A shield tunneling machine intelligent decision method and system

Abstract

This invention belongs to the field of intelligent tunnel boring machine (TBM) construction, and provides an intelligent decision-making method and system for TBMs. It acquires multi-source data and preprocesses each source. A pre-trained spatiotemporal graph convolutional network (SPCB) is used to extract features from the preprocessed multi-source data, obtaining spatiotemporal feature vectors. In the spatial domain, the SPCB aggregates neighborhood features of nodes and dynamically adjusts node interaction weights through an attention mechanism. In the temporal domain, causal convolution with gating mechanisms captures long-period temporal dependencies, and activation functions are used to achieve feature selection and noise suppression. Based on the spatiotemporal feature vectors, a multi-agent reinforcement learning algorithm is used for intelligent decision-making to obtain the equipment parameters for the TBM in the next time period. During the intelligent decision-making process, multi-scale boundary conditions directly constrain the action space of the multi-agents. This invention enables real-time optimization of TBM construction parameters and intelligent risk decision-making.

Description

Technical Field

[0001] This invention belongs to the field of intelligent tunnel boring machine construction, specifically relating to an intelligent decision-making method and system for tunnel boring machines, and more particularly to an intelligent decision-making method and system for tunnel boring machines based on the fusion of a spatio-temporal graph convolutional network (ST-GCN) and multi-scale physical constraints. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Tunnel boring machines (TBMs) are widely used in subway, highway tunnels, water conservancy projects and other fields. Their construction process faces challenges such as complex geological conditions, strong equipment coupling, and difficulty in accurately controlling dynamic parameters.

[0004] Traditional tunnel boring machine operation relies on human experience, which can lead to problems such as delayed decision-making and rough parameter adjustments, and may result in risks such as ground subsidence, cutter wear, and low tunneling efficiency.

[0005] Existing methods typically utilize data-driven intelligent decision-making techniques (such as neural networks and support vector machines) to predict parameters, but these approaches have the following drawbacks: (1) The sensor data of the tunnel boring machine (such as earth pressure, torque and propulsion speed) has spatiotemporal coupling characteristics. Traditional intelligent decision-making methods cannot adaptively capture spatial relationships, cannot avoid the assumption bias of static diagrams, and cannot jointly optimize time and space characteristics and explore deep coupling laws, resulting in a lack of spatiotemporal correlation.

[0006] (2) Tunnel boring machine construction involves multiple scale physical processes such as macroscopic geological environment, mesoscopic mechanical system, and microscopic cutter-soil interaction. Traditional methods have not established a multi-scale linkage mechanism for macroscopic geological environment (kilometer level), mesoscopic mechanical system (meter level), and microscopic cutter-soil interaction (millimeter level). Moreover, a single model is difficult to optimize globally and difficult to integrate multi-source heterogeneous data. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes an intelligent decision-making method and system for tunnel boring machines (TBMs). By integrating the dynamic feature extraction capabilities of spatiotemporal graph convolutional networks with the prior knowledge constraints of multi-scale physical models, this invention enables real-time optimization of TBM construction parameters and intelligent risk decision-making.

[0008] According to some embodiments, the present invention adopts the following technical solution: A method for intelligent decision-making in tunnel boring machines includes the following steps: Acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. Preprocess the multi-source data separately; A pre-trained spatiotemporal graph convolutional network is used to extract features from preprocessed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, causal convolution with a gating mechanism is used to capture long-term temporal dependencies, and feature selection and noise suppression are achieved by combining activation functions. Based on the spatiotemporal feature vector, a multi-agent reinforcement learning algorithm is used to make intelligent decisions and obtain the equipment parameters of the tunnel boring machine for the next time period. In the process of intelligent decision-making, multi-scale boundary conditions directly restrict the action space of the multi-agent, including geological settlement constraints, hydraulic equation constraints and cutter wear constraints.

[0009] As an alternative implementation, the geological parameters ahead include the compressive strength of the soil and rock ahead of the tunnel, the pore water pressure, and the degree of fracture development. The equipment parameters include propulsion force, cutter head torque, cutter head rotation speed, and propulsion speed; The groundwater parameters include real-time groundwater flow rate, groundwater velocity, and groundwater flow direction; The attitude parameters of the tunnel boring machine include the real-time pitch angle, yaw angle and roll angle of the tunnel boring machine; The environmental parameters include surface subsidence, ambient humidity, and construction chamber temperature.

[0010] As an alternative implementation method, the process of preprocessing multi-source data separately includes: The preprocessing of geological parameters involves Z-score standardization of the acquired geological parameters. High-frequency noise was removed from the equipment parameters using a sliding time window filtering algorithm. The Kalman filter algorithm was used to reduce noise in the groundwater parameters, and the flow velocity and direction were converted to polar coordinates. The attitude parameters of the tunnel boring machine are calculated using quaternions. A complementary filtering algorithm based on gradient descent is used to fuse gyroscope and accelerometer data to update the quaternion attitude. Then, Euler angle transformation is used to convert the quaternions into Euler angles. Accumulated errors are eliminated by zero-speed correction at set intervals. After processing the environmental parameters using a spatial interpolation algorithm, Gaussian smoothing is applied.

[0011] As an alternative implementation, the spatiotemporal graph convolutional network includes an input layer, a graph convolutional layer, and a temporal convolutional layer. The input layer is used to receive preprocessed multi-source data and convert it into a node feature matrix. The graph convolutional layer is used to capture the spatial correlation between data, and the temporal convolutional layer is used to capture the trend of feature changes over time.

[0012] As a further implementation, the graph convolutional layer employs a Chebyshev polynomial approximation graph filter to extract the k-hop neighborhood features of nodes, and introduces a multi-head attention mechanism to dynamically adjust the importance weights of different physically associated edges to obtain a spatial graph matrix.

[0013] As a further implementation, the temporal convolutional layer achieves feature selection and noise suppression through gated linear units, and uses an activation function to perform nonlinear activation on the gated path results to obtain a temporal graph matrix.

