Intelligent control method for pressure in shield cutter compartment

Abstract

The application belongs to the field of intelligent control, and particularly relates to a shield cutter bin internal pressure intelligent control method, which comprises a data acquisition module, a data processing center and a control execution module; the data acquisition module is used for collecting various data related to the cutter bin internal pressure and transmitting the data to the data processing center; the data processing center pre-processes the collected data, adopts a multivariate time series prediction and adaptive control algorithm to perform model training, and generates a control instruction; and the control execution module accurately controls the injection and discharge valve opening degree of compressed air or mud according to the control instruction, and adjusts the cutter bin internal pressure.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) construction technology, and in particular to an intelligent control system and method for pressure inside the cutterhead chamber of a TBM. The aim is to precisely adjust the pressure inside the cutterhead chamber of a TBM through innovative algorithm models and intelligent control strategies, so as to ensure the safe and efficient progress of TBM construction and effectively cope with complex and ever-changing geological conditions and construction conditions. Background Technology

[0002] During shield tunnel construction, pressure control within the cutterhead chamber is crucial. Appropriate and stable pressure within the cutterhead chamber balances the external water and soil pressure, preventing engineering accidents such as ground collapse, water inrush, and sand inrush. It also facilitates the smooth removal of soil cut by the cutterhead and the stable advancement of the tunnel boring machine. Traditional methods for controlling cutterhead chamber pressure primarily rely on experience-based settings and simple pressure feedback adjustments. Construction personnel estimate the ground water and soil pressure based on geological survey reports, manually set the target pressure value within the cutterhead chamber, and then maintain pressure stability by controlling the injection and discharge of compressed air or slurry. This approach has several shortcomings: First, geological conditions are complex and variable. Parameters such as permeability, porosity, and groundwater level can vary significantly in different regions and even at different depths within the same region. Experience-based pressure settings are difficult to accurately adapt to actual working conditions, easily leading to problems such as excessive pressure causing ground heave or insufficient pressure causing ground subsidence. Second, simple pressure feedback adjustment systems have slow response times and significant lag, making it difficult to respond promptly to pressure fluctuations caused by changes in cutting speed and abrupt changes in soil properties during tunnel boring machine (TBM) advancement, thus failing to ensure real-time pressure stability. Third, traditional methods lack in-depth mining and analysis of large amounts of construction data, making it impossible to fully utilize historical experience and real-time data to optimize control strategies and achieve intelligent, adaptive pressure control.

[0003] With the continuous development of automation control technology and intelligent algorithms, some new technologies have been gradually applied to the field of tunnel boring machine (TBM) construction. However, many challenges remain in controlling the pressure inside the cutterhead chamber. Most existing intelligent control methods fail to fully consider the nonlinearity, time-varying nature, and strong coupling of the TBM cutterhead chamber pressure control system, resulting in limited accuracy and reliability of the control model. Some studies focus only on the analysis and control of single factors (such as pressure data), neglecting the comprehensive impact of other relevant parameters (such as cutterhead rotation speed, propulsion speed, and soil improvement) on the pressure inside the cutterhead chamber, making it difficult to achieve comprehensive and precise pressure regulation. Furthermore, current control algorithms lack adaptability and robustness when dealing with sudden abnormal situations (such as sudden pressure changes due to cutter breakage or pressure runaway caused by blockage in the air supply pipeline), failing to effectively ensure the safety and stability of TBM construction. Therefore, an innovative intelligent control system and method for the pressure inside the TBM cutterhead chamber is urgently needed to overcome the shortcomings of existing technologies and improve the quality and safety of TBM construction. Summary of the Invention

[0004] This invention provides an intelligent control method for pressure inside the cutterhead chamber of a tunnel boring machine (TBM). The method includes collecting various types of data related to the pressure inside the cutterhead chamber and transmitting them to a data processing center. The data processing center preprocesses the collected data and uses a multivariate time series prediction and adaptive control algorithm to train a model and generate control commands. The control execution module precisely controls the opening of the compressed air or mud injection and discharge valves according to the control commands to adjust the pressure inside the cutterhead chamber.

[0005] Preferably, the method further includes performing time series processing on the data to construct a multivariate time series dataset; then inputting the dataset into the LSTM-AM model, wherein the LSTM network part is set with 3 hidden layers, each containing 64 LSTM units, for learning the temporal features of the data; the attention mechanism part automatically focuses on the key variable features by calculating the importance weight of each variable to stress prediction at different times.

[0006] Preferably, it also includes constructing an adaptive fuzzy control model, and dynamically adjusting the control strategy through a fuzzy inference system based on the pressure prediction results and the current pressure state information to achieve precise control of the pressure inside the cutterhead chamber.

[0007] Preferably, an intelligent control system for the pressure inside the cutterhead chamber of a tunnel boring machine (TBM) is used, comprising a data acquisition module, a data processing center, and a control execution module. The data acquisition module is used to collect various types of data related to the pressure inside the cutterhead chamber and transmit them to the data processing center. The data processing center preprocesses the collected data and uses a multivariate time series prediction and adaptive control algorithm to train a model and generate control commands. The control execution module precisely controls the opening of the compressed air or mud injection and discharge valves according to the control commands to adjust the pressure inside the cutterhead chamber.

[0008] Preferably, the data acquisition module includes a pressure sensor, a flow sensor, a torque sensor, a speed sensor, a displacement sensor, a velocity sensor, and a geological parameter detection sensor installed in the cutterhead chamber of the tunnel boring machine.

[0009] Preferably, the pressure sensor, flow sensor, torque sensor, speed sensor, displacement sensor, velocity sensor, and geological parameter detection sensor are used to measure the pressure inside the chamber, the flow rate of the medium, the cutterhead torque, the cutterhead speed, the tunnel boring machine's advancing displacement, the advancing speed, and the geological characteristics of the strata, respectively.

[0010] Preferably, the data acquisition module sets the corresponding sampling frequency according to the variation characteristics of different parameters. Pressure data and flow data are collected 5 times per second, cutterhead torque, rotation speed and shield machine propulsion data are collected once per second, and geological parameter data are collected once every 0.5 meters of propulsion.

[0011] Preferably, the data processing center performs preprocessing on the collected data, including data cleaning, outlier detection and removal, and data standardization.

