A method for predicting TBM stable phase performance parameters based on rising phase data

Abstract

The present invention provides a method for predicting performance parameters of a TBM stable segment based on rising segment data, comprising the following steps: collecting TBM tunnel construction data sets under different rock grades and extracting effective excavation cycle data therefrom; dividing the effective excavation cycle data into stages, extracting characteristic data sets of the rising segment and the stable segment, and further extracting an experimental data set for training a preset performance parameter predictor; preprocessing the experimental data set, and obtaining a TBM stable segment performance parameter prediction model through training and optimization based on the preprocessed experimental data set; and inputting the collected operating parameter data of the rising segment during the TBM excavation process into the TBM stable segment performance parameter prediction model to obtain predicted operating parameters of the stable segment. The present invention can predict important parameters of the subsequent stable segment based on the collected rising segment data of the TBM excavation cycle, thereby improving the excavation efficiency and safety of the TBM in tunnel construction.

Description

Technical Field

[0001] The present invention relates to the field of drilling technology, and in particular to a method for predicting performance parameters of a TBM stable section based on rising section data. Background Art

[0002] Tunnel boring machines (TBMs) are among the most advanced equipment used in tunnel construction, primarily used in hard rock tunnel construction, particularly deep and long tunnels. Due to their high tunneling efficiency, operational safety, and environmental friendliness, TBMs have replaced traditional drill and blast methods and become the leading method of tunnel construction.

[0003] However, in traditional TBM operations, operators must assess rock conditions through slow trial excavation phases and then repeatedly adjust tunnel parameters to achieve a stable state for subsequent tunneling. This process relies heavily on worker experience, resulting in low construction efficiency and poor matching of tunnel parameters with rock conditions. Furthermore, when stratum conditions undergo drastic changes, effective real-time perception of surrounding rock parameters is lost, resulting in tunnel parameters being unable to adapt to the complex and changing geological environment. This situation can easily lead to engineering accidents and even casualties, resulting in significant social and economic losses.

[0004] Predicting TBM performance in a given rock mass has been a long-standing research topic. Each TBM's driving cycle can be defined as the process from start to stop, which can be divided into a dry run phase, an ascending phase, and a stable phase. The ascending phase, closely related to the stable phase, contains a wealth of information about the TBM's progress and can reflect the interaction between the TBM and the surrounding rock. Therefore, to ensure safe and efficient tunnel construction, developing methods to adaptively adjust TBM driving parameters based on this ascending phase information is a reliable approach. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a TBM stable segment performance parameter prediction method based on rising segment data. The method can predict important parameters of the subsequent stable segment through the collected rising segment data of the TBM excavation cycle, thereby improving the excavation efficiency and safety of TBM in tunnel construction.

[0006] An embodiment of the present invention provides a method for predicting TBM stable phase performance parameters based on rising phase data, comprising the following steps:

[0007] Collecting a TBM tunnel construction data set under different rock grades, and extracting effective tunneling cycle data from the TBM tunnel construction data set; wherein the TBM tunnel construction data set includes construction data every second during the TBM tunnel construction;

[0008] Dividing the effective excavation cycle data into stages, extracting characteristic data sets of the ascending stage and the stable stage respectively, and extracting experimental data sets for training a preset performance parameter predictor from the characteristic data sets of the ascending stage and the stable stage;

[0009] Preprocessing the experimental data set, training a preset performance parameter predictor based on the preprocessed experimental data set, and obtaining a TBM stable segment performance parameter prediction model through optimization; wherein the preprocessed experimental data set includes a training set, a validation set, and a test set;

[0010] In actual construction, the collected operating parameter data of the rising section during the TBM excavation process are input into the TBM stable section performance parameter prediction model to obtain the predicted operating parameters of the stable section.

[0011] Furthermore, extracting effective excavation cycle data from the TBM tunnel construction data set specifically includes:

[0012] Arranging the construction data in the TBM tunnel construction dataset in chronological order in seconds to obtain a plurality of parameter sequences; wherein one of the parameter sequences includes values ​​of a plurality of operation parameters, the operation parameters including total thrust, cutterhead torque, cutterhead power, cutterhead speed, and penetration;

[0013] The effective excavation cycle in the parameter sequence is determined according to the cutterhead rotation speed, and all working data sets within the effective excavation cycle in the several parameter sequences are extracted respectively as the effective excavation cycle data.

[0014] Preferably, determining the effective excavation period in the parameter sequence according to the cutterhead rotation speed specifically includes:

[0015] When there is a continuous parameter sequence corresponding to a continuous period, and in the continuous parameter sequence the cutter head rotation speed is continuously greater than zero and the duration is greater than a preset cycle threshold, the continuous period is determined to be a valid excavation cycle.

[0016] Furthermore, the effective excavation cycle data is divided into stages, and characteristic data sets of the rising stage and the stable stage are extracted respectively, specifically including:

[0017] In each effective excavation cycle, the ascending section starting point, the ascending section ending point, the stable section starting point and the stable section ending point are determined according to the total thrust; wherein the ascending section ending point coincides with the stable section starting point;

[0018] Extracting a working data set between the starting point of the ascending segment and the ending point of the ascending segment as a feature data set of the ascending segment;

[0019] A working data set between the starting point of the stable segment and the ending point of the stable segment is extracted as a feature data set of the stable segment.

