Hard rock TBM shield tunneling machine auxiliary tunneling parameter dynamic adaptation regulation and control system

Abstract

The invention relates to the technical field of tunnel engineering intelligent construction, and discloses a hard rock TBM shield tunneling machine auxiliary tunneling parameter dynamic adaptation regulation and control system, which comprises a high-precision multi-source sensing network, a high-precision multi-source sensing network, a high-precision multi-source sensing network, a high-precision multi-source sensing network, a high-precision multi-source sensing network, a high-precision multi-source sensing network and a high-precision multi-source sensing network, the TBM-geological environment digital twin predicts the tunneling short-term trend based on the high-fidelity physical simulation and data assimilation technology; and the multi-modal deep learning collaborative decision-making module deeply fuses real-time and prediction data and generates an optimal parameter solution set through a network trade-off tunneling multi-conflict target based on Pareto optimization. And the system executes a decision and forms closed-loop feedback through a parameter dynamic adaptation and adaptive learning module, and continuously optimizes a self model and a knowledge base. According to the method, passive response of TBM tunneling is converted into active pre-judgment, the decision accuracy, the construction safety and the comprehensive tunneling efficiency under the complex working condition are remarkably improved, and the method has the sustainable evolution capacity.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent construction of tunnel engineering, and in particular to a dynamic adaptive control system for auxiliary excavation parameters of a hard rock TBM shield machine. Background Art

[0002] As a critical piece of equipment in the construction of large-scale tunnels and underground projects, the tunneling efficiency and safety of hard rock tunnel boring machines (TBMs) are directly related to the progress and cost of the entire project. TBM operation control is essentially a process of dynamically adjusting tunneling parameters (such as main thrust, cutterhead torque, and cutterhead speed) based on the ever-changing and complex geological environment.

[0003] In existing technical practices, the setting of TBM excavation parameters relies heavily on the operator's personal skills and field experience. Operators manually adjust various parameters by observing real-time operating data displayed on the dashboard and combining their own judgment. This approach is sufficient for uniform and stable geological conditions. However, its limitations become apparent when TBMs encounter complex working conditions such as uneven hardness, fault fracture zones, dense joints, or water-rich areas common in hard rock formations. Due to the sudden and unpredictable nature of geology, passive and responsive control methods that rely on human experience often lag behind, making it difficult to predict risks and make optimal adjustments in advance. This can easily lead to adverse engineering events such as machine jamming, abnormal tool wear or chipping, and face instability, seriously impacting construction safety and efficiency.

[0004] To aid decision-making, some advanced TBM systems are equipped with multiple sensors to obtain richer information. However, this multi-source, heterogeneous data is often presented in isolation, lacking an effective fusion analysis mechanism to deeply explore the underlying correlations and trends. Existing technologies fail to systematically and collaboratively analyze real-time equipment operating parameters, tool status information, and forward geological survey data, resulting in a blind spot in the comprehensive understanding of the current tunneling status.

[0005] Furthermore, TBM excavation itself is a complex optimization problem with multiple conflicting objectives. Simply pursuing excavation speed can dramatically increase tool wear and pose safety risks; overly conservative parameter settings can sacrifice valuable construction time. Existing technologies lack scientific quantitative methods to dynamically balance multiple, mutually constrained objectives such as excavation efficiency, equipment lifespan, construction safety, and cost, making it impossible to generate a globally optimal parameter combination. Furthermore, the valuable excavation experience and troubleshooting knowledge accumulated across various tunnel projects often exists in fragmented reports or in the minds of experts, making it difficult to systematically inherit and reuse them. This results in a high learning cost for each new project facing similar challenges. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a dynamic adaptive control system for the auxiliary excavation parameters of a hard rock TBM shield machine, which solves the problems of low decision-making efficiency and poor safety under complex geological conditions due to the lack of active prediction capabilities, multi-objective collaborative optimization mechanism and cross-project adaptive learning capabilities in the existing TBM excavation control methods.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system, comprising: High-precision multi-source sensing network for real-time acquisition of hard rock TBM operating parameters, tool status parameters, and geological environment parameters; The TBM-geological environment digital twin is used to receive data collected by the high-precision multi-source sensing network and, based on a built-in high-fidelity physical simulation model, to perform real-time deduction and prediction of short-term trends in TBM excavation, outputting prediction data including predicted excavation effects and potential risks; A multimodal deep learning collaborative decision-making module is used to receive the predicted data output by the TBM-geological environment digital twin and the real-time data collected by the high-precision multi-source perception network, and generate a set of dynamically adaptive tunneling parameter combinations by performing feature fusion and multi-objective optimization on the multimodal data; The parameter dynamic adaptation and adaptive learning module is used to select an optimal parameter combination from the dynamically adapted tunneling parameter combination to control the TBM in real time, and based on the actual tunneling effect, feedback updates the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module to form a closed-loop adaptive learning.

[0008] Preferably, the high-fidelity physical simulation model built into the TBM-geological environment digital twin is constructed based on finite element analysis and discrete element method, and can simulate the TBM tool rock breaking process, cutterhead force distribution and tool wear evolution; and, the TBM-geological environment digital twin uses the real-time data collected by the high-precision multi-source perception network to perform online correction and status update of the high-fidelity physical simulation model.

[0009] Preferably, the multimodal data fused by the multimodal deep learning collaborative decision-making module includes at least: structured and serialized prediction data from the TBM-geological environment digital twin, and image data and unstructured text data from the high-precision multi-source perception network.

