Real-time intelligent prediction method for tunneling thrust of shield tunneling machine based on time sequence deep learning neural network

Abstract

The invention relates to the technical field of shield tunneling machine tunneling thrust prediction, in particular to a shield tunneling machine tunneling thrust real-time intelligent prediction method based on a time sequence deep learning neural network. The method comprises the steps that tunneling parameters, geomechanical parameters and tunnel geometric parameters are collected in real time through a multi-source sensor; performing denoising preprocessing on the monitoring data by adopting wavelet transform, retaining a large-scale low-frequency coefficient, and performing threshold quantization reconstruction on a high-frequency coefficient; a graph convolutional neural network and long and short term memory network fusion model is constructed, a multi-source parameter dynamic heterogeneous graph is established by using GCN to extract spatial features, and a time sequence rule is mined through an LSTM gating mechanism; predicting the thrust of the next time step by using the time step sequence, carrying out incremental training every 24 hours, and optimizing by using a root-mean-square error as a loss function through an Adam algorithm; real-time reasoning output is performed through edge computing equipment, and dynamic display is performed through a visual interface. According to the method, the accuracy, the real-time performance and the system generalization ability of thrust prediction under complex geological conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine thrust prediction technology, and more specifically to a real-time intelligent prediction method for tunnel boring machine thrust based on a temporal deep learning neural network. Background Technology

[0002] In the field of shield tunnel construction, the precise dynamic control of the tunnel boring machine's thrust is directly related to construction safety, tunneling efficiency, and tunnel formation quality. However, under complex and variable geological conditions, the shield tunneling process faces multiple extreme challenges, including: the real-time dynamic redistribution of ground stress during tunneling, the strong nonlinear coupling characteristics of the interaction between the soil and rock mass and the cutterhead-shield system, and the complexity of the multi-parameter spatiotemporal coupling relationship between the propulsion system and the geological environment.

[0003] As a core indicator of the tunnel boring machine's (TBM) operational status, accurate prediction of tunneling thrust is a prerequisite for intelligent optimization of construction parameters, proactive avoidance of major risks such as shield jamming and machine shutdown, and soil chamber gushing and depressurization. Insufficient thrust will directly lead to tunneling stagnation or even the cutterhead being locked by the strata, causing serious delays in the construction period and huge cost overruns; excessive thrust can easily cause a chain of engineering accidents such as structural damage to the segment ring, joint sealing failure, excessive surface settlement, and even overload and collapse of the hydraulic propulsion system. Especially when traversing complex geological units such as alternating soft and hard strata, fault fracture zones, highly permeable sand and gravel layers, or water-rich soft strata, the drastic spatial variability and temporal unpredictability of the physical and mechanical properties of the soil and rock make the interaction between the cutterhead-shield system and the surrounding rock exhibit strong state dependence and path memory effects, resulting in thrust signals showing significant non-stationary and non-Gaussian fluctuation characteristics.

[0004] Traditional empirical formulas and static mechanical models reveal fundamental flaws under such dynamic working conditions.

[0005] First, these methods rely excessively on idealized assumptions, such as uniform and continuous strata and constant advance speed, completely ignoring the complex spatiotemporal synergistic effects of multiple parameters, including the real-time evolution of geomechanical parameters, the dynamic response of rock and soil under cutterhead cutting disturbance, the hysteresis effect of hydraulic system pressure transmission, and the feedback mechanism of synchronous grouting pressure on strata reaction force.

[0006] Secondly, the raw data collected by the densely deployed sensor network on site is severely contaminated by high-intensity background noise, such as high-frequency noise from hydraulic pulsation, rock-breaking impact vibration from the cutterhead, and electromagnetic radiation interference from high-power motors. Traditional filtering methods are unable to effectively separate noise from real signals, resulting in severe distortion of key dynamic characteristics.

[0007] Even more challenging is that tunnel boring machine (TBM) construction is a continuous process of spatiotemporal evolution, with geological conditions changing in real time as the tunneling progresses. However, existing prediction models generally lack online adaptability and cannot respond promptly to sudden geological anomalies such as lithological changes, intrusion of isolated rock clusters, or unidentified faults, resulting in a precipitous decline in prediction accuracy as the tunneling progresses.

[0008] Numerous engineering cases have shown that thrust prediction errors not only directly induce engine jamming accidents, but are also more likely to trigger secondary disasters such as surface subsidence, rupture of adjacent pipelines, and instability of building foundations through a chain reaction of ground disturbances.

[0009] Current shield tunneling thrust prediction technology suffers from multiple systemic bottlenecks, severely limiting its engineering applicability and reliability. At the data sensing and preprocessing level, existing sensor deployment schemes exhibit significant structural flaws: Hydraulic thrust monitoring relies heavily on single-point pressure transmitters, which cannot fully capture the dynamic pressure distribution gradient and synergistic mechanism of the six propulsion hydraulic cylinders under non-uniform loads, and are easily affected by random noise interference generated by strong construction vibrations. The torque measurement of the cutter head generally adopts the motor current inversion method, which is easily affected by the harmonic pollution of the frequency converter switch and the electromagnetic crosstalk of the cable, resulting in systematic drift of the torque signal; The spatial coverage density of pressure monitoring points in the soil chamber is insufficient, making it difficult to accurately reflect the three-dimensional non-uniform distribution characteristics and local abnormal fluctuations of the pressure field inside the soil chamber.

[0010] The main technical problem with existing technologies is that the strong vibration and electromagnetic interference in the shield tunneling environment will generate a large amount of non-stationary and non-Gaussian noise. Traditional filtering methods (such as Kalman filtering or moving average) are seriously inadequate in suppressing the non-stationary and non-Gaussian noise generated under the strong vibration environment of shield tunneling, causing the feature extraction results to deviate from the physical essence.