[0012] Preferably, the data processing center uses a multivariate time series prediction model based on long short-term memory network and attention mechanism to predict pressure change trends, and constructs an adaptive fuzzy control model based on the prediction results, and dynamically adjusts the injection and discharge of compressed air or mud through a fuzzy inference system.

[0013] Preferably, the control execution module adopts a high-precision electric or pneumatic regulating valve with a control accuracy of ±0.5%, has a real-time feedback monitoring function, and feeds back the control execution results to the data processing center in real time. It is equipped with a safety protection mechanism, which immediately activates the emergency braking device when the pressure is detected to rise or fall abnormally beyond the set safety threshold.

[0014] Beneficial technical effects of the invention:

[0015] A comprehensive and high-precision data acquisition system was designed, capable of simultaneously acquiring multi-source data such as pressure, flow rate, cutterhead operating parameters, tunnel boring machine propulsion parameters, and geological parameters within the cutterhead chamber, providing a rich information foundation for pressure control within the cutterhead chamber. A deep learning-based outlier detection and repair algorithm was employed to accurately identify and process outliers in the data, preventing them from interfering with pressure prediction and control model training. Furthermore, the multi-source data was standardized, enabling fusion and analysis within the same mathematical framework. This provided high-quality data input for subsequent model training, significantly improving the model's prediction accuracy and control performance.

[0016] This paper proposes for the first time to combine Long Short-Term Memory (LSTM) networks with an attention mechanism for multivariate time-series prediction of pressure within the cutterhead chamber of a tunnel boring machine (TBM). This approach effectively handles the temporal and multivariate correlations of the data, and automatically focuses on key variable features through the attention mechanism, improving the accuracy of pressure trend prediction. An adaptive fuzzy control model is constructed. Based on the pressure prediction results and current pressure status information, the control strategy is dynamically adjusted through a fuzzy inference system, achieving precise control of the pressure within the cutterhead chamber. The fuzzy control rules are optimized based on expert experience and a large amount of historical data, enabling them to adapt to different construction conditions and pressure variations. This overcomes the lag and limitations of traditional control methods, improving the flexibility and intelligence of the control.

[0017] A real-time feedback monitoring mechanism was established between the control execution module and the data processing center. This mechanism enables timely feedback of control execution results and the actual operating status of the tunnel boring machine (TBM) to the data processing center, achieving closed-loop optimization of the pressure control process within the cutterhead chamber. The data processing center continuously adjusts and optimizes the control model based on the feedback information, ensuring the control strategy remains optimal and effectively improving control stability and reliability. The control execution module is equipped with a safety protection mechanism. When an abnormal pressure change exceeds a safety threshold, the emergency braking device is immediately activated, and emergency measures are taken to ensure the safety of the TBM construction. This safety mechanism effectively reduces the risks during TBM construction and improves the safety and continuity of the construction process. Attached Figure Description

[0018] Figure 1 Flowchart. Detailed Implementation

[0019] Example 1

[0020] Multiple high-precision sensors are installed in the cutterhead chamber and related parts of the tunnel boring machine to comprehensively collect various data related to the pressure inside the cutterhead chamber. A pressure sensor array is distributed at different locations within the cutterhead chamber, accurately measuring pressure values ​​at various points within the chamber with a measurement accuracy of ±0.01 kPa to obtain pressure distribution information. Flow sensors are installed in the injection and discharge pipelines of compressed air or slurry to monitor flow rates with an accuracy of ±0.1 m³ / h, used to determine the amount of medium entering and exiting. Torque and speed sensors installed in the cutterhead drive system measure the cutting torque (accuracy ±0.5 kN·m) and speed (accuracy ±0.1 r / min) of the cutterhead, respectively, as the working state of the cutterhead directly affects the pressure within the chamber. Displacement and velocity sensors are installed in the tunnel boring machine's propulsion system to measure the tunnel boring machine's propulsion displacement (accuracy ±1 mm) and speed (accuracy ±0.01 m / min) to analyze the correlation between the propulsion process and the pressure within the chamber. In addition, geological parameter detection sensors, such as resistivity sensors and acoustic sensors, are equipped to detect the geological characteristics of the strata in front of the tunnel boring machine in real time, including information on strata hardness, porosity, and water content, and convert this information into digital data to provide geological basis for pressure control.

[0021] The data acquisition module sets corresponding sampling frequencies according to the changing characteristics of different parameters. Pressure and flow data are collected at a high frequency of 5 times per second to quickly capture changes in pressure and flow; cutterhead torque, rotation speed, and tunnel boring machine (TBM) advance data are collected once per second; geological parameter data are collected every 0.5 meters of TBM advance to ensure that the data can reflect changes in construction conditions in a timely manner. The collected data is transmitted to the data processing center via high-speed data transmission lines. Data encryption and verification technologies are used during transmission to ensure the integrity and accuracy of the data.

[0022] Data Processing Center: Equipped with high-performance computing equipment, possessing powerful data storage, processing, and analysis capabilities. The collected data undergoes preprocessing, including data cleaning, outlier detection and removal, and data standardization. Deep learning-based outlier detection algorithms are employed, such as constructing a deep autoencoder network to learn and encode features of normal data. The data to be detected is then input into the network for decoding and reconstruction, and the reconstruction error is calculated. If the reconstruction error exceeds a set threshold, the data is considered an outlier, and a repair method based on data distribution characteristics is used, such as replacing it with the weighted average of neighboring normal data points. All sensor data is standardized to a mean of 0 and a variance of 1 to facilitate subsequent machine learning model processing. Innovative multivariate time series forecasting and adaptive control algorithms are used for model training. A multivariate time series forecasting model (LSTM-AM) based on a Long Short-Term Memory (LSTM) network and attention mechanism is proposed to predict the changing trend of pressure within the cutterhead chamber. This model can simultaneously process time-series data of pressure data and other relevant parameters (such as cutterhead rotation speed, feed rate, geological parameters, etc.). It learns the temporal characteristics of the data through an LSTM network, while an attention mechanism automatically focuses on variables that significantly impact pressure prediction, improving prediction accuracy. Based on the prediction results, an adaptive fuzzy control model is constructed. This model dynamically adjusts the injection and discharge of compressed air or slurry through a fuzzy inference system based on the predicted pressure change trend and current pressure deviation and rate of change, achieving precise control of the pressure within the cutterhead chamber. The fuzzy control rules are optimized and adjusted based on expert experience and extensive historical construction data, enabling the control strategy to adapt to different construction conditions and pressure variations.

