Deep foundation deformation prediction method, device and medium based on space-time fusion attention

Abstract

This invention provides a method, device, and medium for predicting deep foundation pit deformation based on spatiotemporal fusion attention, belonging to the field of intelligent prediction technology for foundation pit engineering. The method includes acquiring multi-source spatiotemporal fusion data of the construction area to construct a multi-source spatiotemporal fusion dataset; preprocessing the multi-source spatiotemporal fusion dataset to obtain one-dimensional time-series data; converting the one-dimensional time-series data into time-series samples using a sliding window method; inputting the time-series samples into a pre-trained deformation prediction model to obtain deformation increments; acquiring the actual monitored deformation value at the current moment, and superimposing the deformation increments with the actual monitored deformation value at the current moment to obtain the final predicted deformation value. This invention sets a working condition-aware temporal attention layer in the model, maps working condition encoding information to working condition adjustment factors, and dynamically corrects the basic attention score, enabling the model to adaptively increase the feature weights of corresponding time steps for key construction stages such as accelerated earthwork excavation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction technology for foundation pit engineering, and specifically relates to a method, equipment and medium for predicting deep foundation pit deformation based on spatiotemporal fusion attention. Background Technology

[0002] With the continuous expansion of urban underground space development in my country, and the large-scale construction of urban rail transit, underground complexes, and transportation hubs, the number of ultra-large and deep foundation pit projects in cities has increased significantly. Deep foundation pits are characterized by large excavation depths, long construction periods, and complex surrounding environments. Excavation and unloading can easily disrupt the stress balance of the strata, inducing risks such as displacement of the retaining structure, surface settlement, and additional settlement of surrounding buildings. Especially under complex conditions such as earthwork excavation, support replacement, and groundwater disturbance, foundation pit deformation exhibits nonlinear, long-term dependent, and staged abrupt changes, posing a severe challenge to safety control.

[0003] Deformation monitoring of foundation pits is a crucial means of construction safety control. Both the "Technical Specification for Monitoring of Urban Rail Transit Engineering" (GB50911-2013) and the "Technical Standard for Monitoring of Building Foundation Pit Engineering" (GB50497-2019) require full-process dynamic monitoring and risk early warning for deep foundation pits. Therefore, constructing a high-precision, robust deformation prediction model to achieve dynamic risk assessment and advanced early warning under complex working conditions is of great significance for engineering safety.

[0004] Existing prediction methods mainly include empirical formulas, numerical simulations, and machine learning. Traditional empirical models have limited generalization ability and struggle to characterize complex nonlinear time-series features; finite element numerical simulations are parameter-sensitive and computationally intensive, making them unsuitable for real-time prediction. In recent years, LSTM and its improved models have been introduced into foundation pit deformation prediction, but existing methods mostly rely on single monitoring data, failing to fully utilize heterogeneous information such as geological surveys, construction conditions, environmental disturbances, and multi-dimensional monitoring. Furthermore, they lack attention mechanisms for key construction stage characteristics, and the prediction method of directly outputting cumulative deformation values ​​is prone to error accumulation and magnitude drift over long-term rolling processes. These combined shortcomings result in insufficient deformation prediction accuracy of existing models under complex construction conditions, delayed identification of abrupt inflection points, and poor reliability of long-term prediction results. Summary of the Invention

[0005] To address the problems in the background technology, this invention proposes a method, equipment, and medium for predicting deep foundation pit deformation based on spatiotemporal fusion attention.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, this invention proposes a method for predicting the deformation of deep foundation pits based on spatiotemporal fusion attention, comprising: Acquire multi-source spatiotemporal fusion data of the construction area and construct a multi-source spatiotemporal fusion dataset; The multi-source spatiotemporal fusion dataset is preprocessed to obtain one-dimensional time-series data; The one-dimensional time series data is transformed into time series samples using the sliding window method; The time-series samples are input into a pre-trained deformation prediction model to obtain the deformation increment; The actual deformation value at the current moment is obtained, and the deformation increment is superimposed with the actual deformation value at the current moment to obtain the final predicted deformation value.

[0007] Preferably, the multi-source spatiotemporal fusion data includes: geological survey data, construction condition data, environmental disturbance data, and multi-dimensional deformation monitoring data; wherein, the multi-dimensional deformation monitoring data includes at least one of: horizontal displacement of the retaining structure, deep displacement of the retaining structure, surface settlement outside the pit, settlement of surrounding buildings, axial force of the support, groundwater level, pore water pressure, and soil pressure.