[0010] Preferably, the multimodal deep learning collaborative decision-making module realizes the feature fusion through a multimodal feature fusion network. The multimodal feature fusion network adopts a multi-branch deep neural network architecture and uses an attention mechanism to perform weighted fusion of features of different modalities to form a comprehensive high-dimensional feature representation.

[0011] Preferably, the multimodal deep learning collaborative decision-making module generates the dynamically adaptive excavation parameter combination through target conflict resolution and multi-objective optimization network. The network is based on the Pareto optimization idea, which is used to weigh multiple conflicting objectives and output a set of Pareto optimal excavation parameter solutions.

[0012] Preferably, the system further comprises a fault mode map, which stores the fault types, associated geological conditions, excavation parameters, sensor abnormal modes and treatment measures in the historical excavation process in a graph structure.

[0013] Preferably, the multimodal deep learning collaborative decision-making module is also used to perform adaptive decision topology reconstruction based on the reasoning results of the fault mode map when an abnormal mode similar to any fault mode in the fault mode map is identified, so as to dynamically adjust its decision weight and parameter adjustment priority.

[0014] Preferably, the system also includes a cross-project knowledge sharing and privacy protection mechanism based on federated learning, which is used to continuously optimize the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module through distributed collaborative training while protecting the privacy of the original data of each TBM project.

[0015] Preferably, the cross-project knowledge sharing and privacy protection mechanism based on federated learning is specifically implemented as follows: each TBM project acts as a client and uses its private data to train the model locally, and only uploads the updated model parameters to the central server for aggregation to generate a global model. The central server then distributes the global model to each client to update its local model.

[0016] Preferably, the specific method for the parameter dynamic adaptation and adaptive learning module to perform feedback update based on the actual excavation effect is: feeding back the actual excavation effect data to the TBM-geological environment digital twin to correct its high-fidelity physical simulation model, and feeding back to the multimodal deep learning collaborative decision-making module to optimize its decision-making ability, and when a new fault event occurs, updating the event and its handling method to the fault mode map.

[0017] The present invention provides a dynamic adaptive control system for auxiliary tunneling parameters of a hard rock TBM shield machine. It has the following beneficial effects: 1. This invention utilizes a TBM-geological environment digital twin, leveraging its built-in high-fidelity physical simulation model to deduce and predict short-term TBM excavation trends in real time. This allows foreseeing potential risks such as increased tool wear, abnormal TBM vibration, or machine jamming due to geological changes ahead. This enables the system to proactively optimize parameters based on these predictions, proactively avoiding risks and prioritizing excavation. This shifts from the passive model of traditional technologies that rely on real-time monitoring data for delayed adjustments. This effectively improves overall excavation efficiency while ensuring construction safety.

[0018] 2. This invention utilizes a multimodal deep learning collaborative decision-making module to deeply integrate heterogeneous data from multiple sources, including images, sequences, and structured and unstructured text, to form a comprehensive and profound understanding of the current tunneling conditions. This overcomes the limitations of traditional methods that rely solely on a subset of numerical parameters for decision-making. Furthermore, its built-in conflict resolution and multi-objective optimization network, based on the Pareto optimization concept, scientifically balances multiple conflicting objectives, such as tunneling efficiency, tool life, and construction safety, to generate an optimal equilibrium solution. This makes tunneling decisions more accurate and intelligent, and adapts to complex and changing engineering requirements.

[0019] 3. This invention innovatively introduces a cross-project knowledge sharing and privacy protection mechanism based on federated learning, addressing the technical challenge of cross-project knowledge sharing in the TBM industry due to data sensitivity. Through distributed collaborative training, the system can aggregate excavation experience from multiple projects to optimize the global model without accessing the original data from each project. This greatly enriches the diversity of the model's training data, enabling it to learn more universal patterns. This significantly improves the system's adaptability and robustness when facing different geological conditions or new projects, enabling continuous model evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a block diagram of the overall system architecture of the present invention; Figure 2 This is the core working flow diagram of the system of the present invention; Figure 3 This is a schematic diagram of the deployment of high-precision multi-source sensing network sensors of the present invention; Figure 4 This is a structural block diagram of the multimodal deep learning collaborative decision-making module of the present invention; Figure 5 It is a schematic diagram of the failure mode spectrum structure of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the specification of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] Please see the attached Figure 1 -Attached Figure 5 The embodiment of the present invention provides a hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system, comprising: Please see the attached Figure 3 , a high-precision multi-source sensing network for real-time acquisition of hard rock TBM operating parameters, tool status parameters, and geological environment parameters; This network is the perception foundation of the entire system. The comprehensiveness, real-time nature, and accuracy of its data collection directly determine the accuracy of subsequent predictions from the TBM-geological environment digital twin and the effectiveness of decisions made by the multimodal deep learning collaborative decision-making module. The network aims to build a comprehensive and reliable data collection and processing system for the system through an integrated and collaborative approach.

[0023] In this embodiment, the high-precision multi-source sensing network 100 achieves comprehensive coverage of tunneling site information by integrating multiple types of sensors. These sensors work together to form an organic data sensing system.

[0024] For the operation parameters of TBM, the system deploys conventional industrial sensors such as hydraulic sensors, encoders, flow meters, etc. They are used to obtain the core excavation status data of TBM in real time, such as main thrust , cutter head torque , Cutter head speed , penetration and grouting pressure Etc. These data are collected continuously in the form of time series.

