Shield machine oil cylinder pressure accurate prediction method and system based on trusted artificial intelligence

Abstract

The application belongs to the technical field of shield construction, and particularly discloses a shield machine oil cylinder pressure precise prediction method and system based on trusted artificial intelligence, which comprises the following steps: collecting and obtaining experimental data and simulation data of various dynamics and oil cylinders of a continuous tunneling shield machine; based on the experimental data and simulation data, analyzing and deducing the physical law in oil cylinder driving force estimation, wherein the pressure constraint boundary of the hydraulic oil cylinder is constructed in combination with Newtonian mechanics and the Lagrange multiplier method, and is converted into a loss function; taking a multi-head attention deep neural network as a basic framework for oil cylinder pressure prediction, based on the loss function, an oil cylinder pressure prediction model of the continuous tunneling shield machine is constructed, and the oil cylinder pressure is predicted based on the oil cylinder pressure prediction model. The application introduces a physical information neural network, combines an advanced deep learning method, realizes precise stress prediction of the continuous tunneling shield machine pipe piece installation process, and reduces dangerous working conditions in the construction process.

Description

Technical Field

[0001] This invention belongs to the field of tunnel boring machine (TBM) construction technology. Specifically, it relates to a method and system for accurate prediction of hydraulic cylinder pressure of a TBM based on reliable artificial intelligence, and in particular, it relates to a method and system for real-time prediction and control of hydraulic cylinder pressure in the TBM segment installation scenario based on a reliable multi-head attention deep neural network method guided by physical information. Background Technology

[0002] As a type of underground construction equipment, tunnel boring machines (TBMs) have attracted widespread attention due to their high operating efficiency, good integration, and high degree of automation. To improve construction efficiency and shorten the construction cycle, many researchers have begun to develop various types of TBMs capable of continuous, uninterrupted operation; these are also known as continuous tunneling TBMs. Through their special mechanical structure, continuous tunneling TBMs can simultaneously perform front-end soil cutting and rear-end segment assembly, improving construction efficiency by 30-50%. However, the synchronicity of multiple actions in continuous tunneling TBMs presents certain challenges to the stability and precise control of the construction process. In particular, how to rationally match the front-end cutterhead excavation system and the rear-end segment assembly system to prevent stress imbalance and stress concentration within the TBM is a crucial scientific issue. Therefore, accurately characterizing the dynamic parameters of continuous tunneling TBMs to ensure their safe and stable operation is a vital research topic.

[0003] Data-driven models are widely used in the prediction, classification, and early warning of tunnel boring machine (TBM) parameters, achieved through the specific training of relevant machine learning models. Complex machine learning and deep learning models can capture interaction rules across different datasets. While there is considerable research on TBM parameter prediction, research on hydraulic system pressure prediction is scarce. Hydraulic cylinder pressure prediction is a critical issue in the construction of continuously tunneling TBMs, crucial for reducing shield instability and stress concentration. Furthermore, since the input to hydraulic system pressure prediction is typically the force and torque transmitted from the front end, and the output is the pressure of multiple hydraulic cylinders, hydraulic cylinder pressure prediction is a high-dimensional, underconstrained problem. Due to the limited input dimension of the prediction model, usually lower than the output dimension, traditional data-driven methods often fail to achieve satisfactory prediction results. Therefore, it is necessary to explore deep learning methods guided by physical information to ensure accurate prediction of hydraulic cylinder pressure and achieve stable and safe TBM construction.

[0004] Based on the above, this research on the hydraulic cylinder pressure analysis and modeling of continuous tunneling shield machine focuses on how to analyze the stress state of the hydraulic cylinder during the segment installation process and establish a physical model, how to construct simulation and experimental scenarios of the segment assembly process to obtain overall data information, and how to design a physical-guided deep learning neural network to achieve high-precision prediction of hydraulic cylinder pressure. These are the key issues for achieving accurate prediction and control of hydraulic cylinder pressure. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for accurate prediction of hydraulic cylinder pressure in tunnel boring machines (TBMs) based on reliable artificial intelligence. Addressing the challenges of stress concentration, shield swaying, and high precision requirements for segment installation during continuous tunneling TBM construction, this invention introduces the key concept of physical information neural networks and combines it with advanced deep learning methods to achieve accurate stress prediction during segment installation in continuous tunneling TBMs, thereby reducing dangerous working conditions during construction.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for accurate prediction of hydraulic cylinder pressure in a tunnel boring machine based on reliable artificial intelligence is proposed, comprising the following steps:

[0007] Step 1: Collect and obtain experimental and simulation data on various dynamics and hydraulic cylinders of the continuous tunneling shield machine;

[0008] Step 2: Based on the experimental and simulation data, the physical laws in the estimation of the hydraulic cylinder driving force are analyzed and derived. In this step, the pressure constraint boundary of the hydraulic cylinder is constructed by combining Newtonian mechanics and the Lagrange multiplier method, and then transformed into a loss function.