[0008] Preferably, the preprocessing method is to sequentially perform missing value interpolation repair, 3σ criterion outlier removal, time series alignment, working condition category encoding, and Min-Max normalization.

[0009] Preferably, the deformation prediction model includes a serially connected input layer, a two-layer LSTM encoding layer, a condition-aware temporal attention layer, a fully connected prediction layer, and a residual increment output layer; wherein: The input layer is used to verify and adapt the window length and feature dimensions of the time series samples. The dual-layer LSTM encoding layer is used to mine the long-term temporal dependencies in the time series samples layer by layer through a gating mechanism, capture the evolution law and change trend of the foundation pit deformation, and output a high-dimensional feature sequence containing deep temporal information. The working condition-aware temporal attention layer is used to dynamically correct the attention weights by combining the working condition coding information in the temporal samples, and apply the corrected weights to the high-dimensional feature sequence. The fully connected prediction layer is used to perform a nonlinear transformation on the weighted high-dimensional feature sequence, mapping it to an intermediate output value of the prediction target dimension; The residual increment output layer is used to integrate the intermediate output values ​​into a deformation increment of the future time relative to the current time.

[0010] Preferably, training the deformation prediction model specifically includes: Obtain geological survey data, construction condition data, environmental disturbance data, and multidimensional deformation monitoring data throughout the entire historical construction cycle, and construct a historical multi-source training benchmark dataset; The missing value interpolation repair, 3σ criterion outlier removal, time series alignment, working condition category encoding and Min-Max normalization are sequentially performed on the historical multi-source training benchmark dataset to obtain the normalized historical time series dataset. The normalized historical time series dataset is converted into supervised learning training sample pairs by the sliding window method. The supervised learning training sample pairs include the input sequence and the real deformation increment label corresponding to the next time step. The supervised learning training sample pairs are divided into training set, validation set and test set, and input into Attention-LSTM network for iterative training to obtain a trained deformation prediction model.

[0011] Preferably, in the process of training the deformation prediction model, the mean squared error (MSE) is used as the loss function, and its expression is:

[0012] In the formula, This represents the total number of samples within the batch. The label is the actual deformation increment corresponding to the sample. This represents the predicted deformation increment output by the model.

[0013] Preferably, the expression for calculating the final predicted deformation value is:

[0014] In the formula, To accurately monitor deformation values ​​at the current moment, To predict the deformation increment for the model, This is to predict the final deformation value for the next moment.

[0015] Preferably, after obtaining the actual monitored deformation value at the current moment and superimposing the deformation increment with the actual monitored deformation value at the current moment to obtain the final predicted deformation value, the method further includes: Calculate the predicted deformation rate; wherein the predicted deformation rate is the ratio of the difference between the final predicted deformation values ​​at two adjacent time points to the corresponding time step; The risk level is divided into multiple levels, and a corresponding cumulative deformation threshold and deformation rate threshold are set for each level. Determine whether the final predicted deformation value is greater than the cumulative deformation threshold of the corresponding level, and determine whether the predicted deformation rate is greater than the deformation rate threshold of the corresponding level; If either the final predicted deformation value or the predicted deformation rate exceeds the threshold of the corresponding level, an early warning for the corresponding level will be triggered.

[0016] In a second aspect, the present invention proposes a device including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and the processor executes the spatiotemporal fusion attention-based deep foundation pit deformation prediction method described in the first aspect when executing the computer instructions.

[0017] Thirdly, the present invention proposes a computer-readable storage medium storing computer instructions thereon, which, when executed, can realize the deep foundation pit deformation prediction method based on spatiotemporal fusion attention described in the first aspect.

[0018] The beneficial effects of this invention are: 1. This invention sets up a working condition-aware temporal attention layer in the deformation prediction model, maps the working condition encoding information to working condition adjustment factors, and dynamically corrects the basic attention score. This enables the model to adaptively increase the feature weights of the corresponding time steps for key construction stages such as accelerated earthwork excavation, support replacement, sudden changes in groundwater, and abrupt deformation changes. This solves the problems of insufficient response to key construction stage features, lagging identification of deformation abrupt change inflection points, and overly smooth prediction results in existing deep learning methods, and significantly improves the prediction response speed and tracking at abrupt change moments.

