Shield tunneling machine tunneling prediction method and system based on data and physical fusion driving and storage medium

Abstract

The invention provides a shield tunneling machine tunneling prediction method based on data and physical fusion driving. The shield tunneling machine tunneling prediction method comprises the following steps: firstly, carrying out real-time learning prediction on historical geology-control-tunneling big data by using a Transform network; performing off-line simulation on typical working conditions, and establishing a look-up table; and then according to the geological similarity between the current ring and the historical ring, adaptively fusing the two prediction results. The method depends on learning when data is sufficient and depends on simulation when samples are scarce, 92.9% of prediction precision and millisecond response are achieved, the problems that pure data model samples are unbalanced and poor in generalization are solved, the defects that pure physical simulation is large in calculation amount and difficult in real time are overcome, and the method is suitable for large-scale popularization and application. The prediction accuracy and the engineering applicability of the shield tunneling speed and the cutterhead torque under the complex geological condition are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel engineering, in particular to a shield tunneling prediction method and system based on data and physical fusion driving and a storage medium. BACKGROUND

[0002] Shield method has become the mainstream construction method for urban subway, railway tunnel and cross-river and sea channel due to its safety, high efficiency and small disturbance to the ground. In the process of tunneling, the key operating parameters such as thrust, cutterhead speed and penetration degree need to be adjusted in real time to adapt to the complex and changeable geological conditions, so as to ensure the tunneling efficiency and reduce the risk of abnormal wear of equipment. Tunneling speed and cutterhead torque are the most direct and important two indicators to evaluate the tunneling state of the shield machine, and their accurate prediction has great engineering significance for construction parameter optimization, risk warning and cost control.

[0003] In the prior art, the prediction of the tunneling performance of the shield machine mainly develops along two technical routes: (1) Data-driven route The patent with publication number CN202310397499A proposes a "shield machine posture and tunneling parameter multi-task prediction method", which takes the measured parameters such as tunneling speed and cutterhead torque as input, constructs a long short-term memory (LSTM) model, and realizes the synchronous prediction of the shield machine posture angle, total thrust and roll angle.

[0004] The patent with publication number CN202110939424A obtains the physical characteristics such as rock slag particle shape, mineral composition and water content in real time through optical and pressure sensors, and uses a gradient boosting regression tree (GBRT) model to optimize the TBM tunneling parameters. The patent with publication number CN202211237102A collects cutter wear samples, establishes a support vector regression (SVR) model, and realizes the prediction of cutter wear. The patent with publication number CN202110577145A also uses a machine learning method, trains the model through historical monitoring data of the shield machine, predicts the tunneling speed, and optimizes the parameters combined with a genetic algorithm.

[0005] The above data-driven methods perform well when the samples are sufficient and the working conditions are evenly distributed, but have the following shortcomings: (1) In actual engineering, the shield machine passes through complex and changeable strata, and the number of samples of different geological types differs significantly, resulting in a significant increase in prediction error of the model under rare or sudden geological conditions; (2) Deep neural networks have very high requirements for the quality and coverage of training data, while the collected data on site is affected by sensor drift, sudden changes in working conditions and construction downtime, and has the problems of large noise and many missing values, which further aggravates the defect of insufficient model generalization; (3) Machine learning model is a "black box" model, which lacks explicit description of the shield-geology interaction mechanism and is difficult to explain the physical rationality of the prediction results.

[0006] (B) Physical driving route

[0007] Shield tunneling simulation based on discrete element method (DEM), finite element method (FEM) or computational fluid dynamics (CFD) can explicitly simulate the dynamic contact between cutter head and soil, the flow of muck and the formation mechanism of cutter head torque. For example, commercial software such as EDEM and PFC3D have been used to study the influence of cutter head opening rate, cutter arrangement and stratum parameters on tunneling performance. However, this method has the following limitations: (1) Large amount of calculation: In order to ensure accuracy, the number of grids or particles usually reaches millions, and the simulation of a single working condition takes several hours to several days, which is difficult to meet the real-time decision-making needs of the site; (2) Difficulty in parameter calibration: The microscopic parameters of soil (particle friction coefficient, cohesive strength, etc.) need to be inverted through a large number of experiments, and the high uncertainty of parameters caused by limited sampling on site; (3) Limited working condition coverage: Due to the limitation of computing resources, only a small number of typical geological-operation combinations can be simulated, and it is impossible to exhaust the vast number of working conditions in actual construction.