[0025] To accurately monitor subtle changes in the tool's state under complex geological conditions, in some embodiments, a fiber Bragg grating (FBG) sensor array is embedded within the cutterhead's key cutters or their supports. Because the resonant wavelength of a fiber Bragg grating drifts with stress and temperature, monitoring this wavelength drift with a demodulator can accurately calculate the dynamic axial force the cutter is experiencing during rock breaking. and radial force , and the local strain At the same time, a set of high-frequency accelerometers will be installed near key vibration sources such as the TBM cutterhead and main drive to collect multi-channel vibration signals. ,These signals are crucial for analyzing the global vibration mode of the TBM and the abnormal ,condition of the hob (such as chipping and wear).

[0026] In one possible implementation, to provide visibility into the unknown geological environment ahead, the system integrates an advanced geological prediction system based on seismic wave reflection (e.g., horizontal seismic profiling (HSP) or continuous forward focusing (CFC)). This system periodically generates seismic waves near the tunnel face or utilizes the TBM's own excavation vibrations. Highly sensitive geophone arrays positioned on the shield receive these reflected signals at different geological interfaces (e.g., interfaces between soft and hard rock, fault fracture zones, and water-rich zones). By performing specialized migration imaging and velocity inversion on the collected seismic wave signals, three-dimensional geological image data is generated that intuitively reflects the stratigraphic structure ahead, the distribution of discontinuities, or the extent of water-rich areas.

[0027] In addition, several acoustic emission (AE) sensors can be attached to the shield. These sensors are very sensitive to transient elastic waves released by micro-fractures inside the material and can capture micro-fracture signals of the surrounding rock under stress, thereby being used to evaluate the stability of the surrounding rock in real time.

[0028] Generally speaking, raw data streams collected directly from various sensors inevitably contain random noise, signal drift, or outliers. If used directly without processing, they will seriously affect the accuracy of subsequent analysis. Therefore, the high-precision multi-source sensing network 100 in this embodiment also includes a data preprocessing unit.

[0029] This preprocessing unit uses a multi-stage process to process data. For vibration or acoustic emission signals mixed with high-frequency noise, wavelet transform can be used for denoising. This method decomposes the original signal into different scales, processes the detail coefficients representing noise by setting an appropriate threshold, and then uses the processed coefficients for wavelet reconstruction, thereby filtering out noise while preserving the main signal characteristics.

[0030] For TBM operating parameter time series that exhibit random fluctuations and slow drift, filtering algorithms can be applied for smoothing. Alternatively, the system can employ an exponentially weighted moving average (EWMA) filter, which assigns greater weight to recent data while balancing signal smoothness with sensitivity to change. The calculation formula is as follows: ; in: Represents the time step The filtered output value of Represents the time step The original observation value of Represents the time step The filtered output value of Represents the smoothing factor, which is a constant between 0 and 1. The value of determines the weight relationship between new observations and historical filtered values. A larger value results in a faster response to new data, while a smaller value results in a better smoothing effect. This value can be set empirically or adjusted adaptively based on the noise level of the specific signal.

[0031] For data cleaning, the system identifies and marks outliers in the data stream by setting reasonable thresholds based on physical constraints or using statistical methods (such as the 3-sigma rule). For flagged outliers or missing values ​​due to signal transmission interruptions, linear interpolation, spline interpolation, or regression models based on adjacent time steps can be used to fill in missing values, depending on the data characteristics, to ensure data integrity.

[0032] As the foundation for data fusion applications, all preprocessed data streams are ultimately time-aligned and synchronized based on a unified, high-precision system timestamp. This ensures that at any given moment, the system can obtain a multidimensional state vector with a completely consistent time section, providing high-quality, highly reliable data input for the subsequent TBM-geological environment digital twin and multimodal deep learning collaborative decision-making modules.

[0033] Sensor data can be transmitted using industrial Ethernet protocols such as PROFINET or EtherCAT to ensure real-time and highly reliable data transmission. Edge computing nodes can be deployed on the TBM equipment to perform the aforementioned data preprocessing tasks. This not only reduces the subsequent computing pressure on the cloud or central server, but also effectively reduces the bandwidth required for long-distance data transmission.

[0034] The TBM-geological environment digital twin is used to receive data collected by the high-precision multi-source sensing network and, based on a built-in high-fidelity physical simulation model, to perform real-time deduction and prediction of short-term trends in TBM excavation, outputting prediction data including predicted excavation effects and potential risks; This module receives sensory data as input and aims to build a virtual replica that is synchronized with the physical tunneling system in real time and can perform advance deduction, thereby transforming tunneling control from a passive response mode to an active prediction mode.

[0035] The TBM-geological environment digital twin 200 is a complex system that integrates multi-physics coupled simulation, real-time data assimilation, and dynamic environmental modeling. At its core is a high-fidelity physical simulation model that virtually replicates the complete TBM excavation process in an underground environment.

[0036] Typically, the construction of this high-fidelity physical simulation model involves coupling multiple numerical simulation methods. Specifically, the system uses the discrete element method (DEM) to model the microscopic interaction between the cutter and the rock on the cutterhead. This method discretizes the rock mass into a collection of particles with specific physical and mechanical properties. By calculating the contact mechanics model between the particles (such as the Hertz-Mendes contact model), the rock fragmentation, spalling, and crack propagation processes under the cutter's cutting are simulated. Furthermore, the cutter wear process can be modeled based on Archil's wear law, taking into account factors such as the contact stress, relative sliding velocity, and material hardness of the cutter and rock.

