Real-time intelligent shield attitude control method based on machine learning

Abstract

The invention belongs to the field of intelligent control systems, and particularly relates to a shield attitude real-time intelligent control method based on machine learning, according to the method, a data acquisition module can more comprehensively and accurately reflect the actual condition of shield construction, and in the control process, key parameters such as the working state of a jack and the pressure of a hydraulic system are monitored in real time, so that the control accuracy is improved. Abnormal conditions can be found in time, an emergency plan can be started, and the safety of shield construction is guaranteed. According to the real-time monitoring and feedback optimization mechanism, the safety and reliability of shield construction are effectively improved, and the construction risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of shield construction, and in particular to a real-time intelligent control system and method for shield posture based on machine learning, which aims to analyze and process various data in the shield construction process through advanced machine learning algorithms, achieve high-precision control of the shield posture, and improve the safety, efficiency and quality of shield construction. Background Art

[0002] With the acceleration of urban underground space development, shield construction has been widely used in subway, tunnel and other engineering construction. Accurate control of shield posture is crucial to ensure the smooth progress of tunnel construction. However, the shield construction process faces complex geological conditions, changing construction environment and many uncertain factors, making shield posture control a very challenging task.

[0003] Traditional shield attitude control methods are mainly based on empirical formulas and simple sensor feedback control. Construction personnel correct the shield attitude by adjusting the thrust of the jack according to the preset tunnel design axis and the current position of the shield machine. This method has many limitations: first, it relies on empirical parameters and is difficult to adapt to changes in different geological conditions and construction conditions. When encountering complex strata, it is often necessary to manually adjust the control parameters frequently, resulting in low construction efficiency; second, the traditional control method uses sensor data in a relatively simple way, and can only provide simple feedback on the current position of the shield machine, and cannot effectively predict the trend of the change of the shield attitude. It is easy to have control lag problems, which in turn affects the forming quality of the tunnel, such as causing excessive deviation of the tunnel axis and difficulty in assembling the segments; third, due to the lack of in-depth mining and analysis of a large amount of historical construction data, it is difficult for traditional methods to summarize general laws and optimization strategies, and it is impossible to achieve intelligent and refined control of the shield attitude.

[0004] In recent years, the application of machine learning technology in the industrial field has made significant progress. Through learning and training of a large amount of data, machine learning models can automatically extract features and patterns in the data, establish a complex mapping relationship between input and output, and thus achieve prediction and decision-making of unknown data. In the field of shield attitude control, the introduction of machine learning technology is expected to overcome the shortcomings of traditional methods and improve the accuracy and adaptability of control. However, the application of machine learning to shield attitude control is still in the exploratory stage, and there are still some problems in the existing related research and applications: for example, some studies only use machine learning for simple data analysis, fail to deeply integrate with the shield attitude control system, and cannot directly achieve real-time control of the shield attitude; some machine learning models have problems such as difficulty in training, slow convergence speed, and poor generalization ability when facing high-dimensional and nonlinear data in shield construction; moreover, the existing methods are insufficient in detecting and processing abnormal situations in the shield attitude control process, making it difficult to ensure the safety of construction. Therefore, there is an urgent need for an innovative real-time intelligent control system and method for shield attitude based on machine learning to solve the above problems and promote the intelligent development of shield construction technology. Summary of the invention

[0005] The present invention provides a real-time intelligent control system and method for shield machine posture based on machine learning, which effectively solves the problems in the prior art by designing a unique algorithm model and innovative data processing flow.

[0006] A real-time intelligent control system for shield machine attitude based on machine learning, including:

[0007] The data acquisition module is used to collect various data during the shield construction process through various sensors installed on the shield machine, including using a high-precision laser total station to measure the absolute position coordinates and attitude angle of the shield machine, using a pressure sensor to measure the thrust of the jack, using an earth pressure sensor to monitor the pressure of the soil on the shield machine, and using a geological radar to detect the geological conditions in front of the shield machine. The collected data is transmitted to the data processing center through a high-speed data transmission line, and data encryption and verification technology is used during the transmission process;

[0008] The data processing center is used to pre-process the collected data, including data cleaning, outlier detection and elimination, data standardization, and feature extraction using a combination of principal component analysis (PCA) and deep belief network (DBN). Based on the extracted feature vectors, a support vector machine regression (SVR) model is trained for shield posture prediction and a decision tree model is used for shield posture control strategy generation.