[0011] At the level of predictive model architecture, mainstream methods have fundamental limitations. While physical models based on formation resistance theory are mechanistically interpretable, their accuracy is highly dependent on the completeness of advanced geological exploration data. However, drilling is costly and it is difficult to update geological parameters in real time as tunneling progresses. Statistical learning models can only establish static mapping relationships and cannot capture the significant time lag effect (typically a 3-5 second phase difference) between cutterhead torque changes and thrust response, nor can they model the long-range temporal dependencies between parameters. Even when using time-series models such as basic long short-term memory networks, although they can handle the historical evolution of thrust itself, they completely sever the inherent spatial topological relationships between tunneling parameters, such as the cross-system dynamic coupling of the spatial distribution of soil chamber pressure and synchronous grouting pressure, and the instantaneous spatial transmission effect of geological parameter abrupt changes on mechanical parameters. When the tunnel traverses soft and hard composite strata, the sudden cross-scale interaction between geological parameters (such as a sudden drop in cohesion) and mechanical parameters (such as an emergency adjustment of cutterhead speed) will directly trigger the collapse of model generalization performance because it is not effectively encoded.

[0012] At the system integration and real-time response levels, the existing architecture struggles to support closed-loop predictive control. Limited edge computing resources necessitate the remote transmission of massive amounts of raw data to the cloud for processing, with data transmission delays in long-distance tunnels far exceeding the millisecond-level response threshold for thrust control. The rigid model update mechanism is severely out of sync with the rapid time-varying characteristics of tunneling parameters, making online dynamic optimization of model parameters impossible. Visual interaction is equally weak; the control room only displays instantaneous scalar thrust values, lacking spatiotemporal comparative analysis of predicted trends and historical data. This forces drivers to rely on experience and intuition for decision-making in situations involving sudden changes in rock formations or obstacles, resulting in extremely high risks of delayed response.

[0013] Therefore, it is urgent to overcome the bottlenecks of multi-source heterogeneous perception, the obstacles of spatiotemporal correlation modeling, and the barriers to real-time system performance, and to build a shield tunneling thrust intelligent prediction technology system that integrates high-fidelity data acquisition, intelligent noise suppression, spatiotemporal feature fusion, edge real-time computing, and immersive visualization interaction. Summary of the Invention

[0014] In view of this, the present invention provides a real-time intelligent prediction method for tunnel boring machine thrust based on a temporal deep learning neural network. It aims to solve the problems of low prediction accuracy, poor generalization ability and weak engineering applicability caused by low perception quality of multi-source heterogeneous monitoring data, serious noise interference, insufficient model spatiotemporal correlation modeling ability, and lack of system real-time and adaptive capabilities in existing tunnel boring machine thrust prediction technology. This improves the prediction accuracy, real-time performance and reliability of tunnel boring machine thrust under complex geological conditions, and provides effective support for tunnel construction safety and intelligent control.

[0015] To achieve the above objectives, the present invention adopts the following technical solution: A real-time intelligent prediction method for tunnel boring machine thrust based on temporal deep learning neural networks includes the following steps: Real-time acquisition of multi-source heterogeneous monitoring data during the tunnel boring machine excavation process, including excavation parameters, geomechanical parameters, and tunnel geometric parameters; Wavelet transform is used to perform denoising preprocessing on the multi-source heterogeneous monitoring data to obtain denoised feature data. The denoised feature data is input into a trained deep learning model; the deep learning model is a model based on the fusion of graph convolutional neural network and long short-term memory network, used to mine the spatiotemporal correlation features between multi-source parameters; the trained deep learning model is constructed and trained in the following way: A heterogeneous graph model of shield tunneling multi-source parameters is constructed using a graph convolutional neural network. The heterogeneous graph model represents the tunneling parameters, geomechanical parameters, and tunnel geometric parameters as nodes, and constructs an association network between the parameters through edge connections. The graph convolutional neural network extracts high-order feature vectors that reflect the spatial association between the multi-source parameters by aggregating and transforming node features. The high-order feature vector sequence is input into a long short-term memory network, and its gating mechanism is used to mine the temporal variation pattern in order to predict the tunneling thrust at future moments. Based on the output of the deep learning model, the predicted value of the tunneling thrust of the tunnel boring machine at future moments is obtained; Output the predicted value of the tunneling thrust.

[0016] In one specific implementation, the step of using wavelet transform for denoising preprocessing includes: Wavelet decomposition was performed on the multi-source heterogeneous monitoring data to obtain wavelet coefficients at each scale; Threshold quantization is performed on wavelet coefficients at high resolution at various scales. Wavelet coefficients with amplitudes below the preset threshold are set to zero, while wavelet coefficients with amplitudes above the preset threshold are retained or shrunk. The denoised feature data is reconstructed by using the wavelet coefficients retained after threshold quantization and the low-frequency coefficients under low resolution to perform inverse wavelet transform.

[0017] In a specific implementation scheme, the threshold quantization process adopts a soft threshold function. The processing rule of the soft threshold function is as follows: if the amplitude of the wavelet coefficient is not less than the preset threshold, the wavelet coefficient is processed in the manner of "the amplitude of the coefficient minus the threshold is consistent with the sign of the coefficient"; if the amplitude of the wavelet coefficient is less than the preset threshold, the wavelet coefficient is set to zero.

[0018] In one specific implementation scheme, the tunneling parameters include thrust, cutterhead torque, cutterhead rotation speed, tunneling speed, soil chamber pressure, grouting pressure, auger discharger rotation speed, and auger discharger pressure; the geomechanical parameters include density, water content, internal friction angle, Poisson's ratio, and porosity; and the tunnel geometric parameters include tunnel diameter and burial depth.

[0019] In one specific implementation scheme, the deep learning model is trained using supervised learning, with historical construction data as training samples, root mean square error as the loss function, and parameters optimized using the Adam optimization algorithm.