[0037] Finite element analysis (FEA) is used to model the stress and deformation responses of the TBM's macrostructures, including the cutterhead, main drive, and shield. By dividing the TBM's key structural components into a finite number of elements, a mechanical response model is established under complex loads (such as cutterhead torque, thrust, and surrounding rock pressure). The DEM and FEA models are bidirectionally coupled via a load transfer interface, fully simulating the force-work transmission chain from rock crushing to the overall structural response of the machine.

[0038] A static simulation model cannot cope with the dynamic changes of a construction site. Therefore, the digital twin in this embodiment has the ability to map data and correct models in real time. It continuously "feeds" the high-fidelity physical simulation model with real-time data streams from the high-precision multi-source perception network 100 through data assimilation technology. As an option, the system uses a Kalman filter algorithm framework to continuously correct the state vector of the simulation model online to eliminate model errors and uncertainties and ensure high consistency between the virtual model and the physical entity state. The core equation for its state update can be expressed as: ; in: Represents the system at time step The optimal estimation vector of the real state, which may include multiple state variables such as the cutter position, speed, force, TBM posture, etc. Represents time-step-based The state of the current time step A priori predicted value of the state, which is derived from a high-fidelity physical simulation model; Represents the time step The actual observation vector provided by the high-precision multi-source perception network 100 includes real measurement values ​​such as TBM operating parameters and tool forces; Represents the observation matrix, which transforms the state vector Linearly map from state space to observation space so that the predicted value can be compared with the actual observed value Make comparisons; Represents the Kalman gain matrix, whose core role is to weigh the prior prediction value The reliability and actual observation value When the observation noise is small, the gain The larger the value, the more the system trusts the actual observations; otherwise, the more it trusts the model's predictions.

[0039] In one possible implementation, the dynamic changes in the geological environment are modeled by integrating geological sensor data. Specifically, forward geological image data acquired by the High Prediction System (HSP / CFC) is processed and interpreted to construct and update a three-dimensional geological attribute model ahead of the TBM excavation face in real time. This model incorporates information such as rock mass structure, faults, fissures, and interfaces between soft and hard rock. Simultaneously, data on surrounding rock microfracture activity detected by acoustic emission (AE) sensors serves as a dynamic indicator for evaluating the real-time stability of the surrounding rock and is fed back into the simulation model to adjust its settings for surrounding rock mechanical parameters and boundary conditions.

[0040] Through this approach, the TBM-geological environment digital twin 200 first synchronizes with the physical world and calibrates its model during each computational cycle. It then utilizes its built-in high-fidelity physical simulation model to perform high-speed simulations. This simulation process aims to predict the TBM's potential excavation performance and potential risks over a short period of time or excavation distance, given the currently configured excavation parameters. Ultimately, the module outputs a structured set of prediction data, including but not limited to: predicted penetration rate PR,predPR,pred, predicted tool wear rate Wtool,predWtool,pred, predicted TBM vibration energy level VTBM,predVTBM,pred, and a quantified jam risk index Rjam,predRjam,pred. This forward-looking prediction data serves as key input to the subsequent multimodal deep learning collaborative decision-making module 300, providing a scientific basis for its intelligent decision-making.

[0041] As an extension of this technology, given the computational complexity of high-fidelity physical simulation models, they can be deployed on dedicated high-performance computing (HPC) servers or cloud platforms. Data assimilation calculations can be performed on edge computing nodes close to the TBM, forming a "cloud-edge collaborative" computing architecture that balances computational accuracy with real-time response requirements.

[0042] Please see the attached Figure 4 , a multimodal deep learning collaborative decision-making module is used to receive the prediction data output by the TBM-geological environment digital twin and the real-time data collected by the high-precision multi-source perception network, and generate a set of dynamically adaptive excavation parameter combinations by performing feature fusion and multi-objective optimization on the multimodal data.

[0043] After the TBM-geological environment digital twin 200 completes its forward-looking predictions of the tunneling process, the system obtains valuable structured information about future trends. However, predictive data alone is insufficient for optimal decision-making; it requires deep integration and collaborative analysis with real-time, multi-dimensional sensory data from the high-precision multi-source perception network 100. This critical task is handled by the multimodal deep learning collaborative decision-making module 300. This module serves as the system's "decision-making brain," designed to learn complex mapping relationships from massive, heterogeneous data, resolve inherent conflicts between multiple objectives during tunneling, and ultimately generate scientific and dynamic tunneling parameter combinations.

[0044] In this embodiment, the core functions of the multimodal deep learning collaborative decision-making module 300 are jointly implemented by two closely connected sub-networks: a multimodal feature fusion network, and a target conflict resolution and multi-objective optimization network.

[0045] First, the data is fed into a multimodal feature fusion network. This network is designed to effectively integrate data from diverse sources and formats to form a unified, high-dimensional feature representation that comprehensively and accurately describes the current tunneling conditions. Typically, this network employs a parallel multi-branch deep neural network architecture, with each branch specialized for processing a single data modality.

[0046] Specifically, one branch uses a convolutional neural network (CNN) to process image data. For example, the forward geological profile obtained from the Advanced Geological Prediction System (HSP / CFC) and rock mass images obtained from the tunneling face vision sensor Through multi-layer convolution and pooling operations, the spatial structural features contained therein, such as stratum boundaries, crack directions and rock integrity, can be effectively extracted.

