A data-driven deep learning tunnel shield behavior prediction method and system

Abstract

The present disclosure belongs to the technical field of tunneling engineering, and specifically provides a tunnel shield behavior prediction method and system based on data-driven deep learning, wherein the method comprises: using an improved dynamic prediction model-subset simulation (SuS)-bidirectional long short-term memory method (Bi-LSTM) (namely, uS-Bi-LSTM), and incorporating subset simulation into Bi-LSTM to realize automatic parameter search. In addition, the framework also integrates wavelet packet transform as a filter to remove background noise recorded together with time series data in the tunnel excavation process. As a comparison, an improved dynamic prediction model SuS-LSTM is also constructed. Taking a large-diameter underwater tunnel project of a slurry balance shield tunneling machine as an example, the effectiveness and feasibility of the framework are verified, and the prediction accuracy can be improved to better avoid tunnel excavation risks.

Description

Technical Field

[0001] This disclosure relates to the field of tunnel boring engineering technology, and in particular to a method and system for predicting tunnel shield behavior based on data-driven deep learning. Background Technology

[0002] Tunnel boring machines (TBMs) are widely used in tunnel excavation projects, effectively addressing various challenges posed by complex geological conditions during shield tunneling. Uncertainties in geological conditions along the excavation direction can lead to problems such as cutterhead wear, positional deviations, and tunnel collapses. TBM operators must determine the TBM's advance speed (AS) based on potential geological conditions ahead. Real-time dynamic prediction of AS helps reduce excavation risks and provides TBM operators with reasonable recommendations. Due to the massive amount of data generated during tunnel excavation, machine learning (ML) algorithms have been widely applied to predict TBM performance during tunnel excavation. Koopialipoor et al. attempted to utilize genetic algorithms (GA) to improve the prediction accuracy of artificial neural networks (ANNs) in order to optimize over-blasting during tunnel blasting operations.

[0003] Currently, numerous models have been proposed using deep learning methods to analyze the tunnel boring machine (TBM) process. For example, Elbaz et al. proposed a deep learning model that combines convolutional neural networks with Long Short-Term Memory (LSTM) to optimize the energy cost of cutterhead drive during tunnel excavation. However, these comparisons neglect the impact of hyperparameter selection; for different machine learning and deep learning algorithms, hyperparameters are directly given without optimization. Therefore, as the application of deep learning in tunnel boring machines becomes increasingly widespread, a hyperparameter search method should be proposed.

[0004] The choice of hyperparameters significantly impacts the effectiveness and predictive performance of machine learning algorithms. Currently, hyperparameter optimization has become a crucial step in most machine learning algorithms, and proposed hyperparameter optimization methods can be categorized into two types: manual hyperparameter optimization and automatic hyperparameter optimization. Manual hyperparameter optimization methods, such as grid search and expert selection based on empirical rules, are always time-consuming and require substantial computational resources. Summary of the Invention

[0005] This disclosure aims to address at least one of the technical problems existing in the prior art, and proposes a data-driven deep learning-based method and system for predicting tunnel shield behavior.

[0006] In a first aspect, this disclosure provides a data-driven deep learning-based method for predicting tunnel shield behavior, including: Data from the earth pressure balance shield tunneling process is acquired in real time, input into a trained dynamic prediction model, and the tunneling path is obtained. The training process of the dynamic prediction model includes: Multiple sets of samples following a truncated Gaussian distribution were obtained through Monte Carlo simulation and input into the bidirectional long short-term memory model for training. The sample sets included: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

[0007] Preferably, the acquisition of raw data collected during the earth pressure balance shield tunneling process specifically includes: Export the raw data collected during the earth pressure balance shield tunneling process and filter out low-quality data points according to standards; Wavelet transform is used to denoise the data, and the denoised time series parameters are obtained. The time series parameters are sorted by importance, and low-correlation parameters are filtered out using LGBM to obtain the preprocessed dataset.

[0008] Preferably, the step of sorting the time series parameters according to importance and filtering out low-relevance parameters using LGBM to obtain a preprocessed dataset, wherein the preprocessed dataset is sorted from high to low importance, specifically including: The current cutting wheel drive CCWD, total thrust AF, total dry mass DM, working chamber pressure PWC, slurry pipeline flow rate FSL, attitude angle, feed pipeline flow rate FFL, axial direction, working chamber filling liquid level FLWC, output material volume OM, and feed bentonite FB are measured.