[0023] Control Execution Module: Closely connected to the compressed air or slurry supply system of the tunnel boring machine (TBM), this module precisely controls the opening of the compressed air or slurry injection and discharge valves based on control commands generated by the data processing center, thereby regulating the pressure within the cutterhead chamber. This module employs high-precision electric or pneumatic regulating valves, enabling rapid response to control commands and precise adjustment of valve openings with a control accuracy of ±0.5%. For example, when the control model at the data processing center determines that an increase in pressure within the cutterhead chamber is needed, the control execution module will increase the opening of the compressed air injection valve or decrease the opening of the slurry discharge valve, gradually raising the pressure to the target value. It features real-time feedback monitoring, transmitting control execution results (such as actual pressure changes and valve openings) to the data processing center in real time. The data processing center then optimizes and adjusts the control model online based on the feedback information, ensuring that the control strategy always adapts to various changes during the TBM construction process. Meanwhile, the control execution module is also equipped with a safety protection mechanism. When an abnormal increase or decrease in pressure is detected that exceeds the set safety threshold (such as a sudden increase in pressure exceeding ±10 kPa), the emergency braking device is immediately activated to stop the tunnel boring machine's advance and take corresponding emergency measures to ensure the safety of tunnel construction.

[0024] II. Method and Flow

[0025] Data Acquisition and Preprocessing Steps: The data acquisition module collects multi-source data in real time, including pressure, flow rate, cutterhead rotation speed, propulsion speed, and geological parameters within the shield cutterhead chamber, according to a set sampling frequency. This data is then transmitted to the data processing center via a high-speed data transmission network. At the data processing center, a deep learning-based outlier detection algorithm is first used to clean and repair the data. For example, if the pressure value collected by the pressure sensor deviates significantly from the data at previous and subsequent moments (the reconstruction error calculated by the deep autoencoder network exceeds a set threshold), the data is considered potentially anomaly. Further analysis is conducted, considering the cutterhead's operating status, the shield machine's propulsion, and geological conditions. If an outlier is confirmed, a repair method based on data distribution characteristics is used, replacing the outlier with a weighted average of nearby normal data points to ensure data continuity and accuracy.

[0026] All sensor data after cleaning were standardized to have a mean of 0 and a variance of 1. Different standardization methods were used for different types of data. For example, pressure data were linearly standardized based on its range and statistical characteristics; qualitative descriptive information in geological data (such as stratigraphic type) was converted into numerical data using one-hot encoding before standardization, ensuring that all data could be used for subsequent feature fusion and model training within the same mathematical framework.

[0027] The pressure prediction and control model construction steps are as follows: A multivariate time series prediction model (LSTM-AM) based on a Long Short-Term Memory (LSTM) network and an attention mechanism is used to predict pressure change trends. Taking data on cutterhead chamber pressure, cutterhead rotation speed, propulsion speed, and geological parameters as examples, these data are first processed into time series data to construct a multivariate time series dataset. This dataset is then input into the LSTM-AM model. The LSTM network has three hidden layers, each containing 64 LSTM units, used to learn the temporal features of the data. The attention mechanism automatically focuses on key variable features by calculating the importance weight of each variable for pressure prediction at different times. For example, when the tunnel boring machine enters hard rock strata and the cutterhead rotation speed is high, the attention mechanism will pay more attention to cutterhead torque and geological parameter data, as these data have a greater impact on the pressure inside the cutterhead chamber. The LSTM-AM model is trained using a large amount of historical construction data to optimize the model's weight parameters, enabling the model to accurately predict the pressure change trend inside the cutterhead chamber.

[0028] Based on the prediction results, an adaptive fuzzy control model is constructed. The predicted pressure change trend, current pressure deviation, and pressure change rate are used as inputs to the fuzzy control model. The input and output linguistic variables of the fuzzy control model are fuzzified; for example, the pressure deviation is divided into fuzzy subsets such as "negative large (NB)," "negative medium (NM)," "zero (ZO)," "positive medium (PM)," and "positive large (PB)," and the output variable (the adjustment amount of compressed air or mud flow rate) is similarly fuzzified. A fuzzy control rule table is formulated based on expert experience and a large amount of historical construction data. For example, when the pressure deviation is "positive large" and the pressure change rate is "positive large," a larger increase in compressed air flow rate or a decrease in mud flow rate is output. Through a fuzzy inference system and a declarative method, the fuzzy control output is converted into actual control quantities, namely the adjustment amount of the opening of the compressed air or mud injection and discharge valves, achieving precise control of the pressure inside the cutterhead chamber.

[0029] Pressure Control and Optimization Steps: During shield tunneling, the data acquisition module continuously collects data and transmits it to the data processing center. The data processing center uses a trained LSTM-AM model to predict the pressure change trend inside the cutterhead chamber. Then, based on the adaptive fuzzy control model, it generates control commands. The control execution module precisely controls the opening of the compressed air or slurry injection and discharge valves according to the commands to regulate the pressure inside the cutterhead chamber. For example, when the LSTM-AM model predicts that the pressure inside the cutterhead chamber will drop by 0.5 kPa in the next 10 seconds due to the shield machine about to pass through a sand layer and the cutterhead rotation speed will increase, the adaptive fuzzy control model calculates the need to increase the amount of compressed air injected based on the current pressure deviation and rate of change. The control execution module then correspondingly increases the opening of the compressed air injection valve to keep the pressure inside the cutterhead chamber stable.

[0030] At regular construction intervals (e.g., 3 meters) or time intervals (e.g., 5 minutes), the data processing center optimizes the control model online based on the actual control results fed back from the control execution module and the latest operating status data of the tunnel boring machine (TBM). Newly collected construction data is added to the training dataset to retrain the LSTM-AM model and the adaptive fuzzy control model, adjusting the model parameters and fuzzy control rules to adapt to the influence of factors such as geological conditions, construction conditions, and changes in the TBM's own performance. For example, if a deviation is found between the actual pressure control effect of the TBM and the model prediction in a certain construction section, the data processing center analyzes the causes of the deviation (e.g., sudden changes in geological conditions, decreased sensor accuracy, etc.) and optimizes the model accordingly to improve its accuracy and adaptability, ensuring that the pressure control within the cutterhead chamber is always at its optimal state.