[0009] The control execution module is used to control the jack action of the shield machine to adjust the shield posture according to the model prediction results and control strategies of the data processing center. It has real-time monitoring and feedback functions, can feed back the control execution results to the data processing center, and monitor the working status of the jack and the pressure of the hydraulic system during the control process, and activate the emergency plan when an abnormal situation is found.

[0010] Furthermore, the measurement accuracy of the laser total station in the data acquisition module can reach ±1mm and ±0.1°, the accuracy of the pressure sensor is ±0.5kN, and the measuring range of the soil pressure sensor is 0-1MPa with an accuracy of ±0.01MPa.

[0011] Furthermore, the position and posture data in the data acquisition module are collected once per second, the pressure data are collected five times per second, and the geological radar data are collected once at a certain distance according to the shield advancement speed.

[0012] Furthermore, the data processing center uses the 3σ criterion based on statistical principles to detect abnormal values ​​in the pressure data, and regards data beyond the range of ±3 times the standard deviation of the mean as abnormal values ​​and removes them.

[0013] Furthermore, the PCA of the data processing center determines the number of principal components to be retained according to the cumulative variance contribution rate reaching more than 90%, and the DBN is formed by stacking multiple restricted Boltzmann machines (RBMs).

[0014] Furthermore, the control execution module is connected to the hydraulic system of the shield machine and can accurately control the extension and thrust of each jack.

[0015] On the other hand, the present invention provides a real-time intelligent control method for shield posture based on machine learning, comprising the following steps:

[0016] Data collection and preprocessing steps: the data collection module collects shield construction data according to the set frequency and method and transmits it to the data processing center, which performs integrity check, data cleaning, outlier detection and standardization processing;

[0017] In the feature extraction and model training steps, a feature extraction method combining PCA and DBN is used. First, PCA is used to reduce the dimension of high-dimensional sensor data, and then the principal components are input into DBN to learn deep-level features. Based on the extracted feature vectors, the SVR model is trained for shield posture prediction and the decision tree model is used to generate control strategies.

[0018] Attitude control and optimization steps: During the shield construction process, the data acquisition module collects data in real time, the data processing center uses the SVR model to predict the trend of shield attitude changes, the decision tree model generates the control strategy, and the control execution module adjusts the jack thrust to achieve real-time control. The newly collected data is regularly added to the training data set, the model is retrained and the control strategy is updated, and the model and strategy are corrected according to the feedback information.

[0019] Furthermore, in the data collection and preprocessing step, the formula for standardizing the cleaned sensor data is: standardized value = (original value - mean) / standard deviation.

[0020] Furthermore, in the feature extraction and model training steps, the SVR model optimizes the model parameters by adjusting the kernel function parameters and the penalty factor and adopting the cross-validation method.

[0021] Furthermore, when the construction distance or time interval in the posture control and optimization step reaches a set value, the model and control strategy are optimized.

[0022] Beneficial technical effects of the invention:

[0023] Compared with the traditional data acquisition method, the data acquisition module of the present invention can more comprehensively and accurately reflect the actual situation of shield construction, and improve the reliability of control. The feature extraction method combining PCA and DBN can effectively process high-dimensional and nonlinear shield construction data, extract more representative and sensitive feature vectors, and provide high-quality data input for subsequent machine learning model training. For the first time, the support vector machine regression (SVR) model is used for shield posture prediction, and the control strategy is generated in combination with the decision tree model to realize the intelligent control of shield posture. Through the training and optimization of the machine learning model, the system can automatically learn the optimal control strategy under different geological conditions and construction conditions, without the need for frequent manual adjustment of control parameters, which greatly improves the efficiency of shield construction. At the same time, with the continuous accumulation of construction data, the prediction accuracy and control effect of the model will continue to improve, realizing the self-learning and self-adaptation of shield posture control. The control execution module has real-time monitoring and feedback functions, and can timely feedback the control execution results to the data processing center, realizing closed-loop monitoring of the shield posture control process. According to the feedback information, the data processing center can timely discover problems and deficiencies in the control process, optimize and adjust the model and control strategy, and improve the accuracy and stability of control. During the control process, the working status of the jack, the pressure of the hydraulic system and other key parameters are monitored in real time, abnormal situations can be discovered in time and emergency plans can be activated to ensure the safety of shield construction. This real-time monitoring and feedback optimization mechanism effectively improves the safety and reliability of shield construction and reduces construction risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the process. DETAILED DESCRIPTION

