Deep foundation pit ground surface settlement prediction method based on residual bidirectional gate recurrent unit

Abstract

The deep foundation pit ground surface settlement prediction method based on the residual bidirectional gate recurrent unit of the present application comprises the following steps: collecting accumulated ground surface settlement time series data of monitoring points around the deep foundation pit; using a sliding window method to convert normalized one-dimensional time series data into supervised learning sample pairs; constructing a residual bidirectional gate recurrent unit model; using an Adam optimizer combined with a cosine annealing learning rate scheduling strategy to update parameters; using the trained residual bidirectional gate recurrent unit model for test set prediction, performing inverse normalization on the prediction results, and outputting the final ground surface settlement prediction value. The present application is simple and efficient, has few model parameters, and has a fast training speed. The residual connection mechanism can still ensure that the prediction value does not deviate from the true order of magnitude under conditions of data scarcity or high noise, thereby enhancing the robustness of the model and facilitating online deployment and real-time early warning in actual engineering.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering deformation monitoring and deep learning technology, specifically to a method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units. Background Technology

[0002] Surface settlement is one of the most direct and important deformation monitoring indicators in deep foundation pit engineering, comprehensively reflecting the dynamic response of foundation pit excavation and the degree of damage to the soil and rock mass. Traditional surface settlement prediction methods mainly include Peck's empirical formula method, finite element numerical simulation method, and elasticity theory analytical method. The empirical formula method relies on a large number of engineering analogies for parameter acquisition, and its applicability is significantly limited by geological conditions; the numerical simulation method involves a large amount of modeling work, high uncertainty in soil and rock parameter input, low computational efficiency, and is difficult to meet the needs of real-time monitoring.

[0003] In recent years, data-driven methods, represented by machine learning and deep learning, have made significant progress in the field of deformation prediction in geotechnical engineering due to their strong ability to fit nonlinear relationships. Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Random Forests (RF) have been successfully applied to land subsidence prediction with good results. However, these methods have limited ability to capture the dynamic evolution patterns and long-range dependencies of time series data.

[0004] Recurrent Neural Networks (RNNs) are naturally suited for time-series data modeling tasks. Long Short-Term Memory (LSTM) networks effectively solve the gradient vanishing problem through gating mechanisms and have demonstrated strong time-series modeling capabilities in engineering predictions such as tunnel construction settlement and dam deformation. Gated Recurrent Units (GRUs), as a simplified variant of LSTMs, have fewer parameters and higher computational efficiency. Bidirectional Gated Recurrent Units (BiGRUs) further introduce a backpropagation path on top of the standard GRU, enabling the model to utilize both forward and backward contextual information of the sequence, theoretically providing stronger feature extraction and pattern recognition capabilities.

[0005] However, existing settlement prediction methods based on BiGRU or GRU directly output absolute settlement values. When the settlement time series shows a slow accumulation and the change between adjacent time steps is much smaller than the absolute order of magnitude, the model is prone to over-smoothing and slow response to abrupt changes. Especially during the construction of deep foundation pits, the surface settlement curve exhibits obvious stage-by-stage abrupt changes—an initial slow settlement period, an accelerated settlement period, a fluctuating stabilization period, and a re-settlement period alternately appear, causing the standard BiGRU and GRU models that directly predict absolute values ​​to have significantly increased prediction errors during periods of abrupt rate changes.

[0006] Therefore, the existing technology has the following shortcomings: First, the standard BiGRU / GRU model directly predicts the absolute settlement value, and the inertial effect of the historical hidden state leads to a sluggish response to sudden changes in settlement rate; second, the existing method does not effectively utilize the accurate observation value at the current moment as the prediction benchmark, resulting in a systematic shift in the predicted value in terms of magnitude; third, there is a lack of a special modeling scheme for the staged non-steady settlement characteristics of deep foundation pits, which makes it difficult to meet the real-time early warning needs of engineering projects. Summary of the Invention