[0020] Preferably, determining the ascending segment start point, the ascending segment end point, the stabilization segment start point, and the stabilization segment end point according to the total thrust specifically includes:

[0021] Assume k0, k1, k2, k3, and k4 are the starting point of the cycle, the starting point of the rising segment, the starting point of the stable segment, the end point of the stable segment, and the end point of the cycle respectively, and F is the total thrust. Then the expressions of k1, k2, and k3 are:

[0022]

[0023] Among them, F i is the total thrust of the TBM at time i, ΔF i F i The first-order difference of , and ΔF i =F i+1 -F i ; Δ 2 F i F i The second-order difference of , and Δ 2 F i =ΔF i+1 -ΔF i ; argmax means taking the maximum value, and argmin means taking the minimum value.

[0024] Preferably, the experimental data set for training the preset performance parameter predictor is extracted from the characteristic data sets of the rising segment and the stable segment, specifically including:

[0025] Using the Pearson correlation coefficient, the correlation coefficient between each characteristic parameter in the characteristic data set of the ascending section and the parameter to be predicted is calculated; wherein the parameter to be predicted includes total thrust, cutterhead torque, cutterhead power, and penetration;

[0026] Selecting a parameter sequence within a preset time period after the start of the rising segment as a sampling sequence, and selecting a characteristic parameter whose correlation coefficient is greater than a preset correlation threshold as a sampling parameter, and extracting the sampling parameters in the sampling sequence as a sample set;

[0027] The average values ​​of the sampling parameters in the characteristic data set of the stable segment are calculated respectively as labels of the sample set, which together with the sample set constitute the experimental data set.

[0028] Furthermore, the preprocessing of the experimental data set specifically includes:

[0029] The experimental data set was normalized using the min-max normalization method;

[0030] The experimental data set is classified according to the rock grade, and a fixed number of sample data are randomly extracted from the sample data corresponding to different rock grades to form the preprocessed experimental data set, and the preprocessed experimental data set is divided into a training set, a validation set and a test set.

[0031] Furthermore, the preset performance parameter predictor is trained based on the preprocessed experimental data set, and optimized after training to obtain a TBM stable segment performance parameter prediction model, specifically including:

[0032] The preset performance parameter predictor includes an input layer, an RNN layer, a fully connected layer, and an output layer; wherein the RNN layer is composed of a combination of several RNN-based predictors;

[0033] Inputting the experimental data set into the preset performance parameter predictor, and training the preset performance parameter predictor through a preset training step; wherein the preset training step includes forward propagation, back propagation, and weight update;

[0034] The trained performance parameter predictor is optimized by layer normalization and Bayesian optimization algorithm to obtain the TBM stable segment performance parameter prediction model.

[0035] Preferably, the layer normalization is used to optimize the input data of the performance parameter predictor, specifically including:

[0036] Normalize the input data to obtain a standardized input vector;

[0037] Performing a linear transformation on the normalized input vector to obtain a linear output vector;

[0038] Normalizing the linear output vector to obtain a standardized output vector;

[0039] The standardized output vector is subjected to a nonlinear transformation through an activation function to obtain optimized input data.

[0040] Preferably, the Bayesian optimization algorithm is used to optimize the hyperparameters of the performance parameter predictor, specifically including:

[0041] Based on the validation set, the hyperparameters are automatically optimized by introducing the Optuna hyperparameter optimization framework. Several optimizations are performed with the mean absolute percentage error of the validation set as the optimization target, and the hyperparameters corresponding to the minimum mean absolute percentage error among the several optimizations are taken as the optimal hyperparameters.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] By taking advantage of the close correlation between the rising section and the stable section during TBM excavation, and inputting the collected rising section excavation information into a pre-trained and optimized TBM performance parameter prediction model, the predicted parameters of the TBM stable section can be quickly obtained, providing a strong reference for the driver to operate the TBM control parameters in the operator's cab, no longer relying solely on manual experience, thereby improving the TBM excavation efficiency and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of a method for predicting TBM stable segment performance parameters based on rising segment data provided by one embodiment of the present invention.

[0045] Figure 2 A visualization diagram of six key parameters of a TBM within 90,000 seconds is provided in accordance with an embodiment of the present invention.

[0046] Figure 3 A schematic diagram of a complete TBM excavation cycle provided by one embodiment of the present invention.

[0047] Figure 4 A schematic diagram illustrating the three operation phases of a TBM provided in one embodiment of the present invention.

[0048] Figure 5 A framework diagram of a preset performance parameter predictor provided by one embodiment of the present invention.

[0049] Figure 6 This is an optimization result of the hyperparameters of an LSTM algorithm provided by one embodiment of the present invention.

[0050] Figure 7 A schematic diagram of the test results of an optimized model provided by one embodiment of the present invention on a test set.