[0020] In a specific feasible implementation, the deep learning model performs incremental training at a preset period, using new data to fine-tune the model to achieve online parameter updates. The preset period is 24 hours, that is: the model performs incremental training at a preset period of 24 hours to achieve online parameter updates; the incremental training utilizes new shield tunneling construction data within the past 24 hours to adapt to the dynamic changes in geological conditions as the tunneling mileage changes.

[0021] In one specific implementation scheme, the step of obtaining the predicted tunneling thrust is performed by an edge computing device deployed at the tunnel boring machine site to achieve real-time inference. In the edge computing device, wavelet denoising preprocessing is performed on the collected monitoring data. The preprocessing includes dynamic threshold calculation. The dynamic threshold calculation refers to: dynamically calculating a universal threshold based on the noise standard deviation of the monitoring data, and quantizing the wavelet coefficients based on the threshold to eliminate high-frequency noise from hydraulic pulsation, rock-breaking impact vibration noise from the cutterhead, and electromagnetic radiation interference from the motor, thereby ensuring the purity of the data input to the deep learning model.

[0022] In a specific feasible implementation, the output step includes: displaying the predicted tunneling thrust value and the historical measured data sequence in real time on a visualization interface, the display supporting interactive query and data backtracking; the historical measured data sequence is the historical measured data sequence of the most recent 10 cycles; the data refresh cycle of the visualization interface is 500 milliseconds; the response latency is no more than 100 milliseconds, and it supports gesture zoom interactive query and full-cycle data backtracking.

[0023] In one specific implementation scheme, the method further includes the steps of: connecting the visualization interface corresponding to the predicted tunneling thrust value to the shield PLC control system via an industrial Ethernet connection based on the TCP / IP protocol (the connection is a bidirectional communication connection), transmitting data using the OPC UA protocol, and synchronizing PLC alarm signals in real time.

[0024] In one specific implementation scheme, a real-time intelligent prediction system for tunnel boring machine thrust for implementing any of the methods described above includes: The tunneling parameter sensing unit is used to collect real-time data on the operating status of the tunnel boring machine, geomechanical parameters, and tunnel geometric parameters. A data preprocessing unit, connected to the tunneling parameter sensing unit, is used to perform noise reduction processing on the acquired data using wavelet transform. An edge computing device, connected to the data preprocessing unit, is used to run a deep learning model to generate predicted tunneling thrust values ​​in real time. A central computing device is communicatively connected to the edge computing device and is used to build and train the deep learning model, and to distribute the trained model to the edge computing device. A visualization system, connected to the edge computing device, is used to display the thrust prediction value.

[0025] In one specific implementation scheme, the tunneling parameter sensing unit includes a piezoresistive hydraulic cylinder pressure sensor array, a flange-type strain torque sensor, a Hall speed sensor, a laser Doppler velocimeter, a vibrating wire earth pressure gauge array, a piezoelectric pressure sensor, a magnetoelectric speed encoder, and a diffused silicon pressure transmitter.

[0026] In one specific implementation scheme, the system further includes a power supply system, which adopts a dual-path redundancy architecture to power the edge computing device; the main path is a three-phase 380V AC power supply that outputs 24V DC power through an AC / DC conversion module, and the backup path is a 48V / 80Ah lithium battery pack that achieves seamless switching through a bidirectional DC / DC converter, with a switching response time of no more than 15 milliseconds; the lithium battery pack integrates an intelligent management system that monitors the voltage balance of individual cells in real time and performs over-temperature / over-current protection.

[0027] In one specific implementation, the central computing device is deployed on a high-performance computing cluster for building and training a deep learning model that integrates a graph convolutional neural network and a long short-term memory network.

[0028] In one specific implementation scheme, the visualization system includes an industrial-grade touch screen deployed in the tunnel boring machine's operating room and / or a large-screen display system deployed in the project control center; the industrial-grade touch screen is an explosion-proof panel that meets IP66 certification and has an operating temperature range of -10℃ to 50℃; the LED large-screen display system receives data through an industrial ring network and a 5G dual link.

[0029] In one specific implementation scheme, the system further includes a data storage device; the data storage device is deployed in a dedicated computer room of the project management center outside the tunnel, adopts a 24-bay rack-mount disk array, has 20 built-in 10TB monitoring-grade mechanical hard drives and forms a RAID60 redundant storage system; the data storage device supports parallel writing of 256 sensor data streams, with a continuous read / write bandwidth of not less than 1.8GB / s, saves raw data in 1-minute granularity, retains historical data for not less than 3 years, and has data encryption function.

[0030] Compared with existing technologies, the present invention provides a real-time intelligent prediction method for tunnel boring machine thrust based on temporal deep learning neural networks. This method is used for dynamic thrust prediction during tunnel construction. By constructing a multi-source heterogeneous sensor collaborative sensing system and a high-reliability power supply architecture, and combining wavelet transform-based adaptive signal denoising technology with a deep learning model that integrates graph convolutional neural networks and long short-term memory networks, the method achieves multi-dimensional spatiotemporal feature extraction and real-time intelligent prediction of tunneling thrust. This effectively improves the accuracy, real-time performance, and system generalization ability of thrust prediction under complex geological conditions, and has the following beneficial effects: 1. This invention employs a deep learning model that combines Graph Convolutional Neural Network (GCN) and Long Short-Term Memory Network (LSTM). By constructing a heterogeneous graph model with multiple types of nodes, including tunneling parameters, geomechanical parameters, and tunnel geometric parameters, it utilizes adjacency matrix self-loops and degree matrix normalization to process the spatial correlation features between tunneling parameters and captures long-term temporal dependencies through a gating mechanism. This significantly enhances the model's adaptability to complex working conditions and its prediction accuracy.