[0007] The deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic unit proposed in this invention can at least solve one of the technical problems in the background art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units includes the following steps: S100. Collect cumulative surface settlement time-series data from monitoring points around the deep foundation pit, perform Min-Max normalization on the time-series data, and use the sliding window method to transform the normalized one-dimensional time-series data into supervised learning sample pairs. ; Settlement data from k consecutive time steps are used as model input, and the settlement value at the next time step is used as the prediction target. Input-output sample pairs are constructed, as shown below: in, The input sample sequence is constructed using a sliding window. This corresponds to the predicted target value; This represents the normalized settlement data at time step t; This represents the normalized settlement data at time step t+1; t represents the current time step; k represents the sliding window length. S200. Construct a residual bidirectional gated cyclic unit model, which includes three parts: a bidirectional GRU encoder, an incremental prediction head, and residual superimposed output; the incremental prediction head structure includes: stitching and hiding the output of the bidirectional GRU encoder. After Dropout regularization, the data is fed into a two-layer fully connected network connected by the GELU activation function, outputting the predicted settlement increment value. The calculation formula is as follows:

[0009] in, , This is the weight matrix of the fully connected layer. =64 represents the hidden layer dimension. , For the corresponding bias vector; The Gaussian error linear unit activation function is used; the incremental prediction head only outputs the settlement change between adjacent time steps, not the absolute settlement value. S300: The sample pairs obtained in S100 are divided into training set and test set in chronological order. The residual bidirectional gated recurrent unit model constructed in S200 is trained under supervision. The mean square error is used as the loss function, and the Adam optimizer combined with the cosine annealing learning rate scheduling strategy is used to update the parameters. S400. The trained residual bidirectional gated recurrent unit model is used for prediction on the test set. The prediction results are denormalized to restore the original values, and the final predicted surface settlement value is output. RMSE, MAE, and RMS values ​​are then used as indicators. 2 Three indicators are used to comprehensively evaluate the accuracy of the prediction.

[0010] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic units described in this invention, wherein: the Min-Max normalization process in S100 is used to eliminate the difference in settlement magnitude between different measuring points, and the sliding window method constructs the input sequence and target output pair according to the chronological order; the Min-Max normalization formula in S100 is: ; in, 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 this monitoring point in the training set, respectively; the normalization parameters are fitted only by the training set, and the same parameters are used to transform the test set.

[0011] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic unit described in this invention, the structure of the bidirectional GRU encoder in S200 includes: a forward GRU layer processing the input sequence in forward time order, and a backward GRU layer processing the input sequence in reverse time order, with the calculation formulas as follows:

[0012] in, This represents the hidden state output by the feedforward GRU network at time step t; This represents the normalized settlement data at time step t; This indicates the forward hidden state of the previous time step; This indicates the calculation process of the gated loop unit; t represents the current time step; This represents the hidden state output by the feedforward GRU network at time step t. This indicates the backward hidden state at the next time step; At each time step, the forward and backward hidden states are concatenated to obtain the complete hidden representation at that moment:

[0013] The bidirectional GRU encoder comprises a two-layer stacked structure with 64 hidden units per layer. Dropout regularization with a ratio of 0.2 is applied between the two layers. The spliced ​​hidden state at the last time step of the sequence is taken and fed into the subsequent incremental prediction head. The bidirectional GRU encoder of the residual bidirectional gated cyclic unit model can simultaneously capture the forward temporal dependency and backward context information of the time series data.

[0014] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic unit described in this invention, the GRU further includes two core gating mechanisms: a reset gate and an update gate. At each time step, given the input vector Hidden state from the previous moment The calculations for each gate and hidden state are as follows:

[0015]

[0016]

[0017]

[0018] in, To reset the inner vector, control the contribution of the hidden state in the previous time step to the candidate state; To update the inner vector, determine the proportion of the old state that is retained; This is the candidate hidden state; Hide the current state; This is the hidden state from the previous moment; The input weight matrix is ​​used to reset the gate; To update the gate input weight matrix; Input a weight matrix for the candidate states; To reset the weight matrix of the hidden state of the door; To update the hidden state weight matrix of the gate; The candidate state and hidden state weight matrix; These are the reset gate, update gate, and candidate hidden state bias vector, respectively. Use the Sigmoid activation function; This represents element-wise product.