[0051] Figure 8 A schematic diagram showing the visualization of the test results of an optimized model on a test set provided in one embodiment of the present invention.

[0052] Figure 9 A schematic diagram showing the comparison results of a TBM stable segment performance parameter prediction model provided by one embodiment of the present invention and six other classic machine learning methods. DETAILED DESCRIPTION

[0053] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0054] It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0055] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0056] Reference Figure 1 , which is a flow chart of a method for predicting TBM stable segment performance parameters based on rising segment data provided by one embodiment of the present invention, comprising the following steps:

[0057] S1: Collecting a TBM tunnel construction dataset under different rock grades and extracting effective tunneling cycle data from the TBM tunnel construction dataset; wherein the TBM tunnel construction dataset includes construction data every second during the TBM tunnel construction;

[0058] S2: Dividing the effective excavation cycle data into stages, extracting feature data sets of the ascending stage and the stable stage respectively, and extracting experimental data sets for training a preset performance parameter predictor from the feature data sets of the ascending stage and the stable stage;

[0059] S3: Preprocessing the experimental data set, training a preset performance parameter predictor based on the preprocessed experimental data set, and obtaining a TBM stable segment performance parameter prediction model through optimization; wherein the preprocessed experimental data set includes a training set, a validation set, and a test set;

[0060] S4: In actual construction, the collected operating parameter data of the ascending section during the TBM excavation process is input into the TBM stable section performance parameter prediction model to obtain the predicted operating parameters of the stable section.

[0061] Regarding step S1, specifically, extracting effective excavation cycle data from the TBM tunnel construction data set includes:

[0062] Arranging the construction data in the TBM tunnel construction dataset in chronological order in seconds to obtain a plurality of parameter sequences; wherein one of the parameter sequences includes values ​​of a plurality of operation parameters, the operation parameters including total thrust, cutterhead torque, cutterhead power, cutterhead speed, and penetration;

[0063] The effective excavation cycle in the parameter sequence is determined according to the cutterhead rotation speed, and all working data sets within the effective excavation cycle in the several parameter sequences are extracted respectively as the effective excavation cycle data.

[0064] Preferably, determining the effective excavation period in the parameter sequence according to the cutterhead rotation speed specifically includes:

[0065] When there is a continuous parameter sequence corresponding to a continuous period, and in the continuous parameter sequence the cutter head rotation speed is continuously greater than zero and the duration is greater than a preset cycle threshold, the continuous period is determined to be a valid excavation cycle.

[0066] In a preferred embodiment, excavation data from the Yinsong Water Supply Tunnel Project was used as a demonstration study. During the excavation process, TBM operational data was collected on a second-by-second basis, yielding 4.08 billion data sets. These data included 199 operational parameters, such as cutterhead torque and speed. The TBM's operating conditions and other actual performance parameters under different conditions were also recorded. The TBM operated for 728 days, with 4 to 10 excavation cycles per day. Daily data was stored in text format, totaling 134GB.

[0067] In order to better understand the TBM excavation process, some key TBM parameters are visualized, such as Figure 2 As shown. It can be found that the five TBM parameters (total thrust F, cutterhead torque T, cutterhead speed n, cutterhead power P, and penetration rate PPR) all show a sawtooth-like pattern. The reason is that the excavation of the TBM is cyclical, and the TBM uses the excavation cycle as a working unit. In addition to the regular excavation period, it also includes other operating periods, such as tool replacement, unexpected downtime, and machine maintenance. In addition to the regular excavation period, it also includes other operating periods, such as tool replacement, unexpected downtime, and machine maintenance. Therefore, the collected data contains a large amount of invalid data. These invalid data are useless for building a predictor and may even reduce the prediction quality. Therefore, it is necessary to extract valid excavation cycle data from the original data and eliminate invalid data.

[0068] Analysis of the raw data reveals that the cutterhead speed exhibits significant periodic variation, consistent with engineering practice. Therefore, we select cutterhead speed n as the basis for extracting the effective tunneling cycle. A day contains 86,400 seconds of data, and each second contains the values ​​of 199 operating parameters, as shown in the following formula:

[0069]

[0070] Where i represents time (seconds) and j represents a certain operation parameter at the i-th second. Assuming that the index among the 199 TBM parameters is 121, the TBM operation parameter 121 at the i-th second is represented as In daily data There exists a number of seconds between K1 and K2 {0≤K1≤K2≤86400}, which satisfies the following conditions:

[0071]

[0072] The excavation cycle is divided according to the above formula. That is, when the cutterhead speed is not zero for 300 seconds within a period of time, this period is considered to be a valid excavation cycle, and the data therein is extracted, as shown in the following formula:

[0073]

[0074] According to the above steps, valid excavation cycle data can be extracted from the original data, invalid data can be eliminated, and valid excavation cycle extraction can be completed.