[0008] In summary, the single data-driven method has poor robustness in the case of insufficient or unevenly distributed samples; the single physical driving method is limited by computational efficiency and parameter uncertainty, making it difficult to achieve real-time prediction. Therefore, a new prediction technology that can combine the advantages of efficient data learning and physically interpretable simulation is urgently needed to solve the "unbalanced samples-simulation lag" dual bottleneck problem faced by shield machines in actual construction. SUMMARY

[0009] The present application provides a shield tunneling prediction method based on data and physical fusion driving, which solves the problems of low data standardization and governance efficiency and the inability to comprehensively cover the whole process from data governance to intelligent analysis in the prior art. The technical scheme of the present application is as follows: A shield tunneling prediction method based on data and physical fusion driving, comprising the following steps: Step S1: Real-time prediction of shield tunneling target parameters by a data-driven model; Step S2: Offline prediction of simulation prediction of the tunneling target parameters by a physical driving model and construction of a simulation database; Step S3: Calculate the similarity R of the current geological working condition and the historical working condition; Step S4: Dynamically fuse the data-driven prediction results and the physical driving prediction results to generate a fusion prediction value ; The fusion formula satisfies:

[0010] in, represents the learned prediction of the prediction loop, represents the simulation prediction of the prediction loop, represents the optimal estimate of the prediction ring.

[0011] Furthermore, the data-driven model is a deep neural network based on the Transformer architecture, which includes an input embedding layer, a Transformer encoding layer, and an output layer. The input features are the combined vector of shield machine operating parameters and geological parameters, and the output is the tunneling speed and cutterhead torque. Through training and evaluation of the shield machine tunneling data set, a multivariate regression model of the shield machine is formed to realize data-driven real-time prediction of shield machine tunneling.

[0012] Furthermore, when training the deep neural network, we set smooth L1 as the loss function, Adam as the optimizer, and set the learning rate to 10e -3 , and use ReduceLROnPlateau to schedule the learning rate; and introduce a regularization process with a Dropout rate of 30%.

[0013] Furthermore, the shield machine tunneling data set was processed through resampling, correlation analysis and empirical judgment steps to form 11-dimensional sample data including geology, thrust, cutterhead speed, penetration, tunneling speed and cutterhead torque.

[0014] Furthermore, the number of sample data is at least 1600 groups; and all features are normalized and scaled. The scaling method is as follows:

[0015] in, is the characteristic variable The maximum value of is the characteristic variable The minimum value of is the normalized value after scaling, is the number of feature samples.

[0016] Furthermore, in step S2, simulation prediction refers to the prediction of the tunneling speed and cutterhead torque of the shield machine based on physical drive. The discrete element physical simulation method is used to perform offline tunneling simulation on typical geological conditions, gradually forming a series of prediction results, and storing them in a list form to form a simulation prediction query table.

[0017] Furthermore, in step S3, the geological condition similarity is calculated by uniformly scaling the characteristic parameters St of the predicted ring and comparing them with the historical adjacent N ring samples St-1 The matching is performed, the Euclidean distance of the feature vectors between the current ring and the historical ring is calculated, and the similarity of the geological conditions of the two is defined as R: .

[0018] Further, the typical geological conditions are single stratum types, including sandy clay layer, fully weathered mixed granite layer, soil-like strongly weathered mixed granite layer, block-like strongly weathered mixed granite layer, moderately weathered mixed granite layer, and slightly weathered mixed granite layer.

[0019] A shield tunneling prediction system based on data and physical fusion driving, comprising: A learning prediction module, configured to provide a current ring data-driven shield tunneling speed and cutterhead torque prediction result in real time according to the geological conditions and shield machine operating parameters of the current ring; A simulation prediction module, configured to query the most similar geological conditions in a simulation prediction list offline according to the geological conditions and shield machine operating parameters of the current ring, and provide a current ring physical-driven shield tunneling speed and cutterhead torque prediction result; A fusion prediction module, configured to calculate the similarity of the geological conditions between the current ring and the historical ring, and fuse the two prediction results according to the similarity.

[0020] A computer storage medium, in which a computer program executable by a processor, a shield tunneling database, and a simulation prediction query list are stored, the computer program executes a shield tunneling prediction method based on data and physical fusion driving.