[0014] As an alternative implementation, the construction process of the spatiotemporal graph convolutional network includes: defining an attribute vector containing geological nodes, including each set unit of soil and rock elements, based on geological parameters; Define device nodes based on device parameters; Define environment nodes based on environmental parameters; The physical interaction strength between the cutterhead and the soil element is calculated based on Hertzian contact theory, and physical association edges are constructed. The dynamic time warping algorithm is used to quantify the correlation of parameter sequences and generate spatiotemporal proximity edges. The spatial distribution of geological nodes is updated every set time interval, and the connection relationship of equipment nodes is dynamically adjusted according to the real-time construction progress.

[0015] As an alternative implementation, the multi-scale boundary conditions are three-level constraints, where the macro-scale constraint is geological subsidence. Based on the model-predicted geometric parameters, baseline geometric parameters, and weighting coefficients, a surface subsidence physical loss function is constructed. During the training phase, the predicted subsidence value is compared with the measured value. When the deviation is greater than a set value, a backpropagation loss operation is performed. At the mesoscale, hydraulic equations are used as constraints. For shield grouting and hydraulic systems, the relationship between grouting pressure and flow rate and mortar mix ratio is established based on Bernoulli's equation or generalized fluid dynamics equations as additional constraints. In the decision-making and reasoning stage, the predicted propulsion speed is verified. When the model's predicted value deviates from the calculated value of the equation by more than the set value, physical correction is triggered to perform forced alignment. At the microscale, there is a constraint on tool wear. Based on the correlation model between the cutter head wear rate and the tool material and soil, the functional relationship between tool wear and excavation time is determined and embedded into the training process. In the data preprocessing stage, the predicted wear rate is corrected by using the measured rock compressive strength.

[0016] As an alternative implementation method, the process of making intelligent decisions using multi-agent reinforcement learning algorithms includes: using multiple agents, each agent independently learning to control one or more parameters of the tunnel boring machine, each agent continuously interacting with the virtual simulation environment through reinforcement learning algorithms, and optimizing the overall construction parameters according to the reward and punishment function, so that the tunnel boring machine construction efficiency and construction safety reach the optimal balance.

[0017] A tunnel boring machine intelligent decision-making system includes: The data acquisition module is configured to acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. The data preprocessing module is configured to preprocess multi-source data separately. The spatiotemporal graph construction module is configured to use a pre-trained spatiotemporal graph convolutional network to extract features from pre-processed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, it uses causal convolution with a gating mechanism to capture long-term temporal dependencies and combines activation functions to achieve feature selection and noise suppression. The intelligent decision-making module is configured to make intelligent decisions based on the spatiotemporal feature vector using a multi-agent reinforcement learning algorithm to obtain the equipment parameters of the tunnel boring machine for the next time period. The multi-scale physical constraint module is configured to directly restrict the action space of multiple agents in the intelligent decision-making process using multi-scale boundary conditions, including geological subsidence constraints, hydraulic equation constraints, and tool wear constraints.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts the spatiotemporal characteristics of the tunnel boring machine and the geological environment through a spatiotemporal graph convolutional network, and integrates multi-scale physical constraints (such as surface settlement loss function, hydraulic dynamics equation, and energy conservation correction model) of macroscopic geomechanics, mesoscopic equipment dynamics, and microscopic cutter-soil interaction. Combined with the reward mechanism of multi-agent reinforcement learning, it realizes dynamic optimization of construction parameters and risk classification and early warning, solves the problem of traditional models deviating from physical laws, and significantly improves construction efficiency and safety.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a flowchart of a method according to one embodiment; Figure 2 This is a schematic diagram of the structure and process of a spatiotemporal graph convolutional neural network according to one embodiment; Figure 3 This is a schematic diagram illustrating the construction of a spacetime graph in one embodiment. Figure 4 This is a flowchart of a multi-scale constraint optimization method for shield tunneling construction, as shown in one embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0026] Example 1 A method for intelligent decision-making in tunnel boring machines, such as Figure 1 As shown, it includes the following steps: Acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. Preprocess the multi-source data separately; A pre-trained spatiotemporal graph convolutional network is used to extract features from preprocessed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, causal convolution with a gating mechanism is used to capture long-term temporal dependencies, and feature selection and noise suppression are achieved by combining activation functions. Based on the spatiotemporal feature vector, a multi-agent reinforcement learning algorithm is used to make intelligent decisions and obtain the equipment parameters of the tunnel boring machine for the next time period. In the process of intelligent decision-making, multi-scale boundary conditions directly restrict the action space of the multi-agent, including geological settlement constraints, hydraulic equation constraints and cutter wear constraints.

[0027] In this embodiment, the geological parameters ahead include the compressive strength of the soil and rock ahead of the tunnel, the pore water pressure, and the degree of fracture development. The equipment parameters include propulsion force, cutter head torque, cutter head rotation speed, and propulsion speed; The groundwater parameters include real-time groundwater flow rate, groundwater velocity, and groundwater flow direction; The attitude parameters of the tunnel boring machine include the real-time pitch angle, yaw angle and roll angle of the tunnel boring machine; The environmental parameters include surface subsidence, ambient humidity, and construction chamber temperature.

[0028] In this embodiment, the process of preprocessing multi-source data includes: The preprocessing of geological parameters involves Z-score standardization of the acquired geological parameters. High-frequency noise was removed from the equipment parameters using a sliding time window filtering algorithm. The Kalman filter algorithm was used to reduce noise in the groundwater parameters, and the flow velocity and direction were converted to polar coordinates. The attitude parameters of the tunnel boring machine are calculated using quaternions. A complementary filtering algorithm based on gradient descent is used to fuse gyroscope and accelerometer data to update the quaternion attitude. Then, Euler angle transformation is used to convert the quaternions into Euler angles. Accumulated errors are eliminated by zero-speed correction at set intervals. After processing the environmental parameters using a spatial interpolation algorithm, Gaussian smoothing is applied.

[0029] In this embodiment, as Figure 2 As shown, the spatiotemporal graph convolutional network includes an input layer, a graph convolutional layer, and a temporal convolutional layer. The input layer is used to receive preprocessed multi-source data and convert it into a node feature matrix. The graph convolutional layer is used to capture the spatial correlation between data. The temporal convolutional layer is used to capture the trend of feature changes over time.