[0025] Example 1

[0026] 1. System Architecture

[0027] Data acquisition module: Various sensors installed on the shield machine are used to collect various data during the shield construction process. They include high-precision laser total stations, which are used to measure the absolute position coordinates (X, Y, Z coordinates) and attitude angles (pitch, yaw, and roll angles) of the shield machine in real time, with a measurement accuracy of ±1mm and ±0.1°; pressure sensors, which are distributed at various jack positions of the shield machine, measure the thrust of the jacks, with an accuracy of ±0.5kN; soil pressure sensors, which are installed around the cutterhead and shield of the shield machine, monitor the pressure of the soil on the shield machine, with a range of 0-1MPa and an accuracy of ±0.01MPa; and geological radars, which are used to detect the geological conditions in front of the shield machine, such as stratum type, rock and soil characteristics, groundwater level and other information, and convert them into digital geological feature data.

[0028] The frequency of sensor data collection is set according to the actual needs of shield construction. Generally, the frequency of position and posture data collection is once per second, the frequency of pressure data collection is 5 times per second, and the geological radar data is collected every certain distance (such as 0.5 meters) according to the shield advancement speed. The collected data is transmitted to the data processing center through a high-speed data transmission line. Data encryption and verification technology is used during the transmission process to ensure the integrity and accuracy of the data.

[0029] Data processing center: Equipped with high-performance computer server clusters, with powerful data storage and computing capabilities. Preprocess the collected data, including data cleaning, outlier detection and elimination, data standardization and other operations. For example, the 3σ criterion based on statistical principles is used to detect outliers in pressure data, and data that exceeds the mean ±3 times the standard deviation is considered as an outlier and eliminated; all sensor data are standardized so that the mean of the data is 0 and the variance is 1, so as to facilitate subsequent machine learning model processing.

[0030] Machine learning algorithms are used to extract features and train models for preprocessed data. A method combining principal component analysis (PCA) and deep belief network (DBN) is used for feature extraction. PCA can reduce the dimensionality of high-dimensional sensor data, extract the main characteristic components, and remove redundant information in the data; DBN further mines the deep features and nonlinear relationships in the data, and constructs a feature vector that can accurately reflect the complex relationship between the shield posture and various influencing factors through layer-by-layer unsupervised pre-training and back propagation fine-tuning. Based on the extracted feature vectors, a variety of machine learning models are trained, such as the support vector machine regression (SVR) model for the prediction of shield posture, and the decision tree model for the generation of shield posture control strategy.

[0031] Control execution module: According to the model prediction results and control strategy of the data processing center, the jack action of the shield machine is controlled to adjust the shield posture. This module is connected to the hydraulic system of the shield machine and can accurately control the extension and thrust of each jack. For example, when the SVR model predicts that the shield machine will have an upward deviation trend, the control execution module increases the thrust of the lower jack of the shield machine and reduces the thrust of the upper jack according to the control strategy generated by the decision tree, so that the shield machine posture is restored to the vicinity of the design axis. It has real-time monitoring and feedback functions, and can feed back the control execution results to the data processing center for further optimization of the model and control strategy. At the same time, during the control process, the working status of the jack, the pressure of the hydraulic system, etc. are monitored. Once an abnormal situation is found (such as jack overload, hydraulic system leakage, etc.), the emergency plan is immediately activated to ensure the safety of shield construction.

