A shield cutting pile foundation thrust force intelligent prediction system

Abstract

The application discloses a shield cutting pile foundation thrust intelligent prediction system, which adopts the architecture of sequentially connecting data acquisition modules, prediction model modules, physical constraint modules and result output modules along the data flow direction; the prediction model is a hybrid time series prediction model with cascaded convolutional neural network and bidirectional long short-term memory network, which has local feature extraction and time series dependence capturing capabilities; the physical constraint sequentially applies non-negative constraint, recursive adjacent time thrust change rate constraint and shield equipment thrust distance constraint. The application can improve the accuracy and stability of thrust prediction under the cutting pile foundation working condition, and the output result conforms to the engineering practice and equipment characteristics, and has good on-site engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control in shield tunnel construction, and specifically to an intelligent prediction system for the thrust of shield tunnel cutting pile foundation. Background Technology

[0002] With the increasing density of urban underground space development, the number of engineering scenarios involving tunnel boring machines (TBMs) directly cutting through existing reinforced concrete pile foundations is growing. In this condition, the tunneling face is a heterogeneous composite medium of the stratum and the pile foundation concrete. The thrust is influenced by multiple factors, including earth pressure, cutterhead condition, stratum strength, and burial depth, resulting in large fluctuations and strong abrupt changes. Improper control can easily lead to safety accidents such as cutterhead breakage, TBM jamming, and excessive surface settlement. Therefore, accurate prediction of thrust under this condition is of significant engineering importance.

[0003] Currently, the setting and adjustment of shield thrust at the engineering site mainly rely on the driver's practical experience for manual control, or are supplemented by empirical formulas for preliminary estimation. Under the complex working conditions of cutting pile foundations, it is difficult to respond quickly to the violent fluctuations in thrust, and the control accuracy and timeliness are insufficient, which cannot meet the precise control requirements in this scenario.

[0004] Existing research on shield tunneling parameter prediction mainly focuses on conventional homogeneous strata, with less attention paid to the special working conditions of shield tunneling directly cutting pile foundations. The prediction models do not fully consider the load mutation mechanism and the coupling effect of multiple influencing factors during the cutting process, and the prediction accuracy and adaptability are insufficient when directly applied to the working conditions of cutting pile foundations.

[0005] Furthermore, existing deep learning-based intelligent thrust prediction methods generally prioritize numerical accuracy as the core optimization objective, neglecting the engineering physical constraints and on-site usability of the prediction results. Purely data-driven models are prone to outputting results that do not conform to engineering realities, such as negative thrust, step-like abrupt changes, and exceeding the equipment's rated range. These results cannot match the execution characteristics of the tunnel boring machine's hydraulic system and are difficult to use directly to guide construction. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide an intelligent prediction system for shield tunneling thrust in pile foundation cutting, which solves the problems of existing shield tunneling thrust prediction relying on human experience, having insufficient prediction accuracy, and poor practicality in pile foundation cutting conditions.

[0007] Technical Solution: The intelligent thrust prediction system for shield tunneling pile foundations according to the present invention includes a data acquisition module, a prediction model module, a physical constraint module, and a result output module connected sequentially along the data flow direction. The data acquisition module is used to collect tunneling parameters and historical thrust data under shield tunneling pile foundation working conditions, providing multi-dimensional input basis for thrust prediction. The prediction model module incorporates a hybrid time-series prediction model composed of a convolutional neural network and a bidirectional long short-term memory network, possessing both local feature extraction and time-series dependency capture capabilities, used to output preliminary thrust prediction values ​​based on the input data. The physical constraint module is used to sequentially apply non-negativity constraints, adjacent-time thrust change rate constraints, and shield equipment thrust range constraints to the preliminary prediction values, eliminating abnormal outputs that do not conform to engineering reality, and obtaining the final thrust prediction value. The result output module is used to output the final thrust prediction value.

[0008] Furthermore, a data preprocessing module is provided between the data acquisition module and the prediction model module. This module cleans and standardizes the acquired raw data before inputting it into the prediction model module, thereby eliminating the impact of missing data and dimensional differences on prediction accuracy.

[0009] Furthermore, a time series sample construction module is provided between the data preprocessing module and the prediction model module. This module is used to construct fixed-length time series prediction samples from the preprocessed data and input them into the prediction model module, thereby converting the continuously running data into an input format that can be recognized by the hybrid time series model.

[0010] Furthermore, the data acquisition module interfaces with the shield machine's sensor interface and collects parameters including earth pressure, cutterhead temperature, torque, stratum strength, shield depth, and other tunneling parameters, as well as historical thrust data, covering the main influencing factors of thrust changes under the condition of cutting pile foundations in multiple dimensions.