[0031] 2. This invention constructs a preprocessing method for multi-source heterogeneous monitoring data based on wavelet transform. By decomposing, thresholding, and reconstructing the original signal, it effectively suppresses non-stationary and non-Gaussian noise interference in the sensor signal, concentrates the signal energy in the effective coefficients, provides high-purity input data for the model, and improves data quality and feature reliability.

[0032] 3. This invention integrates edge computing and central computing collaborative architecture, combines 5G and industrial Ethernet to achieve low-latency data transmission and dynamic model updates, and displays thrust prediction values ​​and historical sequences in real time through a multi-terminal visualization system, supports human-computer interaction and multi-screen collaborative operation, and comprehensively enhances the system's real-time response capability and operational reliability in actual construction environments. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0034] Figure 1 This is a system architecture diagram for implementing the prediction method.

[0035] Figure 2 This is a flowchart of the training and inference process for the tunneling thrust prediction model.

[0036] Figure 3 Flowchart of wavelet denoising preprocessing for monitoring data.

[0037] Figure 4 Visualize and interact with the prediction results using data flow diagrams. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] The present invention discloses a real-time intelligent prediction method for tunnel boring machine thrust based on temporal deep learning neural network, which is applied to the real-time intelligent prediction of tunnel boring machine thrust in shield tunnel construction. To realize the above-mentioned real-time intelligent prediction method for tunnel boring machine thrust based on temporal deep learning neural network, the system adopted by the present invention includes a tunneling parameter sensing unit, a power supply system, a data storage device, a central computing device, an edge computing device, a data transmission system based on 5G network and Ethernet, and corresponding connecting parts.

[0040] The tunneling parameter sensing unit is used to collect real-time data on the operating status of the tunnel boring machine, geomechanical parameters, and tunnel geometric parameters, providing raw data input for the prediction model.

[0041] The power supply system provides a stable and reliable power supply to all sensing, computing and transmission devices in the system, ensuring their continuous normal operation in the complex industrial environment of high-intensity vibration, high humidity and electromagnetic interference of the tunnel boring machine.

[0042] The data storage device is used to securely and efficiently store massive amounts of raw monitoring data and processing results, and supports historical data backtracking and incremental model learning.

[0043] The central computing device is responsible for building and training complex deep learning models. It has powerful parallel computing capabilities and is the core of model iterative optimization.

[0044] The edge computing device is deployed on-site at the tunnel boring machine and is responsible for receiving pre-processed data and running inference models to achieve real-time prediction of tunneling thrust, meeting the millisecond-level response requirements of the control system.

[0045] The data transmission system based on 5G network and Ethernet constitutes a high-speed, low-latency, and highly reliable data transmission network covering the underground, responsible for connecting devices on the edge side and the central side, and ensuring bidirectional real-time communication of data and commands.

[0046] The corresponding connecting parts include various armored shielded cables, industrial connectors, optical cables, etc., used to connect the above-mentioned units and equipment to form a complete system.

[0047] like Figure 1 As shown, data is first collected in real time through the tunneling parameter sensing unit 11. A piezoresistive hydraulic cylinder pressure sensor array is embedded in the side wall of the piston chamber of the propulsion hydraulic cylinder to monitor the dynamic changes in total thrust 111. A flange-type strain torque sensor is rigidly coupled to the main drive output shaft flange to capture the torque load fluctuations of the cutterhead 112. A Hall effect speed sensor is embedded in the stator winding slot of the cutterhead drive motor to calculate the real-time speed of the cutterhead 113. A laser Doppler velocimeter is fixed to the reference positioning plate on the central axis of the shield to track the propulsion speed vector 114. A vibrating wire soil... A pressure gauge array is welded to the soil chamber partition to monitor the internal pressure of the soil chamber 115. A piezoelectric pressure sensor is integrated into the grouting pipeline to collect the grouting pressure pulsation characteristics 116. A magnetoelectric speed encoder is directly connected to the output shaft end of the screw excavator reducer to monitor the speed of the soil discharge system 117. A diffused silicon pressure transmitter is fixed to the screw conveyor cylinder to detect the internal conveying pressure fluctuations 118. At the same time, various geomechanical parameters 12 are connected, including but not limited to key indicators such as density and water content, as well as tunnel geometric parameters 13, which mainly involve basic parameters such as diameter and burial depth.

[0048] The acquired data is transmitted via armored twisted-pair shielded cables and undergoes wavelet transform denoising. This involves wavelet decomposition of the original signal to retain large-scale low-frequency coefficients, while high-frequency coefficients at each scale are quantized using threshold settings (below the threshold are set to zero, above the threshold are retained or reduced). The signal is then reconstructed through inverse transform. The processed data is transmitted via a dedicated underground optical cable to data storage equipment 14 at the project management center outside the tunnel. This equipment includes a 24-bay rack-mount disk array, using 20 10TB hard drives to form a RAID60 redundant storage system, supporting 256 parallel data streams. Raw data is saved in 1-minute granularity, with a retention period of no less than 3 years. After data preprocessing (15), the detection data is transmitted to the central computing device (16) to construct a deep learning model combining a Graph Convolutional Neural Network (GCN) and a Long Short-Term Memory (LSTM) network (17). Fifteen types of parameters (including thrust, cutterhead torque, geological parameters, etc.) are used as nodes to construct a heterogeneous graph model. Spatial correlations are extracted by aggregating node features through graph convolutional layers, and then input into the LSTM module to process temporal features using a gating mechanism. The thrust for the next time step is predicted based on a pre-set time step sequence. Each data acquisition cycle is 24 hours. After each acquisition cycle, the data denoised by wavelet transform is added to the complete dataset as incremental data for the deep learning model. The graph structure of the features is recalculated based on the new dataset without changing the existing number of nodes, resulting in a new adjacency matrix and node feature tensors specific to the shield tunneling characteristics. Since training a high-precision model requires parameter tuning and is time-consuming, an incremental training mode is adopted. This mode uses all network weights and biases of the existing model as initial parameters, and iterative training is performed on a newly constructed dataset. The loss function is the root mean square error (RMSE). The parameters are optimized using the Adam algorithm, with an initial learning rate of 0.01 that decreases by one-tenth every 200 epochs. Specifically, the initial learning rate is set to 0.01, and an epoch-based decay strategy is employed, multiplying the learning rate by a decrease factor of 0.9 every 200 epochs. Given that geological changes are characterized by continuous transitions, the new and old models in adjacent incremental updates exhibit good adaptability and continuity. Considering the core requirement of shield tunneling thrust prediction—that the model only needs to accurately adapt to the tunneling patterns of the short-term construction section without excessive concern about forgetting features from earlier construction sections—iterative updates based on newly collected data can achieve rapid model adaptation to the current construction conditions with lower training costs. The trained model is uploaded to the edge computing device 181 via a 5G network. The device is equipped with an industrial-grade processor to obtain the preprocessed feature vector in real time and generate the future thrust prediction value 19. The edge computing device 181 has a built-in adaptive wavelet threshold denoising module and is equipped with an independent power supply system 182, which is replenished by the tunnel boring machine. The data is transmitted to the visualization system 110 via 5G / Ethernet.

