Shield intelligent tunneling control method and system based on multi-source perception data fusion

Abstract

The application provides a shield intelligent tunneling control method and system based on multi-source perception data fusion, and relates to the technical field of underground tunnel engineering and automatic construction. The method comprises the following steps: acquiring multi-source heterogeneous perception data and preprocessing; combining a shield construction field knowledge graph to perform multi-level data fusion and generate a fusion feature vector; constructing a digital twin, applying hierarchical reinforcement learning and model predictive control algorithms to generate optimal tunneling control instructions; issuing the instructions to each subsystem for execution and performing safety interlocking monitoring; and updating and optimizing the model according to the deviation between the actual tunneling result and the predicted result. The application fuses multi-source heterogeneous perception data of a shield tunneling machine and external environment data, constructs a digital twin driven artificial intelligence decision model, generates and executes a globally optimal tunneling control instruction in real time, and thus realizes safe, efficient and accurate unmanned or few-people intelligent tunneling.

Description

Technical Field

[0001] This application relates to the fields of underground tunnel engineering and automated construction technology, and in particular to a shield tunneling intelligent tunneling control method and system based on multi-source sensing data fusion. Background Technology

[0002] As the mainstream technology for urban tunnel construction, shield tunneling has been widely used in underground engineering projects such as subways, highways, railways, and municipal pipelines due to its advantages of minimal impact on surface traffic and the surrounding environment, high construction speed, and high degree of mechanization. With the continuous expansion of underground space development and the increasing complexity of geological conditions, higher requirements are being placed on the safety, accuracy, and efficiency of shield tunneling.

[0003] Currently, tunnel boring machine (TBM) construction still primarily relies on manual operation, where operators manually control key parameters such as advance speed, cutterhead rotation speed, soil chamber pressure, and grouting volume based on instrument data and personal experience. This operating mode has the following significant drawbacks: First, the tunneling process is highly complex and uncertain. Geological conditions (such as soft soil, sand, hard rock, and composite strata) are spatially uneven and dynamically change over time. Traditional manual operation relies on the operator's subjective experience and judgment, making it difficult to respond promptly and accurately to complex and changing geological conditions. When encountering sudden geological changes, operators often struggle to make optimal decisions in a short period, easily leading to construction accidents such as excessive surface subsidence or uplift, shield machine attitude deviation, abnormal wear of cutterhead tools, or even machine jamming, seriously affecting project safety and quality.

[0004] Secondly, a tunnel boring machine (TBM) is a complex electromechanical system comprising multiple subsystems, including a propulsion system, a cutterhead system, a screw conveyor system, and a grouting system. In traditional operation modes, data from each subsystem is isolated, lacking effective collaborative analysis and global optimization. For example, when adjusting the propulsion speed, operators may fail to simultaneously consider its impact on soil chamber pressure, excavated soil volume, and surface settlement, leading to unreasonable parameter matching and triggering a chain reaction of problems. The independent control mode of each subsystem makes it difficult to achieve optimal balance in overall performance.

[0005] Secondly, existing control methods are mostly reactive, meaning they only address problems after they occur. For example, operators only increase grouting volume or adjust propulsion parameters after surface settlement monitoring values ​​exceed standards. This delayed control method lacks the ability to adaptively predict unknown strata and is unable to proactively take preventative measures before risks occur; it is essentially a passive response rather than proactive avoidance.

[0006] To address the aforementioned issues, the industry has undertaken some automation attempts. For example, some tunnel boring machines (TBMs) are equipped with PLC-based automatic control systems, capable of simple logic control based on preset thresholds (such as automatic speed reduction when thrust exceeds limits); some research has also attempted closed-loop control based on a single parameter (such as solely relying on soil chamber pressure). However, these existing automation solutions are mostly based on simple rules or single parameters, failing to achieve deep fusion and intelligent decision-making from multi-source heterogeneous data. They struggle to capture the complex nonlinear relationships between geological conditions and tunneling parameters, lack the ability to collaboratively optimize multiple objectives (safety, quality, efficiency, cost), and lack self-learning and knowledge accumulation capabilities. Therefore, when facing highly nonlinear and uncertain complex geological environments, the adaptability and decision-making level of existing technologies remain significantly limited.

[0007] Therefore, how to achieve deep fusion of multi-source sensing data during shield tunneling, construct a high-fidelity model that can simulate the interaction between the shield and the strata, and on this basis realize intelligent decision-making and control with adaptive, forward-looking and self-evolving capabilities has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] This application provides a shield tunneling intelligent tunneling control method and system based on multi-source sensing data fusion. By integrating multi-source heterogeneous sensing data of the shield machine with external environmental data, a digital twin-driven artificial intelligence decision-making model is constructed to generate and execute globally optimal tunneling control commands in real time, thereby achieving safe, efficient, and precise unmanned or minimally manned intelligent tunneling.

[0009] In a first aspect, this application provides a shield tunneling intelligent control method based on multi-source sensing data fusion, the method comprising: Acquire multi-source heterogeneous sensing data during the tunnel boring machine excavation process, and preprocess the multi-source heterogeneous sensing data to obtain standardized fused sensing data; Feature extraction is performed on the fused sensing data, and multi-level data fusion is carried out in combination with knowledge of shield tunneling construction to generate a fused feature vector representing the current tunneling status. A digital twin of the tunnel boring machine process is constructed, the fused feature vector is input into the digital twin, and an artificial intelligence decision-making algorithm is applied to generate the optimal tunneling control command. The optimal tunneling control command is sent to the controllers of each subsystem of the tunnel boring machine for execution, and the execution process is monitored by safety interlocks. Based on the deviation between the actual tunneling results and the prediction results of the digital twin, the artificial intelligence decision-making algorithm and / or the model of the digital twin are updated and optimized.

[0010] In one possible design, the multi-source heterogeneous sensing data includes geological survey data, tunnel boring machine equipment status data, and construction environment monitoring data; preprocessing of the multi-source heterogeneous sensing data includes: An edge-cloud collaborative sensing system is constructed, in which an industrial sensor network deployed on the shield machine body at the end is used to collect raw sensing data in real time, an edge computing gateway on the edge is used to perform preliminary filtering and caching of the raw sensing data and execute millisecond-level real-time response control, and a project server in the cloud is used to aggregate data from the entire line and establish a tunneling big data lake. The original sensed data is sequentially subjected to spatiotemporal alignment, missing data repair, and noise filtering. Spatiotemporal alignment includes using a high-precision time synchronization service to unify sensor timestamps and combining the shield machine motion model to uniformly convert the data to a geodetic coordinate system. Missing data repair uses a spatiotemporal graph neural network to fill in missing data by utilizing the spatial correlation between sensors and the physical and logical correlation of parameters. Noise filtering uses adaptive methods such as wavelet transform or empirical mode decomposition to separate noise from the true feature signals for high-frequency signals.

[0011] In one possible design, features are extracted from the fused sensing data, and combined with knowledge from the tunnel boring machine (TBM) construction field, multi-level data fusion is performed to generate a fused feature vector representing the current tunneling state, including: Multi-dimensional feature extraction is performed on standardized fusion sensing data, including low-level numerical features calculated based on time-domain statistics and frequency-domain energy distribution; Standardized fusion sensing data is analyzed for working condition characteristics. Clustering algorithms or classifiers are used to automatically identify tunneling working conditions, and corresponding combinations of working condition characteristics are selected for different identified working conditions. By integrating geological survey reports, equipment parameters, and expert experience rules, a knowledge graph for shield tunneling construction is constructed, which includes entity relationships and causal logic rules. High-level semantic features related to the current tunneling status are extracted from the knowledge graph. Advanced composite indices are constructed based on standardized fusion sensing data. These advanced composite indices include the cutterhead load balance index calculated based on the thrust symmetry of the left and right partitions, the propulsion system efficiency coefficient calculated based on the ratio of actual power to theoretical power, and the shield tail sealing risk index calculated based on the weighted average of grouting pressure gradient and shield tail gap change. The underlying numerical features, the combination of operating condition features, the high-level semantic features, and the high-level composite indicators are fused together, and an attention mechanism is used to dynamically weight the various features to generate the fused feature vector.