[0047] The other branch uses network models that can process sequence data, such as long short-term memory networks (LSTM), gated recurrent units (GRU), or transformer models, to process time series data. For example, the high-frequency vibration signal sequence of a tool Harmony acoustic emission signal sequence ,Through this branch, we can capture its dynamic evolution pattern and contextual dependencies in the time and frequency domains.

[0048] At the same time, a branch consisting of a multi-layer perceptron (MLP) is responsible for processing structured numerical data. This data includes real-time TBM operating parameters (such as etc.), and prediction data from the TBM-geological environment digital twin 200 (e.g. wait).

[0049] In one possible implementation, in order to make full use of expert knowledge in historical documents such as geological survey reports, the system may also include a branch based on a pre-trained natural language processing (NLP) model (such as BERT) to extract deep semantic features from text descriptions.

[0050] The modality-specific feature vectors extracted by each branch are not simply concatenated. Alternatively, this embodiment employs an attention mechanism to adaptively weight these features. This mechanism dynamically learns and assigns different weights to different data modalities based on the current working conditions, highlighting key information and suppressing interference from irrelevant or noisy information. The attention-weighted fusion process can be described by the following formula: ; in: Represents the high-dimensional feature vector generated by the final fusion; Represents the set of all data modalities, e.g. ; Representative The feature vector extracted by the modal branch network; Representatives assigned to The attention weight of each modal feature is calculated based on all modal features and is calculated as follows: ; ; in, It is The attention scores of the modalities, is a trainable weight matrix, is the bias vector, is a trainable context vector. In this way, the network can independently determine whether the features of the geological image are more important or the features of the vibration sequence are more indicative at the current moment, thereby achieving highly intelligent information screening and fusion.

[0051] In order to obtain high-dimensional fusion features that fully characterize the working conditions This feature is then input into a target conflict resolution and multi-objective optimization network. TBM excavation is inherently a multi-objective optimization problem. For example, simply pursuing high excavation efficiency (high penetration rate) can increase tool wear, energy consumption, and even lead to safety risks such as machine jams. This network is designed to resolve these conflicts and trade-offs among multiple objectives.

[0052] Generally speaking, the network is built based on the concept of Pareto Optimization. Its goal is not to find a "unique solution" that is optimal in all objectives (which usually does not exist), but to learn a solution that is optimal for all objectives based on the current working conditions. To a Pareto Optimal Set Any solution (i.e., a set of tunneling parameters) in this solution set cannot improve the performance of a certain objective without compromising the performance of at least one other objective.

[0053] In one possible implementation, the network can be implemented using a multi-task learning architecture. It then uses a shared input to generate multiple parallel output heads, each corresponding to a specific optimization objective (e.g., predicting tunneling efficiency, tool life, safety index, etc.). By simultaneously optimizing the loss functions for all tasks during training, the network learns the inherent connections and constraints between these different objectives.

[0054] As an extended technical content, this multi-objective optimization problem can also be modeled as a multi-objective reinforcement learning (MORL) problem. In this framework, the system regards the TBM as an agent, whose action is to select the excavation parameters, and the state is the fused features. , while the reward is a vector that encompasses multiple objectives (such as efficiency, loss, and safety). By training a policy network, the agent learns which actions to take in different states to maximize the long-term accumulated vectorized reward. The optimal policy learned naturally corresponds to the Pareto optimal solution.

[0055] Ultimately, the network outputs not a single tunneling parameter setting, but a set of Pareto optimal tunneling parameter combinations. , each of which Each of these is an effective parameter solution that has been scientifically weighed under the current working conditions. This solution set is then passed to the parameter dynamic adaptation and adaptive learning module 600 for final selection and execution based on actual engineering preferences.

[0056] Please see the attached Figure 5 , a fault mode map, which stores the fault types, associated geological conditions, excavation parameters, sensor abnormal modes and treatment measures in the historical excavation process in a graph structure.

[0057] When the multimodal deep learning collaborative decision-making module 300 makes intelligent decisions based on a data-driven approach, its decision-making process primarily relies on implicit associations learned from the data. To further enhance the robustness and interpretability of system decisions and incorporate long-term accumulated engineering experience and expert knowledge, the present invention introduces a failure mode map 400. This module, as an explicit knowledge base, complements the deep learning model and can provide critical prior knowledge guidance, particularly when dealing with low-probability, high-risk abnormal operating conditions, enabling knowledge-based risk intervention.

[0058] In this embodiment, the failure mode graph 400 is a structured, dynamically evolving knowledge network. It is stored and managed in the form of a graph database, aiming to transform the fragmented, unstructured failure experience and causal relationships during TBM excavation into a format that is understandable and inferential for machines.

[0059] Specifically, the graph consists of different types of nodes and edges. Nodes represent various entity concepts during tunneling. For example, node types may include: fault type nodes (e.g., "machine jam," "cutter blade breakage," "shield tail water leakage"); geological condition nodes (e.g., "uneven soft and hard strata," "high quartz content rock mass," "water-rich fault zone"); sensor anomaly mode nodes (e.g., "periodic sharp increase in cutterhead torque," "characteristic frequency in the main drive vibration spectrum," "sudden drop in tunnel face pressure"); tunneling parameter status nodes (e.g., "combination of high thrust and low speed," "excessive overshoot"); and recommended action nodes (e.g., "immediately reduce thrust and reverse the cutterhead," "adjust synchronous grouting parameters").