[0009] Step 3: Using a multi-head attention deep neural network as the basic framework for cylinder pressure prediction, and based on the loss function, construct a cylinder pressure prediction model for a continuous tunneling shield machine, and predict cylinder pressure based on this model.

[0010] As a further preferred option, step one includes the following steps:

[0011] Collect various dynamic data and hydraulic cylinder data of the continuous tunneling shield machine and use them as experimental data;

[0012] A simulation model was constructed based on the geometric model of a continuous tunneling shield machine. Mesh independence tests were conducted in conjunction with experimental data to obtain simulation data on various dynamics and hydraulic cylinders.

[0013] As a further preferred embodiment, the acquisition of various dynamic data and hydraulic cylinder data of the continuous tunneling shield machine includes:

[0014] (111) Reset all experimental parameters and system variables. The assembly hydraulic cylinder and the excavation hydraulic cylinder are both locked, and the test bench is in the reset state.

[0015] (112) Activate the cutterhead cutting mode and realize the cutterhead cutting process and the tunneling machine forward process by controlling the excavation hydraulic cylinder parameters;

[0016] (113) After the rock and soil cutting of one ring is completed, the synchronous excavation and segment installation mode is started. The front-end excavation hydraulic cylinder continues to cut the rock and soil, and the rear-end assembly hydraulic cylinder starts the segment installation.

[0017] (114) Select the corresponding segment, retract the corresponding hydraulic cylinder, and the segment installation machine will move the segment to the corresponding position to complete the segment installation. Repeat this operation until all segments of the current ring are installed.

[0018] (115) After the hydraulic excavation segment is assembled in the entire ring, start excavating the next ring. Repeat steps (111) to (114) to complete the assembly of multiple ring segments and record the relevant experimental data.

[0019] (116) After the experiment was completed, the cylinder was reset and the experimental data was collected.

[0020] As a further preferred option, step two includes the following steps:

[0021] (21) Based on Newtonian mechanics, construct the dynamic equation of the assembly hydraulic cylinder during the segment assembly process;

[0022] (22) The pressure constraint boundary of the hydraulic cylinder is constructed using the Lagrange multiplier method. Since the hydraulic cylinders are uniformly distributed along the circumference, the force of each hydraulic cylinder is obtained according to the lever arm relationship of the hydraulic cylinders.

[0023] (23) Based on the dynamic equation and the force structure loss function of each hydraulic cylinder.

[0024] As a further preferred embodiment, in step (22), the forces acting on each hydraulic cylinder are:

[0025]

[0026] In the formula, F i For the forces in the hydraulic cylinder, F and M x M y These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively; n represents the number of hydraulic cylinders; r is the radius from the hydraulic cylinder to the center of the shield; and x... i and y i These represent the lever arms of the hydraulic cylinder in the horizontal and vertical directions, respectively.

[0027] As a further preferred embodiment, in step (23), the loss function includes:

[0028]

[0029] In the formula, P ij For the predicted hydraulic cylinder pressure, The actual hydraulic cylinder pressure is given by: i is the hydraulic cylinder serial number, j is the sample serial number, λ is the physical information proportionality coefficient (λ is a constant), M represents the sample size, n is the number of hydraulic cylinders, S is the cross-sectional area of ​​the hydraulic cylinder, r is the radius from the hydraulic cylinder to the shield center, and F... j M xj M yj M represents the force, vertical bending moment, and horizontal bending moment transmitted from the intermediate shield to the j-th sample, respectively. x M y These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively. ij and y ij Let represent the lever arm of the j-th sample hydraulic cylinder in the horizontal and vertical directions, respectively.

[0030] As a further preferred embodiment, in step three, the prediction model includes an input part, an output part, an encoder, and a decoder. The input part is used to input experimental data and / or simulation data, the output part is used to output the pressure of each hydraulic cylinder, the encoder is used to embed relevant relationships with historical information, and the decoder is used to receive information of the current hydraulic cylinder.

[0031] The encoder and decoder consist of multi-head attention mechanism sub-layers and deep neural network sub-layers. Each sub-layer has LayerNormalization and residual connections. The execution process of the multi-head attention mechanism includes:

[0032]

[0033] Q i =QW i Q

[0034] K i =KW i K

[0035] V i =VW i V

[0036] head i =Attention(Q) i ,K i V i )

[0037] MultiHeadAttention(Q,K,V)=Concact(head1,head2,…,head i )

[0038] Where Q, K, and V are respectively composed of input data and W Q WK W V Multiplying them together, we get d k Let W be the dimension of the vector in K, W represent the weight matrix, I be the number of positive faces, and i be the sample number.

[0039] As a further preferred option, step three also includes: using fixed coefficients, root mean square error, maximum absolute error, and explained variance score to evaluate the output error of the prediction model.

[0040] As a further preferred option, step three also includes: using the SHAP and LIME methods to interpret the features of the prediction model's output.