[0032] 2. Methodology

[0033] Data collection and preprocessing steps: The data collection module collects shield construction data according to the set frequency and method, and transmits it to the data processing center. In the data processing center, the data is first checked for integrity to ensure that no data is lost or damaged. Then, the data cleaning algorithm is used to remove noise and interference information in the data, such as filtering the jitter noise in the laser total station measurement data. The distance-based outlier detection method is used to mark and further analyze the data that differs too much from adjacent data points to determine whether it is an outlier and process it.

[0034] The sensor data after cleaning is standardized so that data of different types and ranges can be analyzed and processed on the same scale. For example, for pressure data, its original value is standardized and converted according to the formula, where is the original pressure value, is the mean of the pressure data, and is the standard deviation of the pressure data.

[0035] 1. Feature extraction and model training steps:

[0036] A feature extraction method combining PCA and DBN is used. First, PCA is used to reduce the dimensionality of high-dimensional sensor data to determine the number of principal components to be retained, which is generally determined based on the cumulative variance contribution rate reaching more than 90%. For example, for the original 20-dimensional sensor data, 5-8 principal components may be retained after PCA processing. Then, the principal components extracted by PCA are used as the input of DBN, which is composed of multiple restricted Boltzmann machines (RBMs) stacked together. Through unsupervised layer-by-layer training, deep feature representations in the data are learned.

[0037] Based on the extracted feature vectors, the SVR model is trained for shield posture prediction. The feature vectors in the historical construction data are used as input, and the corresponding shield posture data (such as position deviation, angle deviation, etc.) are used as output. By adjusting the kernel function parameters of SVR (such as the parameters of the radial basis function) and the penalty factor, the model parameters are optimized using the cross-validation method so that the model can accurately predict the changes in the shield posture. At the same time, the decision tree model is trained to generate the shield posture control strategy. According to different posture prediction results and construction conditions, the decision tree model determines the thrust adjustment plan for each jack.

[0038] 2. Posture control and optimization steps:

[0039] During the shield construction process, the data acquisition module collects data in real time, and the data processing center uses the trained SVR model to predict the change trend of the shield posture. According to the prediction results, the decision tree model generates the corresponding control strategy, and the control execution module adjusts the thrust of the jack according to the control strategy to achieve real-time control of the shield posture.

[0040] The model and control strategy are optimized at certain construction distances (such as 10 meters) or time intervals (such as 1 hour). The newly collected construction data are added to the training data set, the SVR model and decision tree model are retrained, and the model parameters and control strategy are updated to adapt to changes in geological conditions and construction conditions. At the same time, the model and control strategy are revised based on the information fed back by the control execution module to improve the accuracy and stability of the control.

[0041] Beneficial effects of the invention:

[0042] 1. Innovative data collection and processing system:

[0043] A comprehensive and high-precision data acquisition system has been built, integrating a variety of advanced sensors, which can obtain real-time data information on the position, posture, pressure, geology, etc. during the shield construction process, providing a rich data basis for shield posture control. Compared with traditional data acquisition methods, the data acquisition module of the present invention can more comprehensively and accurately reflect the actual situation of shield construction and improve the reliability of control.

[0044] In the data processing center, the feature extraction method combining PCA and DBN can effectively process high-dimensional, nonlinear shield construction data, extract more representative and sensitive feature vectors, and provide high-quality data input for subsequent machine learning model training. This feature extraction method overcomes the limitations of traditional feature extraction methods in processing complex data and improves the training efficiency and prediction accuracy of the model.

[0045] 2. Intelligent control strategy based on machine learning:

[0046] For the first time, the support vector machine regression (SVR) model was used for shield posture prediction, and combined with the decision tree model to generate control strategies, realizing intelligent control of shield posture. The SVR model can accurately predict the changing trend of shield posture based on historical construction data, take control measures in advance, and avoid the lag of traditional control methods; the decision tree model can flexibly generate targeted control strategies based on different prediction results and construction conditions, improving the adaptability and refinement of control.