[0012] In one possible design, the methods for constructing a digital twin of the tunnel boring machine (TBM) process include: A digital twin including a multiphysics coupling model is constructed to simulate the physical processes of cutterhead-soil interaction, spoil remediation and discharge, and grout diffusion and solidification during tunnel boring machine (TBM) excavation. The multiphysics coupling model includes a cutterhead-soil interaction model described by discrete element-finite element coupling and Mohr-Coulomb constitutive model and Hertz-Mindlin contact model, a spoil remediation and spiral discharge process model simulated by a fluid dynamics model, and a grout diffusion and solidification model described by Bingham fluid constitutive relation. Using real-time acquired tunneling data, the equivalent geomechanical parameters of the current tunnel face are dynamically inverted through Kalman filtering or particle filtering algorithms. The equivalent geomechanical parameters include cohesion, internal friction angle, elastic modulus, Poisson's ratio, permeability coefficient, and dilatation angle. The multiphysics coupling model is updated in real time based on the parameters obtained from the inversion, so that the deviation between the simulation output of the digital twin and the actual tunneling conditions is kept within a preset range.

[0013] In one possible design, the application of artificial intelligence decision-making algorithms to generate optimal tunneling control commands specifically includes employing a hierarchical reinforcement learning architecture, wherein: The upper strategy layer adopts a deep deterministic strategy gradient or near-end strategy optimization algorithm to formulate and output a macro-tunneling strategy in the form of normalized multi-dimensional action vectors based on the fused feature vector and the simulation data of the digital twin. The lower execution layer receives the macro-tunneling strategy and incorporates it into the optimization objective. It then performs rolling time-domain optimization in conjunction with the updated digital twin to calculate the specific control sequence for each actuator.

[0014] In one possible design, the method further includes: A dynamic weighting function is designed for four optimization objectives: safety, quality, schedule, and cost. The dynamic weighting function automatically adjusts the weight coefficients of each optimization objective based on the current construction stage of the tunnel boring machine and the identified risk level. The adjusted weight coefficients are input into the hierarchical reinforcement learning architecture, serving as the optimization guide for the upper policy layer to formulate macro-mining strategies and as a component of the objective function for the lower execution layer to perform rolling temporal optimization.

[0015] In one possible design, the optimal tunneling control commands are issued to the controllers of each subsystem of the tunnel boring machine for execution, and the execution process is monitored by safety interlocks, including: Each subsystem of the tunnel boring machine, including propulsion, cutterhead, auger, and grouting, is set as an independent intelligent agent. While receiving the optimal tunneling control command, each intelligent agent adaptively fine-tunes the received command parameters based on real-time operating data collected by its own configured local sensors. The fine-tuning range is limited to a preset safety boundary. Each intelligent agent establishes a communication mechanism through a data distribution service protocol or a message queue telemetry transmission protocol. A consensus protocol is used to periodically confirm the consistency of the collaborative state of each subsystem, so that the fine-tuning actions of each intelligent agent can be adapted to each other, and multi-agent collaborative control can be achieved. Based on multi-agent collaborative control, a human-machine interactive co-driving mode is introduced. The recommended operating range is provided to the operator based on the current tunneling conditions and the confidence level of the decision model composed of the digital twin and the artificial intelligence decision-making algorithm. A risk warning is issued when the operating parameters approach the preset warning value. When the confidence level of the decision model is continuously higher than the preset standard, the system automatically switches to full control mode, and the operator becomes a monitoring role. When the confidence level of the decision model is detected to be lower than the preset threshold or when the operator's active intervention instruction is received, control is smoothly transferred to manual operation.

[0016] In one possible design, a three-level safety interlock mechanism is established throughout the entire process of collaborative control and human-machine interaction. Three levels of thresholds are set, which correspond to issuing prompts, limiting the parameter adjustment rate, and triggering emergency shutdown, respectively, to ensure that all operations of each subsystem are always within the safety threshold range.

[0017] In one possible design, a lifelong learning mechanism is adopted to trigger the updating and optimization of the artificial intelligence decision-making algorithm and / or the model of the digital twin based on the deviation between the actual tunneling results and the prediction results of the digital twin. When the prediction error continues to be large or new working conditions are encountered, the model is fine-tuned using new data. And / or, using a federated learning architecture, tunnel boring machines from multiple projects train and update models locally, upload encrypted model parameters to the cloud center for aggregation, generate an updated global model, and distribute it to each project.

[0018] Secondly, this application provides a shield tunneling intelligent control system based on multi-source sensing data fusion, used to implement the methods described in the first aspect and various possible designs of the first aspect, the system comprising: The perception layer is used to collect multi-source heterogeneous perception data, including sensor networks deployed on the tunnel boring machine itself, edge computing gateways deployed on the construction site, and server clusters deployed in the cloud. The fusion layer is used to preprocess and fuse the data collected by the perception layer, and combine it with the knowledge graph of the shield tunneling construction field to generate a fused feature vector. The decision-making layer includes a digital twin and an intelligent decision engine. The digital twin is used to simulate the tunnel boring machine (TBM) excavation process, and the intelligent decision engine is used to generate optimal tunneling control commands based on the fused feature vector and the simulation results of the digital twin. The control layer, connected to the controllers of each subsystem of the tunnel boring machine, is used to receive and execute the optimal tunneling control commands and to perform safety interlock monitoring. An evolution layer is used to update and optimize the model of the digital twin and / or the intelligent decision engine based on the deviation between the actual tunneling results and the predicted results.

[0019] The intelligent tunneling control method and system based on multi-source sensing data fusion provided in this application have at least the following beneficial effects: This application's method realizes a closed-loop process from multi-source data acquisition, fusion processing, intelligent decision-making, collaborative control to model self-evolution. The system is data-driven, constrained by physical laws, uses artificial intelligence as its core algorithm, and employs a digital twin as a simulation and verification platform, aiming to improve the safety and efficiency of tunnel boring machine (TBM) construction.

[0020] Specifically, this method constructs a knowledge graph for tunnel boring machine (TBM) construction and combines it with multi-level data fusion technology to achieve deep integration and feature extraction of multi-source heterogeneous information, enabling the system to accurately perceive and characterize the complex and ever-changing tunneling environment. By constructing a high-fidelity digital twin and combining it with hierarchical reinforcement learning and model predictive control algorithms, the system can dynamically invert geological parameters and generate optimal tunneling control commands that take into account multiple objectives, thereby transforming the traditional experience-based post-adjustment mode into an intelligent decision-making mode with adaptive predictive capabilities. Simultaneously, through multi-agent collaborative control and safety interlocking mechanisms, coordinated operation of various subsystems and real-time risk monitoring are achieved. Furthermore, the system's built-in lifelong learning and federated learning mechanisms enable it to continuously utilize new data for model updates and optimization, realizing the digital accumulation and inheritance of construction experience.

[0021] Compared with existing technologies, this application transforms shield tunneling from manual operation to intelligent control, significantly improving the adaptability to complex geological environments and the level of multi-objective collaborative optimization. At the same time, by constructing an evolvable industrial knowledge platform, it provides key technical support for the intelligent construction and full life-cycle operation and maintenance of tunnel engineering. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1A flowchart illustrating a shield tunneling intelligent control method based on multi-source sensing data fusion, provided for embodiments of this application; Figure 2 A flowchart illustrating the implementation of a shield tunneling intelligent control method based on multi-source sensing data fusion in a subway shield tunneling project, as provided in this application embodiment. Figure 3 This is a structural diagram of the intelligent tunneling control system for shield tunneling based on multi-source sensing data fusion, provided in an embodiment of this application.

[0024] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with some aspects of this application as detailed in the appended claims.

[0026] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0027] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0028] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0029] As the mainstream technology for urban tunnel construction, shield tunneling has significant shortcomings under traditional manual operation: it relies on operator experience, making it difficult to respond in real time to complex and changing geological conditions, easily leading to problems such as surface subsidence, attitude deviation, and cutterhead wear; data from various subsystems are isolated, lacking collaborative analysis and global optimization; and control methods are mostly post-event adjustments, lacking adaptive prediction for unknown strata. Existing automation attempts are mostly based on simple rules or single parameters, failing to achieve deep integration of multi-source data and intelligent decision-making, and are unable to cope with the high nonlinearity and uncertainty of complex geological environments.