[0019] 2. The accuracy of this invention is achieved by constructing a multi-source spatiotemporal fusion dataset that integrates four types of heterogeneous data: geological exploration, construction conditions, environmental disturbance, and multi-dimensional deformation monitoring. This enables a comprehensive characterization of the inherent laws governing the deformation evolution of foundation pits, solving the problem of insufficient prediction accuracy caused by single or incomplete information input in existing methods. It significantly improves the breadth of perception and depth of characterization of factors influencing foundation pit deformation.

[0020] 3. This invention employs a residual increment prediction mechanism, which enables the model to output only the deformation increment of the next time step relative to the current time step. This value is then superimposed with the actual monitoring value at the current time step to obtain the final predicted value. By using the actual monitoring value at each time step as a benchmark anchor point, the propagation chain of error in the time dimension is effectively cut off. This solves the problem that in traditional absolute prediction methods, the prediction error of the previous time step is continuously accumulated as the input of the next time step during long-term rolling, leading to a continuous amplification of prediction deviation. This significantly improves the prediction stability and long-term accuracy of the model in long-term series.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart of the spatiotemporal fusion attention-based deep foundation pit deformation prediction method of the present invention is shown; Figure 2 A schematic diagram of the Attention-LSTM prediction model structure of the present invention is shown; Figure 3 A schematic diagram of the working condition perception temporal attention mechanism of the present invention is shown; Figure 4 A schematic diagram of the residual increment prediction mechanism of the present invention is shown; Figure 5 A schematic diagram illustrating the dynamic risk warning and BIM linkage of the present invention is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Reference Figure 1 As shown, a method for predicting the deformation of deep foundation pits based on spatiotemporal fusion attention specifically includes the following steps: S10. Obtain multi-source spatiotemporal fusion data of the construction area and construct a multi-source spatiotemporal fusion dataset; The multi-source spatiotemporal fusion data comprises four categories: geological exploration data, construction condition data, environmental disturbance data, and multi-dimensional deformation monitoring data. These four categories correspond to the geological background, construction-driven factors, environmental disturbances, and state response of the foundation pit deformation, respectively, comprehensively covering the internal and external factors influencing the evolution of foundation pit deformation and providing comprehensive feature support for subsequent deformation prediction. Among them: Geological survey data is the inherent geological foundation data of the construction area. It is obtained through the results of geological survey work carried out before the project construction. Geological survey data contains basic information such as site strata conditions and rock and soil mechanical properties, and is the underlying geological constraint basis for the deformation and evolution law of the foundation pit.

[0026] Construction condition data is obtained in real time through process control records, construction progress ledgers, and condition information collection channels at the construction site. It also collects dynamic condition information such as the type of construction process, work progress, and support installation and dismantling status at the current moment. This data is used to characterize the internal driving effect of construction activities on the deformation of the foundation pit and to support the subsequent deformation prediction model in identifying the deformation characteristics of key construction stages.

[0027] Environmental disturbance data is collected in real time by environmental monitoring equipment (such as temperature and humidity sensors, vibration monitors, etc.) deployed in the construction area and the affected area. Simultaneously, dynamic fluctuation information of the surrounding environment is collected to characterize the external interference effect of external environmental changes on the deformation of the foundation pit and to help the model adapt to the deformation fluctuations caused by environmental disturbances.

[0028] Multidimensional deformation monitoring data is obtained by periodically measuring real-time values ​​at each monitoring point using corresponding monitoring components and acquisition equipment (such as total stations and water level gauges) deployed in the foundation pit retaining structure, surrounding ground surface, adjacent buildings, support system, and underground soil at a uniform time step. Specifically, this data includes at least one of the following: horizontal displacement of the retaining structure, deep displacement of the retaining structure, ground surface settlement outside the pit, settlement of surrounding buildings, axial force of the support, groundwater level, pore water pressure, and earth pressure. Multidimensional deformation monitoring data is the core dynamic data reflecting the real-time deformation and stress state of the foundation pit.

[0029] After completing the above four types of data collection, the monitoring points are used as spatial matching identifiers and the unified time step is used as the time alignment benchmark. Data from different sources and with different initial formats are collected into the same spatiotemporal coordinate system and integrated to form a structured multi-source spatiotemporal fusion dataset.