[0041] According to another aspect of the present invention, a shield tunneling machine hydraulic cylinder pressure accurate prediction system based on reliable artificial intelligence is also provided, comprising:

[0042] The first main control module is used to collect and acquire various dynamic and hydraulic cylinder experimental data and simulation data of the continuous tunneling shield machine;

[0043] The second main control module is used to analyze and derive the physical laws in the estimation of the hydraulic cylinder driving force based on the experimental data and simulation data. In this process, the pressure constraint boundary of the hydraulic cylinder is constructed by combining Newtonian mechanics and the Lagrange multiplier method, and then converted into a loss function.

[0044] The third main control module is used to use a multi-head attention deep neural network as the basic framework for cylinder pressure prediction. Based on the loss function, it constructs a cylinder pressure prediction model for a continuous tunneling shield machine and predicts cylinder pressure based on this model.

[0045] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0046] 1. This invention achieves accurate prediction and hazard control of hydraulic cylinder pressure by collecting experimental simulation data, analyzing the physical laws of tunnel boring machine operation scenarios, and using deep learning models guided by physical information.

[0047] 2. This invention integrates physical models, simulation models, and physics-guided deep learning neural networks to achieve accurate prediction of hydraulic cylinder pressure in continuously tunneling shield machines. The physical model combines Newton's laws of motion with the Lagrange multiplier method, aiming to find the physical laws governing the relationship between hydraulic cylinder pressure and tunneling mechanics. A simulation model is established using the finite element method, and reasonable boundary conditions and mesh parameters are designed to ensure consistency between simulation and experimental results. The physics-guided deep learning neural network is based on a multi-head attention deep neural network (MD), and a loss function is designed according to physical laws to achieve cylinder pressure prediction.

[0048] 3. The physical information-guided deep learning model established in this invention achieves high prediction accuracy for hydraulic cylinders, with R², RMSE, MAE, and EVS of 0.977, 7.302, 5.299, and 0.978, respectively. Furthermore, by incorporating simulation data and physical laws, this invention improves R², RMSE, MAE, and EVS by 9.04%, 53.96%, 55.91%, and 9.89%, respectively. In addition, this invention combines a data-driven method for cylinder pressure prediction, achieving accurate prediction and hazard control of hydraulic cylinder pressure. Finally, through model interpretation, this invention demonstrates that the importance of network structure and loss function is higher than that of input features, illustrating the importance of physical information and network structure for deep learning models. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method for accurately predicting the hydraulic cylinder pressure of a tunnel boring machine based on trusted artificial intelligence, according to an embodiment of the present invention.

[0050] Figure 2 This is a framework diagram of a method for accurate prediction of shield machine cylinder pressure based on trusted artificial intelligence, which is involved in an embodiment of the present invention.

[0051] Figure 3 Image (a) is a side view of the continuous tunneling shield machine structure. Figure 3 Image (b) is a rear view of the continuous tunneling shield machine structure. Figure 3 (c) in the diagram is a schematic diagram of the rear end structure of the hydraulic cylinder of a continuous tunneling shield machine. Figure 3 (d) in the diagram is a schematic diagram of the front end structure of the hydraulic cylinder of a continuous tunneling shield machine. Figure 3 (e) in the diagram is a schematic diagram of the segment assembly machine of a continuous tunneling shield machine;

[0052] Figure 4 (a) in the diagram is a mesh diagram of the simulation model of a continuous tunneling shield machine. Figure 4 (b) in the figure is the stress cloud diagram of the simulation model of the continuous tunneling shield machine;

[0053] Figure 5 (a) in the figure is the mesh independence test diagram of the simulation model of the continuous tunneling shield machine. Figure 5 (b) in the figure is the mesh reliability test diagram of the simulation model of the continuous tunneling shield machine;

[0054] Figure 6 This is a mechanical scenario diagram of segment assembly involved in an embodiment of the present invention;

[0055] Figure 7 (a) in the figure is the overall framework diagram of the information-guided deep learning model framework involved in the embodiments of the present invention. Figure 7 (b) in the diagram is a deep neural network framework diagram. Figure 7(c) in the diagram is the framework diagram of the multi-head attention layer;

[0056] Figure 8 This is a comparison chart of the number of hidden elements in the deep neural network involved in the embodiments of the present invention;

[0057] Figure 9 This is a schematic diagram illustrating the number of convergence iterations under different learning rates in embodiments of the present invention;

[0058] Figure 10 This is a comparison diagram of the physical information scaling factor λ involved in the embodiments of the present invention;

[0059] Figure 11 This is a prediction accuracy diagram of 16 hydraulic cylinders involved in an embodiment of the present invention;

[0060] Figure 12 This is a schematic diagram of the feature importance ranking structure of the prediction model of this invention;

[0061] Figure 13 (a) in the figure shows the feature interpretation results of the prediction model using SHAP. Figure 13 (b) in the figure is the result of using LIME to interpret the features of the prediction model. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0063] like Figure 1 and Figure 2 As shown, this invention proposes a hydraulic cylinder pressure prediction method and system integrating physical models, simulation models, and physics-guided deep learning neural networks. The physical model combines Newton's laws of motion with the Lagrange multiplier method, aiming to find the physical laws governing hydraulic cylinder pressure and tunneling mechanics. A simulation model is established using the finite element method, and reasonable boundary conditions and mesh parameters are designed to ensure consistency between simulation and experimental results. The physics-guided deep learning neural network is based on a multi-head attention deep neural network (MD), and a loss function is designed according to physical laws to achieve cylinder pressure prediction. Specifically, it includes the following steps:

[0064] Step 1: Data collection.