[0011] Furthermore, the hybrid time-series prediction model, composed of a convolutional neural network and a bidirectional long short-term memory network cascaded together, includes, along the data flow direction, a sequence input layer, a one-dimensional convolutional layer, an activation layer, a max pooling layer, a first dropout layer, a bidirectional long short-term memory layer, a second dropout layer, and a fully connected layer. The one-dimensional convolutional layer is used to extract local features of the input sequence and capture the short-term abrupt changes in thrust. The bidirectional long short-term memory layer is used to capture the bidirectional long-term dependencies of time-series data and grasp the overall evolution trend of thrust. The two dropout layers are used to reduce the risk of model overfitting and improve the generalization ability of the working conditions.

[0012] Furthermore, the key hyperparameters of the hybrid time-series prediction model are determined through Bayesian optimization. These key hyperparameters include the length of the historical sequence, the number of units in the bidirectional long short-term memory network, the dropout rate, and the initial learning rate. By automatically searching for the optimal parameter combination through Bayesian optimization, the model can adapt to the characteristics of limited sample size and strong thrust fluctuations in cutting pile foundation conditions, thereby improving prediction stability and accuracy.

[0013] Furthermore, the physical constraint module first performs denormalization processing on the normalized preliminary prediction value output by the prediction model module to restore it to the original physical dimensions, and then applies the three constraints in sequence to ensure that each engineering constraint is effective based on the real physical magnitude and that the constraint results have engineering significance.

[0014] Furthermore, the thrust change rate constraint between adjacent time moments is implemented in a recursive manner. Based on the final thrust value after constraint correction at the previous time moment, the change range of the thrust value at the current time moment is limited to a preset threshold range, ensuring that the output thrust curve is smooth and continuous, adapting to the response rate of the shield hydraulic system, and avoiding step thrust commands that the equipment cannot execute.

[0015] Furthermore, the data preprocessing module fills in missing values ​​in the original data and uses a minimum-maximum normalization method to map each feature to a preset numerical range, while saving the normalization parameters of each feature for subsequent inverse normalization; while unifying the data units to improve the model convergence efficiency, it ensures that the final prediction result can be restored to the real physical units.

[0016] Furthermore, the time series sample construction module uses the sliding window method to construct time series samples from the preprocessed multi-feature sequences. It slides to construct historical time series samples of a fixed length according to a set step size, which are used to predict the thrust value at the next moment. High-density sample pairs are generated to fully explore the thrust change pattern in the time series data.

[0017] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: it is highly adaptable to the special working conditions of shield tunneling and pile foundation cutting, can significantly improve the thrust prediction accuracy in this complex scenario, can effectively respond to the drastic fluctuations in thrust, has both stable prediction performance and excellent dynamic following ability, and the prediction results are consistent with the actual engineering and equipment operation characteristics, thus having good practicality in field engineering. Attached Figure Description

[0018] Figure 1 The overall system structure flowchart; Figure 2 Results comparison curves. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Figure 1As shown in the figure, the intelligent prediction system for shield tunneling pile foundation thrust described in this embodiment is composed of a data acquisition module, a data preprocessing module, a time series sample construction module, a prediction model module, a physical constraint module, and a result output module connected sequentially along the data flow direction.

[0020] The data acquisition module connects to the onboard sensor interface of the tunnel boring machine (TBM) to collect five types of tunneling parameters in real time under the condition of cutting pile foundation: earth pressure, cutterhead temperature, torque, stratum strength, and shield depth. At the same time, it reads the shield thrust value at the current moment. The above six-dimensional data are packaged into feature vectors according to the sampling time sequence and sent to the data preprocessing module. The historical value of shield thrust is also used as one of the input features of the prediction model.

[0021] The data preprocessing module has a built-in missing value processor and normalizer. First, the missing value processor scans the original data column by column. When missing data is detected in parameters such as formation strength, linear interpolation is used to fill the gap by taking the linear average of the two adjacent valid values ​​before and after the missing position. Then, the normalizer uses the minimum-maximum normalization method, and the calculation formula is: In the formula, The normalized values ​​are given by x, which is the initial eigenvalue. The minimum value of the corresponding feature. The maximum value of the corresponding feature is 1; all six-dimensional features are mapped to the interval [-1, 1], and the minimum and maximum values ​​of each dimension are stored in an internal register for subsequent inverse normalization calls.