[0019] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic unit described in this invention, wherein: the residual superposition output step in S200 includes: using the normalized settlement value at the end of the input window. Using the residual baseline, the predicted settlement increment output by the incremental prediction head is used as the settlement increment prediction value. By superimposing the residual baseline, the final predicted value for the next time step is obtained. :

[0020] The residual mechanism allows the model to fit only a small increment distribution during the training phase, enhancing the model's ability to respond to local abrupt changes in sedimentation acceleration.

[0021] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic unit described in this invention, wherein: in S300, the training set and test set are divided according to time sequence to preserve the sequential relationship of time series data and avoid data leakage.

[0022] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated recurrent units described in this invention, the training strategy steps in S300 include: updating parameters using the Adam optimizer, with an initial learning rate of 0.001 and a weight decay coefficient of 1×10⁻⁶. -3 The learning rate scheduling uses a cosine annealing strategy, with the minimum learning rate decaying to 1×10. -5 The maximum number of training epochs is 2000; gradient clipping is used during training, with a maximum norm of 1.0 to prevent gradient explosion; mean squared error is used as the loss function.

[0023] in, This represents the batch sample size. These are the model's predicted values. This represents the actual settlement value.

[0024] As a preferred embodiment of the deep foundation pit surface settlement prediction method based on residual bidirectional gated cyclic units described in this invention, wherein: in step S400, the model output is inversely normalized to obtain the settlement prediction value at the actual physical magnitude, and the predicted value is obtained by measuring RMSE, MAE, and R... 2 Three indicators are used to comprehensively evaluate the accuracy of the prediction; The formulas for calculating accuracy evaluation indicators include: Root mean square error:

[0025] Mean absolute error:

[0026] Coefficient of determination R 2 :

[0027] in, These are the model's predicted values. This represents the actual settlement value. This represents the true average settlement. R is the sample size; 2 The closer to 1, the better the fit.

[0028] A deep foundation pit surface settlement prediction system based on residual bidirectional controlled cyclic unit includes: The data acquisition and preprocessing module is used to collect cumulative land subsidence time series data and perform Min-Max normalization and sliding window sample construction. The model building module is used to build a residual bidirectional gated cyclic unit model that includes a bidirectional GRU encoder, an incremental prediction head, and residual superposition output. The model training module is used to divide the dataset in chronological order and train it using the MSE loss function, Adam optimizer, and cosine fire learning rate. The prediction and evaluation module is used to predict, inverse normalize, and evaluate test data using RMSE, MAE, and R... 2 The accuracy of the indicators is evaluated.

[0029] The beneficial effects of this invention are: This invention introduces a residual incremental learning mechanism to decompose the prediction target from absolute settlement into the sum of the current observation value and the incremental prediction value, fundamentally eliminating the response lag problem caused by the hidden state inertia effect and significantly improving the model's response capability to stage-specific abrupt changes such as accelerated settlement transitions. This invention employs a bidirectional GRU coding backbone, which can simultaneously utilize the forward trend features and backward dependency features of settlement time series. Compared with unidirectional GRU, it has a stronger ability to extract time series features. In synergy with the residual increment mechanism, it shows outstanding advantages in the settlement time series of deep foundation pits that are non-stationary and undergo stage-specific changes. The method of this invention is simple and efficient, with few model parameters and fast training speed. The residual connection mechanism can still ensure that the predicted value does not deviate from the true magnitude under the condition of scarce data or high noise, which enhances the robustness of the model and facilitates online deployment and real-time early warning in actual engineering. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall flow of the residual bidirectional gated cyclic unit model of the present invention; Figure 2 This is a schematic diagram of the internal structure of a GRU unit; Figure 3 This is a schematic diagram of the BiGRU bidirectional structure; Figure 4 This is a diagram of the residual bidirectional gated cyclic unit network structure. Detailed Implementation

[0031] 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 some embodiments of the present invention, but not all embodiments.