[0047] Through the training and optimization of machine learning models, the system can automatically learn the best control strategies under different geological conditions and construction conditions, without the need for frequent manual adjustment of control parameters, greatly improving the efficiency of shield construction. At the same time, with the continuous accumulation of construction data, the prediction accuracy and control effect of the model will continue to improve, realizing the self-learning and self-adaptation of shield attitude control.

[0048] 3. Real-time monitoring and feedback optimization mechanism:

[0049] The control execution module has real-time monitoring and feedback functions, and can promptly feed back the control execution results to the data processing center, realizing closed-loop monitoring of the shield attitude control process. Based on the feedback information, the data processing center can promptly discover problems and deficiencies in the control process, optimize and adjust the model and control strategy, and improve the accuracy and stability of the control.

[0050] During the control process, the working status of the jack, the pressure of the hydraulic system and other key parameters are monitored in real time, abnormal situations can be discovered in time and emergency plans can be activated to ensure the safety of shield construction. This real-time monitoring and feedback optimization mechanism effectively improves the safety and reliability of shield construction and reduces construction risks.

[0051] Example 2

[0052] Data collection and preprocessing: Data collection modules are installed on urban subway shield machines, and laser total stations are installed at the rear of the shield machines to measure the position and posture data of the shield machines in real time; pressure sensors are installed at 32 jack positions to monitor the thrust of the jacks; soil pressure sensors are evenly distributed around the cutterhead and the shield body, with a total of 16 installed; geological radars are installed at the head of the shield machine to detect the geological conditions ahead. The data collection frequency is carried out according to the set requirements, with position and posture data collected once per second, pressure data collected 5 times per second, and geological radar data collected every 0.5 meters. After the collected data is transmitted to the data processing center, data cleaning and preprocessing are performed. For example, during a construction process, it was found that some data points in the pressure sensor data were significantly higher than the normal range. After being determined as abnormal values ​​by the 3σ criterion test, the average value of adjacent data points was used to replace them. All data are standardized so that the data can be analyzed at the same scale.

[0053] 1. Feature extraction and model training:

[0054] A feature extraction method combining PCA and DBN is used. For the collected raw data, after PCA processing, 8 principal components are extracted from the 50-dimensional data. These principal components contain key information of shield posture control, such as the overall stress of the shield machine, the impact of geological conditions on the shield machine, etc. These 8 principal components are input into DBN, which is composed of 3 RBM stacks. After layer-by-layer unsupervised pre-training and back propagation fine-tuning, a 12-dimensional feature vector is obtained.

[0055] Based on the extracted feature vectors, the SVR model and decision tree model are trained. The feature vectors in the historical construction data and the corresponding shield posture data are used as training sets. The parameters of the SVR model are optimized through cross-validation. The parameters are set to 0.1 and the penalty factor is set to 100. The trained SVR model can accurately predict the changes in the shield posture. For example, when predicting the vertical position deviation of the shield machine, the prediction error is within ±2mm. The decision tree model generates a thrust adjustment strategy for the jack based on different posture prediction results and construction conditions. For example, when it is predicted that the shield machine has a tendency to yaw to the left, the decision tree model determines the specific value of the thrust increase of the right jack.

[0056] 2. Posture control and optimization:

[0057] During the shield construction process, the data acquisition module collects data in real time, and the data processing center uses the trained SVR model to predict the change trend of the shield posture. For example, in a certain construction section, the SVR model predicts that the shield machine will have a slight increase in pitch angle. Based on this prediction result, the decision tree model adjusts the thrust of the front and rear jacks of the shield machine to restore the pitch angle to normal.

[0058] Every 10 meters of excavation, the newly collected construction data is added to the training data set, the SVR model and decision tree model are retrained, and the model parameters and control strategies are updated. After multiple optimizations, the prediction accuracy and control effect of the model are continuously improved, and the control accuracy of the shield posture is stabilized within the design requirements, effectively ensuring the construction quality and progress of the urban subway tunnel.