[0031] Beneficial technical effects of the present invention:

[0032] Innovative data acquisition and processing strategies:

[0033] A comprehensive and high-precision data acquisition system was designed, capable of simultaneously collecting multi-source data such as pressure and flow rate within the cutterhead chamber, cutterhead operating parameters, tunnel boring machine propulsion parameters, and geological parameters, providing a rich information foundation for pressure control within the cutterhead chamber. Compared with traditional data acquisition methods, the data acquisition module of this invention can more comprehensively and accurately reflect various factors affecting the pressure within the cutterhead chamber during tunnel boring machine construction, effectively improving the reliability and adaptability of control.

[0034] Employing a deep learning-based outlier detection and repair algorithm, this approach accurately identifies and processes outliers in the data, preventing them from interfering with stress prediction and control model training. Simultaneously, standardizing multi-source data allows for fusion and analysis within the same mathematical framework, providing high-quality data input for subsequent model training and significantly improving the model's prediction accuracy and control performance.

[0035] Innovation in intelligent control models based on LSTM-AM and adaptive fuzzy control:

[0036] This paper proposes for the first time to combine Long Short-Term Memory (LSTM) networks with attention mechanisms for multivariate time series prediction of pressure inside the cutterhead chamber of a tunnel boring machine. This approach can effectively handle the temporal and multivariate correlations of the data, and automatically focus on key variable features through the attention mechanism, thereby improving the accuracy of pressure change trend prediction.

[0037] An adaptive fuzzy control model was constructed. Based on pressure prediction results and current pressure status information, the control strategy was dynamically adjusted through a fuzzy inference system, achieving precise control of the pressure inside the cutterhead chamber. The fuzzy control rules were optimized based on expert experience and a large amount of historical data, enabling them to adapt to different construction conditions and pressure changes. This overcomes the lag and limitations of traditional control methods, improving the flexibility and intelligence of the control.

[0038] Real-time feedback optimization and security mechanism innovation:

[0039] A real-time feedback monitoring mechanism was established between the control execution module and the data processing center. This mechanism enables timely feedback of control execution results and the actual operating status of the tunnel boring machine to the data processing center, achieving closed-loop optimization of the pressure control process within the cutterhead chamber. Based on the feedback information, the data processing center continuously adjusts and optimizes the control model, ensuring the control strategy remains at its optimal state and effectively improving the stability and reliability of the control.

[0040] The control and execution module is equipped with a safety protection mechanism. When an abnormal pressure change is detected that exceeds the safety threshold, the emergency braking device can be activated immediately and emergency measures can be taken to ensure the safety of the tunnel boring machine (TBM) construction. This safety mechanism effectively reduces the risks during TBM construction and improves the safety and continuity of the construction process.

[0041] Example 2

[0042] Data Acquisition and Preprocessing: Data acquisition modules were installed in the cutterhead chamber and related areas of the tunnel boring machine (TBM) for mountainous railway tunnels. Ten pressure sensors with a measurement accuracy of ±0.01 kPa were deployed in the cutterhead chamber to comprehensively monitor the pressure at different locations within the chamber. Flow sensors with an accuracy of ±0.1 m³ / h were installed in the compressed air and mud pipelines. Torque sensors (accuracy ±0.5 kN·m) and speed sensors (accuracy ±0.1 r / min) were installed in the cutterhead drive system. Displacement sensors (accuracy ±1 mm) and velocity sensors (accuracy ±0.01 m / min) were installed in the TBM propulsion system. Geological parameter detection sensors, such as ground-penetrating radar, were used to collect geological characteristic data of the strata ahead every 0.4 meters of advancement. Pressure and flow data were collected 5 times per second, while cutterhead-related data and propulsion data were collected once per second. After the collected data was transmitted to the data processing center, it was cleaned and repaired using a deep learning-based outlier detection algorithm. For example, when the pressure value at a certain moment differed significantly from the data at previous and subsequent moments, it was identified as an outlier by a deep autoencoder network. If outliers are caused by sensor vibration, a weighted average of nearby normal data points should be used as a replacement. All data should be standardized to ensure analysis is conducted within the same mathematical framework.

[0043] Pressure Prediction and Control Model Construction: A multivariate time series prediction model based on LSTM-AM is used to predict pressure change trends. Taking the pressure inside the cutterhead chamber, cutterhead rotation speed, propulsion speed, and ground-penetrating radar data as examples, these data are processed into a time series dataset to construct a multivariate time series dataset. The dataset is input into the LSTM-AM model, with the LSTM network having three hidden layers, each containing 64 LSTM units. The attention mechanism automatically adjusts the degree of attention given to key variables according to different construction stages. For example, when the tunnel boring machine passes through rock strata and the cutterhead rotation speed decreases, the attention mechanism will pay more attention to geological parameters and cutterhead torque data. The model is trained with a large amount of historical construction data to optimize model parameters and improve the accuracy of pressure prediction. Based on the prediction results, an adaptive fuzzy control model is constructed. The predicted pressure change trend, current pressure deviation, and pressure change rate are used as inputs. The input and output linguistic variables are fuzzified, such as dividing the pressure deviation into fuzzy subsets like "negative large," "negative medium," "zero," "positive medium," and "positive large," and the output variables (the adjustment amount of compressed air or mud flow rate) are processed similarly. Based on expert experience and historical construction data, a fuzzy control rule table is developed. When pressure deviation and rate of change are in a specific state, the corresponding control quantity is output. Through a fuzzy inference system and a definitive method, the fuzzy control output is converted into the actual valve opening adjustment.