[0030] Based on this, embodiments of this application provide a shield tunneling intelligent control method based on multi-source sensing data fusion, such as... Figure 1 As shown, this control method can be implemented through the following steps S1-S5.

[0031] S1: Acquire multi-source heterogeneous sensing data during the tunnel boring machine's excavation process, and preprocess the data to obtain standardized fused sensing data.

[0032] Step S1 first acquires multi-source heterogeneous sensing data generated during the tunnel boring machine's excavation process. This data covers multiple dimensions, including geological survey information, equipment operating status, and construction environment monitoring. Subsequently, preprocessing operations such as spatiotemporal alignment, missing data repair, and noise filtering are performed on the raw data. Missing data is repaired with high precision using a spatiotemporal graph neural network that leverages sensor spatial correlation and parametric physical logic. Noise is filtered out using adaptive methods such as wavelet transform or empirical mode decomposition. This step transforms the chaotic multi-source data into high-quality fused sensing data through standardization, laying a reliable data foundation for subsequent analysis and decision-making, and solving the problems of data isolation and inconsistent quality in traditional methods.

[0033] In some embodiments, step S1 acquires multi-source heterogeneous sensing data during the tunnel boring machine excavation process and preprocesses the data to obtain standardized fused sensing data. Specifically, the real-time acquisition of multi-source heterogeneous sensing data is completed by constructing an edge-cloud collaborative sensing system, and then multi-stage preprocessing operations are performed on the acquired raw data to achieve data standardization. The specific implementation process is as follows: steps S101 to S107.

[0034] S101: A highly reliable industrial sensor network is deployed on the tunnel boring machine (TBM) body. This industrial sensor network includes various types of sensors such as fiber optic earth pressure gauge arrays, high-precision tilt and gyroscopes, multispectral vision sensors, acoustic emission sensors, and laser scanners. Each type of sensor collects raw sensing data related to soil chamber pressure, machine posture, muck properties, cutterhead cutter status, and tail shield clearance during the TBM's tunneling process, achieving comprehensive acquisition of multi-dimensional physical quantities during TBM tunneling.

[0035] S102: Deploy an edge computing gateway on the field server. This edge computing gateway receives raw sensing data transmitted in real time from the industrial sensor network on the tunnel boring machine (TBM) side, and performs preliminary filtering and caching operations on the raw sensing data. Simultaneously, the edge computing gateway has built-in logical judgment rules. For example, when the soil chamber pressure instantaneously exceeds the limit or the cutterhead torque abnormally increases, it achieves millisecond-level real-time response control of the TBM tunneling process based on preset rules, such as triggering an emergency stop command or a command to reduce the advance speed, thus completing rapid intervention in the equipment's operating status.

[0036] S103: By aggregating the sensing data uploaded by the field servers corresponding to all tunnel boring machines along the entire line through the project server, a tunneling big data lake is established based on the project server, which provides a unified data storage and computing foundation for subsequent model training, digital twin simulation and cross-project knowledge transfer, and completes the centralized aggregation and storage of multi-source heterogeneous sensing data.

[0037] S104: High-precision time synchronization services based on GPS or networks are used to perform unified time dimension processing on the sensing data collected by each sensor, achieving consistency in the timestamps of all sensing data. At the same time, combined with the tunnel boring machine's motion model, spatial coordinate transformation operations are performed on the sensing data collected by the moving sensors, uniformly transforming the sensing data in the local coordinate system of the tunnel boring machine to the geodetic coordinate system, completing the standardized alignment processing of the spatiotemporal dimensions of the sensing data.

[0038] S105: A spatiotemporal graph neural network is used to process the spatiotemporally standardized sensing data. This network first constructs an adjacency matrix based on the physical distance and signal correlation between sensors. Then, spatial correlation features are extracted through graph convolutional layers, followed by temporal evolution features extracted through temporal convolutional layers. Utilizing the spatial correlation between sensors and the physical logical correlation between parameters, high-precision repair is performed on missing sensing data caused by sensor failures or transmission interruptions, thus completing the missing sensing data.

[0039] S106: For high-frequency signals such as vibration and acoustic signals in the sensing data, adaptive processing methods such as wavelet transform or empirical mode decomposition are used to separate the high-frequency signals. For example, the vibration signal is decomposed into 5 levels using the db8 wavelet basis, and the wavelet coefficients are processed by a soft threshold function to filter out noise components, thereby effectively distinguishing the noise part of the signal from the real characteristic signal reflecting the actual operating state of the equipment, and completing the noise filtering of the sensing data.

[0040] S107: After completing the entire process of spatiotemporal alignment, missing data repair, and noise filtering, the perceptual data from each dimension is integrated and encapsulated to obtain standardized fused perceptual data. This standardized fused perceptual data can be directly used for subsequent feature extraction and multi-level data fusion steps.

[0041] S2: Extract features from the fused sensing data and combine them with knowledge from the shield tunneling construction field to perform multi-level data fusion and generate a fused feature vector that represents the current tunneling status.

[0042] Step S2 extracts features from the preprocessed data, calculating not only low-level indicators such as time-domain statistics and frequency-domain features, but also constructing high-level composite indicators such as the cutterhead load balance index and the propulsion system efficiency coefficient. Simultaneously, a knowledge graph from the shield tunneling construction field is introduced. This graph integrates structured knowledge and logical rules from geological conditions, equipment parameters, and expert experience, from which high-level semantic features are extracted. Finally, an attention mechanism is used to dynamically weight and fuse the low-, mid-, and high-level features, generating a fused feature vector that comprehensively represents the current tunneling status. This step achieves deep integration of multi-source heterogeneous information, enabling the system to accurately perceive and understand the complex and ever-changing geological environment, providing high-quality input for intelligent decision-making.

[0043] In some embodiments, step S2 performs feature extraction and multi-level data fusion processing on the standardized fusion perception data obtained in step S1. By constructing a knowledge graph in the shield tunneling construction field, logical guidance is provided for data fusion. At the same time, advanced feature engineering is combined to complete the extraction and screening of multi-dimensional features, and finally, multi-level fusion of perception data is achieved. The specific implementation process is as follows: steps S201 to S204.

[0044] S201: Integrating structured prior knowledge such as geological survey reports, design axes, segment models, and equipment nameplate parameters, while also collecting expert experience rules summarized during shield tunneling construction, a knowledge graph for the shield tunneling construction field is constructed based on the above content. This knowledge graph includes entity relationships and causal logic rules between geological conditions, equipment status, and process parameters. In the subsequent feature fusion stage, the logical rules in this knowledge graph are embedded into the loss function of the neural network in the form of constraint terms. For example, when the system identifies that it is currently in a weak clay layer through visual or torque features, the pre-set rule in the knowledge graph, "weak strata require special attention to soil chamber pressure," is activated. This rule is transformed into a gain coefficient for the weights of soil chamber pressure-related features, thereby achieving enhanced filtering of corresponding associated features.

[0045] S202: Perform condition feature analysis on standardized fused sensing data. Unsupervised clustering algorithms or supervised learning-based classifiers are used to process the sensing data. Different working conditions during the tunnel boring machine (TBM) excavation process are automatically identified through data features, enabling automatic differentiation of various working conditions such as normal excavation, attitude correction, crossing risk sources, and equipment maintenance. For each identified excavation condition, sensing data features adapted to that condition are extracted. For example, in the case of crossing risk sources, the focus is on extracting surface settlement-related features and grouting pressure features; while in the case of hard rock excavation, the focus is on extracting cutterhead torque and vibration features. This allows for the selection of feature combinations that match the working condition, ensuring that the extracted features accurately reflect the TBM's excavation status and equipment operation under the corresponding working condition.