[0059] Example 3

[0060] Data acquisition and preprocessing: Due to the more complex construction environment of the cross-river tunnel, the data acquisition module installed on the shield machine has been optimized in terms of sensor type and quantity. In addition to conventional sensors, water flow sensors have been added to monitor the impact of the flow rate and pressure of the river water on the shield machine. A total of 4 water flow sensors with a measurement accuracy of ±0.05m / s and ±0.02MPa have been installed; high-precision gyroscopes have also been added to assist in measuring the attitude angle of the shield machine, with a measurement accuracy of up to ±0.05°. The data acquisition frequency has also been increased accordingly, with position and attitude data collected twice per second, pressure data collected eight times per second, geological radar data collected every 0.3 meters, and water flow sensor data collected once per second.

[0061] When the data processing center pre-processes the collected data, a special filtering algorithm is used to process the interference data generated by the river water flow. For example, the high-frequency fluctuations in the water flow sensor data are removed by low-pass filtering to extract the main trend of the river water's impact on the shield machine. After all the data are standardized, a feature extraction method combining PCA and DBN is used. Due to the higher dimensionality of the data, 10 principal components are extracted from the 60-dimensional data after PCA processing, and then input into the DBN composed of 4 RBM stacks to obtain a 15-dimensional feature vector.

[0062] Feature extraction and model training: The SVR model and decision tree model are trained based on the extracted feature vectors. In the construction of cross-river tunnels, the geological conditions are complex and changeable. In order to improve the adaptability of the model, a weighted SVR model is adopted to assign different weights according to the importance of different geological condition data. For example, different weight coefficients are assigned to the data of sand and rock layers, so that the model can more accurately predict the changes in the shield posture in different strata. When generating control strategies, the decision tree model fully considers the combined effects of river water pressure, flow rate and geological conditions. For example, when the shield machine is excavating in the sand layer near the bottom of the river and the river water pressure is high, the decision tree model will carefully adjust the thrust of the jack to prevent the shield machine from losing control of its posture.

[0063] Attitude control and optimization: During the shield construction of the cross-river tunnel, the data acquisition module collects data in real time, and the data processing center uses the trained SVR model to predict the changing trend of the shield attitude. For example, when crossing the fault zone at the bottom of the river, the SVR model predicts in advance that the shield machine will be subjected to large uneven geological forces. The decision tree model finely adjusts the thrust of each jack based on this prediction result, so that the shield machine can smoothly cross the fault zone. Every 5 meters of excavation or every 30 minutes, the newly collected construction data is added to the training data set, the SVR model and decision tree model are retrained, and the model parameters and control strategies are updated. Through continuous optimization, in the construction of the cross-river tunnel, the control accuracy of the shield attitude has reached an extremely high level, effectively avoiding safety accidents such as tunnel leakage and collapse caused by the loss of control of the shield attitude, and ensuring the smooth progress of the cross-river tunnel project.

[0064] Example 4

[0065] (I) Data acquisition module

[0066] Attitude data collection uses high-precision attitude measurement sensors installed on the shield machine to collect attitude data such as pitch angle, roll angle, horizontal deviation, vertical deviation, etc. in real time. These sensors have high-resolution and high-frequency data collection capabilities to ensure the timeliness and accuracy of attitude data.

[0067] Construction parameter collection collects relevant parameters during shield construction, including propulsion force, propulsion speed, cutter head speed, torque, grouting pressure and flow rate, etc. These parameters directly affect the posture change of the shield machine, and their values ​​are obtained through the sensor network and construction control system.

[0068] Geological data collection combines geological exploration data and geological detection equipment equipped on the shield machine (such as advanced geological prediction system) to obtain geological information in front of the shield, including stratum type and rock and soil parameters (such as hardness, density, water content, etc.). Geological data is of great significance for predicting changes in the shield posture.