[0009] Preferably, the training stops after the training conditions are met, and the training conditions are: (6-8); RMSE q For the first q Group sample set.

[0010] Preferably, the distribution range of each parameter in the hyperparameter sample is as follows: Batch size, hidden layer size, and dropout rate [100,200], [50,1000], [0.01,0.5].

[0011] Preferably, dividing the original data into a training set and a test set specifically includes: The original data was divided into training and test sets, with a ratio of 80% and 20% respectively.

[0012] Preferably, stopping training after the training conditions are met specifically includes: Set a preset quantity value i, and calculate the i-th RMSE value of the test set; Sort the RMSE values ​​in ascending order and find the minimum value.

[0013] This invention also provides a data-driven deep learning-based method for predicting the behavior of tunnel boring machines (TBMs). The system can be used to implement the aforementioned data-driven deep learning-based TBM behavior prediction method. The system includes: The data acquisition module is configured to acquire data in real time during the earth pressure balance shield tunneling process; The input / output module is configured to input data into a trained dynamic prediction model to obtain the tunneling path. The training module is configured to obtain multiple sets of samples that follow a truncated Gaussian distribution through Monte Carlo simulation and input them into the bidirectional long short-term memory model for training. The sample sets include: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

[0014] The present invention also provides an electronic device, comprising: One or more processors; Memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the data-driven deep learning-based tunnel shield behavior prediction method. Attached Figure Description

[0015] Figure 1 A flowchart of a data-driven deep learning-based method for predicting tunnel shield behavior is provided in this embodiment of the present disclosure. Figure 2 A graph showing the trend of RMSE values ​​as the number of subset simulation layers (i) increases, provided for embodiments of this disclosure. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this disclosure, the disclosure will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Unless otherwise defined, the technical or scientific terms used in this disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms “first,” “second,” and similar terms used in this disclosure are not intended to indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an,” “a,” or “the,” and similar terms are not intended to limit the quantity, but rather to indicate the presence of at least one. The terms “comprising,” “including,” or “including,” and similar terms mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0018] In the various figures, the same elements are represented by similar reference numerals. For clarity, not all parts in the figures are drawn to scale. Furthermore, some well-known parts may not be shown in the figures.

[0019] Many specific details of this disclosure, such as the structure, materials, dimensions, processing methods, and techniques of the components, are described below to provide a clearer understanding of the disclosure. However, as those skilled in the art will understand, this disclosure may be implemented without following these specific details.

[0020] like Figure 1 As shown, this embodiment of the invention provides a data-driven deep learning-based method for predicting tunnel shield behavior, characterized by comprising: Data from the earth pressure balance shield tunneling process is acquired in real time, input into a trained dynamic prediction model, and the tunneling path is obtained. The training process of the dynamic prediction model includes: Multiple sets of samples following a truncated Gaussian distribution were obtained through Monte Carlo simulation and input into the bidirectional long short-term memory model for training. The sample sets included: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

[0021] This paper explores the feasibility of using a Bi-LSTM deep learning neural network to predict the performance of a tunnel boring machine (TBM) during tunneling, based on deep learning methods. Multiple dynamic prediction models incorporating LSTM, Bi-LSTM neural networks, LGBM, subset simulation, and wavelet transform are proposed to obtain the optimal solution for predicting the TBM's forward path. The model uses subset simulation to optimize the prediction hyperparameters and LGBM to filter important input parameters. The proposed dynamic prediction model is validated in a large-scale engineering project under complex underwater construction conditions, using 408,688 data points as the training set and 102,172 data points as the test set for big data analysis. The predictive capabilities of the LSTM-based and Bi-LSTM-based dynamic prediction models are compared. The proposed Bi-LSTM-based dynamic prediction model can reflect the complexity and nonlinearity of the modeling problem during TBM tunneling and can effectively predict the performance parameters of a slurry balance TBM, thus preventing operational errors by the TBM operator during tunnel excavation. Furthermore, it can achieve real-time prediction of operating parameters based on time, thereby better realizing the goal of intelligent operation.