[0065] A geometric model and test bench for a continuous tunneling shield machine were constructed, the scenarios of segment installation and cutterhead excavation were determined, the on-site sampling steps were set up, and various dynamic data and hydraulic cylinder data were collected.

[0066] Step 2: Simulation data acquisition.

[0067] A simulation model was constructed based on the geometric model of a continuous tunneling shield machine. Boundary conditions were modified in conjunction with experimental data constraints. Further mesh independence tests were conducted to ensure that the simulation data met the actual construction conditions, and the collection of simulation data was completed.

[0068] Step 3: Analysis of the physical laws governing the hydraulic cylinder pressure prediction process.

[0069] For the scenario of hydraulic cylinder pressure prediction, a pressure constraint boundary for the hydraulic cylinder was constructed by combining Newtonian mechanics and the Lagrange multiplier method, and then transformed into a loss function, which provides support for the design of subsequent physical information-guided deep learning models.

[0070] Step 4: Construction and prediction of deep learning models guided by physical information.

[0071] This invention selects a multi-head attention deep neural network as the basic framework for hydraulic cylinder pressure prediction. By integrating relevant loss functions derived from physical laws, it achieves reliable hydraulic cylinder pressure prediction driven by a mixture of physical and data information.

[0072] More specifically, refer to Figure 2 As shown, this embodiment provides a real-time prediction method for hydraulic cylinder pressure in a tunnel segment installation scenario based on a reliable multi-head attention deep neural network guided by physical information, including the following steps:

[0073] Step 1: Description of experimental data acquisition methods;

[0074] To highlight the importance of hydraulic cylinder pressure prediction and the versatility of the pressure prediction framework, this invention develops a novel continuous tunneling shield machine. Compared to conventional shield machines that operate intermittently, continuous tunneling shield machines require simultaneous excavation and segment assembly, resulting in a more complex dynamic environment and necessitating more precise control schemes. Because the hydraulic cylinders of continuous tunneling shield machines must simultaneously withstand forces transmitted from both ends, they are more sensitive to pressure changes and possess greater research value.

[0075] A continuous tunneling shield machine mainly consists of several parts, including an excavation hydraulic system, an assembly hydraulic system, a front shield, a cutterhead, a segment frame, and a middle shield. Its structure is as follows: Figure 3As shown in Table 1, the rear-end assembly hydraulic system, which is in direct contact with the tunnel segment, has the most significant impact on the stability of the entire excavation and segment assembly process. Furthermore, because some assembly hydraulic cylinders need to retract during segment installation to create space, it is exposed to overall instability and stress concentration. Therefore, the cylinder driving force of the assembly hydraulic system requires special attention. Accurately estimating the driving force and pressure of the hydraulic cylinders can effectively avoid extreme working conditions and ensure the stability of segment installation.

[0076] Table 1. Modules and Main Functions of Continuous Tunneling Shield Machine

[0077]

[0078] To accurately estimate the driving force of the hydraulic cylinder and provide data support for the deep learning model, this invention conducted experimental data collection based on an experimental platform. The experimental platform is as follows: Figure 3 As shown, the experimental process is as follows: (1) Before the experiment begins, all experimental parameters and system variables are reset. The assembly hydraulic cylinder and the excavation hydraulic cylinder are both locked, and the test bench is in the reset state. (2) The cutterhead cutting mode is turned on, and the cutterhead cutting process and the tunneling machine forward process are realized by controlling the parameters of the excavation hydraulic cylinder. (3) After the rock and soil cutting of one ring is completed, the synchronous excavation and segment installation mode is started. The front excavation hydraulic cylinder continues to cut the rock and soil, and the rear assembly hydraulic cylinder starts segment installation. (4) Select the corresponding segment and retract the corresponding hydraulic cylinder. The segment installation machine moves the segment to the corresponding position and completes the segment installation. Repeat this operation until all parts of the current ring are installed. (5) After the front hydraulic excavation cylinder completes the assembly of the entire ring segment, it begins to excavate the next ring. Repeat steps (1) to (4) to complete the assembly of multiple ring segments and record the relevant experimental data. (6) After the experiment is completed, the cylinder is reset. The relevant sensors collect experimental data and then immediately transmit it to the control terminal.