[0021] Compared with the prior art, the present scheme has the following beneficial effects: (1) The prediction accuracy is significantly improved. By adaptively fusing a data-driven model (Transformer) and a physical simulation-driven model, two complementary prediction sources, the data learning is mainly used when the geological samples are sufficient, and the physical simulation is mainly used when the samples are scarce. The overall prediction accuracy can reach 92.9%, and the mean square error is reduced to 0.011; (2) Solving the sample imbalance problem. When the historical data lacks similar geological conditions, the system can automatically increase the weight of the physical simulation result, avoiding the "overfitting" or "extrapolation failure" of the traditional pure data model due to insufficient samples; (3) Ensuring real-time decision-making capability. The data-driven module requires only millisecond-level time for one forward inference; the physical driving part adopts an "offline simulation + online query table" mode, which does not need to repeat the calculation on site, thus retaining the explainability of the physical mechanism and meeting the real-time requirements on site; (4) Reduce the cost of calculation and test. Compared with full-condition online simulation, the scheme only performs finite off-line simulation on typical geology-operation combination, significantly reduces the calculation resource and the workload of geotechnical parameter calibration, and reduces the overall implementation cost; (5) Improve the robustness and scalability of the system. The Transformer network introduces regularization strategies such as Dropout and learning rate scheduling to improve the noise resistance of the model; at the same time, the system adopts modular design, and can be conveniently updated incrementally with new data or new geological types without overall retraining; (6) The engineering application value is outstanding. Through accurate prediction of the driving speed and cutter torque, the key operation parameters such as thrust and speed can be optimized in advance, reducing the risk of abnormal wear of the cutter, machine jamming and ground subsidence, and directly bringing economic benefits such as shortening of construction period, reduction of energy consumption and reduction of maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 is a flowchart of the shield tunneling prediction method based on data and physical fusion driving of the embodiment of the present application; Figure 2 is a data-driven Transformer deep neural network structure diagram of the embodiment of the present application; Figure 3 is a data-driven Transformer learning prediction result of the embodiment of the present application; Figure 4 is a finite element simulation prediction result of the physical driving embodiment of the present application; Figure 5 is a finite element simulation prediction database of the physical driving embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solutions of the present application will be described in detail below with reference to the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] The present application provides a shield tunneling prediction method based on data and physical fusion driving, referring to Figure 1 , comprising the following steps: S1, correlation analysis of shield tunneling historical data (geology, control, tunneling parameters), extraction of key feature parameters, creation of shield tunneling dataset, and establishment of deep neural network, through training and testing, to realize real-time prediction of shield tunneling speed and cutter torque based on data-driven; S2, a finite element method is used to establish a shield tunneling simulation model to simulate the tunneling process of the shield machine under typical geological conditions, to realize offline prediction of shield tunneling speed and cutter torque based on physical driving, and to create a database to store simulation prediction results in list form for subsequent queries; S3, based on the above two prediction methods, according to the similarity of the geological conditions of the prediction ring and the historical ring, the prediction confidence of the two prediction methods is calculated; S4, the prediction confidence is converted into the corresponding fusion probability, and the prediction results of the two different methods are probability fused to realize the optimal estimation of shield tunneling prediction under different geological conditions.

[0026] Further, in step S1, the key feature parameters are obtained by XGBoost correlation analysis, mainly referring to the geological condition parameters that have a significant impact on the tunneling speed and cutter torque, including total thrust, cutter speed, penetration, and geological conditions. Among them, the geological conditions involve 6 types of strata, namely sandy clay layer, fully weathered mixed granite layer, soil-like strong weathered mixed granite layer, block-like strong weathered mixed granite layer, medium weathered mixed granite layer, and slightly weathered mixed granite layer. The 6 types of geological conditions are composed of each tunneling ring in a certain proportion, for example, assuming that a certain ring is composed of 30% sandy clay layer and 90% fully weathered mixed granite layer, then the geological conditions of the ring are defined as the vector [0.3, 0.7, 0, 0, 0, 0].

[0027] Further, in step S1, the created shield tunneling dataset is composed of historical shield tunneling data, and each sample is an 11-dimensional vector, including total thrust, cutter speed, penetration, geological conditions, tunneling speed and cutter torque. The number of samples at least 1600, for subsequent deep neural network model training and evaluation. Considering the different dimensions of the features, feature scaling is required, specifically, a certain feature variable The scaling can be calculated by the normalization formula:

[0028] wherein, is the maximum value of the feature variable is the minimum value of the feature variable is the normalized value after scaling, ​​is the number of feature samples.

[0029] Among them, in step S1, the deep neural network constructed is designed as a Transformer regression prediction model, such as Figure 2 As shown in the figure, it includes an input embedding layer, a Transformer encoding layer, and an output layer. The input embedding layer is a linear transformation responsible for adjusting the dimension of the input features and performing positional encoding. The Transformer encoder is composed of L multi-head attention layers, a normalization layer, a feedforward neural network layer, and a normalization layer. Generally, L = 3, that is, the encoder has 3 layers, including 3 multi-head attention layers, which are responsible for extracting contextual information from the feature sequence. The output layer is formed by stacking a Flaten layer and 3 fully connected layers, responsible for converting the high-dimensional features output by the Transformer into the final prediction value.