[0075] Regarding step S2, specifically, dividing the effective excavation cycle data into stages and extracting characteristic data sets of the rising stage and the stable stage respectively includes:

[0076] In each effective excavation cycle, the ascending section starting point, the ascending section ending point, the stable section starting point and the stable section ending point are determined according to the total thrust; wherein the ascending section ending point coincides with the stable section starting point;

[0077] Extracting a working data set between the starting point of the ascending segment and the ending point of the ascending segment as a feature data set of the ascending segment;

[0078] A working data set between the starting point of the stable segment and the ending point of the stable segment is extracted as a feature data set of the stable segment.

[0079] Preferably, determining the ascending segment start point, the ascending segment end point, the stabilization segment start point, and the stabilization segment end point according to the total thrust specifically includes:

[0080] Assume k0, k1, k2, k3, and k4 are the starting point of the cycle, the starting point of the rising segment, the starting point of the stable segment, the end point of the stable segment, and the end point of the cycle respectively, and F is the total thrust. Then the expressions of k1, k2, and k3 are:

[0081]

[0082] Among them, F i is the total thrust of the TBM at time i, ΔF i F i The first-order difference of , and ΔF i =F i+1 -F i ; Δ 2 F i F i The second-order difference of , and Δ 2 F i =ΔF i+1 -ΔF i ; argmax means taking the maximum value, and argmin means taking the minimum value.

[0083] In a preferred embodiment, the normal excavation cycle of a TBM can be defined as the process from the start to the stop of the equipment. The data change process is regular, where the data distribution of n, T, F and PPR in a normal excavation cycle is as follows: Figure 3 shown.

[0084] TBM uses the excavation cycle as a working unit, and each cycle consists of three stages: the idle push stage, the rising stage, and the stable stage. The idle push stage is the period from the start of the TBM to the contact between the cutter head and the rock. During this period, the cutter head speed n reaches the set value; the rising stage is when the cutter head contacts the rock, and the TBM excavation parameters PPR, F, and T rise rapidly to a stable value; the stable stage is the period when the tunnel parameters remain stable and have little fluctuation. At this time, the TBM excavates in a safe, fast, and high-quality stable state. By visually analyzing the original data, such as Figure 3 shown.

[0085] It can be found that the transition between different stages of the tunneling cycle is closely related to the change of the total thrust F, so the thrust is selected as the stage division standard of each cycle. Let k0, k1, k2, k3, and k4 be the starting point of the cycle, the starting point of the rising section, the starting point of the stable section, the end point of the stable section, and the end point of the cycle respectively, and let F i is the total thrust of the TBM at time i, ΔF i F i The first-order difference of 2 F i F i The second-order difference of ΔF i and Δ 2 F i Satisfy the following formula:

[0086]

[0087] The three working stages are divided according to the first-order and second-order differences of the total thrust parameters of the cutter head. Figure 4 , it can be found that the sign of the beginning of the rising section is that the slope of the total thrust F increases sharply and the second-order differential reaches a peak; the position of the end point of the rising section and the beginning point of the stable section are both at the mean value of △F of the rising section and the stable section; the sign of the end of the stable section k3 is that the total thrust of the cutter head suddenly drops to near 0 and basically does not change. Therefore, the expressions of k1, k2, and k3 are:

[0088]

[0089] Where argmax and argmin represent the maximum and minimum values.

[0090] According to the above formula, k1, k2 and k3 of each normal excavation cycle are calculated. The data between k0 and k1 are defined as the empty push section data; the data between k1 and k2 are defined as the rising section data; and the data between k2 and k3 are defined as the stable section data.

[0091] Regarding step S2, specifically, extracting an experimental data set for training a preset performance parameter predictor from the feature data sets of the rising segment and the stable segment includes:

[0092] Using the Pearson correlation coefficient, the correlation coefficient between each characteristic parameter in the characteristic data set of the ascending section and the parameter to be predicted is calculated; wherein the parameter to be predicted includes total thrust, cutterhead torque, cutterhead power, and penetration;

[0093] Selecting a parameter sequence within a preset time period after the start of the rising segment as a sampling sequence, and selecting a characteristic parameter whose correlation coefficient is greater than a preset correlation threshold as a sampling parameter, and extracting the sampling parameters in the sampling sequence as a sample set;

[0094] The average values ​​of the sampling parameters in the characteristic data set of the stable segment are calculated respectively as labels of the sample set, which together with the sample set constitute the experimental data set.

[0095] In a preferred embodiment, after obtaining the characteristic data sets of the rising segment and the stable segment, further processing is required to obtain the experimental data set ultimately used for training the model.

[0096] First, using the Pearson correlation coefficient, we calculated the correlation coefficient between each feature parameter in the feature dataset of the rising segment and the parameter to be predicted. Based on the ranking of the parameter correlations, we balanced the predictor's time consumption and accuracy, and selected the top 25 feature parameters with the highest correlations as sampling parameters. The sample data in the sample set consisted of the top 25 feature data with the highest correlations in the first 30 seconds of the rising segment. The mean of each sampled parameter in the stable segment was taken as the label for the sample set. Finally, based on the sample set and labels, we constructed the dataset used to train the predictor in this paper.