[0079] During segment installation, data from continuously tunneling tunneling machines (MTBF) is limited, particularly regarding the surrounding environment and hydraulic system parameters, which are difficult to collect. Furthermore, the observed data is uncertain due to varying worker preferences and geological conditions. Accurately predicting the pressure of each hydraulic cylinder cannot be achieved solely based on engineering construction data. Therefore, this invention, based on a data-driven approach, constructs relevant simulation scenarios and adds relevant physical information to ensure the realism and reliability of the prediction model.

[0080] Step 2: Description of simulation data acquisition method;

[0081] The simulation model mainly includes a 3D model, mesh design, boundary conditions, and finite element solution. The 3D model is simplified based on the structural model in step 1. A hybrid mesh of tetrahedrons and hexahedrons is used to effectively ensure mesh continuity. The boundary conditions are set consistent with the actual experiment, with the hydraulic cylinder receiving the force and torque transmitted from the front end. The relevant simulation model and hydraulic cylinder pressure results are as follows: Figure 4 As shown.

[0082] Since the number of mesh elements directly affects simulation accuracy, mesh independence testing is crucial for ensuring the reliability of simulation results. Too many mesh elements increase computational costs, while too few mesh elements reduce simulation accuracy. This invention selects a certain range of mesh numbers and uses the maximum and minimum cylinder pressures as evaluation indicators. Figure 5 As shown in (a) of the figure, when the number of grids is greater than 500,000, the fluctuation of the simulation results is less than 1%. Therefore, this invention selects 500,000 grids as the basis for grid division. Simultaneously, the simulation results are compared with the experimental results, and the comparison results are as follows: Figure 5 As shown in (b) above, the reliability of the simulation model is further verified. Relevant data are shown in Table 2.

[0083] Table 2. Main data types acquired in experiments and simulations.

[0084]

[0085]

[0086] Step 3: Analysis of the physical laws governing the hydraulic cylinder pressure prediction process;

[0087] Through data acquisition from experiments and simulations, key data for the dynamic analysis of the hydraulic cylinder were obtained. This step analyzed and derived the physical laws governing the estimation of the cylinder's driving force. Force constraint boundaries for the hydraulic cylinder were constructed using Newtonian mechanics and the Lagrange multiplier method, and then transformed into a loss function. This provides support for the design of a physics-guided deep learning model.

[0088] To clearly explore the physical laws governing the segment assembly process, this step constructs a dynamic model. The dynamic state is as follows: Figure 6 As shown. For assembled hydraulic cylinders, the pressure mainly comes from the force and torque transmitted by the intermediate guard. The basic dynamic equations are as follows:

[0089]

[0090]

[0091]

[0092] In the formula, Pi F represents the pressure of the hydraulic cylinder. i Let P be the force of the hydraulic cylinder, and S be the contact area of ​​the hydraulic cylinder. Since the hydraulic cylinders of this invention are of the same type, P... i With F i They are positively correlated. N represents the number of hydraulic cylinders, which is 16 in this invention. F, M x M y These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively. i and y i These represent the lever arms of the hydraulic cylinder in the horizontal and vertical directions, respectively.

[0093] Because this is a high-dimensional underconstrained problem (i.e., 16 variables and 3 constraint equations), it is impossible to obtain an analytical value for the pressure of a single hydraulic cylinder. To ensure the stability of the hydraulic cylinder, optimization equations need to be designed:

[0094]

[0095] The pressure boundary of the hydraulic cylinder can be obtained using the Lagrange multiplier method. Based on the above equations, the Lagrange equations are constructed as follows:

[0096]

[0097] Where λ1, λ2, and λ3 are scaling factors.

[0098] Taking the partial derivatives of the constructed Lagrange equation with respect to Fi and λ respectively, we get:

[0099]

[0100]

[0101]

[0102]

[0103] Meanwhile, since the hydraulic cylinders are evenly distributed circumferentially, the lever arms of the hydraulic cylinders have the following relationship:

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] By solving the above system of equations, we can obtain the forces acting on each hydraulic cylinder as follows:

[0110]

[0111] The loss function can be constructed based on the above formula:

[0112]

[0113] Among them, P ij This represents the hydraulic cylinder pressure predicted by the model. The actual hydraulic cylinder pressure is given. i is the hydraulic cylinder serial number, and j is the sample serial number. λ is the physical information scaling factor, ranging from 0 to 1. M represents the sample size. N is the number of hydraulic cylinders. S is the cross-sectional area of ​​the hydraulic cylinder. r represents the radius from the hydraulic cylinder to the shield center. F j M x M yj Let these represent the force, vertical bending moment, and horizontal bending moment transmitted from the intermediate shield to the j-th sample, respectively. Figure 6 As shown.

[0114] Step 4: Construction of a deep learning model guided by physical information;

[0115] To construct a physics-guided deep learning model, this invention selects a multi-head attention deep learning network as the basic framework, such as... Figure 7 As shown. Due to their complex network structure and massive model parameters, deep learning models perform well in prediction problems. This step describes the basic framework of physics-guided deep learning models and provides relevant model evaluation and interpretation methods.