[0032] A method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units includes the following steps: S100. Collect cumulative surface settlement time-series data from monitoring points around the deep foundation pit, perform Min-Max normalization on the time-series data, and use the sliding window method to transform the normalized one-dimensional time-series data into supervised learning sample pairs. ; Settlement data from k consecutive time steps are used as model input, and the settlement value at the next time step is used as the prediction target. Input-output sample pairs are constructed, as shown below: in, The input sample sequence is constructed using a sliding window. This corresponds to the predicted target value; This represents the normalized settlement data at time step t; This represents the normalized settlement data at time step t+1; t represents the current time step; k represents the sliding window length. S200. Construct a residual bidirectional gated cyclic unit model, which includes three parts: a bidirectional GRU encoder, an incremental prediction head, and residual superimposed output. The structure of the incremental prediction head includes: stitching the output of the bidirectional GRU encoder to hide the state. After Dropout regularization, the data is fed into a two-layer fully connected network connected by the GELU activation function, outputting the predicted settlement increment value. The calculation formula is as follows:

[0033] in, , This is the weight matrix of the fully connected layer. =64 represents the hidden layer dimension. , For the corresponding bias vector; The Gaussian error linear unit activation function is used; the incremental prediction head only outputs the settlement change between adjacent time steps, not the absolute settlement value. S300: The sample pairs obtained in S100 are divided into training set and test set in chronological order. The residual bidirectional gated recurrent unit model constructed in S200 is trained under supervision. The mean square error is used as the loss function, and the Adam optimizer combined with the cosine annealing learning rate scheduling strategy is used to update the parameters. S400. The trained residual bidirectional gated recurrent unit model is used for prediction on the test set. The prediction results are denormalized to restore the original values, and the final predicted surface settlement value is output. RMSE, MAE, and RMS values ​​are then used as indicators. 2 Three indicators are used to comprehensively evaluate the accuracy of the prediction.

[0034] The specific implementation examples are as follows: Example 1: like Figure 1 As shown, the present invention provides a method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units, comprising the following steps: S100 data acquisition and normalization; S200 constructing a residual bidirectional gated cyclic unit model; S300 model training; S400 prediction output and accuracy evaluation.

[0035] Cumulative surface settlement time-series data were collected from multiple monitoring points around the deep foundation pit at fixed time intervals. The data included the monitoring date and the corresponding cumulative settlement amount (unit: mm). To eliminate dimensional differences between different monitoring points and accelerate model convergence, the original settlement time-series data were processed using Min-Max normalization.

[0036] in, This represents the original cumulative settlement value. The normalized settlement value. and These are the maximum and minimum values ​​of the time-series data for this monitoring point in the training set, respectively, which linearly map the original settlement sequence to the [0,1] interval. It should be noted that the normalization parameters are fitted only by the training set samples; the same parameters are used for the test set to avoid data leakage leading to distorted test set evaluations.

[0037] A sliding window is used to transform normalized one-dimensional time-series data into supervised learning samples. Let the window length be... (This invention takes) =4), with the first To the The input vector consists of normalized settlement values ​​at each time step. , No. The normalized settlement value of the step is the prediction target. Construct sample pairs :

[0038] The sample pairs were divided into a training set (first 70%) and a test set (last 30%) in chronological order, without random shuffling, in order to preserve the temporal dependence of the sedimentation time sequence.

[0039] The steps for constructing the residual bidirectional gated cyclic unit model in S200 are as follows: like Figure 4 As shown, the network structure of the residual bidirectional gated cyclic unit consists of three parts: a bidirectional GRU encoder, an incremental prediction head, and a residual superposition output. Given a length of... The input window sequence and the calculation process of the model to predict the settlement value at the next moment are completed in three steps.

[0040] Step 1: Bidirectional GRU Encoding. The input sequence is processed by a bidirectional GRU encoder to extract temporal features. For example... Figure 2 As shown, the GRU unit contains two core gating mechanisms: the ResetGate and the UpdateGate. At each time step, given the input vector Hidden state from the previous moment The calculations for each gate and hidden state are as follows:

[0041]

[0042]

[0043]

[0044] in, To reset the inner vector, control the contribution of the hidden state in the previous time step to the candidate state; To update the inner vector, determine the proportion of the old state that is retained; This is the candidate hidden state; Hide the current state; This is the hidden state from the previous moment; The input weight matrix is ​​used to reset the gate; To update the gate input weight matrix; Input a weight matrix for the candidate states; To reset the weight matrix of the hidden state of the door; To update the hidden state weight matrix of the gate; The candidate state and hidden state weight matrix; These are the reset gate, update gate, and candidate hidden state bias vector, respectively. Use the Sigmoid activation function; This represents element-wise product.