[0069] (II) Data preprocessing module

[0070] Remove outliers, noise, and erroneous data from the collected data. For example, identify and remove obviously unreasonable posture data or construction parameter data caused by sensor failure or communication interference. At the same time, use appropriate interpolation methods (such as linear interpolation, polynomial interpolation, etc.) to fill in the missing values ​​in the data. Standardize different types of data so that they have the same dimension and numerical range. Commonly used standardization methods include Z-score standardization and Min-Max standardization. Through standardization, improve the training efficiency and stability of machine learning models. Extract and construct features that have an important impact on the shield posture from the original data. For example, calculate the rate of change of posture data, the relationship characteristics between construction parameters, and the correlation characteristics between geological data and posture data. Through feature engineering, the expressiveness of data is enhanced, which helps machine learning models to explore potential laws in the data.

[0071] (III) Machine Learning Model Module

[0072] Model selection and construction According to the characteristics of the shield attitude control problem, select a suitable machine learning model. For the time series prediction of attitude data, you can choose a recurrent neural network model such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU); for the problem of predicting attitude changes based on geological and construction parameters, you can choose a model such as a support vector machine (SVM) or a random forest (RF), or build a combination of models such as a multi-layer neural network (MLP).

[0073] Model training uses historical shield construction data (including preprocessed posture data, construction parameter data, and geological data) to train the selected machine learning model. The data set is divided into training set, validation set, and test set, and the parameters of the model are adjusted through optimization algorithms (such as gradient descent algorithm) to minimize the prediction error of the model on the training set and validation set while preventing overfitting.

[0074] Model evaluation and optimization uses a variety of evaluation indicators (such as mean square error (MSE), mean absolute error (MAE), accuracy, etc.) to evaluate the trained model on the test set. According to the evaluation results, the model is optimized, such as adjusting the model structure (increasing or decreasing the number of neurons, the number of layers, etc.), adjusting the training parameters (such as learning rate, batch size, etc.), using regularization technology, etc., to improve the prediction performance of the model.

[0075] (IV) Attitude Control Decision Module

[0076] Posture prediction uses a trained machine learning model to predict the posture change of the shield machine at the next moment based on the construction parameter data and geological data collected in real time. Accurate posture prediction results are obtained through time series analysis of posture data and comprehensive prediction of multiple factors.

[0077] The control strategy is generated based on the attitude prediction results and the preset shield attitude control target (such as keeping it within a certain range of the design axis) to formulate the corresponding control strategy. The control strategy includes adjusting the thrust of each cylinder of the propulsion system, adjusting the cutter head speed and torque, controlling the grouting parameters, etc., to achieve real-time adjustment of the shield attitude.

[0078] The control command output converts the control strategy into a specific control command and sends it to the propulsion control system and other related actuators of the shield machine. The control command is output in the form of a digital signal or an analog signal to ensure that the shield machine can adjust its posture according to the predetermined control strategy.

[0079] 4. Real-time intelligent control method of shield machine attitude based on machine learning

[0080] 1. Data collection and organization

[0081] Develop a data collection plan Based on the process and characteristics of shield construction, develop a detailed data collection plan. Determine the frequency of data collection, the location of collection points, and the collection methods of different types of data. Ensure that data collection is comprehensive and systematic, covering data from different geological conditions and construction stages.

[0082] Data integration and storage Integrate the collected attitude data, construction parameter data and geological data to establish a unified database. Use a suitable database management system (such as MySQL, MongoDB, etc.) to store data, and design a reasonable data structure and index to facilitate data query, retrieval and subsequent processing.

[0083] (II) Data preprocessing process

[0084] Data cleaning rule formulation: Specific data cleaning rules are formulated for different types of data. For example, for posture data, a reasonable threshold is set to determine abnormal values; for construction parameter data, erroneous data is identified based on the normal working range and experience value of the equipment. At the same time, the processing method and filling rules for missing data values ​​are clarified.

[0085] Data Standardization Operation Steps The data standardization operation steps are described in detail, including calculating the mean and standard deviation (for Z-score standardization) or the maximum and minimum values ​​(for Min-Max standardization) of the data. The data is converted according to the standardization formula, and the standardization parameters are recorded so that they can be restored in subsequent data use.

[0086] Feature Engineering Implementation Methods Explain how to implement feature engineering, such as calculating the rate of change of attitude data through differential operations, using correlation analysis methods to determine the correlation characteristics between construction parameters, and using cluster analysis and other methods to mine the intrinsic relationship between geological data and attitude data. Describe the process of feature construction and selection in detail to improve data quality and model performance.