[0046] S203: Based on standardized fusion sensing data, multiple advanced composite indicators are constructed to form a composite index that can intuitively reflect the health status of the tunnel boring machine (TBM) system. The constructed advanced composite indicators include a cutterhead load balance index calculated based on the thrust symmetry of the left and right zones, a propulsion system efficiency coefficient calculated based on the ratio of actual input power to theoretical propulsion power, and a tail shield sealing risk index calculated based on the weighted average of grouting pressure gradient and tail shield gap change. Each indicator reflects the operating status of key components and systems of the TBM from different dimensions such as load distribution, energy conversion, and sealing reliability. The combination of multi-dimensional indicators achieves a comprehensive reflection of the overall health status of the TBM.

[0047] S204: Based on the logical guidance and physical constraints of the knowledge graph in the shield tunneling construction field, combined with the combination of working condition features obtained from the working condition division and the constructed advanced composite indicators, multi-level data fusion processing is carried out on the extracted multi-dimensional features. An attention mechanism is used to dynamically weight and fuse various features, where the attention weight is jointly determined by the current working condition type, the strength of the rules activated by the knowledge graph, and the statistical significance of the feature itself. This dynamically adjusts the weights of various features, strengthening the role of key features while weakening the influence of irrelevant features. Through this method, effective fusion of multi-source features is achieved, ultimately generating a 256-dimensional fused feature vector, which can be directly used in subsequent digital twin-driven artificial intelligence decision-making steps.

[0048] S3: Construct a digital twin of the tunnel boring machine process, input the fused feature vector into the digital twin, and apply artificial intelligence decision-making algorithms to generate the optimal tunneling control commands.

[0049] Step S3 first constructs a digital twin of the tunnel boring machine (TBM) excavation process. This twin integrates multiple physical models, including a geomechanical interaction model, a spoil remediation model, and a grouting diffusion model. Using algorithms such as Kalman filtering, it dynamically inverts the equivalent geomechanical parameters of the tunnel face based on real-time tunneling data, continuously updating the model to approximate the real physical world. Based on this, a hierarchical reinforcement learning architecture is applied. The upper-layer policy network uses deep reinforcement learning algorithms to formulate macroscopic tunneling strategies, while the lower-layer execution layer uses model predictive control combined with the high-fidelity digital twin for rolling temporal optimization, calculating the specific control sequence of the actuators. This step realizes the mapping and interaction from the physical world to the digital space, enabling the system to have adaptive prediction capabilities for unknown strata and multi-objective collaborative optimization capabilities.

[0050] In some embodiments, based on the fused feature vector obtained in step S2, artificial intelligence decision processing driven by digital twin is carried out. A digital twin model that fits the actual tunneling working conditions is constructed through refined modeling of the digital twin. Intelligent decision-making is completed by combining a hierarchical reinforcement learning architecture. At the same time, a dynamic weight function is designed to realize multi-objective optimization and control. Finally, a globally optimal tunneling control command that adapts to the actual construction needs is generated. The specific implementation process is as follows: steps S301 to S306.

[0051] S301: Conduct refined modeling of the digital twin, first building the multi-physics coupled model required for the digital twin. A discrete element-finite element coupled model is used to simulate the interaction between the cutterhead and the soil. The finite element domain uses the Mohr-Coulomb constitutive model to describe the mechanical behavior of the continuous medium, while the discrete element domain uses the Hertz-Mindlin contact model to simulate the contact mechanical behavior between soil particles. The coupling solution of the two domains is achieved through bidirectional transmission of force and displacement data at the interface. A fluid dynamics model is used to simulate the entire process of slag improvement and spiral soil removal. A Bingham fluid constitutive relation grouting diffusion and solidification model is used to predict and analyze the filling effect behind the tunnel lining segments. The basic physical model of the digital twin is constructed through the collaborative building of multiple models.

[0052] S302: Utilizing real-time acquired tunnel boring machine (TBM) excavation data, the equivalent geomechanical parameters of the current tunnel face are dynamically inverted using Kalman filtering or particle filtering algorithms. The state vector includes six geological parameters: cohesion, internal friction angle, elastic modulus, Poisson's ratio, permeability coefficient, and dilatation angle. The observation vector includes four measurable variables: total thrust, cutterhead torque, surface settlement, and soil removal rate. Based on the inverted parameters, the established multiphysics coupling model is updated and adjusted in real time, ensuring that the digital twin's model state continuously approximates the actual TBM excavation conditions, guaranteeing the simulation accuracy of the digital twin and its matching degree with actual working conditions.

[0053] S303: A hierarchical reinforcement learning architecture is built on the digital twin that has completed refined modeling and is updated in real time. This architecture consists of an upper policy layer and a lower execution layer. In the upper policy layer, a deep deterministic policy gradient or proximal policy optimization algorithm is used. Combining the 256-dimensional fused feature vector obtained in step S2 with the simulation data of the digital twin, a macro-strategy for the tunnel boring machine is formulated. This macro-strategy is output in the form of a normalized eight-dimensional action vector, which corresponds to the macro-control targets of key parameters such as propulsion speed, cutterhead rotation speed, soil chamber pressure, and grouting pressure.

[0054] S304: A model predictive control method is employed at the lower execution layer. It receives the macro-level strategy output from the upper strategy layer and incorporates it into the optimization objective system. A high-fidelity digital twin model is used to perform rolling time-domain optimization calculations. The prediction time domain is set to 20 steps, and the control time domain to 10 steps. The optimization problem is solved using quadratic programming, with the objective function including the weighted sum of squared errors in state tracking and a penalty term for changes in control variables. Through the above calculations and deductions, the optimal control sequence corresponding to each specific actuator of the tunnel boring machine is obtained, realizing the transformation from macro-level tunneling strategy to specific execution commands. This hierarchical architecture simultaneously possesses the global optimization capabilities of artificial intelligence and the precise constraint handling capabilities of model predictive control.

[0055] S305: A dynamic weighting function is designed for the four optimization objectives of safety, quality, schedule, and cost during the tunnel boring machine (TBM) excavation process. This function can automatically adjust the weight coefficients corresponding to each optimization objective based on the actual construction stage of the TBM and the risk level identified on site. Under normal tunneling conditions, the weight vectors are set to 0.3, 0.3, 0.3, and 0.1, corresponding to safety, quality, schedule, and cost, respectively. In construction scenarios where the TBM passes through important buildings, the weight vectors are adjusted to 0.6, 0.3, 0.1, and 0.0 to increase the weight of the safety objective corresponding to minimizing surface settlement. In construction scenarios where the TBM is in homogeneous and stable strata, the weight vectors are adjusted to 0.3, 0.2, 0.4, and 0.1 to increase the weight of the schedule objective corresponding to maximizing tunneling efficiency. Through dynamic adjustment of the weights, the intelligent decision-making process adapts to different construction conditions, ensuring the rationality and optimality of the decision results.

[0056] S306: Based on a multi-objective system with dynamic weight optimization, combined with the computational results of a hierarchical reinforcement learning architecture, a globally optimal tunneling control instruction set adapted to the current construction conditions is generated, providing accurate instruction basis for subsequent tunneling control operations of the tunnel boring machine.

[0057] S4: Send the optimal tunneling control command to the controllers of each subsystem of the tunnel boring machine for execution, and monitor the safety interlocking during the execution process.

[0058] Step S4 sends the generated optimal tunneling control commands to the controllers of each subsystem of the tunnel boring machine, including the propulsion system, cutterhead system, auger system, and grouting system. A multi-agent collaborative control framework is adopted, treating each subsystem as an independent agent. While receiving central commands, each subsystem performs fine-tuning based on local perception and negotiates with each other through communication mechanisms, achieving more compliant and interference-resistant collaborative operation. Simultaneously, a three-level safety interlocking mechanism is established. When key parameters such as earth pressure, torque, or settlement exceed preset thresholds, corresponding protection measures are automatically triggered, with an emergency shutdown response time of less than 100 milliseconds. This step translates intelligent decision-making into precise physical execution, ensuring the stability and safety of the tunneling process through collaborative control and safety monitoring.

[0059] In some embodiments, step S4 sends the optimal control instruction set generated in step S3 to the controllers of each subsystem of the tunnel boring machine. Through multi-agent collaborative control, the adaptation operation of each subsystem is realized. At the same time, the human-machine interaction co-driving mode is combined to realize the flexible switching of control rights. A safety interlocking mechanism is established to ensure the safety of the tunneling process. Finally, the adaptive collaborative control of the tunnel boring machine is completed. The specific implementation process is as follows: steps S401 to S407.