[0044] Pressure Control and Optimization: During tunnel boring machine (TBM) construction, the data acquisition module continuously collects data and transmits it to the data processing center. The data processing center uses a trained LSTM-AM model to predict the pressure change trend within the cutterhead chamber and generates control commands based on an adaptive fuzzy control model. The control execution module precisely controls the valve openings of compressed air or slurry according to these commands, adjusting the pressure within the cutterhead chamber. For example, if the model predicts that the pressure within the cutterhead chamber will rise in the future due to changes in geological conditions, the control execution module will reduce the injection of compressed air or increase the discharge of slurry. At regular construction intervals (e.g., 2 meters) or time intervals (e.g., 4 minutes), the data processing center optimizes the control model online based on feedback from the control execution module and the latest operating status data of the TBM. Newly acquired data is added to the training dataset to retrain the LSTM-AM model and the adaptive fuzzy control model, adjusting model parameters and fuzzy control rules to adapt to various changes during construction. If a deviation is found between the actual pressure control effect and the model prediction, the cause is analyzed, and the model is optimized accordingly to ensure that the pressure control within the cutterhead chamber is always at its optimal state.

[0045] Example 3

[0046] A machine learning-based intelligent control system and method for pressure inside the cutterhead chamber of a tunnel boring machine (TBM) aims to achieve real-time intelligent control of pressure inside the cutterhead chamber through big data analysis and machine learning technology, thereby improving construction quality and safety.

[0047] System Architecture: Data Acquisition Module: Sensors: Various sensors installed inside the tunnel boring machine's cutterhead chamber to collect data such as pressure, temperature, humidity, and gas composition. Data Transmission: Data is transmitted to the central server via wired or wireless networks. Data Preprocessing Module: Data Cleaning: Removes invalid and redundant data, and fills in or deletes missing values. Feature Extraction: Extracts key features, such as pressure, temperature, humidity, and gas composition within the cutterhead chamber. Model Building Module: Solution 1: Time series prediction model based on Long Short-Term Memory (LSTM) network; Solution 2: Multivariate prediction model based on Graph Neural Network (GNN); Solution 3: Adaptive control model based on Bayesian optimization. Pressure Control Module: Real-time Monitoring: Monitors pressure changes within the cutterhead chamber in real time. Pressure Prediction: Predicts future pressure changes within the cutterhead chamber using the model, adjusting control strategies in advance. Control Execution: Adjusts the pressure within the cutterhead chamber based on prediction results and optimization strategies to ensure construction quality and safety. Human-Machine Interaction Module: Visual Interface: Provides a user-friendly visual interface displaying the real-time status and control information within the cutterhead chamber. Operational suggestions: Based on the model's prediction results, operational suggestions are generated to guide on-site engineers in their operations.

[0048] Solution 1: Time Series Prediction Model Based on Long Short-Term Memory Network (LSTM)

[0049] Data preprocessing: The Pandas library was used to clean the data, removing invalid and redundant data, and imputing or deleting missing values. Key features were extracted, such as pressure, temperature, humidity, and gas composition within the cutter head chamber. Model building: Long Short-Term Memory (LSTM) network:

[0050] Gating mechanism:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057] in, It's an input gate. It is the Gate of Oblivion. It's an output gate. It is a candidate cell state. It is a cellular state. σ represents the hidden state, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, and ⊙ represents element-wise multiplication.

[0058] Model Training: Data Labeling: The collected data was labeled to indicate pressure changes within the cutter head chamber. Training Process: The model was trained using the Keras library, and cross-validation was used to evaluate its performance. Hyperparameter Optimization: The model's hyperparameters were optimized using GridSearchCV and RandomizedSearchCV methods to improve its generalization ability.

[0059] Pressure Control: Real-time Monitoring: Monitors pressure changes within the cutterhead chamber in real time. Pressure Prediction: Predicts future pressure changes within the cutterhead chamber using models, allowing for proactive adjustments to control strategies. Control Execution: Adjusts the pressure within the cutterhead chamber based on prediction results and optimized strategies to ensure construction quality and safety.

[0060] Solution 2: Multivariate prediction model based on graph neural networks (GNN)

[0061] Data modeling: The structure and operational data within the tunnel boring machine's cutterhead chamber are abstracted into a graph structure, where nodes represent different sensors and control points, and edges represent the relationships between them. Topological features of the graph are extracted, such as node degree, path length, and connectivity.

[0062] Model building:

[0063] Graph convolution operations:

[0064]

[0065] in, It is the node feature matrix of the l-th layer. It is a normalized adjacency matrix. It is the weight matrix of the l-th layer, and σ is the activation function (e.g., ...). Attention mechanism:

[0066]

[0067]

[0068]

[0069] in, It is the attention score between node i and node j, where a is the attention vector and W is the weight matrix. It is the feature vector of node i. It is the set of the nodes' neighbors, ( It is a normalized attention score. It is the updated node feature vector. It is an activation function (such as ReLU).

[0070] Model Training: Data Labeling: Graph-structured data is labeled to mark pressure changes within the cutterhead chamber. Training Process: The model is trained using the PyTorchGeometric library, and cross-validation is employed to evaluate its performance. Hyperparameter Optimization: The model's hyperparameters are optimized using GridSearchCV and RandomizedSearchCV methods to improve its generalization ability.

[0071] Pressure Control: Real-time Monitoring: Monitors pressure changes within the cutterhead chamber in real time. Pressure Prediction: Predicts future pressure changes within the cutterhead chamber using models, allowing for proactive adjustments to control strategies. Control Execution: Adjusts the pressure within the cutterhead chamber based on prediction results and optimized strategies to ensure construction quality and safety.

[0072] Solution 3: Adaptive Control Model Based on Bayesian Optimization

[0073] Data preprocessing: The Pandas library is used to clean the data, removing invalid and redundant data, and filling or deleting missing values. Key features are extracted, such as pressure, temperature, humidity, and gas composition within the cutter head chamber.

[0074] Model building: Bayesian optimization:

[0075] Gaussian process regression (GPR):

[0076]

[0077] Where p(y|f,X) is the likelihood function. It is a prior distribution.

[0078] Bayesian optimization:

[0079]

[0080] in, It is a get function. It is existing observational data.

[0081] Model Training: Data Labeling: The collected data is labeled to indicate pressure changes within the cutter head chamber. Training Process: The model is trained using the Scikit-Optimize library, and its performance is evaluated using cross-validation. Hyperparameter Optimization: The model's hyperparameters are optimized using Bayesian optimization methods to improve its generalization ability.

[0082] Pressure Control: Real-time Monitoring: Monitors pressure changes within the cutterhead chamber in real time. Pressure Prediction: Predicts future pressure changes within the cutterhead chamber using models, allowing for proactive adjustments to control strategies. Control Execution: Adjusts the pressure within the cutterhead chamber based on prediction results and optimized strategies to ensure construction quality and safety.