[0022] The time-series sample construction module includes a sliding window extractor and a sample storage, employing the sliding window method to construct time-series prediction samples. The sliding window extractor slides sequentially along the time frame with a step size of 1. The window length is provided by the hyperparameter configuration register of the prediction model module, ranging from 40 to 90 time units. At each sliding step, the feature matrix within the window is extracted as an input sample, and the thrust value of the next time step immediately following the window is used as the prediction target label for that sample. The resulting sample pairs are stored in the sample storage. This embodiment uses actual tunneling data from a shield tunneling pile foundation construction section to verify the system performance. All time-series data is divided sequentially, with the first 70% used as the training set and the last 30% as the validation set. The Bayesian optimization process uses the root mean square error on the validation set as the objective function. After determining the optimal hyperparameter combination, model training is completed on the training set, with the same validation set used for monitoring during training.

[0023] The prediction model module incorporates a hybrid temporal prediction model consisting of a convolutional neural network and a bidirectional long short-term memory network, and is equipped with a hyperparameter configuration register, a Bayesian optimizer, and a network trainer, which are used for hyperparameter storage, optimal hyperparameter search, and network training, respectively. The network structure of the hybrid temporal prediction model, along the data flow direction, consists of a sequence input layer, a one-dimensional convolutional layer, a ReLU activation layer, a max pooling layer, a first dropout layer, a bidirectional long short-term memory layer, a second dropout layer, and a fully connected layer.

[0024] The sequence input layer receives a feature matrix of size 6×L, where L is the length of the historical sequence window. The one-dimensional convolutional layer has 64 convolutional kernels with a kernel size of 1, a stride of 1, and no padding. It is used to map the feature dimensions and expand the channels of the input sequence, extract local features of the input sequence, and capture the short-term mutation characteristics of thrust. The ReLU activation layer performs non-linear activation on the convolutional output. The max pooling layer has a pooling window size of 2 and a stride of 2. It is used to compress the temporal length and reduce the subsequent computational load. The first and second dropout layers have the same dropout rate, which is uniformly set by the hyperparameter configuration register. It is used to randomly deactivate some neurons, reduce the risk of model overfitting, and improve the generalization ability under working conditions. The bidirectional long short-term memory layer contains two long short-term memory networks, forward and backward. The output mode retains only the output of the last time step. It is used to capture the bidirectional long-term dependencies in the temporal data and grasp the overall evolution trend of thrust. The fully connected layer outputs a scalar value, namely the normalized preliminary thrust prediction value.

[0025] The key hyperparameters of the hybrid time-series prediction model are automatically determined through Bayesian optimization. These key hyperparameters include the historical sequence length, the number of units in the bidirectional long short-term memory network, the dropout rate, and the initial learning rate. The specific implementation of Bayesian optimization is as follows: a hyperparameter search space is defined, with the historical sequence length being an integer from 40 to 90, the number of bidirectional long short-term memory units being an integer from 48 to 96, the dropout rate being a real number from 0.15 to 0.3, and the initial learning rate being a logarithmic scale from 0.0005 to 0.005. A set of hyperparameters is used to temporarily configure the network. After training on the training set for 60 epochs, the root mean square error (RMSE) of the validation set is calculated as the objective function value. A Gaussian process surrogate model and an expected improvement acquisition function are used. After 25 iterations of evaluation, the hyperparameter combination that minimizes the RMS error of the validation set is selected and stored in the hyperparameter configuration register. The optimal hyperparameter combination obtained in this embodiment is: historical sequence length 40, number of bidirectional long short-term memory units 48, dropout rate 0.2333, and initial learning rate 0.0012461. After obtaining the optimal hyperparameters, the prediction network is trained completely using a network trainer. The total number of training rounds is 100, the minimum batch size is 128, the Adam optimizer is used, the initial learning rate is the optimal value in the register, the sample order is not shuffled during training, and the network weights are fixed after training for subsequent thrust prediction.

[0026] The physical constraint module includes an inverse normalizer, a non-negativity limiter, a rate of change limiter, and a range limiter along the data flow direction, which sequentially correct the initial predicted values ​​output by the model.

[0027] First, using the minimum and maximum thrust dimensions stored in the data preprocessing module, the initial normalized prediction value is restored to the actual physical dimension thrust value through the inverse normalizer. The calculation formula is as follows: In the formula, F is the actual thrust value after restoration, and y is the normalized thrust prediction value output by the model. The maximum value of the thrust in the training set. This represents the minimum thrust in the training set.