[0022] The method involves calculating the root mean square error (RMSE) of each sample set and finding the minimum value. The corresponding sample set is considered the optimal hyperparameter. This method, known as Subset Simulation (SuS), was proposed and applied in the field of geotechnical engineering reliability (Au, 2004; Au, 2005; Straub and Papaioannou, 2015). The core idea of ​​Subset Simulation is to introduce intermediate failure events and represent small failure probabilities with a series of large failure probabilities. Here, the problem of hyperparameter optimization is to find the optimal hyperparameter set, whose output hyperparameter samples enable the deep learning model to have better predictive performance. Therefore, the process of finding the optimal hyperparameter samples is similar to finding the small failure probabilities. Assume F0 is the final optimal sample, F1, F2, …, F… m It is a series of intermediate optimal samples. The relationship between each sample can be expressed by the following formula: (6-1); For each sample F i The definition of F i ={g(x) <g i}, where g(x) represents the probability of the optimal value of the subset occurring, g i Used to determine the critical value of a function, whose magnitude is determined by g1>g2>...>g m The constraint is 0. The final probability of the optimal value occurring can be expressed by the following formula: ; Where n is the number of samples, It is an exponential function. Furthermore, the... i The samples of the +1 subset are the first... i The samples generated are from the range of the optimal solutions of each subset, therefore the following expression holds: ; F i The sample points in the data are generated according to the following formula: ; Conditional probability P i+1 It can be represented as: ; Here, q(x) is the conditional probability density function of the optimal value region.

[0023] The flowchart of the proposed SuS-Bi-LSTM dynamic prediction model is as follows: Figure 1 As shown. In the first layer of the subset simulation, the hyperparameters required to find the optimal solution are determined. In this study, these are the batch size (…). μ ), hidden size ( ) and dropout ratio ( ), forming a set of hyperparameter samples ( Here, n sets were generated through Monte Carlo simulation (MCS). i These follow a truncated Gaussian distribution and are fed into the Bi-LSTM training. Each input group... i Perform i-LSTM training and calculate the i-th RMSE value on the Bi-LSTM test set. To find the optimal sample, calculate the RMSE... q Sort the values ​​in ascending order and find the minimum value. q = 1, 2,..., n). Furthermore, the criterion for the subset simulation to break the cycle is expressed by formula (6-8). If this criterion is not met, 1 / u samples will be simulated from each seed to obtain n samples for the next level subset, and new samples will be generated through Markov Chain Monte Carlo Simulation (MCMCS). The value of u is related to the number of optimal hyperparameter samples output. As the subset simulation continues, the hyperparameter samples will gradually fall into adjacent regions; (6-8); In the proposed SuS-Bi-LSTM dynamic prediction model, the first 80% of the data is used for subset simulation training, and the remaining 20% ​​is used as the validation set. In the Bi-LSTM (or its alternative LSTM) step, the data is divided into training and test sets at 80% and 20% ratios, respectively (referencing the division ratios of Aliyuda and Howell, 2019).