[0083] Performance testing and comparative experiments: Dataset: Publicly available tunnel boring machine (TBM) operation datasets, such as TBM operation data from a specific tunnel project, are used. Simulated TBM operation datasets are constructed, covering various pressure variations and control scenarios. Performance indicators: Control accuracy: The accuracy of pressure control within the TBM cutterhead chamber. Response time: The response time of pressure control within the TBM cutterhead chamber. Stability: The stability of pressure control within the TBM cutterhead chamber. Construction quality: The quality of TBM construction, such as the straightness and smoothness of the tunnel.

[0084] Comparative Experiments: Traditional Method: Using traditional shield tunneling machine cutterhead chamber pressure control methods, such as PID controllers. Machine Learning Method: Using traditional machine learning methods, such as Support Vector Machine (SVM) and Random Forest (RF). Deep Learning Method: Using traditional deep learning methods, such as DNN and CNN.

[0085] Reinforcement learning methods: Traditional reinforcement learning methods such as Q-Learning and DQN are used. Invention method: Three solutions are used, based on LSTM, GNN, and Bayesian optimization.

[0086] Experimental Results: Control Accuracy: The three solutions of this invention significantly outperform traditional methods and machine learning methods in terms of control accuracy. Response Time: The three solutions of this invention significantly outperform traditional methods and machine learning methods in terms of response time. Stability: The three solutions of this invention significantly outperform traditional methods and machine learning methods in terms of stability. Construction Quality: The three solutions of this invention significantly outperform traditional methods and machine learning methods in terms of construction quality.

[0087] A machine learning-based intelligent control system and method for pressure inside the cutterhead chamber of a tunnel boring machine (TBM) is proposed. Through big data analysis and machine learning techniques, real-time intelligent control of the pressure inside the cutterhead chamber is achieved, improving construction quality and safety. Experimental results show that the method of this invention significantly outperforms traditional methods and machine learning methods in terms of control accuracy, response time, stability, and construction quality, demonstrating significant application value and promising prospects for wider application.

[0088] Example 4

[0089] Algorithm Model - Pressure Control Algorithm Based on Deep Reinforcement Learning: This paper designs an algorithm model based on Deep Reinforcement Learning (DRL) to control the pressure inside the cutterhead chamber of a tunnel boring machine (TBM). This algorithm combines the powerful representational capabilities of deep neural networks with the decision-making abilities of reinforcement learning, enabling it to automatically learn the optimal pressure control strategy based on the current construction status. The algorithm architecture includes a deep neural network as the policy network and an environment model. The policy network receives current TBM construction status information, including cutterhead chamber pressure, cutterhead rotation speed, propulsion speed, geological parameters, etc., and outputs control actions, namely the injection and discharge of compressed air or slurry. The environment model simulates the TBM construction process, updating the construction status and returning a new status and reward signal based on the control actions output by the policy network and the current status. The reward signal is determined based on factors such as the accuracy, stability, and construction efficiency of pressure control. For example, a positive reward is given when the pressure is close to the target value and the fluctuation is small, while a negative reward is given when the pressure deviation is large or the fluctuation is severe. The training process employs policy gradient methods in deep reinforcement learning algorithms, such as the Proximal Policy Optimization (PPO) algorithm. By continuously interacting with the environment, the policy network gradually learns the optimal control strategy to maximize long-term cumulative rewards.

[0090] Design Principles and Advantages: Deep reinforcement learning algorithms can automatically adapt to complex and ever-changing construction environments without relying on extensive prior knowledge or human experience. Through continuous interaction and learning with the environment, the algorithm can automatically adjust control strategies according to different geological conditions and construction scenarios, improving the accuracy and adaptability of pressure control. This algorithm can respond to changes in construction status in real time, exhibiting rapid decision-making and responsiveness. Compared to traditional control methods, it can more promptly address pressure fluctuations caused by changes in cutting speed and abrupt changes in soil properties, ensuring real-time pressure stability. Deep reinforcement learning algorithms can simultaneously consider multiple objectives, such as the accuracy, stability, and construction efficiency of pressure control. By optimizing the reward function, a balance and optimization of multiple objectives can be achieved, improving the overall efficiency of tunnel boring machine (TBM) construction.

[0091] Algorithm Model - Pressure Control Algorithm Based on Fuzzy Neural Network: An algorithm model based on a fuzzy neural network (FNN) is developed to control the pressure inside the cutterhead chamber of a tunnel boring machine (TBM). This model combines the fuzzy reasoning ability of fuzzy logic with the learning ability of neural networks, enabling it to handle uncertainties and nonlinear relationships during TBM construction. The fuzzy neural network consists of an input layer, a fuzzification layer, a rule layer, a defuzzification layer, and an output layer. The input layer receives TBM construction status information, such as the pressure inside the cutterhead chamber, cutterhead rotation speed, propulsion speed, and geological parameters. The fuzzification layer fuzzifies the input data, converting it into fuzzy linguistic variables. The rule layer infers based on fuzzy rules to determine the output fuzzy linguistic variables. The defuzzification layer converts the fuzzy output into precise control actions, namely the injection and discharge of compressed air or slurry. The training process employs a combination of backpropagation and genetic algorithms. First, the weights of the neural network are adjusted using the backpropagation algorithm to make the output of the fuzzy neural network approach the target value. Then, the genetic algorithm is used to optimize the fuzzy rules, improving the model's performance and generalization ability.

[0092] Design Principles and Advantages: Fuzzy neural networks can handle uncertainties and fuzziness in shield tunneling construction, such as the complexity and variability of geological conditions and noise in sensor data. Through fuzzy inference, uncertain input data can be converted into fuzzy linguistic variables, fuzzy rule reasoning can be performed to obtain fuzzy output, and then defuzzification can be used to obtain precise control actions, improving the robustness and adaptability of pressure control. This model has strong learning and generalization capabilities. Through neural network learning, fuzzy rules and weights can be automatically adjusted to adapt to different construction conditions and geological conditions. Compared with traditional fuzzy control methods, it can better handle nonlinear relationships and improve the accuracy and stability of pressure control.