[0030] The graph convolutional network (GCN, i.e., the spatial convolutional layer in the graph) uses a Chebyshev multinomial approximation graph filter to extract the k-hop neighborhood features of nodes and introduces a multi-head attention mechanism to dynamically adjust the importance weights of different physically associated edges to obtain a spatial graph matrix.

[0031] The temporal convolutional layer (i.e., the temporal convolutional network in the figure) achieves feature selection and noise suppression through gated linear units, and uses an activation function to perform nonlinear activation on the gated path results to obtain the temporal map matrix.

[0032] There are multiple temporal convolutional layers, which are interconnected by residuals. The units formed by the graph convolutional layers and the temporal convolutional layers (referred to as STGCN in this embodiment) include several units (two in this embodiment). Features are passed between units. The residuals of the graph convolutional layers of each unit are connected to a Long Short-Term Memory (LSTM) network. The last unit passes the features to the LSTM network, and the LSTM network outputs the final temporal features to obtain the prediction result.

[0033] Each unit includes two temporal convolutional layers. The first temporal convolutional layer passes the extracted local temporal features to the spatial convolutional layer, and the spatial convolutional layer passes the aggregated neighbor spatial information to the second temporal convolutional layer. The second temporal convolutional layer and other units form a feature transfer mechanism.

[0034] In this embodiment, as Figure 3 As shown, the construction process of the spatiotemporal graph convolutional network includes: at the geological layer, based on geological parameters, defining geological nodes, including each set unit (e.g., 1m). 3 The attribute vectors of geological units include compressive strength, pore water pressure, and fracture index. The correlation of parameter sequences between units is quantified using a dynamic time warping algorithm. In practice, the address node space distribution is updated every set time interval (e.g., 5 seconds), and the device node connection relationship is dynamically adjusted according to the real-time construction progress.

[0035] At the equipment layer, based on equipment parameters, equipment nodes are defined, including obtaining the state vector of the cutterhead / propulsion system, including propulsion force, propulsion speed, cutterhead matrix, and cutterhead rotation speed, etc. At the environmental layer, environmental nodes are defined based on environmental parameters, including the acquisition of sedimentation, temperature, and humidity. The physical interaction strength between the cutterhead and the soil element is calculated based on Hertzian contact theory, and physical association edges are constructed. The dynamic time warping algorithm is used to quantify the correlation of parameter sequences and generate spatiotemporal proximity edges. The spatial distribution of geological nodes is updated every set time interval, and the connection relationship of equipment nodes is dynamically adjusted according to the real-time construction progress. For example, settlement correlation is established between geological units and environmental layers, and Hertzian contact is formed between geological units and equipment nodes. There is mechanical vibration connection between the propulsion system and the equipment layer.

[0036] In this embodiment, as Figure 4As shown, the multi-scale boundary conditions are three-level constraints. At the macro scale, there is a geological subsidence constraint. Based on the model-predicted geometric parameters, the baseline geometric parameters, and the weight coefficients, a surface subsidence physical loss function is constructed. During the training phase, the predicted subsidence value is compared with the measured value. When the deviation is greater than the set value, a backpropagation loss operation is performed. At the mesoscale, hydraulic equations are used as constraints. For shield grouting and hydraulic systems, the relationship between grouting pressure and flow rate and mortar mix ratio is established based on Bernoulli's equation or generalized fluid dynamics equations as additional constraints. In the decision-making and reasoning stage, the predicted propulsion speed is verified. When the model's predicted value deviates from the calculated value of the equation by more than the set value, physical correction is triggered to perform forced alignment. At the microscale, there is a constraint on tool wear. Based on the correlation model between the cutter head wear rate and the tool material and soil, the functional relationship between tool wear and excavation time is determined and embedded into the training process. In the data preprocessing stage, the predicted wear rate is corrected by using the measured rock compressive strength.

[0037] In this embodiment, the process of making intelligent decisions using a multi-agent reinforcement learning algorithm includes: using multiple agents, each agent independently learning to control one or more parameters of the tunnel boring machine, and each agent continuously interacting with the virtual simulation environment through the reinforcement learning algorithm, optimizing the overall construction parameters according to the reward and punishment function, so that the tunnel boring machine construction efficiency and construction safety reach the optimal balance.

[0038] Example 2 A tunnel boring machine intelligent decision-making system includes: The data acquisition module is configured to acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. The data preprocessing module is configured to preprocess multi-source data separately. The spatiotemporal graph construction module is configured to use a pre-trained spatiotemporal graph convolutional network to extract features from pre-processed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, it uses causal convolution with a gating mechanism to capture long-term temporal dependencies and combines activation functions to achieve feature selection and noise suppression. The intelligent decision-making module is configured to make intelligent decisions based on the spatiotemporal feature vector using a multi-agent reinforcement learning algorithm to obtain the equipment parameters of the tunnel boring machine for the next time period. The multi-scale physical constraint module is configured to directly restrict the action space of multiple agents in the intelligent decision-making process using multi-scale boundary conditions, including geological subsidence constraints, hydraulic equation constraints, and tool wear constraints.

[0039] The following is an introduction to each module.

[0040] The first module is the data acquisition module, which uses the following sensors and algorithms to collect and preprocess data.

[0041] (1) Geological parameters: Data Acquisition Method: In this embodiment, a ground-penetrating radar array is used to scan the rock and soil ahead of the tunneling machine in real time at a resolution of 1m. The data is discretized into a 1m³ cubic unit grid and collected in real time from multiple measurement units during the current construction process of the tunnel boring machine, including parameters such as the compressive strength of the rock and soil ahead of the tunneling machine, pore water pressure, and fracture development degree.

[0042] Preprocessing method: Perform Z-score normalization on the grid data.

[0043] (2) Equipment parameters: Data acquisition method: Equipment operating data parameters such as thrust, cutterhead torque, cutterhead rotation speed and thrust speed are acquired through the tunnel boring machine's PLC main control system and distributed sensors; Preprocessing method: Use sliding time window filtering (window length 5 seconds, Hanning window function) to eliminate high-frequency noise.