[0049] After constructing a deep learning model 17 consisting of a Graph Convolutional Neural Network (GCN) and a Long Short-Term Memory Network (LSTM), the training and prediction flow of the deep learning model is as follows: Figure 2 As shown. After collecting the original detection data 21, wavelet transform 22 is used for denoising: first, the signal is decomposed by wavelet to retain all wavelet coefficients at the large scale, then the high-frequency coefficients at each scale are quantized by setting thresholds, and finally the processed wavelet coefficients and low-frequency coefficients are used to perform inverse transform to complete signal reconstruction, resulting in denoised data 23. Then, the model update judgment stage 24 is entered: if the current time is more than 24 hours since the last model training, the incremental training process 252 is started to iteratively train the deep learning model composed of GCN and LSTM; training data 27 is generated using the historical data 26 of the past 24 hours, which includes tunneling parameters, geomechanical parameters and tunnel geometric parameters. The training data 27 is input into the graph neural network 28 to construct a shield tunneling multi-source parameter heterogeneous graph model containing 15 types of nodes such as thrust, cutterhead speed, soil chamber pressure, geological density, tunnel burial depth, etc. and edge connections. The node features are aggregated and transformed by graph convolutional layers, and the adjacency matrix and degree matrix are normalized by adding self-loops to extract high-order feature vectors 29 that reflect the spatial correlation of parameters. The feature vector sequence 29 is input into an LSTM network 210. A gating mechanism is employed, involving a forget gate to discard historical information, an input gate to update unit states, a unit state to update stored long-term dependency information, and an output gate to generate hidden states. This mechanism uncovers the temporal variation patterns of parameters and alleviates the gradient vanishing problem in long-sequence training. During model training, the loss function RMSE is continuously checked to see if it reaches its minimum value 211. If convergence is not achieved, the gradient is calculated using the backpropagation algorithm, and the network parameters are updated using the Adam optimization algorithm 212. If convergence has occurred, an optimized end-to-end prediction model 213 is generated. If the current time is less than 24 hours since the last model training, the preprocessed feature vector sequence is directly used as the prediction input 251. The prediction model 213 outputs a predicted shield tunneling thrust value 214 for the next time step based on the prediction input 251, achieving real-time intelligent prediction of shield tunneling thrust.

[0050] The specific process of using wavelet transform for noise reduction is as follows: Figure 3As shown, the process begins by acquiring raw monitoring data 31 of the tunnel boring machine through a sensing unit. Then, wavelet decomposition 32 is performed on the data to obtain wavelet coefficients 33 at various scales. Next, threshold quantization is applied to the wavelet coefficients 33. If the amplitude of the wavelet coefficient 33 is lower than a preset threshold, the coefficient is zeroed out 35; if the amplitude is higher than the threshold, it is retained or contracted 36. After processing, inverse wavelet transform 37 is performed using the processed wavelet coefficients to reconstruct the signal, resulting in a reconstructed signal 38. Finally, denoised feature data 39 is obtained. This process achieves efficient denoising of the monitoring data, concentrating signal energy in the wavelet domain at large wavelet coefficients 33, effectively eliminating noise interference, and providing clean feature data for subsequent tunneling thrust prediction.

[0051] The visualization system 110, as shown Figure 4 As shown. Edge computing device 41 transmits thrust prediction value 43 to explosion-proof touch screen 45 in the control room via 5G / Ethernet network 42, displaying thrust prediction value 43 and nearly 10 cyclic measured historical sequences in real time. The historical sequences are interpolated by cubic spline 44, support gesture zooming, refresh every 500ms, and have a response latency of no more than 100ms. PLC alarm signals are synchronized. LED screen 46 in the project control center receives data through industrial ring network and 5G / Ethernet dual links, dynamically displays core indicators, refreshes every 5 seconds, has a latency of no more than 500ms, and supports multi-screen collaboration and data backtracking.

[0052] The following is a specific example.