[0030] S20. Preprocess the multi-source spatiotemporal fusion dataset to obtain one-dimensional time series data; Preprocessing is performed in a fixed order, including missing value interpolation and repair, outlier removal using the 3σ criterion, time series alignment, operating condition category coding, and Min-Max normalization. Specifically: Missing value interpolation repair: For missing data points in the multi-source spatiotemporal fusion dataset caused by acquisition interruption, signal fluctuation, etc., interpolation is used to complete and repair them, ensuring the temporal continuity of each time series data and avoiding data breakpoints from destroying temporal dependencies and affecting subsequent feature extraction results.

[0031] 3σ Criterion Outlier Removal: For time-series data with missing value repair completed, outlier identification and correction are performed. Based on the statistical 3σ criterion, the mean and standard deviation of each individual time-series sequence are calculated. Data exceeding a preset multiple (e.g., ±3 times) of the standard deviation of the mean are identified as outliers. These outliers are generally caused by factors such as momentary equipment failures or on-site interference and do not represent the true state of the foundation pit. After outlier identification, linear interpolation is used to fill in and correct the outlier locations, removing data noise and ensuring data accuracy and stability.

[0032] Time series alignment: Since there are differences in the collection frequency and timestamp benchmark of the four types of data, namely geological exploration, construction conditions, environmental disturbance, and multidimensional deformation monitoring, it is necessary to use a unified time step as the alignment benchmark to match and calibrate the time dimension of all data sequences, so that data from different sources correspond one-to-one at the same time, forming a time series that is synchronized in the time dimension, and ensuring that each set of data input into the model has a consistent spatiotemporal correspondence.

[0033] Working condition category coding: Construction working condition data are discrete categorical features that cannot be directly used in numerical calculations. Therefore, it is necessary to perform numerical coding conversion on different working condition categories, mapping discrete process types, support status and other category information into numerical features that can participate in the calculation, so that construction working condition information can be integrated into the model calculation process, and provide feature input for the working condition perception temporal attention mechanism of the subsequent model.

[0034] Min-Max Normalization: For all numerical time-series data that have undergone the aforementioned processing, the Min-Max normalization method is used to unify the data scale, eliminate differences in dimensions and numerical magnitudes between different monitoring indicators, and improve the stability of model inference. The expression for normalization calculation is:

[0035] In the formula, This represents the original cumulative settlement value. The normalized settlement value. and These are the minimum and maximum values ​​of the time series data for that monitoring point, respectively.

[0036] By eliminating quality defects, format differences, and scale differences in the original data through multi-stage standardized preprocessing, we can output well-organized and usable one-dimensional time series data, providing basic data that conforms to the input specifications for subsequent time series sample construction and model inference.

[0037] S30. Transform one-dimensional time series data into time series samples using the sliding window method; The sliding window length L ranges from 3 to 10, with a preferred value of 4. Specifically, using the normalized time-series data corresponding to the current moment as the end of the window, complete data from L consecutive time steps are extracted along the time axis to form a single time-series sample. The data for each time step includes all normalized features corresponding to the preprocessed geological exploration data, construction condition data, environmental disturbance data, and multi-dimensional deformation monitoring data. This ensures that the single sample fully covers the geological background conditions, construction driving state, degree of environmental disturbance, and deformation evolution trend within the corresponding time window, fully reflecting the temporal variation law of foundation pit deformation.

[0038] In a single prediction scenario, only the data from the L consecutive time steps closest to the current time need to be extracted to generate a time series sample for single inference. If rolling prediction over multiple consecutive time steps is required, the sliding window will slide backward by one time step each time step is advanced, and input samples for the corresponding prediction time will be generated sequentially.

[0039] After transformation using the sliding window method, the one-dimensional time series data, which was originally arranged linearly in time order, is organized into structured time series samples of fixed length, which can be directly input into the pre-trained deformation prediction model.

[0040] S40. Input the time series samples into the pre-trained deformation prediction model to obtain the deformation increment; Reference Figure 2 As shown, the deformation prediction model includes a serially connected input layer, a two-layer LSTM encoding layer, a condition-aware temporal attention layer, a fully connected prediction layer, and a residual increment output layer. Specifically: The time series samples first enter the input layer. The input layer verifies and adapts the window length and feature dimensions of the samples to ensure that the input data format is completely matched with the input specifications during the model training stage. After the adaptation is completed, the data is passed to the two-layer LSTM encoding layer.