[0060] Edges represent specific relationships between these entity concepts. For example, a causes edge might connect the "water-rich fault zone" node with the "machine stuck" node; a manifests_as edge might connect the "machine stuck" node with the "periodic sharp increase in cutterhead torque" node; and a recommends edge might connect the "machine stuck" node with the "immediately reduce thrust and reverse the cutterhead" node. In this way, a complete knowledge subgraph is constructed around a typical fault, encompassing its causes, manifestations, associated parameters, and recommended countermeasures.

[0061] The construction of this failure mode map isn't a one-time process; it's a continuous, iterative one. Its initial knowledge comes from mining historical project data, extracting text from geological survey reports, and collaborating with experienced TBM operators. More importantly, through the subsequent dynamic parameter adaptation and adaptive learning module 600, it automatically or semi-automatically converts new failure events and successful handling experiences during the current tunneling process into new nodes and edges, continuously expanding and updating this knowledge base.

[0062] In one possible implementation, the system uses this graph for real-time risk identification and reasoning. As the multimodal deep learning collaborative decision-making module 300 processes the real-time data stream, the extracted features (such as specific sensor anomaly patterns) are used as queries for pattern matching within the failure mode graph 400. To enable deeper reasoning, such as predicting potential failure chains or identifying root causes, the system can apply a graph neural network (GNN) to learn from the entire graph.

[0063] As an option, the system can deploy a graph convolutional network (GCN) model. This model can aggregate the information within the neighborhood of each node, thereby learning a low-dimensional vector representation (Embedding) rich in structural information for each node in the graph. In a GCN model, from the first Layer to The node feature propagation and update rules of the layer can be expressed as follows: ; in: It is The feature matrix of all nodes in the layer, each row of the matrix is ​​the feature vector of a node in the layer; It is the initial feature of each node in the graph.

[0064] It is A layer is a trainable weight matrix, which is the parameter that the neural network needs to learn.

[0065] is the graph adjacency matrix with self-loops added, is the original adjacency matrix, is the identity matrix. Self-loops are added to consider the node's own information when transmitting information.

[0066] yes The diagonal matrix whose diagonal elements .

[0067] It is an operation to symmetric normalize the adjacency matrix to prevent the scale of node features from changing dramatically during the aggregation process.

[0068] Is a non-linear activation function, such as ReLU (Rectified Linear Unit).

[0069] It is the new node feature matrix obtained by updating after a layer of graph convolution.

[0070] By stacking several such graph convolutional layers, the GNN model can capture complex dependencies. Based on the trained GNN model, the system can efficiently perform fault diagnosis (node ​​classification tasks) or causal inference (link prediction tasks).

[0071] When the GNN inference results indicate that the current operating condition matches a high-risk failure mode in the graph (such as a "stuck machine") to a degree exceeding a preset threshold, the system triggers an adaptive decision topology reconstruction mechanism. This mechanism does not directly change the hardware connections or software code structure, but rather dynamically intervenes in the decision logic of the multimodal deep learning collaborative decision module 300.

[0072] Specifically, the reconstruction can take the following forms: First, directly adjust the weights of the optimization objectives within the target conflict resolution and multi-objective optimization network. For example, when a high risk of machine jamming is identified, the system will automatically and significantly increase the weight of the "construction safety" goal, while correspondingly reducing the weight of the "excavation efficiency" goal, thereby guiding the system to prioritize the generation of excavation parameter combinations that prioritize risk avoidance. Second, the risk type or risk level inferred by the GNN is used as a high-weight "knowledge feature" and is additionally input into the multi-objective optimization network to provide it with clear risk-oriented information. In this way, the system organically combines data-driven deep learning with knowledge-driven expert systems, significantly improving the reliability of decision-making under complex and dangerous working conditions.

[0073] A cross-project knowledge sharing and privacy protection mechanism based on federated learning is used to continuously optimize the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module through distributed collaborative training while protecting the privacy of the original data of each TBM project. The performance and reliability of the aforementioned modules, such as the TBM-geological environment digital twin 200, the multimodal deep learning collaborative decision-making module 300, and the fault mode map 400, are highly dependent on the scale and diversity of the training data. The data accumulated by a single TBM project has inherent limitations in terms of both quantity and the complexity of the working conditions covered, which may lead to insufficient generalization capabilities of the trained model. However, in actual engineering, due to its high commercial sensitivity and security, excavation data between different projects are difficult to share directly in a centralized manner. To address this technical bottleneck, the present invention further provides a cross-project knowledge sharing and privacy protection mechanism 500 based on federated learning, which aims to achieve model collaborative optimization and knowledge sharing among multiple projects without transmitting the original data.

[0074] In this embodiment, the mechanism establishes a distributed collaborative training architecture based on a "central server and multiple clients." The central server maintains and distributes the global model, while each TBM project deploying the system acts as an independent client. The core concept of this mechanism is "data remains static, model remains dynamic." This means that the raw tunneling data remains local to each project, while only the model parameters, stripped of their identity, are transmitted and aggregated across the network.

[0075] A complete cross-project collaborative training and knowledge sharing process can be broken down into the following sequential steps: First, the central server initializes a global model. This global model can be a neural network model in the multimodal deep learning collaborative decision module 300, or a machine learning model used for calibration in the TBM-geological environment digital twin 200. The central server sets the initial parameters of this global model to (in is the current training round) is distributed to all clients (i.e., TBM projects) selected to participate in this round of training.