[0093] Fuzzy neural networks can be trained by combining expert experience and historical data, making full use of prior knowledge and actual construction data to improve the reliability and practicality of the model. Furthermore, the model can learn and update online, continuously adjusting control strategies as construction progresses to adapt to changes in the construction process.

[0094] Algorithm Model - Pressure Control Algorithm Based on Multi-Agent Reinforcement Learning: This paper proposes an algorithm model based on Multi-Agent Reinforcement Learning (MARL) to control the pressure inside the cutterhead chamber of a tunnel boring machine (TBM). This model treats the TBM construction system as a complex system composed of multiple agents, each responsible for controlling a specific subsystem, such as the cutterhead drive system, propulsion system, and pressure control system. Through cooperation and competition among the agents, precise control of the pressure inside the cutterhead chamber is achieved. Each agent is trained using a deep reinforcement learning algorithm to learn the optimal control strategy. Agents exchange information and collaborate through communication and coordination mechanisms to jointly optimize the performance of the entire system. For example, the cutterhead drive agent can adjust the cutting speed and direction of the cutterhead based on the current cutterhead rotation speed and torque information to influence the pressure inside the cutterhead chamber. The propulsion agent can adjust the propulsion force and direction of the TBM based on the current propulsion speed and displacement information to coordinate with the cutterhead cutting and pressure control. The pressure control agent adjusts the injection and discharge of compressed air or slurry based on information such as the pressure inside the cutterhead chamber, the cutterhead rotation speed, and the feed speed, achieving precise pressure control. The training process employs distributed reinforcement learning algorithms, such as Independent Q-Learning or Deep Deterministic Policy Gradient (DDPG). Through parallel training and collaboration among multiple agents, the optimal control strategy is gradually learned to maximize the long-term cumulative reward of the entire system.

[0095] Design Principles and Advantages: Multi-agent reinforcement learning algorithms can better simulate the complexity and dynamics of tunnel boring machine (TBM) construction systems. Decomposing the system into multiple agents allows for optimized control of each subsystem, improving control accuracy and efficiency. Simultaneously, cooperation and competition among agents promote overall system performance optimization, enhancing the stability and reliability of pressure control. The algorithm exhibits strong adaptability and scalability. The number and functions of agents can be flexibly adjusted according to different construction needs and system configurations, adapting to various TBM construction scenarios. Furthermore, with continuous development of construction technology and system upgrades, new agents can be easily added or the control strategies of existing agents adjusted, improving system scalability and adaptability. Agent reinforcement learning algorithms can utilize distributed computing resources for parallel training, improving training efficiency and speed. Distributed training also reduces the training time and computing resource requirements of individual agents, lowering system cost and complexity.

[0096] Beneficial Effects: Three novel algorithm models based on deep reinforcement learning, fuzzy neural networks, and multi-agent reinforcement learning were developed. These algorithms have not yet been widely applied in the field of pressure control within the cutterhead chamber of tunnel boring machines (TBMs), providing innovative ideas and methods for solving existing technical problems. Through precise design of the algorithm architecture and training methods, effective control of complex nonlinear systems during TBM construction was achieved, improving the accuracy, stability, and adaptability of pressure control. All three algorithm models can simultaneously handle multiple factors affecting pressure within the cutterhead chamber, such as cutterhead rotation speed, propulsion speed, and geological parameters. Compared with traditional methods, they more comprehensively consider various factors during TBM construction, improving the overall performance of pressure control. Through comprehensive analysis and processing of multiple factors, the algorithms can better adapt to different construction conditions and geological conditions, achieving intelligent and adaptive pressure control. All three algorithm models have fast decision-making speed and response capabilities, enabling real-time responses to changes in construction status. Compared with traditional pressure feedback regulation systems, they can respond to pressure fluctuations more promptly, ensuring real-time pressure stability. The algorithm models can learn and update online, continuously adjusting control strategies as construction progresses to adapt to changes during the construction process. Real-time optimization improved the accuracy and reliability of pressure control, ensuring the safety and efficiency of tunnel boring machine (TBM) construction.

[0097] Example 5

[0098] Data Acquisition and Preprocessing: Data acquisition modules similar to those used in traditional methods are installed, including pressure sensors, flow sensors, torque sensors, speed sensors, displacement sensors, velocity sensors, and geological parameter detection sensors. Sampling frequencies are set according to the variation characteristics of different parameters to collect multi-source data such as pressure, flow rate, cutterhead speed, propulsion speed, and geological parameters within the shield tunneling cutterhead chamber. A deep learning-based outlier detection algorithm is used to clean and repair the data. The cleaned data is then standardized to ensure it can be used for subsequent model training within the same mathematical framework.

[0099] Pressure Control Model Construction and Training: A pressure control algorithm model based on deep reinforcement learning is constructed, including a policy network and an environment model. The policy network receives current construction state information and outputs control actions. The environment model simulates the shield tunneling process, updates the construction state based on the control actions and the current state, and returns a reward signal. The proximal policy optimization (PPO) algorithm is used for training. Through continuous interaction with the environment, the policy network gradually learns the optimal control policy to maximize long-term cumulative rewards.

[0100] Pressure Control and Optimization: During tunnel boring machine (TBM) construction, the data acquisition module continuously collects data and transmits it to the data processing center. The data processing center uses a trained strategy network to predict the optimal control actions. The control execution module precisely controls the opening of the compressed air or slurry injection and discharge valves based on these actions, adjusting the pressure within the cutterhead chamber. At regular construction intervals or time intervals, the strategy network is optimized online based on feedback from the control execution module and the latest operating status data of the TBM, adjusting model parameters to adapt to changes during construction.

[0101] Example 6

[0102] Data Acquisition and Preprocessing: Similar data acquisition and preprocessing steps as in Example 1 ensure data accuracy and reliability. Pressure Control Model Construction and Training: A pressure control algorithm model based on a fuzzy neural network is constructed, including an input layer, a fuzzification layer, a rule layer, a defuzzification layer, and an output layer. The input layer receives shield tunneling construction status information; the fuzzification layer fuzzifies the input data; the rule layer performs fuzzy rule inference; and the defuzzification layer converts the fuzzy output into precise control actions. Training is performed using a combination of backpropagation and genetic algorithms. First, the weights of the neural network are adjusted using the backpropagation algorithm to make the output of the fuzzy neural network approach the target value. Then, the genetic algorithm is used to optimize the fuzzy rules, improving the model's performance and generalization ability.