[0041] The dual-layer LSTM coding layer receives the adapted time-series samples and mines the long-term temporal dependencies in the time-series data layer by layer through the gating mechanism of the two-layer LSTM network. It captures the evolution law and change trend of the foundation pit deformation in continuous time steps and outputs a high-dimensional feature sequence containing deep temporal information.

[0042] Reference Figure 3 As shown, the working condition-aware temporal attention layer receives the feature sequence output by LSTM encoding and dynamically adjusts the attention weights by combining the working condition encoding information in the samples: the working condition encoding information is mapped through a fully connected layer to obtain the working condition adjustment factor. and with basic attention score After product correction, normalization is performed to obtain the final attention weights. This mechanism can automatically increase the feature weights of the corresponding time steps for key stages such as support replacement, accelerated earthwork excavation, groundwater fluctuations, and sudden deformation changes, thereby strengthening the impact of key disturbance information on the prediction results and making feature extraction more consistent with the actual driving laws of foundation pit deformation.

[0043] The attention-weighted feature sequence is input into the fully connected prediction layer. Through the nonlinear transformation of the fully connected network, the high-dimensional temporal features are mapped to intermediate output values ​​of the prediction target dimension.

[0044] The output of the fully connected prediction layer is integrated by the residual increment output layer to output the deformation increment at future time. This deformation increment is the amount of deformation change at the prediction time relative to the current time, rather than the total cumulative deformation value, which matches the calculation logic of subsequent residual superposition.

[0045] Furthermore, the deformation prediction model is trained to obtain a well-trained deformation prediction model, specifically including: Referring to the data collection and integration method in step S10 of the reasoning and deployment phase, geological survey data, construction condition data, environmental disturbance data and multi-dimensional deformation monitoring data of the target foundation pit throughout its historical construction cycle are collected to construct a historical multi-source training benchmark dataset.

[0046] Following the preprocessing procedure in step S20, the historical multi-source training benchmark dataset is sequentially processed with missing value interpolation and repair, 3σ criterion outlier removal, time series alignment, working condition category encoding, and Min-Max normalization to obtain a normalized historical time series dataset. When performing Min-Max normalization as described in S20, its extreme value parameters (i.e., the extreme values ​​of each monitoring indicator) are... and The fitting source must be strictly limited to the portion of the current historical data to be subsequently assigned to the training set. When transforming the time-series data corresponding to the validation and test sets, the extreme value parameters obtained from the training set must be uniformly normalized; it is strictly prohibited to introduce extreme values ​​from their respective data intervals for independent fitting. This constraint can prevent the leakage of future time-series information to the training process from the data preprocessing source, ensuring the effectiveness of the validation set in monitoring the model's generalization ability and the test set in evaluating the final performance.

[0047] Referring to the sliding window processing logic in step S30, sample transformation is performed on the normalized historical time-series dataset. Unlike the inference phase, the training phase employs an overlapping sliding sampling strategy. Normalized data from L consecutive time steps are used as the input sequence, and the actual deformation increment corresponding to the next time step is used as the supervision label. The entire dataset is traversed to generate pairs of supervised learning training samples, "input sequence - actual label," and the sample size is expanded by overlapping windows. The window length L ranges from 3 to 10, with a preferred value of 4.

[0048] All supervised learning training samples were strictly divided into training, validation, and test sets according to chronological order. There was no overlap or intersection between the training, validation, and test sets in terms of time dimensions to avoid data leakage and ensure that the data distribution conformed to the true temporal evolution of foundation pit deformation. The training set was used for iterative updates of the model weight parameters; the validation set was used to monitor the model's generalization ability in real time during training and to provide a basis for early stopping strategies; and the test set was used to verify the model's final prediction accuracy and generalization performance after training was completed.

[0049] An Attention-LSTM network is constructed, consisting of an input layer, a two-layer LSTM encoding layer, a condition-aware temporal attention layer, a fully connected prediction layer, and a residual incremental output layer connected in series. The network weights are initialized and the training hyperparameters are configured to obtain the network model to be trained.

[0050] The number of hidden units in the two-layer LSTM coding layer ranges from 32 to 128, and the Dropout coefficient ranges from 0.1 to 0.5. The preferred hyperparameter configuration is: 64 hidden units in the two-layer LSTM, 0.2 Dropout coefficient, 16 batch size, and 0.001 initial learning rate.