[0053] The tunneling parameter sensing unit includes a piezoresistive hydraulic cylinder pressure sensor array embedded in the sidewalls of the piston chambers of six propulsion hydraulic cylinders via threaded interfaces. This array is used to monitor the dynamic changes in the total thrust output of the hydraulic system in real time. Armored twisted-pair shielded cables are connected to the CAN bus network after electromagnetic isolation. A flange-type strain torque sensor is rigidly coupled to the main drive output shaft flange via high-strength alloy bolts. This sensor is used to capture the torque load fluctuations during the cutterhead rotation cutting process. A built-in temperature self-compensation module eliminates the influence of mechanical thermal deformation. A Hall effect speed sensor is embedded in the insulation layer of the stator winding slot of the cutterhead drive motor. It senses the rotor position of the motor based on changes in the magnetic field and is used to accurately calculate the real-time speed of the cutterhead. A laser Doppler velocimeter is rigidly fixed to the reference positioning plate of the shield's central axis via a three-dimensional adjustable anti-seismic bracket. The transmitter head is kept ±0 to the tunnel axis. A 1° parallel tolerance is used to continuously track the instantaneous advance speed vector of the tunnel boring machine; a vibrating wire earth pressure gauge array is argon-arc welded to a pre-set stainless steel monitoring base on the soil chamber partition, with the sensing diaphragm in direct contact with the soil, used to sense the spatial gradient distribution of pressure inside the soil chamber; a piezoelectric pressure sensor is integrated into the sealing cavity of the three-way interface of the synchronous grouting pipeline ring distributor, transmitting pressure through hydraulic oil, used to dynamically acquire the working pressure pulsation characteristics of the grouting system; a magnetoelectric speed encoder is directly connected to the keyway at the output shaft end of the auger reducer through an elastic coupling, based on the toothed disc pulse counting principle, used to monitor the stability of the conveying speed of the soil discharge system; a diffused silicon pressure transmitter is vacuum-welded and fixed to the reinforced base of the pressure chamber of the auger cylinder, with a 316L stainless steel isolation diaphragm in direct contact with the soil, used to detect the nonlinear fluctuation process of the conveying pressure inside the auger.

[0054] The main power supply system draws power directly from the three-phase output of the 10kV power bus of the tunnel boring machine (TBM), and distributes it to each monitoring terminal via a 200A industrial circuit breaker group. Point-to-point power supply is achieved through armored flame-retardant cables. The edge computing equipment adopts a dual-path direct power supply architecture. The main path is a three-phase 380V AC power output to 24V DC power via an AC / DC conversion module. The backup path is a lithium battery pack (48V / 80Ah) that achieves seamless switching through a bidirectional DC / DC converter, with a switching response time of no more than 15ms. The battery pack integrates an intelligent management system to monitor the voltage balance of individual cells in real time and implement over-temperature / over-current protection. The central computing equipment is powered by a dedicated line from the substation and is equipped with a modular UPS power supply (40kVA / 0.5 hours) to ensure uninterrupted training tasks. The early warning distribution system draws power directly from the ring network through an industrial PoE switch, and a supercapacitor array (120F / 48V) is deployed at the core node to cope with voltage drops within 50ms. All DC power supply circuits are equipped with dual overcurrent protection devices, integrated into the protection unit of the TBM PLC control cabinet.

[0055] The data storage equipment is deployed in a dedicated computer room at the project management center outside the tunnel. It adopts a 24-bay rack-mount disk array, with 20 built-in 10TB monitoring-grade mechanical hard drives forming a RAID60 redundant storage system. It supports parallel writing of 256 sensor data streams, with a continuous read / write bandwidth of no less than 1.8GB / s. The equipment is directly connected to the edge computing node of the tunnel boring machine via a dedicated underground optical cable to realize real-time transmission of tunneling parameters. The storage cycle saves the original data in 1-minute granularity, and the historical records are retained for no less than 3 years. The equipment has data encryption function to ensure the security and integrity of the monitoring data.

[0056] The central computing equipment is deployed in a high-performance computing cluster at the project management center outside the tunnel, equipped with two NVIDIA A5000 GPU servers and connected to the data storage system via a dedicated fiber optic network. The equipment constructs a deep learning model of graph convolutional neural network (GCN) and long short-term memory network (LSTM) to perform fusion analysis on multi-source heterogeneous monitoring data; it uses sequence data to predict the tunneling thrust for the next time step; the model performs incremental training every 24 hours.

[0057] The edge computing device is deployed in a dedicated seismic-resistant cabinet behind the tunnel boring machine's control panel. It is equipped with an industrial-grade multi-core processor and a large-capacity error-correcting memory. It acquires real-time tunneling parameter sensor data such as hydraulic thrust, cutterhead torque, and soil chamber pressure through a high-precision isolated acquisition module. The device runs a deep learning model issued by the central computing device. The input layer receives the preprocessed feature vector sequence and generates a real-time prediction value of future tunneling thrust. The built-in adaptive wavelet threshold denoising module eliminates sensor signal interference. The dual-redundant power supply system has a main channel connected to the tunnel boring machine's 24V DC bus and a backup channel supported by a wide-temperature-range lithium battery pack, with a power switching latency of less than 15 milliseconds.

[0058] Wavelet transform is used for denoising monitoring data, concentrating the signal energy in a few large wavelet coefficients in the wavelet domain, while the noise energy is distributed throughout the entire wavelet domain. A threshold is applied to retain useful signal coefficients and reduce noise coefficients to zero. The wavelet transform denoising calculation process includes the following steps: (1) Wavelet decomposition: The multi-source heterogeneous monitoring data to be processed is decomposed using wavelet decomposition to retain all wavelet coefficients at the large scale (low resolution). The wavelet transform uses a universal threshold, and the formula can be expressed as: in, N The length of the original signal; The most common method for estimating the noise standard deviation is based on the median absolute deviation (MAD) of the detail coefficients at the first scale. The principle is that under a Gaussian distribution, the relationship between the median absolute deviation (MAD) and the standard deviation is as follows: .

[0059] (2) Dynamic threshold calculation: Based on the above formula, the noise standard deviation σ estimated by the current monitoring data is used to dynamically calculate the noise threshold λ applicable to the current data segment; the threshold λ is adaptively adjusted as the noise level of the input data changes.

[0060] (3) High-frequency coefficient threshold quantization: Set a threshold to perform threshold quantization on the wavelet coefficients at each scale of high resolution obtained by wavelet decomposition: set the wavelet coefficients with amplitudes lower than the threshold to zero; and retain or shrink the wavelet coefficients with amplitudes higher than the threshold.