[0116] An overview of the model is as follows: Figure 7 As shown in (a) above, the model comprises four components: an input section, an output section, an encoder, and a decoder. The input section includes relevant dynamic parameters of the tunnel boring machine, including experimental and simulation data. The output section is the pressure of each hydraulic cylinder. The encoder embeds correlations with massive amounts of historical information, while the decoder receives current hydraulic cylinder information. By leveraging the cross-focus of the encoder and decoder, the latest hydraulic cylinder information is paired with relevant mapping patterns in historical data, and prediction patterns are inferred, improving computational efficiency and reducing redundant historical information. These techniques enable deep learning methods to model non-stationary data with multi-scale stable characteristics and generate prediction results in a generative paradigm, representing an attempt to solve the prediction problem of high-dimensional, massive data.

[0117] The encoding and decoding modules adopt a similar structure, consisting of multi-head attention mechanism sub-layers and deep neural network sub-layers, each with LayerNormalization and residual connections. Figure 7(b) shows how this model processes a deep neural network. A deep neural network, also known as a feedforward neural network, consists of multiple layers. In a deep neural network, each neuron is connected to all neurons in the previous layer. Signals propagate unidirectionally from the input layer to the output layer. It uses a GeLU activation function, which further transforms the features learned in the attention mechanism, as shown below:

[0118]

[0119] Figure 7 (c) in the diagram illustrates the execution process of the multi-head attention mechanism in this model. Attention mechanisms can extract global features from the input sequence. Multi-head attention is an improvement upon the traditional attention mechanism. By forming multiple subspaces, different aspects of the model can be explored. Its computational process is shown below:

[0120]

[0121]

[0122]

[0123]

[0124] head i =Attention(Q) i ,K i V i )

[0125] MultiHeadAttention(Q,K,V)=Concact(head1,head2,…,head i )

[0126] Where Q, K, and V are respectively composed of input data and W Q W K W V Multiplying them together gives d. k is the dimension of the vectors in K, and its function is to obtain a stable gradient through scaling. W represents the weight matrix. I is the number of positive faces.

[0127] Using this model and the data from steps 1 and 2, model training and testing were conducted. To ensure the reliability of the training results, the input features were processed using the Min-Max normalization method. The loss function used was the one from step 3, and the Adam optimizer was employed. The hyperparameters, crucial for model training, were analyzed as follows.

[0128] The hidden size, learning rate, number of iterations, and physical information ratio of the loss function are core hyperparameters of deep learning models. Since this model contains four deep neural networks, each can have a different hidden size. The computational accuracy for different hidden sizes of the four deep neural networks is as follows: Figure 7 As shown, the optimal value is obtained after 10 calculations to obtain the corresponding calculation result. According to... Figure 7 Set the hidden size to 128, 128, 128, and 32 respectively. Figure 9 The number of convergence iterations is shown for different learning rates. Therefore, the learning rate chosen in this invention is 0.01, and the number of iterations chosen in this invention is 10,000. Figure 10 This shows the effect of the physical information scaling factor λ on the model. Figure 9 It can be seen that the convergence effect is best when the value is 0.7.

[0129] To verify the goodness of the model, the output error of the prediction model is mainly measured by model evaluation metrics. Commonly used evaluation metrics include the coefficient of determination (R²). 2 The metrics include root mean square error (RMSE), maximum absolute error (MAE), and explained variance score (EVS). For RMSE and MAE, the smaller the value, the higher the model's predictive accuracy. 2 The closer the value of EVS is to 1, the higher the prediction accuracy of the model.

[0130]

[0131]

[0132]

[0133]

[0134] The results show that a deep neural network guided by physical information can achieve high-precision prediction of hydraulic cylinder pressure. For example... Figure 11 As shown in Table 3, good prediction accuracy can be achieved for all 16 cylinders at different positions. 2 The mean was 0.977, the mean RMSE was 7.302, the mean MAE was 5.299, and the mean EVS was 0.978. 2 The variances of RMSE, MAE, and EVS were 0.011, 1.315, 0.910, and 0.011, respectively.