[0076] Subsequently, each participating client After receiving the global model parameters, it uses its locally stored, private tunneling dataset with privacy The model is trained locally. Local training can be performed for one or more epochs, the purpose of which is to allow the model to learn the unique working condition characteristics and knowledge contained in the project data. After the training is completed, each client You will get a model parameter update after "fine-tuning" the local data During this process, the original data, such as geological radar images and TBM operation parameter sequences, will never leave the local computing environment where the client is located.

[0077] Then, each client updates the model parameters obtained through local training Upload to the central server. It is important to emphasize that the uploaded content is only the model parameters such as weights and biases, and not any raw data related to specific tunneling events.

[0078] After collecting model parameter updates from multiple clients, the central server performs aggregation operations to generate a new global model with better performance and stronger generalization ability. As an option, the system uses the Federated Averaging (FedAvg) algorithm for weighted aggregation. This aggregation process ensures that clients with larger data volumes can make greater contributions to the update of the global model. The aggregation formula is as follows: ; in: Representatives in the The global model parameters are updated after round aggregation; Represents the index of the client; Represents the total number of clients participating in this round of aggregation; Representing the client Model parameters uploaded after local training; Representing the client The number of data samples used for local training; Represents the total number of data samples of all clients participating in this round of aggregation, that is, .

[0079] Finally, the central server aggregates the newly obtained global model parameters The model is then redistributed to each client to replace their local models. This "distribution-local training-upload-aggregation" process is iterated continuously until the global model's performance converges or the preset number of training rounds is reached. In this way, the global model can integrate experience from different geological conditions and construction projects without exposing any party's sensitive data.

[0080] To further enhance privacy protection, this mechanism can also integrate other privacy-enhancing technologies. Specifically, before clients upload model parameter updates, they can apply differential privacy. This technology injects precisely calculated random noise into the model parameters, making it impossible to infer any specific data from any individual client from the aggregated global model, thereby providing mathematically provable privacy guarantees.

[0081] For applications requiring extremely high security, homomorphic encryption or secure multi-party computation protocols can be introduced. In these frameworks, the parameters uploaded by the client are encrypted, allowing the central server to perform aggregation operations directly on the ciphertext without decryption. Only after aggregation is the final global model decrypted using a specific method. This ensures that even the central server cannot spy on the parameter update details of any individual client, achieving a higher level of privacy and security.

[0082] The parameter dynamic adaptation and adaptive learning module is used to select an optimal parameter combination from the dynamically adapted tunneling parameter combination to control the TBM in real time, and based on the actual tunneling effect, feedback updates the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module to form a closed-loop adaptive learning.

[0083] After the system generates a Pareto-optimal set of tunneling parameter solutions through a combination of data-driven and knowledge-guided approaches and continuously optimizes its global model using a federated learning mechanism, a module is ultimately required to execute these decisions and learn from the results, thus forming a complete, self-evolving intelligent closed loop. This critical "execution-feedback-learning" function is handled by the Dynamic Parameter Adaptation and Adaptive Learning Module 600. It serves as the bridge between the system's intelligent decision-making and the actual tunneling in the physical world, and is the ultimate guarantee for the system's continuous evolution.

[0084] In this embodiment, the parameter dynamic adaptation and adaptive learning module 600 receives the Pareto optimal solution set output from the multimodal deep learning collaborative decision module 300 The primary task of this module is to select the optimal combination of tunneling parameters from this set of solutions that include multiple effective trade-offs. For execution.

[0085] Typically, this selection process can be done automatically based on a pre-set engineering strategy. Alternatively, the system can use a utility function For each parameter combination in the solution set Score and select the solution with the highest utility value. The utility function is the weighted sum of multiple optimization objectives and can be defined as: ; in: Represents the Pareto optimal solution set The best excavation parameter combination is finally selected; Representative solution set A specific parameter combination in , such as ; Is a combination of parameters The comprehensive utility score of 、 and fl They are used to evaluate the parameter combination A standardized function of performance on three objectives: tunneling efficiency, construction safety and tool life, whose input is a combination of parameters , the output is the corresponding expected performance score. These functions can be directly provided by the multi-objective optimization network or calculated based on its prediction results; 、 and are weight coefficients corresponding to the three objectives of efficiency, safety and life. These weights can be preset and adjusted by engineers according to different stages of the project (for example, pursuing efficiency in stable formations and focusing on safety before fault zones), and meet the requirements of .

[0086] The system also supports a "human-machine collaborative" decision-making model. This involves visualizing the Pareto front on the user interface (e.g., a 2D or 3D scatter plot), allowing experienced TBM operators to interactively select the optimal solution from this set of solutions based on actual site conditions and their own experience.

[0087] Select the best parameter combination After that, the module parses it into specific control instructions and sends them to the TBM's PLC (programmable logic controller) system through industrial field bus protocols such as PROFINET. It can adjust the TBM's thrust, cutterhead speed, cutterhead torque and other actuators in real time to achieve dynamic adaptive control of the excavation process.

[0088] The core value of this module lies in its closed-loop adaptive learning capabilities. After parameter execution, the module continuously obtains actual tunneling performance data from the high-precision multi-source perception network 100 and uses this data as "true feedback" or "supervision signal" to reversely empower other modules in the system.

[0089] Specifically, the learning loop includes multiple feedback paths.

[0090] The first feedback path is to the TBM-geological environment digital twin 200. The module converts the actual observed tunneling effect (such as the actual penetration) , real tool vibration etc.) and the predicted values ​​previously made by the digital twin ( etc.) and calculate the prediction error This error is used to drive an optimization algorithm (such as gradient descent) to adjust key physical parameters (such as the compressive strength and abrasiveness coefficient of the rock) in the high-fidelity physical simulation model, so that the predictions of the digital twin are increasingly close to the real physical process.