[0097] Furthermore, the preprocessing of the experimental data set specifically includes:

[0098] The experimental data set was normalized using the min-max normalization method;

[0099] The experimental data set is classified according to the rock grade, and a fixed number of sample data are randomly extracted from the sample data corresponding to different rock grades to form the preprocessed experimental data set, and the preprocessed experimental data set is divided into a training set, a validation set and a test set.

[0100] In a preferred embodiment, since different parameters in the experimental dataset have different dimensions and dimensional units, this can affect the results of data analysis and increase the difficulty of model training. Therefore, to eliminate the dimensional influence between indicators, data normalization is required. This can improve the predictor's convergence speed and prediction accuracy. Furthermore, when the predicted value is finally output, it must be denormalized back to the original data size before output.

[0101] The min-max standardization method is used to normalize the data. The specific processing formula is as follows:

[0102]

[0103] Where x represents a TBM excavation state parameter in the original data, x normalized represents the normalized TBM excavation state parameter, x min and z max Respectively represent the minimum and maximum values ​​of the parameter in the data set.

[0104] The dataset then needed to be structured to determine the predictor input and output. The input was determined to be the top 25 correlated feature data from the first 30 seconds of the TBM's ascent phase, and the output was determined to be four key performance parameters during the stable phase (including total thrust F, cutterhead torque T, cutterhead power P, and penetration rate PPR).

[0105] After preprocessing the experimental dataset in this way, a dataset containing 4520 valid samples was obtained, each representing a single excavation cycle. Furthermore, since the surrounding rock mass of the Yinsong Water Supply Tunnel Project is composed of Class II, III, IV, and V rock, to improve the overall adaptability of the developed predictor, 250 samples of each rock class were randomly selected, for a total of 1000 samples constituting the experimental dataset. Of these, 800 samples were randomly selected as the training set, 100 as the validation set, and 100 as the test set.

[0106] Regarding step S3, specifically, the preset performance parameter predictor is trained based on the preprocessed experimental data set, and optimized after training to obtain the TBM stable segment performance parameter prediction model, specifically including:

[0107] The preset performance parameter predictor includes an input layer, an RNN layer, a fully connected layer, and an output layer; wherein the RNN layer is composed of a combination of several RNN-based predictors;

[0108] Inputting the experimental data set into the preset performance parameter predictor, and training the preset performance parameter predictor through a preset training step; wherein the preset training step includes forward propagation, back propagation, and weight update;

[0109] The trained performance parameter predictor is optimized by layer normalization and Bayesian optimization algorithm to obtain the TBM stable segment performance parameter prediction model.

[0110] Preferably, the layer normalization is used to optimize the input data of the performance parameter predictor, specifically including:

[0111] Normalize the input data to obtain a standardized input vector;

[0112] Performing a linear transformation on the normalized input vector to obtain a linear output vector;

[0113] Normalizing the linear output vector to obtain a standardized output vector;

[0114] The standardized output vector is subjected to a nonlinear transformation through an activation function to obtain optimized input data.

[0115] Preferably, the Bayesian optimization algorithm is used to optimize the hyperparameters of the performance parameter predictor, specifically including:

[0116] Based on the validation set, the hyperparameters are automatically optimized by introducing the Optuna hyperparameter optimization framework. Several optimizations are performed with the mean absolute percentage error of the validation set as the optimization target, and the hyperparameters corresponding to the minimum mean absolute percentage error among the several optimizations are taken as the optimal hyperparameters.

[0117] In a preferred embodiment, referring to Figure 5 , a framework diagram of a preset performance parameter predictor provided in one embodiment of the present invention, comprising an input layer, an RNN layer, a fully connected layer, and an output layer. The RNN layer includes three RNN-based predictors: a traditional RNN-based predictor, an LSTM-based predictor, and a GRU-based predictor.

[0118] After the predictor framework is built, it is trained using the prepared experimental data training set. The training process includes multiple steps, including forward propagation, backpropagation, and weight updates. Through these steps, the model continuously adjusts its weights and gradually learns the characteristics of the input samples. During each training cycle, the model is optimized by monitoring its performance. This combination of steps enables the model to automatically learn the characteristics of the data from the training set and perform better in subsequent task processing. Through the training process, the model gradually improves its prediction capabilities for the four important performance parameters of the TBM stable phase, enabling the predictor to more accurately predict future performance parameter changes.

[0119] After training is complete, the predictor needs to be further optimized to obtain the final TBM stable performance parameter prediction model. Optimization methods include Bayesian hyperparameter optimization and layer normalization of RNN units.

[0120] As we all know, hyperparameters in deep learning have a significant impact on model training, and hyperparameters are typically manually tuned. Manual tuning often relies on personal experience, making it difficult to find the optimal hyperparameters. Finding the optimal hyperparameters can be time-consuming and significantly increases the time cost of developing predictors.

[0121] To achieve higher predictor accuracy and shorten the time spent adjusting hyperparameters, we incorporate a hyperparameter optimization algorithm into our predictor. The algorithm automatically searches for optimal hyperparameters, eliminating the need for manual tuning. This not only improves predictor development efficiency but also increases predictor accuracy.