[0028] Next, the non-negative limiter forces thrust values ​​less than 0 after restoration to 0, eliminating physically impossible negative thrust output. Then, the rate-of-change limiter uses a recursive method to constrain the thrust rate of change between adjacent moments. Based on the final thrust value after constraint correction at the previous moment, the change in thrust value at the current moment is limited to a preset threshold range. In this embodiment, the rate-of-change threshold is 0.15, meaning the thrust value at the current moment is clamped between 0.85 and 1.15 times the thrust value at the previous moment, ensuring a smooth and continuous output thrust curve and avoiding step commands that the shield hydraulic system cannot execute. Finally, the range limiter presets a lower and upper thrust limit, clipping the thrust value to this range. In this embodiment, the lower limit is set to 2000 kN and the upper limit to 3000 kN. The thrust value after the above processing is used as the final thrust prediction value and sent to the result output module. In practical applications, the rate-of-change threshold and range upper and lower limits can be configured according to the specific shield machine model and the cutting pile foundation conditions.

[0029] The result output module receives the final thrust prediction value output by the physical constraint module and outputs it to the shield machine's display interface or pushes it to the thrust control system, providing a reference for thrust control during shield cutting of pile foundations.

[0030] The trained hybrid time-series prediction model is used to predict thrust on the test set. After the prediction results are processed by the physical constraint module, they are compared with the measured thrust values. The system automatically calculates and outputs evaluation indicators: the mean absolute error is 20.6678, the root mean square error is 31.0951, and the coefficient of determination R² reaches 0.8880.

[0031] like Figure 2As shown in the comparison curves of the measured thrust and the system's predicted thrust on the test set, the solid line represents the actual thrust value, and the dashed line represents the predicted thrust value output by the system after physical constraint correction. The output results have no negative values ​​or abrupt outliers, and the predicted curve matches the measured curve well, effectively following the thrust in the rapid change segment.

Claims

1. A smart prediction system for the thrust of a shield tunneling machine cutting a pile foundation, characterized in that, The system includes a data acquisition module, a prediction model module, a physical constraint module, and a result output module connected sequentially along the data flow direction. The data acquisition module is used to collect tunneling parameters and historical thrust data under the working condition of shield cutting pile foundation. The prediction model module has a built-in hybrid time-series prediction model composed of a convolutional neural network and a bidirectional long short-term memory network, which is used to output a preliminary thrust prediction value based on the input data. The physical constraint module is used to apply non-negativity constraints, thrust change rate constraints at adjacent time moments, and shield equipment thrust range constraints to the preliminary prediction value in sequence to obtain the final thrust prediction value. The result output module is used to output the final thrust prediction value.

2. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 1, characterized in that, A data preprocessing module is also provided between the data acquisition module and the prediction model module, which is used to clean and standardize the acquired raw data before inputting it into the prediction model module.

3. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 2, characterized in that, A time series sample construction module is also provided between the data preprocessing module and the prediction model module, which is used to construct a fixed-length time series prediction sample from the preprocessed data and then input it into the prediction model module.

4. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 1, characterized in that, The data acquisition module interfaces with the tunnel boring machine's sensor interface and collects parameters including earth pressure, cutterhead temperature, torque, ground strength, shield depth, and other tunneling parameters, as well as historical thrust data.

5. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 1, characterized in that, The hybrid time-series prediction model, which is composed of a convolutional neural network and a bidirectional long short-term memory network, includes, in sequence along the data flow direction, a sequence input layer, a one-dimensional convolutional layer, an activation layer, a max pooling layer, a first dropout layer, a bidirectional long short-term memory layer, a second dropout layer, and a fully connected layer; the one-dimensional convolutional layer is used to extract local features of the input sequence, and the bidirectional long short-term memory layer is used to capture the bidirectional long-term dependencies of the time-series data.

6. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 5, characterized in that, The key hyperparameters of the hybrid time-series prediction model were determined through Bayesian optimization. These key hyperparameters include the length of the historical sequence, the number of units in the bidirectional long short-term memory network, the dropout rate, and the initial learning rate.

7. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 1, characterized in that, The physical constraint module first performs inverse normalization processing on the normalized preliminary prediction value output by the prediction model module to restore it to the original physical dimensions, and then applies the three constraints in sequence.

8. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 7, characterized in that, The thrust change rate constraint between adjacent time moments is implemented in a recursive manner. Based on the final thrust value after constraint correction at the previous time moment, the change range of the thrust value at the current time moment is limited to a preset threshold range.

9. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 2, characterized in that, The data preprocessing module fills in missing values ​​in the original data and uses a minimum-maximum normalization method to map each feature to a preset numerical range. At the same time, it saves the normalization parameters of each feature for subsequent inverse normalization.

10. The intelligent prediction system for shield tunneling pile foundation thrust according to claim 3, characterized in that, The time series sample construction module uses the sliding window method to construct time series samples from the preprocessed multi-feature sequences. It slides to construct historical time series samples of a fixed length according to a set step size, which are used to predict the thrust value at the next moment.

Images