[0051] The training set is input into the network model to be trained, and multiple rounds of iterative training are carried out. The parameters of the entire network are continuously updated through the backpropagation algorithm.

[0052] The training process uses the mean squared error (MSE) as the loss function to measure the deviation between the predicted deformation increment output by the model and the true deformation increment label. Its expression is:

[0053] In the formula, This represents the total number of samples within the batch. The label is the actual deformation increment corresponding to the sample. This represents the predicted deformation increment output by the model.

[0054] Training employs the Adam optimizer to update parameters, with three accompanying strategies to ensure model performance: a cosine annealing learning rate strategy is used to dynamically adjust the learning step size, improving convergence stability and global optimization capability; a gradient pruning strategy is used to constrain gradient magnitude and suppress the risk of gradient explosion; and an early stopping strategy is introduced to continuously monitor the loss value on the validation set, automatically terminating training when the loss does not continue to decrease within a preset number of rounds, thus avoiding overfitting.

[0055] After training is terminated, the network weights corresponding to the lowest loss on the validation set are selected as the optimal parameters, and the trained deformation prediction model is solidified; the final performance is then verified through the test set.

[0056] S50. Obtain the actual monitored deformation value at the current moment, and superimpose the deformation increment with the actual monitored deformation value at the current moment to obtain the final predicted deformation value; Reference Figure 4 As shown, the final predicted deformation value is calculated using the following expression:

[0057] In the formula, To accurately monitor deformation values ​​at the current moment, To predict the deformation increment for the model, This is to predict the final deformation value for the next moment.

[0058] The current real-time monitored deformation value can be taken from the multi-dimensional deformation monitoring data collected in real time in step S10. It is the measured cumulative deformation value of the corresponding monitoring point at the current moment. The current real-time monitored deformation value can accurately reflect the current real deformation state of the foundation pit and serve as the benchmark value for incremental superposition.

[0059] By adopting an incremental superposition prediction method, the long-term cumulative bias that is easily caused by the model directly predicting the cumulative deformation value can be reduced, so that the prediction results are always anchored to the current real deformation state, which is more in line with the evolution law of the gradual accumulation of foundation pit deformation, and improves the accuracy and engineering practicality of the prediction results.

[0060] S60. Early warning based on the final predicted deformation value.

[0061] Reference Figure 5 As shown, in addition to the final predicted deformation value (predicted cumulative deformation value), the predicted deformation rate is calculated simultaneously. The calculation method is as follows: based on the ratio of the difference between the final predicted deformation values ​​at two adjacent moments to the corresponding time step, the deformation change rate per unit time is obtained. The predicted cumulative deformation value reflects the degree of total accumulated deformation, while the deformation rate reflects the severity of deformation development. These two indicators together constitute a risk assessment system, which can simultaneously cover two risk scenarios: total deformation exceeding limits and sudden rate changes.

[0062] The risk levels are divided into five categories: Level I (safe), Level II (lower risk), Level III (medium risk), Level IV (higher risk), and Level V (dangerous). A comprehensive judgment rule combining cumulative deformation threshold and deformation rate threshold is adopted. Corresponding cumulative deformation thresholds and deformation rate thresholds are set for each of the five levels. When any indicator exceeds the threshold for the corresponding level, an early warning for that level is triggered, ensuring the comprehensiveness of risk identification.

[0063] Once an alert is triggered, standardized alert notifications are automatically generated. The notifications include the alert level, corresponding monitoring points, trend of the change, risk source analysis, and on-site handling suggestions, providing on-site management personnel with clear risk warnings and actionable guidelines.

[0064] The final predicted deformation value, predicted deformation rate, risk level and early warning information are synchronized to the BIM visualization platform, and the risk status and deformation trend of each point are intuitively displayed on the three-dimensional model of the foundation pit, realizing the dynamic visualization of deep foundation pit risks and assisting on-site management personnel in making decisions such as adjusting construction procedures and handling risks.

[0065] Based on the same inventive concept as the above method, this invention also proposes a deep foundation pit deformation prediction system based on spatiotemporal fusion attention, comprising: The data acquisition module is used to acquire multi-source spatiotemporal fusion data of the construction area and construct a multi-source spatiotemporal fusion dataset. The preprocessing module is used to preprocess the multi-source spatiotemporal fusion dataset to obtain one-dimensional time series data; The sample construction module is used to transform one-dimensional time series data into time series samples using the sliding window method. The prediction module inputs time-series samples into a pre-trained deformation prediction model to obtain deformation increments; The overlay module is used to obtain the actual monitored deformation value at the current moment, and to overlay the deformation increment with the actual monitored deformation value at the current moment to obtain the final predicted deformation value.