[0135] Table 3. Prediction accuracy for 16 hydraulic cylinders

[0136]

[0137]

[0138] Compared to traditional deep learning models, deep learning models that combine physical laws and simulation data exhibit better model accuracy. As shown in Table 4, when the model does not contain any simulation data or physical laws, its R-value is significantly higher. 2 The R² value is 0.896, RMSE is 15.861, MAE is 12.018, and EVS is 0.89. By incorporating physical laws and simulation data, the prediction accuracy can be improved to varying degrees. When the model only contains physical laws, its R² value is... 2 The R-value is 0.941, RMSE is 11.598, MAE is 8.438, and EVS is 0.935. When only simulation data is added, the model's R-value is... 2 The R-value is 0.953, RMSE is 10.375, MAE is 7.662, and EVS is 0.951. When the model includes both simulation data and physical laws, its R-value is... 2 The R² value is 0.977, RMSE is 7.302, MAE is 5.299, and EVS is 0.978. Compared to the model without added physical information, its R² value is significantly higher. 2 The accuracy of predictions improved by 9.04%, 53.96%, 55.91%, and 9.89%, respectively. The results indicate that both physical laws and simulation data can improve the model's predictive performance. Physical laws primarily improve accuracy by constraining the model's predicted values ​​through the loss function. The physical information within the loss function makes the predictions more realistic and understandable. Simulation data, by supplementing the model with a large number of samples and adding more scene data, further enhances the model's accuracy.

[0139] Table 4. Prediction results of different prediction models

[0140]

[0141]

[0142] To verify the reliability of the prediction results, this invention further conducts model interpretation research. Through model interpretation, the importance of physical information in the network structure is second only to the model structure, and higher than the importance of input features. For example... Figure 12 , Figure 13As shown, the importance of network structure and loss function to model prediction is very significant. The importance scores of the four deep neural network layers are 54.914, 54.470, 46.440, and 46.995, ranking 1st, 2nd, 5th, and 4th respectively. The importance scores of the two multi-head attention mechanisms are 49.608 and 42.224, ranking 3rd and 6th respectively. This indicates that network structure has the greatest impact on model reliability, exceeding the impact of loss function and input features. The importance score of the loss function is 34.920, ranking 7th. This shows that the physical information loss function designed in this invention also has a significant impact. After knowing the network structure and physical information, mechanical parameters such as F, Mx, and My are crucial. To demonstrate the reliability of this invention, Figure 12 The model interpretation results for the SHAP and LIME methods are presented. Comparison with SHAP and LIME shows that the feature interpretation in this study is reliable. Since SHAP and LIME cannot perform importance analysis on network structure, the interpretation method proposed in this invention has broader applicability. Through importance ranking, it is found that the selection of physical information is more critical than feature selection, and designing a good loss function can significantly improve the reliability of the model.

[0143] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for accurately predicting the hydraulic cylinder pressure of a tunnel boring machine based on reliable artificial intelligence, characterized in that, Includes the following steps: Step 1: Collect and obtain experimental and simulation data on various dynamics and hydraulic cylinders of the continuous tunneling shield machine; Step 2: Based on the experimental and simulation data, the physical laws in the estimation of the hydraulic cylinder driving force are analyzed and derived. In this step, the pressure constraint boundary of the hydraulic cylinder is constructed by combining Newtonian mechanics and the Lagrange multiplier method, and then transformed into a loss function. Step 3: Using a multi-head attention deep neural network as the basic framework for cylinder pressure prediction, and based on the loss function, construct a cylinder pressure prediction model for a continuous tunneling shield machine, and predict cylinder pressure based on this cylinder pressure prediction model. Step two includes the following steps: (21) Construct the dynamic equation of the assembly hydraulic cylinder during the segment assembly process based on Newtonian mechanics; (22) The pressure constraint boundary of the hydraulic cylinder is constructed using the Lagrange multiplier method. Since the hydraulic cylinders are uniformly distributed along the circumference, the force on each hydraulic cylinder is obtained according to the lever arm relationship of the hydraulic cylinders. (23) Based on the aforementioned dynamic equations and the force structure loss function of each hydraulic cylinder; In step (22), the forces acting on each hydraulic cylinder are: ； In the formula, For the force of the hydraulic cylinder, , , These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively; n represents the number of hydraulic cylinders; and r is the radius from the hydraulic cylinder to the center of the shield. and These represent the lever arms of the hydraulic cylinder in the horizontal and vertical directions, respectively. In step (23), the loss function includes: ； In the formula, For the predicted hydraulic cylinder pressure, Where is the actual hydraulic cylinder pressure, i is the hydraulic cylinder serial number, j is the sample serial number, λ is the physical information proportionality coefficient (λ is a constant), M represents the sample size, n is the number of hydraulic cylinders, S is the cross-sectional area of ​​the hydraulic cylinder, and r is the radius from the hydraulic cylinder to the shield center. , , Let these represent the force, vertical bending moment, and horizontal bending moment transmitted from the intermediate shield to the j-th sample, respectively. , These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively. and Let represent the lever arm of the j-th sample hydraulic cylinder in the horizontal and vertical directions, respectively.

2. The method for accurate prediction of shield machine cylinder pressure based on reliable artificial intelligence according to claim 1, characterized in that, Step one includes the following steps: (11) Collect various dynamic data and hydraulic cylinder data of the continuous tunneling shield machine and use them as experimental data; (12) Based on the geometric model of the continuous tunneling shield machine, a simulation model was constructed, and mesh independence test was carried out in combination with experimental data to obtain simulation data of various dynamics and cylinders.