[0061] (4) Wavelet Reconstruction: Using the wavelet decomposition coefficients obtained in step (1) and the wavelet coefficients retained after threshold quantization in step (3), an inverse wavelet transform is performed to reconstruct the signal. The inverse wavelet transform is performed on the high-frequency coefficients (after thresholding) and low-frequency coefficients at each scale, and the formula can be expressed as: Where J is the number of decomposition levels; j is the scale index, with a maximum value of the number of decomposition levels J; and k is the position index, which marks the time position of the wavelet coefficients in the original time series signal at the same scale (j is fixed). and These are the reconstruction basis functions for wavelet functions and scaling functions, respectively, which can be selected by engineers based on the actual signal characteristics; These are the wavelet coefficients after thresholding. These are the low-frequency coefficients of the last layer.

[0062] The threshold function defines different processing strategies for wavelet coefficients with amplitudes above and below a threshold, and is key to the threshold denoising method. The soft threshold function used can be expressed as: Where w represents the wavelet coefficients, λ is the calculated threshold, and sng is the sign function.

[0063] A method for predicting tunnel boring machine (TBM) thrust based on graph convolutional neural networks (GNNs) and long short-term memory (LSTM) networks is established. This involves constructing a heterogeneous graph model of TBM multi-source parameters using a GNN, representing various parameters involved in the TBM tunneling process as nodes, forming a node set containing 15 types of nodes. These parameters include thrust, cutterhead torque, cutterhead rotation speed, tunneling speed, soil chamber pressure, grouting pressure, auger speed, and auger pressure, which directly reflect the TBM's operating status; geomechanical parameters such as density, water content, internal friction angle, Poisson's ratio, and porosity, which reflect the geological characteristics of the construction area; and tunnel geometric parameters such as diameter and burial depth, which significantly influence the TBM's tunneling conditions. Granger causality tests are used to analyze the temporal characteristics of the parameters and create a graph structure between parameter nodes. The tests are used to filter parameter associations, determine the direction of influence, and create edge connections for the heterogeneous graph (filtering out unrelated parameter pairs) to reduce computational redundancy. Wavelet denoising removes noise-induced pseudo-non-stationary components, highlighting the intrinsic correlations between parameters. The data exhibits "local relative stationarity" (gradual stratigraphic variation and stable parameter correlation patterns), meeting the basic correlation capture requirements for the test. The Granger causality test is passed. F Test to determine the introduction X t Can past values ​​significantly improve the understanding of Y t Its predictive ability. F The null hypothesis being tested is " X t Not Granger caused Y t The alternative hypothesis is " X t Granger caused Y t ". F The formula for the test is: in, F express F Statistic; RSS r This represents the sum of squared residuals from a univariate autoregressive model. RSS ur This represents the sum of squared residuals in a multivariate regression model. q The number of constraints, i.e., the product of the time step and the number of features; n For sample size; k This represents the number of explanatory variables in the model, here it is 1+ q .if F If the statistic is greater than the critical value, it means... X t Granger caused Y tThis means that there are edges connecting the nodes, thus constructing a complex network of relationships between shield tunneling parameters, which realistically simulates the interaction between various parameters during the shield tunneling process.

[0064] Graph convolutional neural networks (GNNs) consist of graph convolutional layers and multilayer perceptrons. The graph convolutional layer is the core component of the entire network. It overcomes the limitation of traditional convolutional neural networks, which can only process regular grid data, and can directly operate on data with arbitrary graph structures, achieving feature extraction and transformation of graph-structured data. The computational principle of the graph convolutional layer is based on the aggregation and transformation operations of node features, and its calculation formula is as follows: in, It is the feature matrix of the nodes in the l-th layer; when l=0, Here, X is the feature matrix of the original input, which contains the initial parameter information of 15 types of nodes. This represents the node feature matrix output after the l-th layer of processing; it is the result of transforming the input feature matrix. Where I is the identity matrix, this operation means It is an adjacency matrix with added self-loops. The introduction of self-loops allows each node to consider its own feature information when aggregating features, enhancing the network's ability to learn the characteristics of its own nodes. yes The degree matrix is ​​used to normalize the adjacency matrix, ensuring that nodes of different degrees receive reasonable weights during feature aggregation. This is the trainable weight matrix of the l-th layer. The essence of GCN lies in aggregating the features of each node and its connected nodes through the node connection relationships defined by the adjacency matrix. During the aggregation process, the features of different nodes are assigned different weights based on their connection strength and correlation, thereby achieving graph learning based on the relationships between nodes. In this way, graph convolutional neural networks can fully explore the potential correlations between multi-source parameters in tunnel boring machine (TBM) construction, extract high-order features that reflect the complex relationships between parameters, and provide more representative and discriminative feature representations for subsequent tunneling thrust prediction.