[0044] (2) Groundwater parameters: Data acquisition method: Real-time groundwater flow rate, groundwater velocity, groundwater flow direction and other dynamic parameters are acquired through a distributed hydrological sensor array (sampling rate 10Hz).

[0045] Preprocessing method: Kalman filtering is used for noise reduction. Groundwater sensors (velocity and direction) on the tunnel boring machine are susceptible to mechanical vibration noise and electromagnetic interference in complex construction environments. Optimal estimation of the dynamic system is achieved through a state-space model. The state at the next moment is predicted based on the system dynamics model, and the prediction results are corrected by fusing sensor observations to reduce uncertainty. After Kalman filtering for noise reduction, the velocity and direction are converted to polar coordinates.

[0046] (3) Tunnel boring machine attitude parameters: Data Acquisition Method: An attitude sensing unit, consisting of a high-precision fiber optic gyroscope (FOG) and a three-axis MEMS accelerometer, is installed on the central axis of the tunnel boring machine (TBM) to acquire real-time attitude deviation parameters such as pitch angle, yaw angle, and roll angle. Preprocessing method: First, quaternion calculation is performed, and a complementary filtering algorithm based on gradient descent is used to fuse gyroscope and accelerometer data to update the quaternion attitude. Then, Euler angle transformation is used to convert the quaternion into Euler angles (pitch θ, yaw ψ, roll φ). At the same time, zero-velocity update (ZUPT) is used every 30 minutes to eliminate accumulated errors. Calibration is triggered when the tunnel boring machine stops advancing to ensure the long-term stability of the attitude angles.

[0047] (4) Environmental parameters: Data Acquisition Method: In this embodiment, parameters such as surface subsidence, ambient humidity, and construction chamber temperature are obtained using InSAR satellite data and a hydrostatic level.

[0048] Preprocessing method: Spatial interpolation is used. Due to the differences in spatial resolution of multi-source data in surface subsidence monitoring (InSAR satellite data has wide coverage but low accuracy, while hydrostatic level data has high accuracy but sparse distribution), it is necessary to generate a high-precision continuous subsidence field through interpolation algorithms.

[0049] The preprocessing steps in this embodiment are as follows: InSAR settlement data and hydrostatic level data are unified to the same coordinate system; relative-absolute correction is performed on the InSAR data to eliminate atmospheric delay errors; the hydrostatic level data is directly used as the reference true value. A spatial variogram model is constructed based on the hydrostatic level data using the Ordinary Kriging algorithm. Interpolation parameters are optimized using Leave-One-Out Cross-Validation (LOOCV), and finally, Kriging interpolation is performed to generate a 1m resolution settlement field.

[0050] Then use Gaussian smoothing (to suppress noise and outliers).

[0051] In this embodiment, a discretized Gaussian kernel matrix is ​​defined. The Gaussian kernel is convolved with the sedimentation field matrix Z. Threshold filtering is performed on the smoothed data, and outliers are identified and replaced with the neighborhood median.

[0052] Next, we will introduce the spacetime graph construction module.

[0053] Definition of Spatiotemporal Graph Convolutional Network (ST-GCN): Spatiotemporal Graph Convolutional Network (ST-GCN) is a deep learning architecture that integrates Graph Convolutional Networks (GCN) and Temporal Convolutional Networks (TCN). It is specifically designed for processing dynamic data with spatial topological relationships and temporal evolution, and is also a deep learning framework for safety prediction and equipment status monitoring in tunnel boring machine (TBM) construction. This framework combines the advantages of GCN and TCN, effectively capturing the spatiotemporal correlations within the tunnel structure and construction system, thereby improving the prediction accuracy for key risks such as ground deformation and equipment failure. The core idea of ​​ST-GCN is to view the tunnel construction system as a graph, where each node represents a monitoring point or key equipment component, and edges represent the physical connections or mechanical influence relationships between monitoring points. In ST-GCN, node features represent settlement, stress, or equipment operating parameters of the monitoring point, while edge features represent the structural correlation or interaction strength between monitoring points. Through graph convolution operations, ST-GCN can learn the spatiotemporal coupling relationships at different locations within the construction system and use them to predict future ground deformation trends or equipment status.

[0054] The ST-GCN architecture mainly consists of three parts: an input layer, a graph convolutional layer, and a temporal convolutional layer.

[0055] (1) Input layer The input layer receives sensor data (such as thrust, cutterhead torque, cutterhead rotation speed and thrust speed) from various monitoring points along the tunnel and converts it into a node feature matrix.

[0056] (2) Convolutional layer Graph convolutional layers use Graphical Networks (GCNs) to capture the spatial correlations between monitoring points. In ST-GCNs, multiple graph convolutional layers can be stacked to increase the model's ability to represent complex structural relationships. Specifically, GCNs perform this through the following steps: (a) Represent the monitoring point features as a node feature matrix; (b) Perform graph convolution on the node feature matrix to obtain updated node features that incorporate spatial neighborhood information; (c) Repeat step (b) until the predetermined number of layers is reached or convergence is achieved.

[0057] (3) Temporal convolutional layer Temporal convolutional layers use TCN to capture the changing trends of monitoring point features over time. In ST-GCN, multiple temporal convolutional layers can be stacked to increase the model's ability to capture long-term dependencies. Specifically, TCN performs this through the following steps: (a) Represent the monitoring point features as a three-dimensional tensor, where the dimensions are time, monitoring point, and feature, respectively; (b) Perform temporal convolution on the three-dimensional tensor to extract the temporal evolution pattern in the monitoring point data and obtain the updated feature representation; (c) Repeat step (b) until the predetermined number of layers is reached or convergence is achieved.

[0058] Detailed Explanation of the Spacetime Graph Construction Module: The spatiotemporal graph construction module is a core component of the ST-GCN framework, responsible for abstracting entity relationships in dynamic systems into computable graph structures. This module employs a three-dimensional modeling approach of "entity-relationship-time" to transform spatiotemporal interactions in complex systems into machine-understandable topological representations.