3. The method for accurate prediction of shield machine cylinder pressure based on reliable artificial intelligence according to claim 2, characterized in that, The various dynamic data and hydraulic cylinder data collected from the continuous tunneling shield machine include: (111) Reset all experimental parameters and system variables. The assembly hydraulic cylinder and the excavation hydraulic cylinder are both locked, and the test bench is in the reset state. (112) Activate the cutterhead cutting mode and control the excavation hydraulic cylinder parameters to realize the cutterhead cutting process and the tunneling machine forward process; (113) After the rock and soil cutting of one ring is completed, the synchronous excavation and segment installation mode is started. The front-end excavation hydraulic cylinder continues to cut the rock and soil, and the rear-end assembly hydraulic cylinder starts the segment installation. (114) Select the corresponding segment, retract the corresponding hydraulic cylinder, and the segment installation machine will move the segment to the corresponding position to complete the segment installation. Repeat this operation until all segments of the current ring are installed. (115) After the hydraulic excavation segment is assembled in the entire ring, start excavating the next ring, repeat steps (111) to (114) to complete the assembly of multiple ring segments, and record the relevant experimental data; (116) After the experiment is completed, the cylinder is reset and the experimental data is collected.

4. The method for accurate prediction of shield machine cylinder pressure based on reliable artificial intelligence according to claim 1, characterized in that, In step three, the prediction model includes an input part, an output part, an encoder, and a decoder. The input part is used to input experimental data and / or simulation data, the output part is used to output the pressure of each hydraulic cylinder, the encoder is used to embed relevant relationships with historical information, and the decoder is used to receive the information of the current hydraulic cylinder. The encoder and decoder consist of multi-head attention mechanism sub-layers and deep neural network sub-layers. Each sub-layer has LayerNormalization and residual connections. The execution process of the multi-head attention mechanism includes: ； ； ； ； ； ； Where Q, K, and V are respectively composed of input data and W Q W K W V Multiplying them together, we get d k Let W be the dimension of the vector in K, W represent the weight matrix, I be the number of positive faces, and i be the sample number.

5. The method for accurate prediction of shield machine cylinder pressure based on reliable artificial intelligence according to claim 1, characterized in that, Step three also includes evaluating the output error of the prediction model using fixed coefficients, root mean square error, maximum absolute error, and explained variance scores.

6. The method for accurate prediction of shield machine cylinder pressure based on reliable artificial intelligence according to claim 1, characterized in that, Step three also includes: using SHAP and LIME methods to interpret the features of the prediction model's output.

7. A precise prediction system for tunnel boring machine cylinder pressure based on reliable artificial intelligence, characterized in that, include: The first main control module is used to collect and acquire various dynamic and hydraulic cylinder experimental data and simulation data of the continuous tunneling shield machine; The second main control module is used to analyze and derive the physical laws in the estimation of the hydraulic cylinder driving force based on the experimental data and simulation data. In this process, the pressure constraint boundary of the hydraulic cylinder is constructed by combining Newtonian mechanics and the Lagrange multiplier method, and then converted into a loss function. The third main control module is used to use a multi-head attention deep neural network as the basic framework for cylinder pressure prediction, and to construct a cylinder pressure prediction model for a continuous tunneling shield machine based on the loss function, and to predict the cylinder pressure based on the cylinder pressure prediction model. The second main control module is also used to perform the following steps: (21) Construct the dynamic equation of the assembly hydraulic cylinder during the segment assembly process based on Newtonian mechanics; (22) The pressure constraint boundary of the hydraulic cylinder is constructed using the Lagrange multiplier method. Since the hydraulic cylinders are uniformly distributed along the circumference, the force on each hydraulic cylinder is obtained according to the lever arm relationship of the hydraulic cylinders. (23) Based on the aforementioned dynamic equations and the force structure loss function of each hydraulic cylinder; In step (22), the forces acting on each hydraulic cylinder are: ； In the formula, For the force of the hydraulic cylinder, , , These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively; n represents the number of hydraulic cylinders; and r is the radius from the hydraulic cylinder to the center of the shield. and These represent the lever arms of the hydraulic cylinder in the horizontal and vertical directions, respectively. In step (23), the loss function includes: ； In the formula, For the predicted hydraulic cylinder pressure, Where is the actual hydraulic cylinder pressure, i is the hydraulic cylinder serial number, j is the sample serial number, λ is the physical information proportionality coefficient (λ is a constant), M represents the sample size, n is the number of hydraulic cylinders, S is the cross-sectional area of ​​the hydraulic cylinder, and r is the radius from the hydraulic cylinder to the shield center. , , Let these represent the force, vertical bending moment, and horizontal bending moment transmitted from the intermediate shield to the j-th sample, respectively. , These represent the force, vertical moment, and horizontal moment transmitted by the intermediate shield, respectively. and Let represent the lever arm of the j-th sample hydraulic cylinder in the horizontal and vertical directions, respectively.

Images