[0024] In a preferred embodiment, sensors mounted on the tunnel boring machine automatically collect data every ten seconds. Therefore, a large amount of raw data is collected during the excavation process, which is crucial for deep learning methods. For example, in a large underwater tunneling project, different geological sections were selected. This chapter uses data collected from rings 445 to 501 (section DK0+243~DK1+070), and the geological composition of the relevant sections is shown in Table 6-1. Before inputting the data into the dynamic prediction model, preprocessing is required. The preprocessing workflow used in the proposed model mainly includes three steps: (1) Remove low-quality data points from the raw data recorded during the SPB tunneling process; (2) Background noise is removed by wavelet packet transform to improve the predictive ability of the model; (3) To avoid the input of low-correlation parameters affecting the prediction accuracy of LSTM and Bi-LSTM networks, parameter selection is performed based on LGBM; As mentioned earlier, LGBM is an improved algorithm based on Gradient Boosting Decision Tree (GBDT), a popular machine learning algorithm. Unlike traditional gradient boosting decision trees, LightGBM has the following advantages: (1) fast convergence speed; (2) low memory usage; (3) strong parallel computing capability; and (4) high accuracy. Compared with small-scale datasets, LightGBM performs better with larger datasets (more than 10,000 data entries), which is very consistent with our data scale. In addition, when the data volume is less than 10,000 data entries, the extreme gradient boosting model (XGBoost) may be the preferred choice for prediction. In order to avoid the input of low-correlation parameters affecting the prediction accuracy, in the proposed improved hyperparameter optimization dynamic prediction model, parameter selection is achieved by regression of LGBM. Based on the results of LGBM calculation (see Table 6-2), we discarded the lower-ranked parameters and selected the top 8 parameters to continue training the proposed dynamic prediction model. In a preferred embodiment, although current research focuses on the optimization of Bi-LSTM, this training method can be adapted to any other machine learning method. In other words, the hyperparameter optimization method in the proposed dynamic prediction framework can be used to optimize various machine learning methods. Besides the SuS-Bi-LSTM dynamic prediction model, this paper also compares the SuS-LSTM dynamic prediction model based on the LSTM network training method. It is important to note that in the entire dynamic prediction framework, only the prediction step is replaced; other steps remain unchanged. Table 6-3 lists the hyperparameters determined by SuS and the empirical method. This includes batch size (…). μ ), hidden size ( σ ) and dropout rate ( ρ Three important hyperparameters, including , constitute a set of hyperparameters. Here, each parameter is assumed to follow a truncated Gaussian distribution over different intervals, which are selected through repeated testing to avoid inappropriate values ​​being included in the subset simulation loop. Based on the distribution characteristics of each hyperparameter, in order to find the optimal solution, μ, σ, ρ The distribution ranges were set to [100, 2000], [50, 1000], and [0.01, 0.5], respectively. Furthermore, for comparison, another set of values ​​was used according to Wang et al. (2021a). To determine; Figure 2 The trend of RMSE values ​​is shown, where the mean, minimum, and maximum RMSE values ​​of each subset simulation level are calculated to find the optimal set of hyperparameters for predicting tunneling speed. The RMSE values ​​exhibit a gradually decreasing trend, indicating the convergence of the proposed model. Compared to the mean and minimum values, the maximum RMSE value of subset simulations varies more widely. In this study, the mean and minimum RMSE values ​​demonstrate the convergence stability of subset simulations, reflected in the less pronounced decrease in RMSE values ​​later in model training. Furthermore, SuS-Bi-LSTMDPM requires 13 subset simulations (i.e.,... i = 13) to find the optimal solution, while SuS-LSTM DPM requires 9 subset simulations (i.e. i =9), which proves that the convergence capabilities of SuS-LSTM DPM and SuS-Bi-LSTM DPM are quite similar. Overall, in the process of finding the optimal hyperparameters through subset simulation, the RMSE value of the LSTM prediction model is lower than that of the Bi-LSTM prediction model, which can be seen from... Figure 2This is evidenced by the comparison curves in each group. Therefore, this demonstrates that LSTM, compared to Bi-LSTM, exhibits better dataset adaptability to a certain extent. This slight limitation is reflected in the small fluctuations in both the minimum and average RMSE values ​​of each subset simulation. Furthermore, due to the reasonable selection of the cutoff interval, appropriate data types, and a reliable prediction model, the overall range of RMSE variation is relatively small. Although the above discussion demonstrates the convergence capability of subset simulation and the necessity of parameter optimization, the computational results obtained from training and testing with the final configuration values ​​are more convincing.

[0027] This invention proposes a data-driven improved dynamic prediction model for tunnel boring machine (TBM) excavation behavior. This improved model, based on a deep learning algorithm, is used to predict TBM excavation parameters and evaluate TBM excavation performance. The LSTM and Bi-LSTM algorithms are replaced in the proposed prediction framework, and the prediction results are compared. Furthermore, a novel data-driven improved dynamic prediction model (SuS-Bi-LSTM) SS-Bi-LSTM is developed by integrating wavelet packet transform, LGBM, subset simulation, and Bi-LSTM. Results demonstrate that the proposed improved dynamic prediction model can automatically search for optimal hyperparameter samples based on deep learning algorithms, and the feasibility of the proposed improved model is verified through a practical engineering case study. In addition, the prediction performance of the proposed SuS-Bi-LSTM improved dynamic prediction model and the SuS-LSTM improved dynamic prediction model are compared. The proposed framework effectively improves prediction accuracy: compared to the case of only performing LSTM calculations, the RMSE value of SuS-LSTM is reduced by 29.82%, and R... 2 The value was improved by 7.85%; furthermore, compared with the case of Bi-LSTM computation alone, the RMSE value of the SuS-Bi-LSTM improved dynamic prediction model was reduced by 29.51%, and R... 2 The value increased by approximately 10%.

[0028] Beneficial effects: Building upon the proposed deep learning dynamic prediction model, the optimal selection rules for hyperparameters in deep learning remain unclear. Previous studies have often employed manual hyperparameter search methods, such as grid search, which are both cumbersome and difficult to find the optimal hyperparameter set. Therefore, this embodiment further optimizes the proposed algorithm by incorporating subset simulation in one step. It transforms the hyperparameter selection problem into a probability problem of falling into the failure domain, gradually determining the optimal hyperparameter selection range and achieving hyperparameter optimization. Compared with the unoptimized dynamic prediction model, the prediction accuracy is significantly improved.