[0059] The basic elements defined in this module are as follows: Entity Node: The basic building block of a system, representing an observable object with an independent state.

[0060] Interaction edge: The connection channel between entities, reflecting spatial topological constraints.

[0061] Temporal Feature: The dynamic attribute of an entity's state as it evolves over time.

[0062] This module consists of: (1) Construction of the topological structure of the static space: This is the basic skeleton of the graph, defining the invariant physical or logical connections between entities. First, define the entity node set (Node Set) V: abstracting all entities as a set of nodes. V = {v1, v2, ..., vN}, where N is the total number of nodes. For example, define three types of entity nodes (geology, equipment, environment). Then construct the interaction edges (adjacency matrix A): define the spatial relationships between nodes based on data-driven (calculating the correlation between data through the Pearson correlation coefficient). The output of this step is a static graph G = (V, A), which captures the spatial skeleton of the system. (2) Construction of the feature matrix for dynamic nodes: This part injects dynamic life into the static skeleton, reflecting the constantly changing state of each entity over time. Temporal convolution (TCN) is used to extract temporal features, and the feature matrix (FeatureMatrix) X is defined: This is an N × D × T tensor (3-dimensional array), where: N: Number of nodes (consistent with the dimension of the adjacency matrix).

[0063] D: The feature dimension of each node at each time step. For example, D may include tunneling speed, timestamp, cutterhead rotation speed, etc.

[0064] T: The length of the time step, i.e. the size of the historical time window we observe.

[0065] The output of this step is a dynamic feature sequence that assigns a state value to each node at each time point.

[0066] (3) Integration and representation of spatiotemporal graph: Finally, the static topology and dynamic features are integrated into a complete spatiotemporal graph, which is the standard input for the ST-GCN model.

[0067] This module designs a multi-layered spatiotemporal feature extraction architecture, specifically including: Graph convolutional layer: A Chebyshev polynomial approximation graph filter (order K=3) is used to extract k-hop neighborhood features of nodes. The formula is as follows: ; The meanings of the parameters are shown in Table 1.

[0068] Table 1. Parameter Meaning Table

[0069] This embodiment also introduces a multi-head attention mechanism to dynamically adjust the importance weights of different physically related edges. The formula for calculating the attention weights is as follows: ; The parameters are shown in Table 2.

[0070] Table 2 Parameter Table

[0071] The final result is the spatial graph matrix.

[0072] Temporal convolutional layers: feature selection and noise suppression are achieved through gated linear units (GLUs). The gating coefficients are calculated using the following formula: ; The parameters are shown in Table 3.

[0073] Table 3 Parameter Table

[0074] The final time-map matrix is ​​obtained.

[0075] The construction process of the spatiotemporal topology includes three types of nodes (geology, equipment, and environment) and two types of edges (physically related edges and spatiotemporally adjacent edges), and the topology evolves through a dynamic update mechanism. The weights of the edges between nodes can be determined by the physical-mechanical correlation based on the Hertzian contact model and the temporal correlation obtained through Granger causality analysis.

[0076] Furthermore, the spatial convolution of the spatiotemporal graph convolutional network is approximated using a K-order Chebyshev polynomial; the temporal convolution employs a gated causal one-dimensional convolution kernel of length T; and the multi-head attention mechanism has N heads. The dynamic spatiotemporal topology graph is updated once every time a predetermined distance is advanced or every fixed time interval is elapsed, and the topology graph is reconstructed based on the latest acquired data each time it is updated.

[0077] Specifically, this embodiment includes the following steps: (1) Define the attribute vector (compressive strength, pore water pressure, fracture index) of each 1m³ soil and rock unit containing geological nodes.

[0078] (2) Define the equipment nodes (state vectors of the cutterhead / propulsion system, propulsion force, propulsion speed, cutterhead torque, and cutterhead rotation speed).

[0079] (3) Definition and environmental nodes (settlement, temperature and humidity of surface monitoring points).

[0080] (4) Calculate the physical interaction strength between the cutterhead and the soil unit based on Hertz contact theory, and construct the physical association edge.

[0081] (5) The correlation of parameter sequences is quantified by dynamic time warping algorithm to generate spatiotemporal proximity edges.

[0082] (6) Update the spatial distribution of geological nodes every 5 seconds and dynamically adjust the connection relationship of equipment nodes according to the real-time construction progress.

[0083] The above process requires attention to the construction of physical correlation edges: the cutting force transmission relationship between the cutterhead and the geological unit (modeled based on Hertz contact theory), and the dynamic coupling relationship between the hydraulic system pressure and the propulsion cylinder; Spatiotemporal proximity edge construction: three-dimensional spatial adjacency matrix between geological units (generated by Delaunay triangulation), temporal sliding window correlation of equipment nodes (connection is established when the time delay mutual information coefficient is >0.8).

[0084] Dynamic graph update mechanism: The spatial distribution of geological nodes is updated every 5 seconds based on the tunneling location; the edge weight matrix is ​​calculated in real time (based on the interaction strength between nodes and the causal Granger test).

[0085] The multi-scale physics constraint module is used to achieve deep coupling between physical priors and data-driven approaches, incorporating various physical model constraints, including three levels of constraints: Macroscopic geological constraints (i.e., geological subsidence constraints): Mathematical expressions for short-term surface loss and long-term consolidation settlement are derived based on shield tunneling theory and empirical formulas. The predicted results are compared with actual monitored settlement and used for model correction. Construct the physical loss function for surface subsidence: ; The parameters are shown in Table 4.

[0086] Table 4 Parameter Table

[0087] Predicting settlement values ​​during the training phase Compared with measured values In contrast, when the deviation is greater than 5 mm, a backpropagation loss operation is performed.

[0088] Mesoscopic equipment constraints (i.e., hydraulic equation constraints): For shield tunneling grouting and hydraulic systems, the relationship between grouting pressure and flow rate and mortar mix ratio is established based on Bernoulli's equation or generalized fluid dynamics equations, and these are used as additional constraints. Hydraulic system dynamic equation constraints: ; The parameters are shown in Table 5.