[0066] Based on the same inventive concept as the above method, the present invention also proposes a device including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and the processor executes the above-mentioned method for predicting deep foundation pit deformation based on spatiotemporal fusion attention when executing the computer instructions.

[0067] Based on the same inventive concept as the above method, the present invention also proposes a computer-readable storage medium storing computer instructions thereon, which, when executed, can realize the above-mentioned method for predicting deep foundation pit deformation based on spatiotemporal fusion attention.

[0068] In one specific embodiment, this embodiment is based on the ultra-large deep foundation pit project of a station of Wuxi Metro. The excavation depth of the foundation pit is 16.64m to 18.61m, and a support system of underground continuous wall + multiple steel supports is adopted. The monitoring period is from July 2024 to September 2025, and a total of 149 sets of valid data including surface settlement, retaining displacement, support axial force, groundwater level and construction conditions are collected.

[0069] The method for predicting and warning the deformation of deep foundation pits described in this embodiment follows the same reasoning process as steps S10 to S60 above, and the same model training process as the corresponding training steps above.

[0070] Data is collected via automated sensors and uploaded wirelessly. Following step S10 above, a multi-source spatiotemporal fusion dataset is constructed, and following step S20 above, the entire preprocessing process is completed to obtain normalized one-dimensional time-series data. The normalization parameters are determined only by fitting the training set samples; the same parameters are used in the testing and inference stages to avoid data leakage. Following step S30 above, a sliding window method is used to convert the data into supervised learning samples. The sliding window length L is set to 4, and an overlapping sliding sampling method is used to generate sample pairs. All samples are strictly divided into training, validation, and test sets in a 7:1:2 ratio according to chronological order, without random shuffling, preserving the temporal dependencies of the transformed time series.

[0071] Model training is performed according to the corresponding process described above. The Attention-LSTM network used in this embodiment is specifically configured as follows: a two-layer LSTM structure with 64 hidden units and a Dropout coefficient of 0.2; the batch size is set to 16, and the initial learning rate is set to 0.001; training uses the Adam optimizer, with a cosine annealing learning rate, gradient pruning, and early stopping strategy, and the maximum number of training epochs is set to 2000; after training is terminated, the weights with the lowest loss on the validation set are selected as the optimal model parameters.

[0072] After comparing and validating multiple models on the same dataset, the prediction accuracy metrics of each model are as follows: Table 1 Model Prediction Accuracy Table

[0073] In the two deformation abrupt change intervals of days 45–52 and 98–105, the prediction error of this method was reduced by 31.2% and 28.6% respectively compared with the original Attention-LSTM. It can accurately distinguish the four deformation stages: initial gradual change, accelerated deformation, fluctuation stabilization, and secondary settling, and the ability to identify abrupt change inflection points is significantly stronger.

[0074] During field application, risk level assessment and early warning were conducted according to step S60 above. On day 46, the model predicted that the deformation rate would exceed the 2mm / d threshold within the next 3 days, automatically triggering a Level III medium-risk early warning. An early warning report was simultaneously sent via SMS and to the management platform. The report included the corresponding monitoring points, deformation trend curves, historical deformation data, and suggested remedial measures: increasing monitoring frequency, slowing the excavation rate, and strengthening support inspections. After implementing the corresponding measures on-site, the deformation effectively converged, validating the advanced early warning effect and engineering practical value of this method.

[0075] This embodiment simultaneously implements BIM linkage visualization function. The system synchronously maps the five-level risk level, high-risk monitoring points, and deformation prediction trend to the foundation pit BIM three-dimensional model, highlights the risk areas, realizes the visualization display of construction risks, and provides intelligent auxiliary decision support for on-site construction.

[0076] Any references to memory, storage, database, or other media used in the embodiments provided in this invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0078] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatiotemporal fusion attention deep foundation pit deformation prediction method, characterized in that, include: Acquire multi-source spatiotemporal fusion data of the construction area and construct a multi-source spatiotemporal fusion dataset; The multi-source spatiotemporal fusion dataset is preprocessed to obtain one-dimensional time-series data; The one-dimensional time series data is transformed into time series samples using the sliding window method; The time-series samples are input into a pre-trained deformation prediction model to obtain the deformation increment; The actual deformation value at the current moment is obtained, and the deformation increment is superimposed with the actual deformation value at the current moment to obtain the final predicted deformation value.

2. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, The multi-source spatiotemporal fusion data includes: geological survey data, construction condition data, environmental disturbance data, and multi-dimensional deformation monitoring data; wherein, the multi-dimensional deformation monitoring data includes at least one of: horizontal displacement of the retaining structure, deep displacement of the retaining structure, surface settlement outside the pit, settlement of surrounding buildings, axial force of the support, groundwater level, pore water pressure, and soil pressure.

3. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, The preprocessing method involves sequentially performing missing value interpolation repair, 3σ criterion outlier removal, time series alignment, working condition category encoding, and Min-Max normalization.

4. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, The deformation prediction model comprises a serially connected input layer, a two-layer LSTM encoding layer, a condition-aware temporal attention layer, a fully connected prediction layer, and a residual increment output layer; wherein: The input layer is used to verify and adapt the window length and feature dimensions of the time series samples. The dual-layer LSTM encoding layer is used to mine the long-term temporal dependencies in the time series samples layer by layer through a gating mechanism, capture the evolution law and change trend of the foundation pit deformation, and output a high-dimensional feature sequence containing deep temporal information. The working condition-aware temporal attention layer is used to dynamically correct the attention weights by combining the working condition coding information in the temporal samples, and apply the corrected weights to the high-dimensional feature sequence. The fully connected prediction layer is used to perform a nonlinear transformation on the weighted high-dimensional feature sequence, mapping it to an intermediate output value of the prediction target dimension; The residual increment output layer is used to integrate the intermediate output values ​​into a deformation increment of the future time relative to the current time.

5. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, Training the deformation prediction model specifically includes: Obtain geological survey data, construction condition data, environmental disturbance data, and multidimensional deformation monitoring data throughout the entire historical construction cycle, and construct a historical multi-source training benchmark dataset; The missing value interpolation repair, 3σ criterion outlier removal, time series alignment, working condition category encoding and Min-Max normalization are sequentially performed on the historical multi-source training benchmark dataset to obtain the normalized historical time series dataset. The normalized historical time series dataset is converted into supervised learning training sample pairs by the sliding window method. The supervised learning training sample pairs include the input sequence and the real deformation increment label corresponding to the next time step. The supervised learning training sample pairs are divided into training set, validation set and test set, and input into Attention-LSTM network for iterative training to obtain a trained deformation prediction model.

6. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 5, characterized in that, In training the deformation prediction model, the mean squared error (MSE) is used as the loss function, and its expression is: In the formula, This represents the total number of samples within the batch. The label is the actual deformation increment corresponding to the sample. This represents the predicted deformation increment output by the model.

7. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, The expression for calculating the final predicted deformation value is: In the formula, To accurately monitor deformation values ​​at the current moment, To predict the deformation increment for the model, This is to predict the final deformation value for the next moment.

8. The method for predicting deep foundation pit deformation based on spatiotemporal fusion attention as described in claim 1, characterized in that, After obtaining the actual monitored deformation value at the current moment and superimposing the deformation increment with the actual monitored deformation value at the current moment to obtain the final predicted deformation value, the method further includes: Calculate the predicted deformation rate; wherein the predicted deformation rate is the ratio of the difference between the final predicted deformation values ​​at two adjacent time points to the corresponding time step; The risk level is divided into multiple levels, and a corresponding cumulative deformation threshold and deformation rate threshold are set for each level. Determine whether the final predicted deformation value is greater than the cumulative deformation threshold of the corresponding level, and determine whether the predicted deformation rate is greater than the deformation rate threshold of the corresponding level; If either the final predicted deformation value or the predicted deformation rate exceeds the threshold of the corresponding level, an early warning for the corresponding level will be triggered.

9. A device comprising a memory and a processor, wherein the memory stores computer instructions executable on the processor, characterized in that, When the processor executes the computer instructions, it performs the spatiotemporal fusion attention-based deep foundation pit deformation prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed, the method for predicting the deformation of deep foundation pits based on spatiotemporal fusion attention, as described in any one of claims 1 to 8, can be implemented.

Images