[0060] S401: The globally optimal tunneling control command set generated in step S3 is synchronously distributed to the controllers corresponding to each subsystem of the tunnel boring machine (TBM). After receiving the control commands, each subsystem controller parses and maps the global commands into a parameter format recognizable by its local controller. Simultaneously, it performs preliminary verification of the received commands based on the preset local safe operating range of each subsystem, thus establishing the command foundation for the execution of tunneling control operations of each TBM subsystem. The TBM subsystems include core operational subsystems such as propulsion, cutterhead, auger conveyor, and grouting.

[0061] S402: The tunnel boring machine's propulsion, cutterhead, auger, and grouting subsystems are each configured as independent intelligent agents. While receiving central intelligent decision-making commands, each agent collects local operational data through its own configured sensing components. Based on this local operational data, it autonomously fine-tunes the execution parameters corresponding to the central intelligent decision-making commands, with the adjustment range limited within preset safety boundaries. For example, the grouting agent adaptively corrects the central command's grouting volume target value within a range of ±5% based on measured grouting pressure fluctuations, ensuring that the operating parameters of each subsystem adapt to its own real-time operating status.

[0062] S403: A communication mechanism based on a data distribution service protocol or a message queue telemetry transmission protocol is established between the intelligent agents. Through this communication mechanism, the agents exchange operational data and negotiate work actions. A consensus protocol is used to periodically confirm the consistency of the coordinated state of each subsystem, such as synchronizing the matching relationship between the propulsion speed and the cutterhead rotation speed every 100 milliseconds. This ensures that the operational actions of each subsystem are mutually adapted and coordinated, avoiding the impact of deviations in a single subsystem on the overall tunneling effect. This achieves smooth operation of each subsystem during the tunneling process and improves the anti-interference capability of the overall control process.

[0063] S404: Based on multi-agent collaborative control, a human-machine interactive co-driving mode is introduced. The system provides the operator with recommended operating ranges for each subsystem based on the tunnel boring machine's current tunneling conditions and the confidence level of the decision-making model. The confidence level of the decision-making model is quantified by inverse statistics of the prediction errors of key state quantities using a digital twin; the smaller the prediction error, the higher the confidence level. Simultaneously, the system monitors various operating parameters of the tunnel boring machine in real time, issuing risk warnings to the operator when operating parameters approach preset warning values, providing data and risk references for the operator's on-site decision-making.

[0064] S405: Within the safe operating range provided by the system, the operator can make final decisions and manual fine-tuning of control parameters based on the actual working conditions at the construction site, making control commands more aligned with on-site construction needs. When the system identifies that the tunnel boring machine is in good tunneling condition and the confidence level of the decision model is consistently higher than the preset standard of 0.8, the system automatically switches to full-authority control mode. At this time, the operator only performs monitoring operations on the tunnel boring machine's operating status. The control mode switching process adopts a double-buffer smooth transition strategy, gradually transferring control weight from manual to automatic within a transition time of 5-10 seconds to avoid abrupt changes in control actions.

[0065] S406: When the system detects that the confidence level of the decision model is lower than a preset threshold of 0.5 or receives an active intervention command from the operator, the system smoothly and safely transfers control from automatic to manual control. During the transfer, the system gradually shifts the current control command to the manually set value in a gradual manner, with a transition time of 3 to 5 seconds to avoid abrupt changes in tunneling actions during the control switch. Simultaneously, the system fully records the reason for the control transfer; this recorded information will be used for subsequent model learning and optimization iterations.

[0066] S407: Throughout the entire execution of the multi-agent collaborative control and human-machine interactive co-driving mode, a safety interlocking mechanism is established. This mechanism sets three threshold levels for graded responses: yellow warning value, orange alarm value, and red emergency value. When the operating parameters reach the yellow warning value, the system issues a prompt; when the orange alarm value is reached, the system automatically limits the parameter adjustment rate; and when the red emergency value is reached, the system directly triggers an emergency shutdown command. By dynamically linking the operation execution process of each subsystem with the safe operating parameters of the tunnel boring machine, it ensures that all operations of each subsystem are always within the safe threshold range, achieving adaptive collaborative control and safety interlocking protection for the tunnel boring machine, and guaranteeing the accurate execution of control commands and the safe and stable tunneling process.

[0067] S5: Based on the deviation between the actual tunneling results and the prediction results of the digital twin, trigger the updating and optimization of the artificial intelligence decision-making algorithm and / or the model of the digital twin.

[0068] Step S5 establishes a lifelong learning and federated learning mechanism to continuously monitor the deviation between the actual tunneling results and the digital twin's prediction results. When the prediction error remains large or new working conditions are encountered, the model fine-tuning process is automatically triggered, utilizing new data to quickly adapt without forgetting old knowledge. After updating the locally trained model, the tunnel boring machines for multiple projects only upload the encrypted model parameters to the cloud center for federated aggregation, generating a more powerful global model before distributing it to each project. This step enables the system to have self-evolution capabilities, not only realizing the digital accumulation and inheritance of construction experience, but also continuously improving its adaptability to complex geological environments through the growth of collective wisdom.

[0069] In some embodiments, step S5 is based on the full-process operation data of the tunnel boring machine and the control execution feedback information to carry out closed-loop feedback and model self-evolution processing. By building a lifelong learning mechanism, the decision model is dynamically optimized and iterated globally. At the same time, the structured accumulation and digital delivery of knowledge of the entire construction process are completed, providing support for the iteration of tunnel construction technology and the operation and maintenance of the tunnel throughout its entire life cycle. The specific implementation process is as follows: steps S501 to S506.

[0070] S501: The system establishes a learning buffer for the decision-making model. This learning buffer continuously receives operational data, control execution data, and model prediction data during the tunnel boring machine's excavation process, and monitors the error value between the model's prediction results and the actual construction results in real time. When the prediction error continuously exceeds a preset threshold and the duration reaches a set time, or when the average error of ten consecutive sampling points exceeds the allowable range, the system determines that the triggering condition is met and simultaneously identifies the working condition type that occurs during the tunneling process.

[0071] S502: When the system detects that the model prediction error is consistently high or identifies new working conditions during tunneling, it automatically triggers the model fine-tuning process. It utilizes newly acquired construction data in real time to optimize and train the decision-making model, employs an elastic weight consolidation algorithm to protect key weights corresponding to existing working conditions in the model, and sets up an experience replay buffer to periodically replay historical data. This allows the decision-making model to quickly adapt to new tunneling conditions while retaining its adaptability to existing construction conditions, continuously ensuring the model's decision-making accuracy matches the actual construction conditions.

[0072] S503: Establish a distributed training and aggregation system for shield tunneling construction models at the group level. Each project's shield tunneling machine completes the update and training of the decision-making model locally. All original construction data are stored in the project's local system, and only the model parameters obtained after model training are encrypted.

[0073] S504: The encrypted model parameters are uploaded to the group's cloud center. The cloud center uses a federated average algorithm to aggregate the encrypted model parameters uploaded by each project, performs a weighted average of the weight parameters of the local models of each project, integrates the construction data and model optimization experience of multiple projects, generates a global optimization model that is adapted to shield tunneling construction in multiple scenarios, and then distributes the global optimization model to the shield tunneling machine system of each project to achieve collective intelligent growth of shield tunneling construction technology at the group level, while effectively protecting the privacy of construction data of each project.

[0074] S505: The system structures and organizes various construction-related knowledge accumulated throughout the entire tunneling process of the tunnel boring machine (TBM). This organization includes optimal control strategies for TBM construction, tunneling response patterns in different geological formations, and performance degradation curves of various TBM components. Graph database technology is used to store this knowledge, constructing a knowledge network that includes the relationships between geological conditions, equipment status, control parameters, and construction effects. This structured construction knowledge is then uniformly stored in the project's digital twin asset library, supporting the retrieval and analogical reasoning of similar cases in subsequent construction scenarios, providing data support for the optimization of subsequent TBM construction and decision-making models.