[0122] Commonly used hyperparameter optimization methods include grid search, random search, and Bayesian optimization. Grid search employs an exhaustive search approach, looping through all candidate parameter choices and trying every possible option. The best-performing parameter is the final result, similar to finding the maximum value in an array. However, because it requires trying all possible hyperparameter combinations, it can be computationally expensive, time-consuming, and inefficient when dealing with large amounts of data or complex models. Unlike grid search, random search doesn't try every parameter value, but instead samples a fixed number of parameter settings from a specified distribution. While generally faster than grid search, random search can miss important information and its results are not guaranteed. Bayesian optimization, based on the Bayesian principle, infers the model's performance under a given set of hyperparameters based on existing observations (model performance under one set of hyperparameters). In short, it estimates the entire hyperparameter space based on a small number of existing hyperparameter combinations and selects the most suitable one. This process often employs Gaussian regression for estimation.

[0123] Compared to grid search and random search, Bayesian super-optimization is more effective and efficient. The differences between Bayesian optimization and conventional grid search or random search are: 1) Bayesian parameter tuning uses a Gaussian process, taking into account previous parameter information and continuously updating priors; grid search does not consider previous parameter information; 2) Bayesian parameter tuning requires fewer iterations and is faster; grid search is slower and can easily lead to dimensionality explosion when there are many parameters; 3) Bayesian parameter tuning is robust for non-convex problems; grid search can easily reach local optima for non-convex problems; and 4) Bayesian optimization provides an elegant framework for finding the global minimum in the fewest possible steps.

[0124] Therefore, the present invention incorporates a Bayesian optimization algorithm (BOA) to search for optimal hyperparameters for the predictor. First, the mean absolute percentage error (MAPE) of the prediction validation set is taken as the optimization target. Given initial hyperparameters, the minimum value across n trials is taken as the optimal parameter. Then, the Optuna hyperparameter optimization framework samples the hyperparameters and efficiently prunes unsuccessful trials, achieving efficient hyperparameter optimization.

[0125] Reference Figure 6 , is the optimization result of the hyperparameters of an LSTM algorithm provided by an embodiment of the present invention. Figure 6 It can be seen that the number of optimization trials was set to 50, and the experiment was set to terminate when MAPE < 0.001 and return the final hyperparameters. The best hyperparameter combination was obtained and recorded for later use in training the final model.

[0126] In addition to optimizing hyperparameters using the Bayesian optimization algorithm, the present invention also utilizes layer normalization of the RNN unit to optimize the model. Layer normalization is a technique for normalizing the input of activation functions in neural networks. It can effectively reduce the problems of vanishing and exploding gradients, and improve the convergence speed and training stability of the model. In RNN models, the propagation of gradients over time steps can lead to vanishing or exploding gradients. Layer normalization can help alleviate these problems and improve the model's predictive performance.

[0127] The following is a detailed introduction to the specific implementation of layer normalization in RNN units:

[0128] 1) Input Normalization: First, for each time step of the input vector x, the input normalization technique is applied. Input normalization calculates the mean and standard deviation of each feature and normalizes the input vector. Specifically, for each feature x i , calculate its mean μ(x i ) and standard deviation δ(x i ), and then normalize: This ensures that each feature has a similar scale and reduces the differences between different features.

[0129] 2) Linear Transformation: A linear transformation is performed on the input vector after input normalization. This is achieved by multiplying it by a weight matrix W and adding a bias vector b. Specifically, for an input vector x, the linear transformation can be expressed as y = Wx + b, where W is a weight matrix of dimension (m, n), m is the dimension of the output vector, n is the dimension of the input vector, and b is the bias vector.

[0130] 3) Normalization: For the result y of the linear transformation, the normalization operation is applied. The normalization operation calculates the mean and standard deviation of each feature and normalizes the output vector. Specifically, for each feature y i , calculate its mean μ(y i ) and standard deviation δ(y i ), and then normalize:

[0131] 4) Activation Function: Apply an activation function to the normalized output vector to perform a nonlinear transformation. Common activation functions include sigmoid, tanh, and ReLU. These activation functions can introduce nonlinearity and are widely used in RNN models.

[0132] Through the above steps, the layer normalization of the RNN unit can normalize, linearly transform, normalize and activate the input at each time step, thereby improving the performance and stability of the model. Layer normalization can help alleviate the gradient vanishing and gradient exploding problems and promote the convergence of the model. Therefore, the layer normalization technology in the embodiment of the present invention has a significant optimization effect on the predictor of the four important performance parameters of the stable segment of the RNN-based TBM.

[0133] Specifically, in actual construction, step S4 simply requires inputting the operating parameter data for the first 30 seconds of the TBM's ascending phase into the TBM's stable phase performance parameter prediction model to obtain the predicted operating parameters for the stable phase. These parameters include total thrust F, cutterhead torque T, cutterhead power P, and penetration rate PPR.