[0045] like Figure 3 As shown, BiGRU is composed of forward GRU layers and backward GRU layers stacked in parallel, and their calculation formulas are as follows:

[0046] in, This represents the hidden state output by the feedforward GRU network at time step t. This represents the normalized settlement data at time step t; This indicates the forward hidden state of the previous time step; This indicates the calculation process of the gated loop unit; t represents the current time step; This represents the hidden state output by the feedforward GRU network at time step t. This indicates the backward hidden state at the next time step; At each time step, the forward and backward hidden states are concatenated to obtain the complete hidden representation:

[0047] Extract the concatenated hidden state at the last time step of the sequence (Dimension 2 × hiddensize = 128) is fed into the incremental prediction head. It should be noted that the encoder contains a 2-layer stacked structure, with 64 hidden units in each layer, and a Dropout regularization with a ratio of 0.2 is applied between the two layers to suppress overfitting.

[0048] Step 2: Incremental prediction head. (The text abruptly ends here.) After Dropout regularization, the data is fed into a two-layer fully connected network connected by the GELU activation function, outputting the predicted settlement increment. :

[0049] This mechanism allows the model to fit only a small incremental distribution during the training phase, resulting in a more stable gradient signal and faster convergence. Simultaneously, by using the current real observation value as the residual benchmark, it ensures that the predicted values ​​do not deviate from the true magnitude. Even under conditions of scarce data or high noise, the residual connections guarantee that the predicted values ​​will not undergo a systematic shift in magnitude.

[0050] The model training steps in S300 are as follows: Of the sample pairs obtained from S100, the first 70% are used for model training, and the latter 30% are used for performance evaluation. No random shuffling is performed to preserve temporal dependencies. The model uses the Adam optimizer for parameter updates, with an initial learning rate of 0.001 and a weight decay coefficient of 1×10⁻⁶. -3The learning rate scheduling uses a cosine annealing (LR) strategy, with the minimum learning rate decaying to 1×10⁻⁶. -5 This allows the model to converge smoothly to a better solution in the later stages of training; the batch size is 16; the maximum number of training epochs is 2000; gradient clipping (maximum norm of 1.0) is used during training to prevent gradient explosion; mean squared error (MSE) is used as the loss function.

[0051] The steps for predicting output and evaluating accuracy in S400 are as follows: The trained residual bidirectional gated recurrent unit model was applied to the test set, and the prediction results were inversely normalized to restore the original values, outputting the final cumulative surface settlement prediction values ​​(unit: mm) in the original dimensions. RMSE, MAE, and RMS metrics were used. 2 The three indicators are used for comprehensive evaluation.

[0052] Example 2: Based on the foundation pit project of a certain station on Metro Line 6 in a certain city, this study selected four surface settlement monitoring points (DBC31-2, DBC32-2, DBC33-2, and DBC34-2) around the foundation pit. From July 6, 2024 to September 20, 2025, the cumulative measured settlement data (149 sets of data for each monitoring point) were collected at a monitoring frequency of 3 days to conduct empirical verification. The foundation pit is approximately 16.64 to 18.61 meters deep and adopts a support system of 800mm thick diaphragm walls plus 3 layers of internal bracing. The safety level of the foundation pit is Level 1.

[0053] Monitoring data shows: The surface settlement at each monitoring point exhibited four stages of change over time: Stage I (July 2024 to October 2024) was the initial slow settlement stage, with cumulative settlement within -3mm, reflecting limited disturbance to the surrounding surface in the early stages of foundation pit excavation; Stage II (November 2024 to February 2025) was the accelerated settlement stage, with the settlement rate at each monitoring point increasing significantly as the excavation depth increased, and the cumulative settlement rapidly increased to -5 to -7mm; Stage III (February 2025 to June 2025) was the fluctuating and stabilizing stage, with the settlement rate slowing down, and the curves at each monitoring point becoming extremely flat with slight fluctuations, reflecting the gradual convergence of stratum deformation after the support system came into play; Stage IV (June 2025 to September 2025) was the re-settlement stage, with settlement at each monitoring point accelerating again due to subsequent construction procedures, ultimately reaching a cumulative settlement of -11 to -12mm. This non-stationary, phased, and abrupt temporal characteristic places high demands on the dynamic response capability of the prediction model.