[0075] S506: Digital delivery will be carried out during the tunnel engineering completion phase. In addition to the physical tunnel structure, a comprehensive intelligent operation and maintenance (O&M) model covering the entire lifecycle of the tunnel will be generated and delivered. This intelligent O&M model is an application-layer model formed by encapsulating and engineering-adapting the structured construction knowledge in the project's digital twin asset library. It can accurately adapt to the tunnel's geological characteristics, equipment configuration, and construction techniques. This model provides a comprehensive data and model foundation for subsequent health monitoring, disease early warning, and maintenance decision-making during the tunnel's operation phase, achieving intelligent support throughout the entire lifecycle of tunnel construction and operation.

[0076] The following example of a subway tunnel boring machine project will be used to illustrate in detail the application of the proposed method in actual engineering.

[0077] like Figure 2 As shown, when applying the method of this application to a subway tunnel project, it can be implemented through the following steps S10~S70.

[0078] S10, System overall architecture and deployment, including steps S11 and S12.

[0079] S11, Hardware System Configuration.

[0080] The system constructs a three-layer collaborative perception and control architecture encompassing the "end-edge-cloud." A comprehensive industrial sensor network is deployed on the tunnel boring machine (TBM) itself (end side), including an earth pressure monitoring array composed of 32 fiber optic earth pressure gauges, an attitude monitoring system consisting of a high-precision inclinometer and fiber optic gyroscope, a visual perception system composed of a multispectral visual sensor and an infrared thermal imager, an acoustic monitoring system composed of an acoustic emission sensor and a vibration accelerometer, and an environmental monitoring system composed of a laser scanner and distributed fiber optic sensors. These sensors collect raw data at different sampling frequencies (10Hz-10kHz) and transmit it in real time via industrial protocols such as PROFINET IRT and OPC UA over TSN.

[0081] The edge computing gateways deployed on the side are equipped with the NVIDIA Jetson AGX Orin platform, featuring 256 CUDA cores and 32GB of memory. They are responsible for millisecond-level real-time response control, such as executing simple rules for emergency shutdown and vibration suppression. The cloud server cluster adopts a distributed architecture, including a Hadoop data lake storage system, a real-time stream processing platform, and a model training cluster, providing data aggregation and computing support for multiple tunnel boring machines along the entire line.

[0082] S12. Establish communication and security mechanisms.

[0083] The system establishes a layered communication protocol stack: real-time control data uses the PROFINET IRT protocol with a period of 1ms, used for critical controls such as the propulsion system and cutterhead drive; sensor data uses the OPC UA over TSN protocol with a period of 10-100ms; and video stream data uses the GigE Vision protocol with a bandwidth of 200Mbps. All data transmissions are encrypted using TLS 1.3 and implement role-based access control (RBAC) and audit logging mechanisms.

[0084] S20, Data preprocessing, including steps S21 to S23.

[0085] S21, Spatiotemporal alignment.

[0086] The system employs PTP (Precise Time Protocol) for microsecond-level time synchronization, with the master clock source being a GPS-disciplined clock server. For mobile sensor data, a transformation model was established between the tunnel boring machine's local coordinate system and the UTM (Underground Measurement System) geodetic coordinate system. The transformation matrix consists of a rotation matrix R and a translation vector T. The rotation matrix is ​​calculated based on the pitch, yaw, and roll angles measured by the IMU (Instrument Measured by the IMU), while the translation vector is determined based on the geodetic coordinates measured by RTK-GPS. All sensor data timestamps are uniformly aligned to the master clock, and spatial coordinates are uniformly transformed to the geodetic coordinate system with millimeter-level accuracy.

[0087] S22, Missing data repair.

[0088] The system employs a Spatiotemporal Graph Neural Network (ST-GNN) for missing data repair. This network architecture comprises four layers: the input layer receives historical time-series data (time step 100, number of sensor nodes N, feature dimension D); the spatial graph convolutional layer is implemented using ChebConv, constructing an adjacency matrix based on sensor physical distance and signal correlation, with a Chebyshev polynomial order K=3 and an output dimension of 64; the temporal convolutional layer uses dilated convolutional network (TCN) with a dilation rate sequence of [1,2,4,8], 64 kernels, and a kernel size of 3; the attention fusion module includes spatial attention, temporal attention, and feature attention guided by a knowledge graph; and the output layer generates repaired values ​​and uncertainty estimates through a three-layer fully connected network (256-128-D).

[0089] During training, various missing data patterns (random missing data, block missing data, sensor malfunction) and missing data rates (10%, 30%, 50%) were used to construct the training set. The Huber loss function combined with physical constraint loss was used for optimization. The learning rate was 1e-3, the batch size was 32, and the number of training epochs was 200. An early stopping strategy was employed (patience value 20, minimum improvement 1e-4). The model achieved a repair accuracy of 95.2% on the test set, with a mean squared error of less than 0.05.

[0090] S23, Adaptive signal denoising.

[0091] Specialized denoising algorithms were employed for different signal types: wavelet transform was used for vibration signals, with a 5-level decomposition based on a db8 ​​wavelet basis and a SURE thresholding strategy; empirical mode decomposition (EMD) was used for acoustic emission signals, with IMF components selected based on mutual information; and Kalman smoothing filtering was used for pressure signals. The denoising performance was evaluated using metrics including signal-to-noise ratio improvement (average increase of 15 dB) and feature retention (>90%).

[0092] S30, Knowledge Graph and Feature Fusion, includes the following steps S31 and S32.

[0093] S31. Knowledge Graph Construction.

[0094] The knowledge graph for tunnel boring machine (TBM) construction comprises four layers: the entity layer defines five types of entities—geology, equipment, technology, risk, and parameters—and four types of relationships—physical, causal, control, and temporal; the data layer integrates structured data such as geological survey reports, design documents, and equipment parameters; the rule layer encodes expert experience rules (SWRL format), physical constraint rules, and operating procedures; and the reasoning layer provides rule-based logical reasoning and graph neural network-based semantic reasoning.

[0095] The knowledge graph contains over 100,000 entity nodes and 500,000 relation edges, while the expert rule base contains 328 empirical rules, such as "high torque ∧ sandy strata → require water injection for improvement" and "low earth pressure ∧ shallow overburden → risk of settlement".

[0096] S32, Multi-level feature fusion.

[0097] Feature extraction is divided into three levels: low-level features extract time-domain statistics (mean, standard deviation, skewness, kurtosis, root mean square) and frequency-domain features (energy distribution in key frequency bands); mid-level features construct engineering composite indicators, including cutterhead load balance index (based on left and right thrust symmetry), propulsion system efficiency coefficient (actual power to theoretical power ratio), and shield tail sealing risk index (weighted by grouting pressure gradient and gap variance); high-level features extract semantic features based on knowledge graphs, including expert suggestion embedding vectors, risk level quantification, and constraint violation degree.

[0098] The fusion process employs an attention mechanism for weighting, with attention weights dynamically calculated based on feature importance, current operating conditions, and the knowledge graph. This ultimately generates a 256-dimensional fusion feature vector, which is used as input to the subsequent decision model.

[0099] S40, Digital Twin Modeling, including the following steps S41 and S42.

[0100] S41. Establish a multiphysics coupling model.

[0101] The geological-mechanical interaction was investigated using a discrete-element-finite-element coupled model (DEM-FEM). The finite element domain employed the Mohr-Coulomb constitutive model, with parameters including cohesion c (a function of water content) and internal friction angle. (Particle size function), elastic modulus E (compaction degree function); the discrete element domain adopts the Hertz-Mindlin contact model, with particle radius distribution of 0.1-5mm, friction coefficient of 0.3-0.6, and damping coefficient of 0.05-0.2. The coupled algorithm realizes the bidirectional transmission of stress and contact force through iterative solution, and the convergence condition is that the force residual is less than 1%.

[0102] The soil improvement process employs a computational fluid dynamics (CFD) model to simulate the mixing, transport, and phase change processes of the foaming agent and soil particles. The grouting diffusion model is based on the Bingham fluid constitutive relation, considering the spatiotemporal variations of grouting pressure, grout viscosity, and formation permeability.

[0103] S42, Real-time parameter inversion.