[0134] Therefore, the TBM stable section performance parameter prediction model provides a powerful reference for the driver to operate TBM control parameters in the operator's cab, no longer relying solely on manual experience, thereby improving the TBM's excavation efficiency and safety.

[0135] Finally, the present invention also conducts a comparative evaluation of TBM stable segment performance parameter prediction models, specifically including:

[0136] (1) In order to evaluate the effectiveness of the developed predictor, it is necessary to compare the predicted results with the actual values. The present invention uses two evaluation indicators, mean absolute percentage error (MAPE) and accuracy (Acc), to comprehensively evaluate the performance of the predictor. Acc is used to evaluate the accuracy of the predictor, and MAPE is used to evaluate the error degree of the predictor. The specific calculation formula is as follows:

[0137]

[0138] Among them, p represents one of the four important performance parameters of the TBM stable stage, K represents the total number of test samples, represents the predicted value, and y represents the true value.

[0139] (2) The Bayesian optimization algorithm is used to obtain the best hyperparameter combination to train the final RNN-based predictor. Finally, the trained model is tested on the test set to predict the four important performance parameters of the TBM stable segment of 100 samples in the test set. The prediction results are as follows: Figure 7 and 8 shown.

[0140] As can be seen, the MAPE values ​​of the three RNN-based predictors are 0.117, 0.112, and 0.113, respectively, with an average of 0.114. The prediction accuracy Acc of the three RNN-based predictors is 88.3%, 88.8%, and 88.7%, respectively, with an average prediction accuracy of 88.6%. This performance is comparable to that of the LSTM-based and GRU-based predictors, and both slightly outperform traditional RNN-based predictors. LSTM and GRU networks perform better than traditional RNNs when handling more complex problems or when long-term memory is required.

[0141] In this scenario, the predictor needs to memorize the first 30 seconds of data during the ramp-up phase, which is not a short memory duration. Therefore, the LSTM- and GRU-based predictors perform slightly better than the traditional RNN-based predictor. Furthermore, the average accuracy of the predicted F, T, P, and PPR is 88.9%, 87.8%, 88.5%, and 89.1%, respectively.

[0142] The results show that it is feasible to predict the four key parameters of the stable phase using information from the first 30 seconds of the rising phase, demonstrating the effectiveness of this method. Furthermore, the average time for these three predictors over 800 iterations was 9.864, 11.99, and 11.05 milliseconds, respectively. Due to the relatively simple structure of traditional RNN-based predictors, their computational cost (i.e., time consumption) is relatively low compared to LSTM- and GRU-based predictors. Since the actual prediction task is a single-sample prediction, the proposed model achieves a shorter time-to-performance ratio when applied to real-world projects, fully meeting real-time requirements.

[0143] (3) To highlight the superiority of the present invention, the predictor of the present invention is compared with other classic machine learning models, including two regression models, ElasticNet and Ridge regression, two neural network models based on Levenberg-Marquardt (LM) and conjugate gradient method (SCG), and two ensemble learning models based on light gradient boosting machine (LightGBM) and adaptive boosting (Adaboost). These classic machine learning algorithms each have their own strengths in regression prediction problems.

[0144] Use these six machine learning models to build predictors and make predictions using the test set. Because traditional regression and machine learning methods do not have memory capabilities, it is necessary to transform the input samples. Convert the input and output feature vectors of the samples in the dataset into row vectors, as shown below:

[0145]

[0146] Where x represents the input sample, y represents its corresponding expected output, j represents the input TBM operation parameter characteristics and j∈[1,25], and 30 represents the first 30 seconds.

[0147] The prediction results of six classic machine learning models are compared with three RNN-based predictors, such as Figure 9 The results show that RNN-based predictors significantly outperform machine learning models in predicting F, T, P, and PPR. In particular, for predicting F, the MAPEs of all six machine learning models are above 1, while the MAPEs of the RNN-based predictors are around 0.11. Although the MAPEs of the classic machine learning models for predicting T and P have declined, they are still significantly higher than those of the RNN-based predictors. For predicting PPR, while the Ridge and DNN-LM models achieve similarly good results to the RNN-based predictors, ElasticNet, DNN-SCG, Adaboost, and LightGBM still achieve significantly higher MAPE values.

[0148] Overall, the prediction error of the prediction model based on the RNN method is much lower than that of other classic machine learning methods, which further highlights the superiority of the present invention and successfully verifies the applicability and effectiveness of the present invention.