[0091] The second feedback path points to the multimodal deep learning collaborative decision module 300. The module will complete the "working condition-decision-result" triple, that is, (the fused feature , selected parameters , actual excavation effect ) as a high-quality, labeled new training sample and stored in the system's experience database. These continuously accumulated new samples can be used to periodically incrementally train or fine-tune the deep neural network in the decision-making module, allowing the decision-making model to continuously learn from the latest excavation experience and continuously improve its decision-making accuracy.

[0092] The third feedback path leads to the Failure Mode Spectrum 400. When a new failure event or risk indicator occurs during excavation that is not already recorded in the Spectrum, this module records a complete data snapshot prior to the event, the excavation parameters used at the time, and the operator or system's final, verified effective action. This information is structured and converted into new nodes and relationships, which are added to the Failure Mode Spectrum automatically or after expert confirmation, thus enabling the self-expansion and evolution of the knowledge base.

[0093] As an extended technical feature, the update frequency of this adaptive learning process is adjustable. The system can be configured for online learning mode, where small model parameter updates are performed at each decision cycle for the fastest response time. Alternatively, it can be configured for batch learning mode, where a certain number of new samples are accumulated before a centralized training session is conducted, achieving more stable model convergence and conserving computing resources. Through this complete closed-loop adaptive learning mechanism, the entire system acts like an organism with memory and learning capabilities, continuously accumulating experience and knowledge during the tunneling process, thereby achieving continuous iteration and evolution of performance.

[0094] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system, characterized in that: include: High-precision multi-source sensing network for real-time acquisition of hard rock TBM operating parameters, tool status parameters, and geological environment parameters; The TBM-geological environment digital twin is used to receive data collected by the high-precision multi-source sensing network and, based on a built-in high-fidelity physical simulation model, to perform real-time deduction and prediction of short-term trends in TBM excavation, outputting prediction data including predicted excavation effects and potential risks; A multimodal deep learning collaborative decision-making module is used to receive the predicted data output by the TBM-geological environment digital twin and the real-time data collected by the high-precision multi-source perception network, and generate a set of dynamically adaptive tunneling parameter combinations by performing feature fusion and multi-objective optimization on the multimodal data; The parameter dynamic adaptation and adaptive learning module is used to select an optimal parameter combination from the dynamically adapted tunneling parameter combination to control the TBM in real time, and based on the actual tunneling effect, feedback updates the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module to form a closed-loop adaptive learning.

2. A hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1, characterized in that: The high-fidelity physical simulation model built into the TBM-geological environment digital twin is constructed based on finite element analysis and discrete element method, and can simulate the TBM tool rock breaking process, cutterhead force distribution and tool wear evolution; and the TBM-geological environment digital twin uses the real-time data collected by the high-precision multi-source perception network to perform online correction and status update of the high-fidelity physical simulation model.

3. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1 is characterized in that: The multimodal data fused by the multimodal deep learning collaborative decision-making module includes at least: structured and serialized prediction data from the TBM-geological environment digital twin, and image data and unstructured text data from the high-precision multi-source perception network.

4. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1 is characterized in that: The multimodal deep learning collaborative decision-making module realizes the feature fusion through a multimodal feature fusion network. The multimodal feature fusion network adopts a multi-branch deep neural network architecture and uses an attention mechanism to perform weighted fusion of features of different modalities to form a comprehensive high-dimensional feature representation.

5. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1 is characterized in that: The multimodal deep learning collaborative decision-making module generates the dynamically adaptive tunneling parameter combination through target conflict resolution and multi-objective optimization network. The network is based on the Pareto optimization idea, is used to weigh multiple conflicting objectives, and outputs a set of Pareto optimal tunneling parameter solutions.

6. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1 is characterized in that: The system also includes a fault mode map, which stores fault types, associated geological conditions, excavation parameters, sensor abnormal modes and treatment measures in the historical excavation process in a graph structure.

7. A hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 6, characterized in that: The multimodal deep learning collaborative decision-making module is also used to perform adaptive decision topology reconstruction based on the reasoning results of the fault mode map when an abnormal mode similar to any fault mode in the fault mode map is identified, so as to dynamically adjust its decision weight and parameter adjustment priority.

8. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 1 is characterized in that: The system also includes a cross-project knowledge sharing and privacy protection mechanism based on federated learning, which is used to continuously optimize the TBM-geological environment digital twin and the multimodal deep learning collaborative decision-making module through distributed collaborative training while protecting the privacy of the original data of each TBM project.

9. A hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 8, characterized in that: The specific implementation method of the cross-project knowledge sharing and privacy protection mechanism based on federated learning is as follows: each TBM project uses its private data as a client to train the model locally, and only uploads the updated model parameters to the central server for aggregation to generate a global model. The central server then distributes the global model to each client to update its local model.

10. The hard rock TBM shield machine auxiliary excavation parameter dynamic adaptive control system according to claim 6, characterized in that: The specific method in which the parameter dynamic adaptation and adaptive learning module performs feedback updates based on actual excavation results is as follows: actual excavation effect data is fed back to the TBM-geological environment digital twin to correct its high-fidelity physical simulation model, and fed back to the multimodal deep learning collaborative decision-making module to optimize its decision-making ability. Moreover, when a new fault event occurs, the event and its handling method are updated to the fault mode map.

Images