[0030] Furthermore, in step S1, the deep neural network is trained and evaluated, using Smooth L1 as the loss function, Adam as the optimizer, a learning rate of 10e-3, and ReduceLROnPlateau to schedule the learning rate. During network training, Dropout (set to 30%) is introduced to improve model robustness.

[0031] Among them, in step S2, a shield machine excavation simulation model is established, and EDEM software is used to simulate the interaction between the shield machine and the excavation geological environment. Due to limited computing power, only limited and typical geological conditions can be simulated, and the simulation prediction results are stored in the database in the form of a list for subsequent fusion prediction query.

[0032] Furthermore, limited, typical geological conditions primarily refer to a single geological environment, for example, a 100% sandy clay layer or a 100% earthy, strongly weathered mixed granite layer, for a total of six geological conditions. Similarly, for a uniform sampling of shield machine control parameter ranges, for example, for a total thrust value range of [20,000, 60,000], assuming a value interval of 1,000, 40 samples are taken.

[0033] Furthermore, in step S3, the similarity of geological conditions is calculated and the characteristic parameters of the prediction ring are (including geology, thrust, cutterhead speed, penetration) after unified scaling, adjacent to the history Ring Sample Matching is performed and the Euclidean distance of the feature vectors between the current ring and the historical ring is calculated, which is defined as the similarity of the geological conditions between the two: The higher the similarity, the closer the geological conditions between the two rings are. =6, increase the scope of query samples while ensuring real-time performance.

[0034]

[0035] Further, in step S4, the prediction credibility is defined as the corresponding fusion probability, the higher the credibility, the more credible the result obtained by the corresponding prediction mode, and the greater the probability in the fusion process of the two prediction results. The specific fusion calculation is:

[0036] wherein, represents the learning prediction of the prediction ring, represents the simulation prediction of the prediction ring, represents the optimal estimation of the prediction ring.

[0037] The application also provides a shield tunneling prediction system based on data and physical fusion driving, comprising the following steps: a learning prediction module, configured to provide a current ring data-driven shield tunneling speed and cutter torque prediction result in real time according to the geological conditions and shield machine operating parameters of the current ring; a simulation prediction module, configured to provide a current ring physical-driven shield tunneling speed and cutter torque prediction result by querying the most similar geological working conditions in the simulation prediction list offline according to the geological conditions and shield machine operating parameters of the current ring; a fusion prediction module, configured to calculate the similarity of the geological working conditions between the current ring and the historical ring, and to adaptively fuse the two prediction results according to the similarity.

[0038] The above system works as follows: First, the geological working conditions of the current prediction ring, including the geological conditions, thrust, cutter speed, and penetration, are input into the simulation prediction module and the learning prediction module, respectively, to obtain the corresponding prediction results, i.e., the tunneling speed and the cutter torque, by calculation. The learning prediction module, as shown in Figure 2 , outputs , as shown in Figure 3 ; the simulation prediction module mainly adopts offline query, as shown in Figure 4 , and outputs .

[0039] Then, the geological working condition feature credibility calculation module is responsible for querying the shield tunneling database, as shown in Figure 5 , calculating the feature similarity between the current prediction ring and the historical tunneling ring, and defining the credibility of the learning prediction module.

[0040] Finally, the deep fusion module provides a fusion framework, which fuses and according to the geological working condition credibility, realizes the optimal estimation of the prediction.

[0041] In the embodiment of the present application, the shield tunneling prediction method based on data and physical fusion driving can realize real-time prediction of the shield tunneling speed and cutter torque. When the geological conditions are similar to the historical tunneling data of the shield, that is, the shield AIS has faced similar geological conditions and operating parameters, the credibility of the learning prediction module is higher, and the weight of the fusion process with the simulation prediction result is greater; on the contrary, the smaller the credibility, the greater the weight of the simulation prediction result. The optimal estimate is obtained by fusing the prediction results of the above two parts through the credibility. The prediction accuracy after fusion is 92.9%, and the mean square error is 0.011.