[0089] Table 5 Parameter Table

[0090] During the decision-making and reasoning phase, the predicted propulsion speed is verified. When the model's predicted value deviates from the calculated value by more than 10%, the physics corrector is triggered to force alignment.

[0091] Microscopic tool constraints (i.e., tool wear constraints): Based on the correlation model between cutter head wear rate and cutter material and soil quality, the functional relationship between cutter wear and excavation time is derived and embedded into the training process; Wear rate correction model based on cutting energy conservation: ; The parameters are shown in Table 6.

[0092] Table 6 Parameter Table

[0093] In the data preprocessing stage, the measured rock compressive strength was used. Corrected predicted wear rate .

[0094] At the macro level, the predicted settlement value is compared with the calculation result of the Mohr-Coulomb model, and the geological loss function is triggered when the deviation exceeds 5 mm. At the meso level, the predicted value of the propulsion speed is verified based on the dynamic equation of the hydraulic system, and the result exceeding the theoretical value by ±10% is forcibly aligned. At the micro level, the predicted tool wear rate is corrected by the measured value of rock compressive strength according to the principle of energy conservation, so as to ensure that the multi-scale prediction results conform to physical laws.

[0095] The dynamic decision-making module includes multiple agents, each of which independently learns to control one or more parameters of the tunnel boring machine (such as the attitude correction agent adjusting the shield rotation, the propulsion parameter agent adjusting the propulsion force and propulsion speed, and the slurry control agent adjusting the slurry injection volume). Each agent continuously interacts with the virtual simulation environment through reinforcement learning algorithms (such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO), and optimizes the overall construction parameters according to the reward and penalty function, so that the tunnel boring machine construction efficiency and construction safety reach the optimal balance. The multi-agent reinforcement learning employs a decentralized Actor-Critic algorithm, where each agent has an independent policy network and value network, and the policy is updated in a coordinated manner through the multi-objective reward function.

[0096] This module implements multi-objective parameter optimization and risk control, and its core technologies include: (1) Multi-agent reinforcement learning framework: Define 3 intelligent agents: 1) The propulsion parameter optimization agent takes ST-GCN equipment node characteristics as input and outputs propulsion speed and hydraulic pressure as outputs, with hydraulic equations as constraints. 2) Tool turret control agent: The input source is the tool wear correction value, and the output action space is: speed and torque threshold. The constraint condition is the wear rate safety threshold.

[0097] 3) Soil pressure balance Agent, the input source is the pore water pressure and fracture data of geological nodes, the output action space is: screw conveyor speed and foam injection volume, and the constraint condition is the surface settlement loss.

[0098] The reward function is set as (fusion efficiency, security, and energy consumption): ; The parameters are shown in Table 7.

[0099] Table 7 Parameter Meaning Table Symbols / Items meaning Mathematical definition R Outcome metrics (such as overall performance score) A comprehensive value determined by efficiency, safety, and energy. α,β,γ Weighting coefficients (α>0, β>0, γ>0) These represent the relative importance of efficiency, safety, and energy, respectively. Efficiency tunneling rate index The value range is usually [0,1] or a standardized value. Safety Safety factor (combined score of settlement and tool wear) High security corresponds to high numerical values ​​(e.g., 0.9 indicates a 90% security probability). Energy Energy consumption factors (such as energy intensity and carbon emissions). The larger the value, the more significant the negative impact (such as energy consumption per unit of output). This involves constructing a multi-agent reinforcement learning environment that includes a propulsion agent, a cutterhead agent, and an earth pressure agent, and designing a multi-objective reward function that includes construction efficiency, safety, and energy consumption. The reward function is then used to train each agent through reinforcement learning to obtain an optimized control strategy.

[0100] Specifically, three intelligent agents are constructed: propulsion parameters, cutterhead control, and earth pressure balance. These agents optimize the propulsion speed-hydraulic pressure combination, cutterhead speed-torque threshold, and screw conveyor speed-foam injection volume parameters, respectively. A weighted reward function integrating efficiency, safety, and energy consumption is designed, where the efficiency index is correlated with the matching degree between tunneling rate and planned progress, and the safety index integrates surface settlement and cutter wear rate scores to form a multi-objective optimization decision-making scheme.

[0101] Some embodiments also include a digital twin and risk warning module, the functions of which are: Implement risk warning and early warning strategies: Pre-simulate decision-making schemes in a virtual environment using a digital twin system, and initiate secondary optimization for abnormal operating conditions that may cause equipment overload. Establish a three-level response mechanism based on feature similarity: When the similarity between the ST-GCN output feature vector and the historical abnormal operating condition database is less than 0.6, trigger an interface warning; when the similarity is less than 0.4, automatically reduce speed by 20%; when the similarity is less than 0.2, execute hydraulic power cutoff and emergency support with support shoes.

[0102] Incremental training mechanism: The spatiotemporal graph convolutional network is incrementally trained weekly based on the latest 30% of the data, dynamically updating node attributes and edge weights to improve prediction accuracy. A learning rate decay strategy is adopted to keep the model prediction accuracy decay rate less than 1%, thus realizing the system's self-evolution capability.

[0103] Digital twin verification: The decision-making scheme is first simulated in the digital twin system. The simulation results are fed back to the multi-scale physical constraint module to calibrate the parameters, and then drive the actual equipment to execute. That is, the actual construction data is transmitted back to the spatiotemporal graph construction module in real time, and the node attributes and edge weight matrix are dynamically updated.

[0104] The synergistic effect of the intelligent decision-making module with other modules: (1) Connection with the spatiotemporal graph construction module Data Acquisition: The dynamic decision module receives the spatiotemporal feature vectors extracted by the spatiotemporal graph convolutional network (ST-GCN) (such as the node feature matrix and attention weight matrix output by the convolutional layer, as shown in the figure).

[0105] Function linkage: The feature vectors output by ST-GCN (such as equipment node status and geological unit anomaly indicators) serve as the state input for multi-agent reinforcement learning, providing decision-making reference for the agents. The decision results (such as adjusted propulsion speed and cutterhead rotation speed) are updated with the equipment node attributes of the spatiotemporal map through a data feedback mechanism, achieving closed-loop optimization.

