk areas in advance and improve the controllability and safety of the construction process.
[0060] This invention achieves high-precision, highly physically interpretable spatiotemporal prediction of deep foundation pit wall deformation and surface settlement through the collaborative work of a data acquisition module, a large language model encoding module, a domain adversarial training module, and a physically-enhanced spatiotemporal deformation prediction module. The modules are tightly integrated through data carriers such as semantic embedding, domain-invariant features, and physical constraint loss, forming an end-to-end process from unstructured geological text to interpretable predictive output. This effectively improves the model's generalization ability and prediction reliability across engineering scenarios, providing reliable support for deep foundation pit construction risk assessment that combines data-driven capabilities with physical consistency.
[0061] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.
[0062] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A method for predicting the spatiotemporal deformation of deep foundation pits, characterized in that, include: Step S1: Obtain the geological survey report text and multi-source monitoring time series data for the deep foundation pit project; Step S2: Use a large language model to perform semantic parsing on the geological exploration report text and generate geological semantic embedding vectors; Step S3: Input the geological semantic embedding vector and multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features; Step S4: Input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and embed the physical constraint loss function of the formation loss theory during the training process of the model to jointly optimize the prediction accuracy loss and physical consistency loss. Step S5: Based on the trained model, perform spatiotemporal joint prediction of wall deformation and ground settlement, and output dynamically updated formation loss rate parameters.
2. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, The large language model includes Qwen3-Embedding-4B, which extracts high-dimensional embedding vectors by semantically parsing the stratigraphic descriptions, soil distributions, and rock and soil parameters in the geological exploration report. The embeddings corresponding to the EOS tokens at the end of the report are used as the comprehensive semantic representation of the entire geological exploration report.
3. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, The domain adversarial neural network includes a feature extractor and a domain discriminator, which are optimized collaboratively through an adversarial training mechanism. The feature extractor minimizes the sum of prediction accuracy loss and adversarial loss, while the domain discriminator maximizes domain classification accuracy.
4. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, The adaptive graph convolutional recurrent network includes a node adaptive parameter learning module and a data adaptive graph generation module; the adaptive parameter learning module generates exclusive convolution weights for each monitoring node, and the data adaptive graph generation module dynamically constructs a spatial relationship graph between nodes.
5. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, The physical constraint loss function is based on the formation loss theory. It calculates the predicted formation loss rate by numerically integrating the predicted wall deformation with the surface settlement curve. m pred and compared with the measured value m act By comparing the results, a physical consistency loss is constructed, and it is then weighted and fused with the prediction accuracy loss to form a hybrid training objective.
6. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, A three-stage training strategy is adopted. The first stage involves co-training the domain adversarial neural network and the adaptive graph convolutional recurrent network to achieve domain-invariant feature learning. The second stage freezes the feature extractor and optimizes the adaptive graph convolutional recurrent network only with the loss of accuracy; the third stage introduces physical constraint loss to jointly optimize prediction accuracy and physical consistency.
7. The method for predicting spatiotemporal deformation of deep foundation pits according to claim 1, characterized in that, The multi-source monitoring time series data includes wall horizontal displacement and ground settlement. After engineering-independent preprocessing, it is unified into a fixed dimension. The original monitoring points are curve fitted and resampled at equal intervals along the normalized depth / horizontal axis to generate a virtual monitoring sequence with consistent dimensions.
8. A deep foundation pit spatiotemporal deformation prediction system, characterized in that, include: The data acquisition module is used to acquire the geological survey report text and multi-source monitoring time series data for deep foundation pit projects; The large language model encoding module uses a large language model to perform semantic parsing on the geological exploration report text and generate geological semantic embedding vectors. The domain adversarial training module is used to input the geological semantic embedding vector and the multi-source monitoring time series data into the domain adversarial neural network to generate cross-engineering domain invariant features. The physical information-enhanced spatiotemporal deformation prediction module is used to input the cross-engineering domain invariant features into the decoder constructed by the adaptive graph convolutional recurrent network, and to embed a physical constraint loss function based on the formation loss theory during the model training process to jointly optimize the prediction accuracy loss and physical consistency loss; based on the trained model, spatiotemporal joint prediction of wall deformation and surface settlement is performed, and dynamically updated formation loss rate parameters are output.
9. The deep foundation pit spatiotemporal deformation prediction system according to claim 8, characterized in that, The large language model includes Qwen3-Embedding-4B, which extracts high-dimensional embedding vectors by semantically parsing the stratigraphic descriptions, soil distributions, and rock and soil parameters in the geological exploration report. The embeddings corresponding to the EOS tokens at the end of the report are used as the comprehensive semantic representation of the entire geological exploration report.
10. The deep foundation pit spatiotemporal deformation prediction system according to claim 8, characterized in that, The domain adversarial neural network includes a feature extractor and a domain discriminator, which are optimized collaboratively through an adversarial training mechanism. The feature extractor minimizes the sum of prediction accuracy loss and adversarial loss, while the domain discriminator maximizes domain classification accuracy.