[0042] In summary, the present application proposes a shield tunneling prediction method based on data and physical fusion driving: first, use the Transformer network to learn and predict the historical geological-control-tunneling big data in real time; then use offline simulation of typical working conditions to establish a query table; then according to the geological similarity of the current ring and the historical ring, adaptively fuse the two prediction results. This method relies on learning when data is sufficient, and relies on simulation when sample is scarce, achieving 92.9% prediction accuracy and millisecond-level response, which not only overcomes the problems of sample imbalance and poor generalization of pure data model, but also avoids the defects of large calculation amount and difficulty in real-time of pure physical simulation, significantly improving the prediction accuracy and engineering applicability of shield tunneling speed and cutter torque under complex geological conditions.

[0043] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A shield machine excavation prediction method based on data and physical fusion drive, characterized in that: The following steps are involved: Step S1: Real-time prediction of shield machine excavation target parameters through data-driven model; Step S2: Offline prediction of tunneling target parameters by physical driving model and construction of simulation database; Step S3: Calculate the similarity R between the current geological conditions and the historical conditions, with a value range of [0,1]; Step S4: Dynamically fusion data driven prediction results based on similarity R Prediction results with physical drivers , generate fusion prediction value ; The fusion formula satisfies: in, represents the learned prediction of the prediction loop, represents the simulation prediction of the prediction loop, represents the optimal estimate of the prediction ring.

2. The shield machine excavation prediction method based on data and physical fusion drive according to claim 1 is characterized in that: The data-driven model is a deep neural network based on the Transformer architecture, which includes an input embedding layer, a Transformer encoding layer, and an output layer. The input features are a combination vector of shield machine operating parameters and geological parameters, and the output is tunneling speed and cutterhead torque. Through training and evaluation of shield machine tunneling data sets, a multivariate regression model for shield machines is formed to achieve data-driven real-time prediction of shield machine tunneling.

3. The shield machine excavation prediction method based on data and physical fusion drive according to claim 2 is characterized in that: When training the deep neural network, smooth L1 is set as the loss function, Adam is selected as the optimizer, and the learning rate is set to 10e -3 , and use ReduceLROnPlateau to schedule the learning rate; and introduce a regularization process with a Dropout rate of 30%.

4. The shield machine excavation prediction method based on data and physical fusion drive according to claim 2 is characterized in that: The shield machine excavation data set is processed through resampling, correlation analysis and empirical judgment steps to form 11-dimensional sample data including geology, thrust, cutterhead speed, penetration, excavation speed and cutterhead torque.

5. A shield machine excavation prediction method based on data and physical fusion drive as claimed in claim 4, characterized in that The number of sample data is at least 1600 groups; and all features are normalized and scaled. The scaling method is as follows: in, is the characteristic variable The maximum value of is the characteristic variable The minimum value of is the normalized value after scaling, is the number of feature samples.

6. The shield machine excavation prediction method based on data and physical fusion drive according to claim 1 is characterized in that: In step S2, simulation prediction refers to the prediction of the tunneling speed and cutterhead torque of the shield machine based on physical drive, using the discrete element physical simulation method to perform offline tunneling simulation on typical geological conditions, gradually forming a series of prediction results, and storing them in a list form to form a simulation prediction query table.

7. The shield machine excavation prediction method based on data and physical fusion drive according to claim 1 is characterized in that: In step S3, the geological condition similarity calculation is performed by uniformly scaling the characteristic parameters St of the predicted ring and comparing them with the historical adjacent N ring samples S t-1 Matching is performed and the Euclidean distance of the feature vectors between the current ring and the historical ring is calculated, which is defined as the geological similarity between the two as R: 。 8. The shield machine excavation prediction method based on data and physical fusion drive according to claim 6 is characterized in that: The typical geological conditions are single stratum types, including sandy clay soil layer, fully weathered mixed granite layer, soil-like strongly weathered mixed granite layer, massive strongly weathered mixed granite layer, moderately weathered mixed granite layer and slightly weathered mixed granite layer.

9. A shield machine excavation prediction system for implementing the method according to any one of claims 1 to 8, characterized in that: include: The learning prediction module provides real-time data-driven predictions of the tunnel boring machine's tunneling speed and cutterhead torque based on the geological conditions and shield machine operating parameters of the current ring. The simulation prediction module queries the most similar geological conditions in the simulation prediction list offline based on the geological conditions of the current ring and the shield machine operating parameters, and is used to provide the prediction results of the shield machine tunneling speed and cutterhead torque based on the physical drive of the current ring; The fusion prediction module calculates the similarity of geological conditions between the current ring and the historical ring, and is responsible for fusing the two prediction results based on the similarity.

10. A computer storage medium, characterized in that It stores a computer program that can be executed by a processor, a shield machine excavation database, and a simulation prediction query list. The computer program executes the shield machine excavation prediction method based on data and physical fusion drive as described in any one of claims 1-8.

Images