[0065] The LSTM module predicts subsequent shield thrust by inputting the temporal feature vector output by the GCN, and mitigates the vanishing and exploding gradient problems in long sequence training through a gating mechanism. In LSTM, the gating mechanism is implemented by training fully connected layers. The calculations of the forget gate, input gate, unit state update, and output gate are as follows: In the formula, W and b are the weight and bias matrices of different gated fully connected layers, respectively; Ct is the current cell state; ht is the current hidden state; and xt is the one-dimensional input matrix. tanh and σ are the tanh activation function and the sigmoid activation function, respectively, and their expressions are as follows: By setting a time window t, i.e., inputting feature vectors from the past t time units, we can obtain the input sequence for the LSTM network, represented as: Through the calculation process of the LSTM network described above, the thrust at the monitoring point at the next moment can be calculated: A method for predicting the thrust of a tunnel boring machine (TBM) is established based on graph convolutional neural networks (GCNNs) and long short-term memory (LSTM) networks. The GCNNs and LSTMs are organically integrated to form an end-to-end prediction model. Specifically, firstly, GCNNs are used to model the graph structure and extract features of multi-source parameters during TBM construction, capturing the spatial relationships between parameters. The output features of the GCN corresponding to each time step parameter are integrated using average pooling, transforming features at dimensions [15, n] into features at dimensions [1, n]. The features from t time steps constitute the temporal features at dimensions [t, n]. The resulting temporal feature sequence is then input into the LSTM network to mine the temporal variation patterns of the parameters. Finally, a fully connected layer maps the temporal features output by the LSTM to the thrust prediction value space. The model training process employs supervised learning, using historical TBM construction data as training samples. Each training sample contains a multi-source parameter sequence within a time window and its corresponding actual thrust value. The training objective is to minimize the error between the model's predicted values ​​and the actual thrust values. A commonly used loss function is the root mean square error (RMSE). Where n is the number of training samples. This is the actual thrust value. These are the model predictions. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the parameters are updated using the Adam algorithm. This process is iterated until the model converges.

[0066] The visualization system is deployed on the main control console of the tunnel boring machine's operating room and integrated into an industrial-grade touch screen (explosion-proof panel) to display the predicted tunneling thrust in real time. and the measured thrust history sequence of the most recent 10 cycles Predicted values ​​are rendered as dynamic fonts on top of the tunnel boring machine's 3D model, with the font size automatically scaling. Historical curves are smoothly plotted using cubic spline interpolation. Data is refreshed every 500ms via an industrial ring network, supporting gesture-based zooming for querying any operating condition. The system is IP66 certified, operates at temperatures from -10℃ to 50℃, has a response latency of no more than 100ms, and synchronizes alarm signals in real-time with the tunnel boring machine's PLC. Monitoring data is synchronously deployed on the project control center's LED screen, receiving data from edge devices in real-time via the industrial ring network and 5G dual links, dynamically displaying full-face tunnel boring machine parameters, including core indicator tables such as real-time thrust, predicted values, cutterhead torque, and soil chamber pressure. Data refreshes every 5 seconds with a latency of no more than 500ms, supporting multi-screen collaborative operation and full-cycle data backtracking.

[0067] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A real-time intelligent prediction method for tunneling thrust of a shield machine based on a timing deep learning neural network, characterized in that, The method comprises the following steps: Real-time acquisition of multi-source heterogeneous monitoring data in the tunneling process of a shield tunneling machine, the data comprising tunneling parameters, geomechanical parameters and tunnel geometric parameters; Wavelet transform is used to pre-process the multi-source heterogeneous monitoring data to obtain denoised feature data; The denoised feature data is input into a trained deep learning model; the deep learning model is a model based on the fusion of a graph convolutional neural network and a long short-term memory network, and is used to mine the spatio-temporal correlation features among multi-source parameters; the trained deep learning model is constructed and trained in the following manner: A graph convolutional neural network is used to construct a heterogeneous graph model of shield multi-source parameters; the heterogeneous graph model represents the tunneling parameters, geomechanical parameters and tunnel geometric parameters in the form of nodes, and constructs a correlation network among the parameters through edges; the graph convolutional neural network extracts a high-order feature vector reflecting the spatial correlation among multi-source parameters by aggregating and transforming node features; The high-order feature vector sequence is input into a long short-term memory network to mine the time series variation law through the gating mechanism thereof to predict the tunneling thrust at a future time; Based on the output of the deep learning model, a tunneling thrust prediction value of the shield tunneling machine at a future time is obtained; The tunneling thrust prediction value is output.

2. The method of claim 1, wherein the method is characterized by, The step of pre-processing the data by wavelet transform comprises: Wavelet decomposition is performed on the multi-source heterogeneous monitoring data to obtain wavelet coefficients at each scale; Threshold quantization processing is performed on the high-resolution wavelet coefficients at each scale, wavelet coefficients with an amplitude lower than a preset threshold are set to zero, and wavelet coefficients with an amplitude higher than the preset threshold are retained or subjected to shrinkage processing; Wavelet inverse transform is performed on the retained wavelet coefficients after threshold quantization processing and low-frequency coefficients at low resolution to reconstruct the denoised feature data.

3. The method of claim 2, wherein the method is characterized by, The threshold quantization processing uses a soft threshold function.

4. The method of claim 1, wherein the method is characterized by, The tunneling parameters comprise thrust, cutterhead torque, cutterhead speed, tunneling speed, soil chamber pressure, grouting pressure, spiral earth unloader speed and spiral earth unloader pressure; the geomechanical parameters comprise density, water content, internal friction angle, Poisson's ratio and porosity; and the tunnel geometric parameters comprise tunnel diameter and burial depth.

5. The method of claim 1, wherein the method is characterized by, The training of the deep learning model uses a supervised learning method, historical construction data is used as a training sample, a root mean square error is used as a loss function, and an Adam optimization algorithm is used for parameter optimization.

6. The method of claim 5, wherein the method is characterized by, The deep learning model performs incremental training at a preset period to realize online updating of parameters.

7. The method of claim 1, wherein the method is characterized by, The step of obtaining the tunneling thrust prediction value is performed by an edge computing device deployed on site of the shield tunneling machine to realize real-time inference. 8.The real-time intelligent prediction method of tunneling thrust of a shield machine based on a timing deep learning neural network according to claim 1, characterized in that, The output step comprises real-time display of the tunneling thrust prediction value and a historical measured data sequence on a visual interface, and the display supports interactive query and data backtracking.

9. The method of claim 8, wherein the method is characterized by, The method further comprises the step of connecting the visual interface corresponding to the tunneling thrust prediction value and a shield PLC control system through an industrial Ethernet based on a TCP / IP protocol, and transmitting data using an OPC UA protocol to realize real-time synchronization of PLC alarm signals.

Images