[0029] Furthermore, the proposed improved dynamic prediction model incorporates XGBoost into the model input parameter importance selection process, and the prediction results are compared with traditional machine learning methods (such as support vector machines), demonstrating significantly superior prediction performance. The optimized dynamic prediction model can provide tunnel boring machine (TBM) operators with more accurate predictions of TBM construction parameters, offering more reliable decision support.

[0030] This invention also provides a data-driven deep learning-based method for predicting the behavior of tunnel boring machines (TBMs). The system can be used to implement the aforementioned data-driven deep learning-based TBM behavior prediction method. The system includes: The data acquisition module is configured to acquire data in real time during the earth pressure balance shield tunneling process; The input / output module is configured to input data into a trained dynamic prediction model to obtain the tunneling path. The training module is configured to obtain multiple sets of samples that follow a truncated Gaussian distribution through Monte Carlo simulation and input them into the bidirectional long short-term memory model for training. The sample sets include: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

[0031] This invention also provides an electronic device, characterized in that it includes: One or more processors; Memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the data-driven deep learning-based tunnel shield behavior prediction method as described above.

[0032] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.

Claims

1. A data-driven deep learning-based method for predicting tunnel shield behavior, characterized in that, include: Data from the earth pressure balance shield tunneling process is acquired in real time, input into a trained dynamic prediction model, and the tunneling path is obtained. The training process of the dynamic prediction model includes: Multiple sets of samples following a truncated Gaussian distribution were obtained through Monte Carlo simulation and input into the bidirectional long short-term memory model for training. The sample sets included: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

2. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 1, characterized in that, Before inputting the data into the dynamic prediction model, preprocessing is required, specifically including: Export the raw data collected during the earth pressure balance shield tunneling process and filter out low-quality data points according to standards; Wavelet transform is used to denoise the data, and the denoised time series parameters are obtained. The time series parameters are sorted by importance, and low-correlation parameters are filtered out using LGBM to obtain the preprocessed dataset.

3. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 2, characterized in that, The process of sorting time series parameters by importance and filtering out low-relevance parameters using LGBM to obtain a preprocessed dataset, specifically includes: The current cutting wheel drive CCWD, total thrust AF, total dry mass DM, working chamber pressure PWC, slurry pipeline flow rate FSL, attitude angle, feed pipeline flow rate FFL, axial direction, working chamber filling liquid level FLWC, output material volume OM, and feed bentonite FB are measured.

4. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 1, characterized in that, The training process stops once the training conditions are met. The training conditions are: ； RMSE q For the first q Group sample set.

5. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 1, characterized in that, The distribution range of each parameter in the hyperparameter sample is as follows: Batch size, hidden layer size, and dropout rate [100,200], [50,1000], [0.01,0.5].

6. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 1, characterized in that, The process of dividing the original data into training and testing sets specifically includes: The original data was divided into training and test sets, with a ratio of 80% and 20% respectively.

7. The method for predicting tunnel shield behavior based on data-driven deep learning according to claim 1, characterized in that, The step of stopping training after the training conditions are met specifically includes: Set a preset quantity value i, and calculate the i-th RMSE value of the test set; Sort the RMSE values ​​in ascending order and find the minimum value.

8. A method for predicting tunnel shield behavior based on data-driven deep learning, characterized in that, The system can be used to implement the data-driven deep learning-based tunnel shield behavior prediction method according to any one of claims 1 to 7, and the system includes: The data acquisition module is configured to acquire data in real time during the earth pressure balance shield tunneling process; The input / output module is configured to input data into a trained dynamic prediction model to obtain the tunneling path. The training module is configured to obtain multiple sets of samples that follow a truncated Gaussian distribution through Monte Carlo simulation and input them into the bidirectional long short-term memory model for training. The sample sets include: Once the training conditions are met, training is stopped. The root mean square error of each sample set is calculated and the minimum value is found. The corresponding sample set is the optimal hyperparameter. The raw data collected during the earth pressure balance shield tunneling process is obtained, and the raw data is divided into training set and test set. The bidirectional long short-term memory model with optimal hyperparameters is trained to obtain a dynamic prediction model.

9. An electronic device, characterized in that, include: One or more processors; Memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the data-driven deep learning-based tunnel shield behavior prediction method as described in any one of claims 1 to 7.

Images