[0054] According to the method of this invention, data preprocessing, model training, and prediction validation were performed on the above four monitoring points, and the results were compared with two baseline models: standard BiGRU and GRU. All three models used the same hyperparameter configuration: hidden layer dimension d=64, stacked layers 2, dropout rate 0.2, batch size 16, and maximum training epochs 2000. The experimental results are shown in Table 1. Table 1 Comparison of accuracy evaluation indicators of the three models on the test set of four monitoring points.

[0055] Table 1 shows that the R-value of ResidualBiGRU on the test set at four monitoring points is... 2 The accuracy values ​​reached 0.9925, 0.9771, 0.9917, and 0.9822 respectively, with MAE all below 0.15 mm. All accuracy indicators were better than or equal to the BiGRU and GRU baseline models. The predictive advantage of ResidualBiGRU is mainly reflected in DBC31-2 and DBC332, where the settlement stage abrupt change characteristics are more significant. The difference between ResidualBiGRU and BiGRU is relatively large at these two monitoring points. However, on DBC32-2, where the settlement curve is more stable, the error difference between the three models is relatively narrowed. This is consistent with the application scenario of the residual increment mechanism—its advantage is mainly reflected in non-stationary time series with obvious stage jump characteristics, which is highly consistent with the actual engineering background of frequent changes in construction conditions and strong nonlinearity of stratum response during deep foundation pit exploration.

[0056] It should be noted that, in this document, 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. Unless otherwise specified, 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.

[0057] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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 method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units, characterized in that, Includes the following steps: S100. Collect cumulative surface settlement time-series data from monitoring points around the deep foundation pit, perform Min-Max normalization on the time-series data, and use the sliding window method to transform the normalized one-dimensional time-series data into supervised learning sample pairs. ; Settlement data from k consecutive time steps are used as model input, and the settlement value at the next time step is used as the prediction target. Input-output sample pairs are constructed, as shown below: in, The input sample sequence is constructed using a sliding window. This corresponds to the predicted target value; This represents the normalized settlement data at time step t; This represents the normalized settlement data at time step t+1; t represents the current time step; k represents the sliding window length. S200. Construct a residual bidirectional gated cyclic unit model, which includes three parts: a bidirectional GRU encoder, an incremental prediction head, and residual superimposed output. The incremental prediction head structure includes: stitching and hiding the output of the bidirectional GRU encoder. After Dropout regularization, the data is fed into a two-layer fully connected network connected by the GELU activation function, outputting the predicted settlement increment value. The calculation formula is as follows: in, , This is the weight matrix of the fully connected layer. =64 represents the hidden layer dimension. , For the corresponding bias vector; The Gaussian error linear unit activation function is used; the incremental prediction head only outputs the settlement change between adjacent time steps, not the absolute settlement value. S300: The sample pairs obtained in S100 are divided into training set and test set in chronological order. The residual bidirectional gated recurrent unit model constructed in S200 is trained under supervision. The mean square error is used as the loss function, and the Adam optimizer combined with the cosine annealing learning rate scheduling strategy is used to update the parameters. S400. The trained residual bidirectional gated recurrent unit model is used for prediction on the test set. The prediction results are denormalized to restore the original values, and the final predicted surface settlement value is output. RMSE, MAE, and RMS values ​​are then used as indicators. 2 Three indicators are used to comprehensively evaluate the accuracy of the prediction.

2. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 1, characterized in that: The Min-Max normalization process in S100 is used to eliminate the difference in settlement magnitude between different measuring points. The sliding window method constructs the input sequence and target output pair according to the chronological order. The Min-Max normalization formula in S100 is: ； in, 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 this monitoring point in the training set, respectively; the normalization parameters are fitted only by the training set, and the same parameters are used to transform the test set.

3. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 1, characterized in that: The structure of the bidirectional GRU encoder in S200 includes: a forward GRU layer processes the input sequence in forward time order, and a backward GRU layer processes the input sequence in reverse time order, with the calculation formulas as follows: in, This represents the hidden state output by the feedforward GRU network at time step t; This represents the normalized settlement data at time step t; This indicates the forward hidden state of the previous time step; This indicates the calculation process of the gated loop unit; t represents the current time step; This represents the hidden state output by the feedforward GRU network at time step t. This indicates the backward hidden state at the next time step; At each time step, the forward and backward hidden states are concatenated to obtain the complete hidden representation at that moment: The bidirectional GRU encoder comprises a two-layer stacked structure with 64 hidden units per layer. Dropout regularization with a ratio of 0.2 is applied between the two layers. The spliced ​​hidden state at the last time step of the sequence is taken and fed into the subsequent incremental prediction head. The bidirectional GRU encoder of the residual bidirectional gated cyclic unit model can simultaneously capture the forward temporal dependency and backward context information of the time series data.

4. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 3, characterized in that: The GRU also contains two core gating mechanisms: a reset gate and an update gate. At each time step, given the input vector Hidden state from the previous moment The calculations for each gate and hidden state are as follows: in, To reset the inner vector, control the contribution of the hidden state in the previous time step to the candidate state; To update the inner vector, determine the proportion of the old state that is retained; This is the candidate hidden state; Hide the current state; This is the hidden state from the previous moment; The input weight matrix is ​​used to reset the gate; To update the gate input weight matrix; Input a weight matrix for the candidate states; To reset the weight matrix of the hidden state of the door; To update the hidden state weight matrix of the gate; The candidate state and hidden state weight matrix; These are the reset gate, update gate, and candidate hidden state bias vector, respectively. Use the Sigmoid activation function; This represents element-wise product.

5. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 1, characterized in that: The residual superposition output step in S200 includes: using the normalized settlement value at the end of the input window. Using the residual baseline, the predicted settlement increment output by the incremental prediction head is used as the settlement increment prediction value. By superimposing the residual baseline, the final predicted value for the next time step is obtained. : The residual mechanism allows the model to fit only a small increment distribution during the training phase, enhancing the model's ability to respond to local abrupt changes in sedimentation acceleration.

6. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 1, characterized in that: The S300 uses a time-series approach to divide the training and test sets, preserving the sequential relationship of the time-series data and preventing data leakage.

7. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 6, characterized in that: The training strategy steps in S300 include: updating parameters using the Adam optimizer, with an initial learning rate of 0.001 and a weight decay coefficient of 1×10⁻⁶. -3 The learning rate scheduling uses a cosine annealing strategy, with the minimum learning rate decaying to 1×10. -5 The maximum number of training epochs is 2000; gradient clipping is used during training, with a maximum norm of 1.0 to prevent gradient explosion; mean squared error is used as the loss function. in, This represents the batch sample size. These are the model's predicted values. This represents the actual settlement value.

8. The method for predicting surface settlement of deep foundation pits based on residual bidirectional gated cyclic units according to claim 1, characterized in that: In S400, the model output is inversely normalized to obtain the settlement prediction value at the actual physical magnitude, and then measured using RMSE, MAE, and R... 2 Three indicators are used to comprehensively evaluate the accuracy of the prediction; The formulas for calculating accuracy evaluation indicators include: Root mean square error: Mean absolute error: Coefficient of determination R 2 : in, These are the model's predicted values. This represents the actual settlement value. This represents the true average settlement. R is the sample size; 2 The closer to 1, the better the fit.

9. A deep foundation pit surface settlement prediction system based on residual bidirectional controlled cyclic unit, characterized in that, include: The data acquisition and preprocessing module is used to collect cumulative land subsidence time series data and perform Min-Max normalization and sliding window sample construction. The model building module is used to build a residual bidirectional gated cyclic unit model that includes a bidirectional GRU encoder, an incremental prediction head, and residual superposition output. The model training module is used to divide the dataset in chronological order and train it using the MSE loss function, Adam optimizer, and cosine fire learning rate. The prediction and evaluation module is used to predict, inverse normalize, and evaluate test data using RMSE, MAE, and R... 2 The accuracy of the indicators is evaluated.

Images