[0103] Pressure Control and Optimization: During tunnel boring machine (TBM) construction, the data processing center uses a trained fuzzy neural network model to predict control actions. The control execution module then adjusts the pressure inside the cutterhead chamber based on these actions. The fuzzy neural network is periodically trained and updated online. By incorporating newly acquired data and expert experience, the fuzzy rules and weights are adjusted to improve the model's adaptability and accuracy.

[0104] Example 7

[0105] Data acquisition and preprocessing: The data acquisition and preprocessing method is the same as in Example 1, providing accurate construction status information for multi-agent reinforcement learning.

[0106] Pressure control model construction and training: The tunnel boring machine (TBM) construction system is decomposed into multiple agents, each responsible for controlling a specific subsystem. Examples include a cutterhead drive agent, a propulsion agent, and a pressure control agent. Each agent is trained using a deep reinforcement learning algorithm to learn the optimal control strategy. Agents exchange information and collaborate through communication and coordination mechanisms. Distributed reinforcement learning algorithms, such as independent Q-learning or deep deterministic policy gradient algorithms, are used for training. Through parallel training and collaboration among multiple agents, the optimal control strategy is gradually learned to maximize the long-term cumulative reward of the entire system.

[0107] Pressure Control and Optimization: During tunnel boring machine (TBM) construction, multiple intelligent agents work collaboratively, adjusting their respective control actions based on current construction status information to achieve precise pressure control within the cutterhead chamber. As construction progresses, the agents continuously learn and update online, adjusting control strategies to adapt to changes in the construction process. Simultaneously, based on system performance indicators and feedback information, the communication and coordination mechanisms between the agents are optimized to improve the overall system performance.

[0108] The present invention has been described in detail above with reference to the embodiments. However, those skilled in the art will understand that, without departing from the spirit of the present invention, various specific parameters in the above embodiments can be changed to form multiple specific embodiments, all of which are common variations of the present invention, and will not be described in detail here.

Claims

1. A method for intelligent control of pressure inside the cutterhead chamber of a tunnel boring machine, characterized in that, The system includes collecting various data related to the pressure inside the cutterhead chamber and transmitting them to a data processing center. The data processing center preprocesses the collected data and uses a multivariate time series prediction and adaptive control algorithm to train the model and generate control commands. The control execution module precisely controls the opening of the compressed air or slurry injection and discharge valves according to the control commands to adjust the pressure inside the cutterhead chamber. The system also includes performing time series processing on the data to construct a multivariate time series dataset. This dataset is then input into an LSTM-AM model. The LSTM network has three hidden layers, each containing 64 LSTM units, used to learn the temporal features of the data. The attention mechanism automatically focuses on key variable features by calculating the importance weight of each variable to stress prediction at different times, using a gating mechanism. in, It's an input gate. It is the Gate of Oblivion. It's an output gate. It is a candidate cell state. It is a cellular state. It is in a hidden state. It is the Sigmoid activation function. It is a hyperbolic tangent activation function, where ⊙ represents element-wise multiplication. The collected data is labeled to indicate pressure changes within the cutterhead chamber. The model is trained using the Keras library, and its performance is evaluated using cross-validation. The hyperparameters are optimized using GridSearchCV and RandomizedSearchCV methods to improve the model's generalization ability. The pressure changes within the cutterhead chamber are monitored in real time. The model predicts future pressure changes within the cutterhead chamber, allowing for advance adjustments to the control strategy. Based on the prediction results and the optimized strategy, the pressure within the cutterhead chamber is adjusted to ensure construction quality and safety.

2. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 1, characterized in that, It also includes building an adaptive fuzzy control model, which dynamically adjusts the control strategy through a fuzzy inference system based on pressure prediction results and current pressure status information to achieve precise control of the pressure inside the cutterhead chamber.

3. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to any one of claims 1-2, characterized in that, The shield tunnel cutterhead chamber pressure intelligent control system includes a data acquisition module, a data processing center, and a control execution module. The data acquisition module is used to collect various types of data related to the pressure inside the cutterhead chamber and transmit them to the data processing center. The data processing center preprocesses the collected data and uses a multivariate time series prediction and adaptive control algorithm to train the model and generate control commands. The control execution module precisely controls the opening of the compressed air or mud injection and discharge valves according to the control commands, and adjusts the pressure inside the cutter head chamber.

4. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 3, characterized in that, The data acquisition module includes pressure sensors, flow sensors, torque sensors, speed sensors, displacement sensors, velocity sensors, and geological parameter detection sensors installed in the cutterhead compartment of the tunnel boring machine.

5. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 4, characterized in that, The pressure sensor, flow sensor, torque sensor, speed sensor, displacement sensor, velocity sensor, and geological parameter detection sensor are used to measure the pressure inside the chamber, the flow rate of the medium, the cutterhead torque, the cutterhead speed, the tunnel boring machine's advancing displacement, the advancing speed, and the geological characteristics of the strata, respectively.

6. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 5, characterized in that, The data acquisition module sets the corresponding sampling frequency according to the changing characteristics of different parameters. Pressure data and flow data are collected 5 times per second, cutterhead torque, rotation speed and shield machine propulsion data are collected once per second, and geological parameter data are collected once every 0.5 meters of propulsion.

7. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 5, characterized in that, The data processing center performs preprocessing on the collected data, including data cleaning, outlier detection and removal, and data standardization.

8. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 4, characterized in that, The data processing center uses a multivariate time series prediction model based on long short-term memory network and attention mechanism to predict pressure change trends. Based on the prediction results, an adaptive fuzzy control model is constructed, and the injection and discharge of compressed air or mud are dynamically adjusted through a fuzzy inference system.

9. The intelligent pressure control method for the cutterhead chamber of a tunnel boring machine according to claim 4, characterized in that, The control execution module uses a high-precision electric or pneumatic regulating valve with a control accuracy of ±0.5%. It has a real-time feedback monitoring function, which feeds back the control execution results to the data processing center in real time. It is equipped with a safety protection mechanism, which immediately activates the emergency braking device when the pressure is detected to rise or fall abnormally and exceed the set safety threshold.

Images