[0149] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting TBM stable phase performance parameters based on rising phase data, characterized in that: The following steps are involved: A TBM tunnel construction dataset under different rock grades is collected, and effective excavation cycle data is extracted from the TBM tunnel construction dataset; wherein the TBM tunnel construction dataset includes construction data every second during TBM tunnel construction, and extracting the effective excavation cycle data from the TBM tunnel construction dataset specifically includes: Arranging the construction data in the TBM tunnel construction dataset in chronological order in seconds to obtain a plurality of parameter sequences; wherein one of the parameter sequences includes values ​​of a plurality of operation parameters, the operation parameters including total thrust, cutterhead torque, cutterhead power, cutterhead speed, and penetration; Determining an effective excavation cycle in the parameter sequence according to the cutterhead rotation speed, and extracting all working data sets within the effective excavation cycle in the plurality of parameter sequences as the effective excavation cycle data; The effective excavation cycle data is divided into stages, and characteristic data sets of the ascending segment and the stable segment are extracted respectively, and an experimental data set for training a preset performance parameter predictor is extracted from the characteristic data sets of the ascending segment and the stable segment; wherein, the effective excavation cycle data is divided into stages, and the characteristic data sets of the ascending segment and the stable segment are extracted respectively, specifically including: In each effective excavation cycle, the starting point of the ascending section, the ending point of the ascending section, the starting point of the stable section, and the ending point of the stable section are determined according to the total thrust; wherein the ending point of the ascending section coincides with the starting point of the stable section; the determining of the starting point of the ascending section, the ending point of the ascending section, the starting point of the stable section, and the ending point of the stable section according to the total thrust specifically includes: set up 、 、 、 and are the starting point of the cycle, the starting point of the rising segment, the starting point of the stable segment, the end point of the stable segment, and the end point of the cycle respectively. F is the total thrust, then 、 、 The expression is: in, is the total thrust of the TBM at time i, for The first-order difference of = ; for The second-order difference of = ; argmax means taking the maximum value, argmin means taking the minimum value; Extracting a working data set between the starting point of the ascending segment and the ending point of the ascending segment as a feature data set of the ascending segment; Extracting a working data set between the starting point of the stable segment and the ending point of the stable segment as a feature data set of the stable segment; Preprocessing the experimental data set, training a preset performance parameter predictor based on the preprocessed experimental data set, and obtaining a TBM stable segment performance parameter prediction model through optimization; wherein the preprocessed experimental data set includes a training set, a validation set, and a test set; In actual construction, the collected operating parameter data of the rising section during the TBM excavation process are input into the TBM stable section performance parameter prediction model to obtain the predicted operating parameters of the stable section.

2. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 1 is characterized in that: Determining the effective excavation period in the parameter sequence according to the cutterhead speed specifically includes: When there is a continuous parameter sequence corresponding to a continuous period, and in the continuous parameter sequence the cutter head rotation speed is continuously greater than zero and the duration is greater than a preset cycle threshold, the continuous period is determined to be a valid excavation cycle.

3. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 1 is characterized in that: An experimental data set for training a preset performance parameter predictor is extracted from the characteristic data sets of the rising segment and the stable segment, specifically including: Using the Pearson correlation coefficient, the correlation coefficient between each characteristic parameter in the characteristic data set of the ascending section and the parameter to be predicted is calculated; wherein the parameter to be predicted includes total thrust, cutterhead torque, cutterhead power, and penetration; Selecting a parameter sequence within a preset time period after the start of the rising segment as a sampling sequence, and selecting a characteristic parameter whose correlation coefficient is greater than a preset correlation threshold as a sampling parameter, and extracting the sampling parameters in the sampling sequence as a sample set; The average values ​​of the sampling parameters in the characteristic data set of the stable segment are calculated respectively as labels of the sample set, which together with the sample set constitute the experimental data set.

4. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 1 is characterized in that: The preprocessing of the experimental data set specifically includes: The experimental data set was normalized using the min-max normalization method; The experimental data set is classified according to the rock grade, and a fixed number of sample data are randomly extracted from the sample data corresponding to different rock grades to form the preprocessed experimental data set, and the preprocessed experimental data set is divided into a training set, a validation set and a test set.

5. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 1 is characterized in that: The preset performance parameter predictor is trained based on the preprocessed experimental data set, and optimized after training to obtain a TBM stable segment performance parameter prediction model, specifically including: The preset performance parameter predictor includes an input layer, an RNN layer, a fully connected layer, and an output layer; wherein the RNN layer is composed of a plurality of RNN-based predictors; Inputting the experimental data set into the preset performance parameter predictor, and training the preset performance parameter predictor through a preset training step; wherein the preset training step includes forward propagation, back propagation, and weight update; The trained performance parameter predictor is optimized by layer normalization and Bayesian optimization algorithm to obtain the TBM stable segment performance parameter prediction model.

6. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 5 is characterized in that: The layer normalization is used to optimize the input data of the performance parameter predictor, specifically including: Normalize the input data to obtain a standardized input vector; Performing a linear transformation on the normalized input vector to obtain a linear output vector; Normalizing the linear output vector to obtain a standardized output vector; The standardized output vector is subjected to a nonlinear transformation through an activation function to obtain optimized input data.

7. The method for predicting TBM stable phase performance parameters based on rising phase data according to claim 5 is characterized in that: The Bayesian optimization algorithm is used to optimize the hyperparameters of the performance parameter predictor, specifically including: Based on the validation set, the hyperparameters are automatically optimized by introducing the Optuna hyperparameter optimization framework. Several optimizations are performed with the mean absolute percentage error of the validation set as the optimization target, and the hyperparameters corresponding to the minimum mean absolute percentage error among the several optimizations are taken as the optimal hyperparameters.

Images