[0104] Real-time inversion of geological parameters was performed using an integrated Kalman filter (EnKF). The state vector contains six parameters: cohesion c, internal friction angle, etc. The variables include elastic modulus E, Poisson's ratio ν, permeability coefficient k, and dilatation angle ψ; the observation vector contains four variables: total thrust, cutterhead torque, surface settlement, and soil removal rate. The system integrates 50 members and updates iteratively through a state transition model (single-step calculation of the physical model) and an observation model (observable output of the model).

[0105] The inversion algorithm executes every 5 minutes, with an average convergence time of 30 seconds and parameter estimation uncertainty of less than 15%. The inversion results are used to update the parameters of the digital twin model in real time, reducing the model prediction error from the initial 25% to less than 8%.

[0106] S50, Intelligent Decision-Making, includes the following steps S51-S53.

[0107] S51. Establish a hierarchical reinforcement learning architecture.

[0108] The upper-layer policy network employs the Proximal Policy Optimization (PPO) algorithm. The network structure includes a state encoder (3 fully connected layers, dimensions 256-128), a policy head (outputting the mean and standard deviation of actions), and a value head (outputting state values). The state encoder takes 256-dimensional fused features as input and outputs a 128-dimensional state representation. The policy head outputs an 8-dimensional action vector, normalized to [-1,1], corresponding to macroscopic commands such as propulsion speed, cutterhead rotation speed, soil chamber pressure, and grouting pressure.

[0109] PPO training employs a pruning substitution target with a pruning coefficient ε=0.2, a value loss coefficient of 0.5, and an entropy regularization coefficient of 0.01. Training data is collected from a digital twin simulation environment and actual construction, and the advantage function is calculated using generalized advantage estimation (GAE, λ=0.95). After 100,000 training steps, the policy achieves an average reward improvement of 45% in the simulation environment.

[0110] S52, Model Predictive Control.

[0111] The lower-level controller employs Model Predictive Control (MPC), with 20 steps of prediction in the time domain and 10 steps of control in the time domain. The optimization problem is formalized as a quadratic programming problem, with the objective function including state tracking error (weight matrix Q) and control variable change penalty (weight matrix R). Constraints include upper and lower limits of control variables, state variable safety boundaries, and process sequence constraints.

[0112] The solver uses OSQP, with a solution time of less than 50ms per iteration, meeting real-time requirements. MPC control accuracy: propulsion speed error ±2mm / min, soil chamber pressure error ±0.05bar, grouting volume error ±3%.

[0113] S53, Multi-objective dynamic optimization.

[0114] The system defines four optimization objectives: safety (weight 0.4), quality (0.3), efficiency (0.2), and cost (0.1). The weights are dynamically adjusted based on risk scenarios: normal operating conditions [0.3, 0.3, 0.3, 0.1]; adjacent buildings [0.6, 0.3, 0.1, 0]; hard rock strata [0.3, 0.2, 0.4, 0.1]; soft soil strata [0.5, 0.3, 0.1, 0.1]; emergency state [0.9, 0.1, 0, 0]. Weight adjustments take into account construction progress factors, with the quality weight increasing linearly by 50%-100% in later stages.

[0115] S60, Collaborative control and security monitoring, including the following steps S61 and S62.

[0116] S61, Multi-agent collaboration.

[0117] Each subsystem (propulsion, cutterhead, auger, and grouting) operates as an independent intelligent agent, coordinated using the MADDPG algorithm. Each agent comprises an actor network (a 3-layer MLP with 128-64 action dimensions) and a critic network (centrally trained), achieving distributed decision-making through a consensus protocol. The collaborative control cycle is 100ms, and the consensus convergence threshold is 1e-3.

[0118] Experiments show that, compared with centralized control, multi-agent collaborative control improves control stability by 30% and reduces energy consumption by 15% under disturbed conditions.

[0119] S62, Human-Machine Collaboration and Safety Interlocking.

[0120] The system offers three control modes: fully automatic (confidence > 0.8), auxiliary (confidence 0.5-0.8), and manual (confidence < 0.5). Mode switching employs a double-buffered smooth transition with a transition time of 5-10 seconds. The safety monitoring system defines three threshold levels: warning value (yellow), alarm value (orange), and emergency value (red), corresponding to different intervention measures.

[0121] The safety interlocking mechanism includes: automatic adjustment of the advance speed when earth pressure exceeds the limit; automatic reduction of the cutterhead speed when torque exceeds the limit; and automatic increase of the grouting pressure when settlement exceeds the limit. Emergency stop response time is <100ms.

[0122] S70, Lifelong Learning and Knowledge Management, includes the following steps S71 and S72.

[0123] S71. Establish a federated learning framework.

[0124] Each project's local model is trained regularly (daily) on an edge server, with training data stored locally. Model parameters are encrypted and uploaded to the cloud center, where they are aggregated using the FedAvg algorithm to generate a global model. A differential privacy mechanism is used to add Gaussian noise (scale 0.01, L2 norm pruning 1.0) to ensure data privacy.

[0125] Federated learning is conducted in rounds every 24 hours, with 30% of clients participating. After 50 rounds of federated training, the global model's initial accuracy on new projects is 40% higher than the baseline model.

[0126] S72, Knowledge Accumulation and Digital Delivery.

[0127] The system establishes a structured knowledge base, including: an optimal control strategy library (categorized by geological formation), equipment performance degradation curves (based on operating time), a risk response case library, and an anomaly pattern feature library. The knowledge base is stored using a graph database, supporting similar case retrieval and analogical reasoning.

[0128] Upon project completion, a digital twin asset package will be delivered, including: complete construction process data (TB-level), a pre-trained intelligent operation and maintenance model, equipment health baselines, and a maintenance decision rule base. These assets provide predictive maintenance capabilities for subsequent operation and maintenance, and are expected to reduce operation and maintenance costs by 30% and unplanned downtime by 50%.

[0129] In summary, the method presented in this application, driven by data, constrained by physical laws, centered on artificial intelligence, and utilizing digital twins as a platform, aims for safety and efficiency. It not only achieves a paradigm shift in tunnel boring machine (TBM) excavation from "manual operation" to "intelligent control," but also constructs an intelligent platform capable of continuously evolving, accumulating, and transmitting industrial knowledge related to tunnel construction. This lays a solid technical foundation for completely resolving the uncertainties in TBM construction and moving towards truly intelligent construction.

[0130] This application also provides a shield tunneling intelligent control system based on multi-source sensing data fusion, used to implement the methods described in any of the above embodiments, such as... Figure 3 As shown, the intelligent tunneling control system based on multi-source sensing data fusion includes: The perception layer 301 is used to collect multi-source heterogeneous perception data, including a sensor network deployed on the tunnel boring machine body, an edge computing gateway deployed on the construction site, and a server cluster deployed in the cloud. The fusion layer 302 is used to preprocess and fuse the data collected by the perception layer, and generate a fused feature vector by combining the knowledge graph of the shield tunneling construction field. The decision layer 303 includes a digital twin and an intelligent decision engine. The digital twin is used to simulate the tunnel boring machine process, and the intelligent decision engine is used to generate optimal tunneling control commands based on the fused feature vector and the simulation results of the digital twin. The control layer 304 is connected to the controllers of each subsystem of the tunnel boring machine, and is used to receive and execute the optimal tunneling control command and perform safety interlock monitoring. Evolution layer 305 is used to update and optimize the model of the digital twin and / or the intelligent decision engine based on the deviation between the actual tunneling results and the predicted results.

[0131] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A shield tunneling intelligent control method based on multi-source sensing data fusion, characterized in that, The method includes: Acquire multi-source heterogeneous sensing data during the tunnel boring machine excavation process, and preprocess the multi-source heterogeneous sensing data to obtain standardized fused sensing data; Feature extraction is performed on the fused sensing data, and multi-level data fusion is carried out in combination with knowledge of shield tunneling construction to generate a fused feature vector representing the current tunneling status. A digital twin of the tunnel boring machine process is constructed, the fused feature vector is input into the digital twin, and an artificial intelligence decision-making algorithm is applied to generate the optimal tunneling control command. The optimal tunneling control command is sent to the controllers of each subsystem of the tunnel boring machine for execution, and the execution process is monitored by safety interlocks. Based on the deviation between the actual tunneling results and the prediction results of the digital twin, the artificial intelligence decision-making algorithm and / or the model of the digital twin are updated and optimized.

2. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1, characterized in that, The multi-source heterogeneous sensing data includes geological survey data, tunnel boring machine equipment status data, and construction environment monitoring data; the multi-source heterogeneous sensing data undergoes preprocessing, including: An edge-cloud collaborative sensing system is constructed, in which an industrial sensor network deployed on the shield machine body at the end is used to collect raw sensing data in real time, an edge computing gateway on the edge is used to perform preliminary filtering and caching of the raw sensing data and execute millisecond-level real-time response control, and a project server in the cloud is used to aggregate data from the entire line and establish a tunneling big data lake. The original sensed data is sequentially subjected to spatiotemporal alignment, missing data repair, and noise filtering. Spatiotemporal alignment includes using a high-precision time synchronization service to unify sensor timestamps and combining the shield machine motion model to uniformly convert the data to a geodetic coordinate system. Missing data repair uses a spatiotemporal graph neural network to fill in missing data by utilizing the spatial correlation between sensors and the physical and logical correlation of parameters. Noise filtering uses adaptive methods such as wavelet transform or empirical mode decomposition to separate noise from the true feature signals for high-frequency signals.

3. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1, characterized in that, Feature extraction is performed on the fused sensing data, and multi-level data fusion is conducted in conjunction with knowledge from the shield tunneling construction field to generate a fused feature vector representing the current tunneling state, including: Multi-dimensional feature extraction is performed on standardized fusion sensing data, including low-level numerical features calculated based on time-domain statistics and frequency-domain energy distribution; Standardized fusion sensing data is analyzed for working condition characteristics. Clustering algorithms or classifiers are used to automatically identify tunneling working conditions, and corresponding combinations of working condition characteristics are selected for different identified working conditions. By integrating geological survey reports, equipment parameters, and expert experience rules, a knowledge graph for shield tunneling construction is constructed, which includes entity relationships and causal logic rules. High-level semantic features related to the current tunneling status are extracted from the knowledge graph. Advanced composite indices are constructed based on standardized fusion sensing data. These advanced composite indices include the cutterhead load balance index calculated based on the thrust symmetry of the left and right partitions, the propulsion system efficiency coefficient calculated based on the ratio of actual power to theoretical power, and the shield tail sealing risk index calculated based on the weighted average of grouting pressure gradient and shield tail gap change. The underlying numerical features, the combination of operating condition features, the high-level semantic features, and the high-level composite indicators are fused together, and an attention mechanism is used to dynamically weight the various features to generate the fused feature vector.

4. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1, characterized in that, Methods for constructing a digital twin of the tunnel boring machine (TBM) process include: A digital twin including a multiphysics coupling model is constructed to simulate the physical processes of cutterhead-soil interaction, spoil remediation and discharge, and grout diffusion and solidification during tunnel boring machine (TBM) excavation. The multiphysics coupling model includes a cutterhead-soil interaction model described by discrete element-finite element coupling and Mohr-Coulomb constitutive model and Hertz-Mindlin contact model, a spoil remediation and spiral discharge process model simulated by a fluid dynamics model, and a grout diffusion and solidification model described by Bingham fluid constitutive relation. Using real-time acquired tunneling data, the equivalent geomechanical parameters of the current tunnel face are dynamically inverted through Kalman filtering or particle filtering algorithms. The equivalent geomechanical parameters include cohesion, internal friction angle, elastic modulus, Poisson's ratio, permeability coefficient, and dilatation angle. The multiphysics coupling model is updated in real time based on the parameters obtained from the inversion, so that the deviation between the simulation output of the digital twin and the actual tunneling conditions is kept within a preset range.

5. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1 or 4, characterized in that, The application of artificial intelligence decision-making algorithms to generate optimal tunneling control commands specifically includes employing a hierarchical reinforcement learning architecture, wherein: The upper strategy layer adopts a deep deterministic strategy gradient or near-end strategy optimization algorithm to formulate and output a macro-tunneling strategy in the form of normalized multi-dimensional action vectors based on the fused feature vector and the simulation data of the digital twin. The lower execution layer receives the macro-tunneling strategy and incorporates it into the optimization objective. It then performs rolling time-domain optimization in conjunction with the updated digital twin to calculate the specific control sequence for each actuator.

6. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 5, characterized in that, The method further includes: A dynamic weighting function is designed for four optimization objectives: safety, quality, schedule, and cost. The dynamic weighting function automatically adjusts the weight coefficients of each optimization objective based on the current construction stage of the tunnel boring machine and the identified risk level. The adjusted weight coefficients are input into the hierarchical reinforcement learning architecture, serving as the optimization guide for the upper policy layer to formulate macro-mining strategies and as a component of the objective function for the lower execution layer to perform rolling temporal optimization.

7. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1, characterized in that, The optimal tunneling control commands are issued to the controllers of each subsystem of the tunnel boring machine for execution, and the execution process is monitored by safety interlocks, including: Each subsystem of the tunnel boring machine, including propulsion, cutterhead, auger, and grouting, is set as an independent intelligent agent. While receiving the optimal tunneling control command, each intelligent agent adaptively fine-tunes the received command parameters based on real-time operating data collected by its own configured local sensors. The fine-tuning range is limited to a preset safety boundary. Each intelligent agent establishes a communication mechanism through a data distribution service protocol or a message queue telemetry transmission protocol. A consensus protocol is used to periodically confirm the consistency of the collaborative state of each subsystem, so that the fine-tuning actions of each intelligent agent can be adapted to each other, and multi-agent collaborative control can be achieved. Based on multi-agent collaborative control, a human-machine interactive co-driving mode is introduced. The recommended operating range is provided to the operator based on the current tunneling conditions and the confidence level of the decision model composed of the digital twin and the artificial intelligence decision-making algorithm. A risk warning is issued when the operating parameters approach the preset warning value. When the confidence level of the decision model is continuously higher than the preset standard, the system automatically switches to full control mode, and the operator becomes a monitoring role. When the confidence level of the decision model is detected to be lower than the preset threshold or when the operator's active intervention instruction is received, control is smoothly transferred to manual operation.

8. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 7, characterized in that, Throughout the entire process of collaborative control and human-machine interaction, a three-level safety interlock mechanism is established, with three levels of thresholds set to provide corresponding graded responses such as issuing prompts, limiting parameter adjustment rates, and triggering emergency shutdowns, ensuring that all operations of each subsystem remain within the safety threshold range.

9. The intelligent tunneling control method for shield tunneling based on digital twins and artificial intelligence according to claim 1, characterized in that, Based on the deviation between the actual tunneling results and the prediction results of the digital twin, a lifelong learning mechanism is adopted to trigger the updating and optimization of the artificial intelligence decision-making algorithm and / or the model of the digital twin. When the prediction error continues to be large or new working conditions are encountered, the model is fine-tuned using new data. And / or, using a federated learning architecture, tunnel boring machines from multiple projects train and update models locally, upload encrypted model parameters to the cloud center for aggregation, generate an updated global model, and distribute it to each project.

10. A shield tunneling intelligent control system based on digital twin and artificial intelligence, characterized in that, The system for implementing the method according to any one of claims 1 to 9, the system comprising: The perception layer is used to collect multi-source heterogeneous perception data, including sensor networks deployed on the tunnel boring machine itself, edge computing gateways deployed on the construction site, and server clusters deployed in the cloud. The fusion layer is used to preprocess and fuse the data collected by the perception layer, and combine it with the knowledge graph of the shield tunneling construction field to generate a fused feature vector. The decision-making layer includes a digital twin and an intelligent decision engine. The digital twin is used to simulate the tunnel boring machine (TBM) excavation process, and the intelligent decision engine is used to generate optimal tunneling control commands based on the fused feature vector and the simulation results of the digital twin. The control layer, connected to the controllers of each subsystem of the tunnel boring machine, is used to receive and execute the optimal tunneling control commands and to perform safety interlock monitoring. An evolution layer is used to update and optimize the model of the digital twin and / or the intelligent decision engine based on the deviation between the actual tunneling results and the predicted results.

Images