[0106] (2) Coupling with multi-scale physical constraint modules Constraint embedding: The multi-scale boundary conditions generated by the physical constraint module directly limit the action space of the dynamic decision module; Macro-geological constraints: The predicted value of surface subsidence is incorporated into the reward function (i.e., the Safety term) as a safety indicator. When the deviation is >5mm, backpropagation correction decision is triggered.

[0107] Mesoscopic equipment constraints: The hydraulic equation verifies the predicted propulsion speed, and forced alignment is applied when the deviation is >10% (e.g., limiting the range of oil pressure output by the agent).

[0108] Microscopic tool constraints: The wear rate correction model provides a tool life threshold, which affects the torque decision of the tool turret agent.

[0109] (3) Closed loop with multi-source data acquisition module Data input: Preprocessed data of environmental parameters (surface subsidence, temperature) and equipment parameters (propulsion force, rotational speed) are indirectly input into the decision-making module through the spatiotemporal map construction module.

[0110] Feedback control: The control parameters output by the intelligent agent (such as propulsion speed and foam injection volume) directly drive the shield machine PLC system. At the same time, newly collected data updates the spatiotemporal map nodes in real time, forming a closed loop of "decision-making → execution → monitoring → update".

[0111] In summary, the data acquisition module is used to collect multi-dimensional sensor data on the status of the tunnel boring machine, geological environment, and construction environment in real time; the spatiotemporal graph construction module is used to construct a dynamic spatiotemporal topology graph based on the real-time data collected by the data acquisition module. This topology graph includes geological nodes, equipment nodes, and environmental nodes. The weights of the edges between the nodes are jointly determined by the physical and mechanical correlation based on the Hertzian contact model and the temporal correlation obtained through Granger causality analysis; the spatiotemporal graph construction module is also used to input the dynamic spatiotemporal topology graph into a spatiotemporal graph convolutional network for modeling, and to extract node features using a multi-head attention mechanism and gated causal one-dimensional convolution.

[0112] The multi-scale physical constraint module is used to introduce multi-scale physical constraints during model training, including formation subsidence loss based on the Mohr-Coulomb criterion, tool wear loss based on the principle of energy conservation, and hydraulic system dynamic constraints.

[0113] The intelligent decision-making module is used to construct a multi-agent system consisting of a propulsion agent, a cutterhead agent, and an earth pressure agent, and to configure a multi-objective reward function for the multi-agent system that includes construction efficiency, safety, and energy consumption, so as to optimize the control parameters of the tunnel boring machine through reinforcement learning.

[0114] In some embodiments, a digital twin and risk warning module is also included. This module is used to establish a digital twin model of the shield tunneling construction process for simulation verification, and to trigger a three-level risk warning signal based on the simulation results and preset thresholds. Specifically, it includes a digital twin model for simulating the entire shield tunneling construction process, which outputs simulation data consistent with the actual construction process. The three-level risk warning mechanism includes three levels: low-risk warning, medium-risk warning, and high-risk alarm. When the key construction parameters predicted by the digital twin model exceed the preset threshold, the corresponding level of risk warning signal is triggered.

[0115] Example 3 Example of the application of the system in the tunnel boring machine construction of a subway in a certain city.

[0116] During the tunnel boring machine (TBM) construction of a section of Metro Line 13 in a certain city, a large slurry balance TBM with a diameter of 14 meters was used. The geological conditions in the construction area were complex, traversing multiple strata (including silty clay, silty sand, and some gravel layers). There was a significant risk of ground disturbance during construction, which placed high demands on the intelligent control system.

[0117] To improve tunneling safety and efficiency, the construction unit adopted the intelligent decision-making system for large-diameter tunnel boring machines based on spatiotemporal graph convolution and multi-scale physical constraints of this invention. The system includes the following modules: Data acquisition module: Multi-source sensors (such as cutterhead torque sensor, propulsion pressure sensor, and surface settlement monitoring instrument) installed on the tunnel boring machine collect construction parameter data in real time; Spatiotemporal graph structure construction module: Based on the structural and functional dependencies between different components of the tunnel boring machine, a dynamic graph model is constructed, where nodes represent different equipment units and edges represent physical coupling relationships; Spatiotemporal Graph Convolutional Neural Network Model (ST-GCN): Extracts structural features between devices through graph convolution, and combines the time series LSTM structure to extract temporal dynamic information, thereby enabling prediction of future tunneling status; Multi-scale physical constraint module: Introduces physical rules based on soil deformation theory and shield tunneling mechanics model, which are used as constraints in the neural network training and inference process to improve the physical consistency and generalization ability of the model; Intelligent decision-making module: Based on the above prediction results and constraint output, it automatically provides adjustment suggestions for key parameters such as propulsion speed, cutterhead speed, and earth pressure setting, and can be manually confirmed and executed in conjunction with the expert system.

[0118] During actual tunneling, the system was able to predict the trend of ground disturbance and equipment status fluctuations in advance, providing early warnings of abnormal situations 15-20 minutes in advance, thus avoiding two mudslide accidents and one equipment failure. Compared with the section without this system, construction efficiency increased by 12%, the frequency of incorrect setting of propulsion parameters decreased by 75%, and the maximum surface settlement was reduced by nearly 40%.

[0119] The system can also be applied to composite shield tunneling, slurry-water conversion areas, and high-risk strata (such as sand layers and fault fracture zones). By adjusting the diagram structure modeling method and physical constraint model, it can be widely applied to different types of shield tunneling equipment and geological environments.

[0120] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0121] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart decision-making method for tunnel boring machines, characterized in that, Includes the following steps: Acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. Preprocess the multi-source data separately; A pre-trained spatiotemporal graph convolutional network is used to extract features from preprocessed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, causal convolution with a gating mechanism is used to capture long-term temporal dependencies, and feature selection and noise suppression are achieved by combining activation functions. Based on the spatiotemporal feature vector, a multi-agent reinforcement learning algorithm is used to make intelligent decisions and obtain the equipment parameters of the tunnel boring machine for the next time period. In the process of intelligent decision-making, multi-scale boundary conditions directly restrict the action space of the multi-agent, including geological settlement constraints, hydraulic equation constraints and cutter wear constraints.

2. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The geological parameters ahead include the compressive strength of the soil and rock ahead of the tunnel, the pore water pressure, and the degree of fracture development. The equipment parameters include propulsion force, cutter head torque, cutter head rotation speed, and propulsion speed; The groundwater parameters include real-time groundwater flow rate, groundwater velocity, and groundwater flow direction; The attitude parameters of the tunnel boring machine include the real-time pitch angle, yaw angle and roll angle of the tunnel boring machine; The environmental parameters include surface subsidence, ambient humidity, and construction chamber temperature.

3. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The process of preprocessing multi-source data separately includes: The preprocessing of geological parameters involves Z-score standardization of the acquired geological parameters. High-frequency noise was removed from the equipment parameters using a sliding time window filtering algorithm. The Kalman filter algorithm was used to reduce noise in the groundwater parameters, and the flow velocity and direction were converted to polar coordinates. The attitude parameters of the tunnel boring machine are calculated using quaternions. A complementary filtering algorithm based on gradient descent is used to fuse gyroscope and accelerometer data to update the quaternion attitude. Then, Euler angle transformation is used to convert the quaternions into Euler angles. Accumulated errors are eliminated by zero-speed correction at set intervals. After processing the environmental parameters using a spatial interpolation algorithm, Gaussian smoothing is applied.

4. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The spatiotemporal graph convolutional network includes an input layer, a graph convolutional layer, and a temporal convolutional layer. The input layer is used to receive preprocessed multi-source data and convert it into a node feature matrix. The graph convolutional layer is used to capture the spatial correlation between data. The temporal convolutional layer is used to capture the trend of feature changes over time.

5. The intelligent decision-making method for tunnel boring machines as described in claim 4, characterized in that, The graph convolutional layer uses a Chebyshev multinomial approximation graph filter to extract the k-hop neighborhood features of nodes, and introduces a multi-head attention mechanism to dynamically adjust the importance weights of different physically associated edges to obtain a spatial graph matrix. The temporal convolutional layer achieves feature selection and noise suppression through gated linear units, and uses an activation function to perform nonlinear activation on the gated path results to obtain a temporal graph matrix.

6. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The construction process of the spatiotemporal graph convolutional network includes: defining an attribute vector containing geological nodes and each set unit of soil and rock elements based on geological parameters; Define device nodes based on device parameters; Define environment nodes based on environmental parameters; The physical interaction strength between the cutterhead and the soil element is calculated based on Hertzian contact theory, and physical association edges are constructed. The dynamic time warping algorithm is used to quantify the correlation of parameter sequences and generate spatiotemporal proximity edges. The spatial distribution of geological nodes is updated every set time interval, and the connection relationship of equipment nodes is dynamically adjusted according to the real-time construction progress.

7. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The multi-scale boundary conditions are three levels of constraints. At the macro scale, there is a geological subsidence constraint. Based on the model-predicted geometric parameters, the baseline geometric parameters, and the weighting coefficients, a physical loss function for surface subsidence is constructed. During the training phase, the predicted subsidence value is compared with the measured value. When the deviation is greater than the set value, a backpropagation loss operation is performed. At the mesoscale, hydraulic equations are used as constraints. For shield grouting and hydraulic systems, the relationship between grouting pressure and flow rate and mortar mix ratio is established based on Bernoulli's equation or generalized fluid dynamics equations as additional constraints. In the decision-making and reasoning stage, the predicted propulsion speed is verified. When the model's predicted value deviates from the calculated value of the equation by more than the set value, physical correction is triggered to perform forced alignment. At the microscale, there is a constraint on tool wear. Based on the correlation model between the cutter head wear rate and the tool material and soil, the functional relationship between tool wear and excavation time is determined and embedded into the training process. In the data preprocessing stage, the predicted wear rate is corrected by using the measured rock compressive strength.

8. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, The process of making intelligent decisions using multi-agent reinforcement learning algorithms includes: using multiple agents, each of which independently learns to control one or more parameters of the tunnel boring machine. Each agent continuously interacts with the virtual simulation environment through reinforcement learning algorithms, and optimizes the overall construction parameters according to the reward and penalty function, so that the tunnel boring machine construction efficiency and construction safety reach the optimal balance.

9. The intelligent decision-making method for tunnel boring machines as described in claim 1, characterized in that, It also includes the following steps: The decision-making scheme is simulated in the digital twin system. The simulation results are fed back to the multi-scale physical constraint module to calibrate the parameters, and then drive the actual equipment to execute. The actual construction data is transmitted back to the spatiotemporal graph construction module in real time, and the node attributes and edge weight matrix are dynamically updated.

10. A tunnel boring machine intelligent decision-making system, characterized in that, include: The data acquisition module is configured to acquire multi-source data, including geological parameters, equipment parameters, groundwater parameters, tunnel boring machine attitude parameters, and environmental parameters. The data preprocessing module is configured to preprocess multi-source data separately. The spatiotemporal graph construction module is configured to use a pre-trained spatiotemporal graph convolutional network to extract features from pre-processed multi-source data to obtain spatiotemporal feature vectors. In the spatial domain, the spatiotemporal graph convolutional network aggregates the neighborhood features of nodes and dynamically adjusts the interaction weights of nodes through an attention mechanism. In the temporal domain, it uses causal convolution with a gating mechanism to capture long-term temporal dependencies and combines activation functions to achieve feature selection and noise suppression. The intelligent decision-making module is configured to make intelligent decisions based on the spatiotemporal feature vector using a multi-agent reinforcement learning algorithm to obtain the equipment parameters of the tunnel boring machine for the next time period. The multi-scale physical constraint module is configured to directly restrict the action space of multiple agents in the intelligent decision-making process using multi-scale boundary conditions, including geological subsidence constraints, hydraulic equation constraints, and tool wear constraints.

Images