[0087] 3. Machine Learning Model Training and Optimization Process

[0088] The training data partitioning method explains how to divide the collected shield construction data into training set, validation set and test set. Random partitioning, chronological partitioning and other methods can be used to ensure that the divided data sets are representative and independent. At the same time, considering the imbalance of data (such as the difference in data volume under different geological conditions), appropriate sampling methods (such as oversampling, undersampling, etc.) are used for processing.

[0089] Model training parameter adjustment strategy During the model training process, the strategy for adjusting parameters is elaborated in detail. For different machine learning models, parameters such as learning rate, batch size, and number of iterations are adjusted according to their characteristics. For example, for a neural network model, an adaptive learning rate algorithm (such as Adam, Adagrad, etc.) can be used, and the parameters can be gradually adjusted according to the changes in the loss function during the training process.

[0090] Application of model optimization techniques introduces various techniques applied to model optimization, such as adding regularization terms (L1 regularization, L2 regularization) to prevent model overfitting, using early stopping to stop training in advance based on the performance of the validation set, and using model fusion techniques (such as ensemble learning) to improve the predictive ability of the model.

[0091] 4. Attitude control decision and command execution

[0092] The posture prediction method describes the posture prediction method in detail based on the selected machine learning model and real-time input data. For prediction models based on time series, such as LSTM, it describes how to use historical posture data and current construction parameters and geological data as input to output the posture prediction value at the next moment. For other models, it explains how to use the output results of the model for posture prediction.

[0093] The principles of control strategy formulation are based on the attitude prediction results, and the principles of control strategy formulation are explained. According to the degree of deviation between the shield attitude and the design axis, the deviation direction, and the adjustable range of construction parameters, a reasonable control strategy is formulated. For example, when the shield machine has an upward vertical deviation, the thrust of the lower propulsion cylinder is appropriately increased, and the grouting parameters are adjusted to balance the attitude of the shield machine.

[0094] Control command execution and feedback adjustment describe how to accurately send control commands to the shield machine's actuators and adjust the shield's posture. At the same time, a feedback mechanism is established to monitor the changes in the shield's posture in real time, and feedback and adjust the control strategy based on the actual adjustment effect to ensure that the shield's posture can be stably controlled within the target range.

[0095] 5. Implementation of the system and method and case analysis

[0096] 1. System Implementation Technology

[0097] Software development platform and tools introduce the software platform and development tools used to develop a real-time intelligent control system for shield posture based on machine learning. For example, Python is selected as the main programming language, and deep learning frameworks such as TensorFlow and PyTorch are used to build and train machine learning models. Integrated development environments such as VisualStudioCode and PyCharm are used for code development and debugging.

[0098] Hardware integration and interface design describes the integration method of the system with the existing hardware equipment of the shield machine (such as attitude measurement sensors, construction control systems, actuators, etc.). Design hardware interfaces to ensure accurate data collection and effective transmission of control instructions. For example, communication connections with different devices are achieved through RS-485 interfaces, Ethernet interfaces, etc.

[0099] Example 5

[0100] Considering the particularity of submarine tunnel construction, a more comprehensive data acquisition module was installed on the shield machine. Multiple high-precision fiber grating sensors were installed at key parts of the shield machine to accurately measure the strain of the shield machine structure. A total of 20 fiber grating sensors were installed, and their measurement accuracy can reach ±1με (micro strain). In addition, sonar detectors were installed around the head of the shield machine to detect the geological structure and obstacles in the surrounding sea area, with a collection frequency of once every 0.2 meters. Conventional laser total stations, pressure sensors, earth pressure sensors, etc. were also installed according to the optimized layout. Among them, the laser total station is used to measure the position and posture data of the shield machine in real time, with a collection frequency of 3 times per second; pressure sensors are installed at 40 jack positions to monitor the jack thrust, with a collection frequency of 10 times per second; earth pressure sensors are evenly distributed around the cutterhead and shield body, with a total of 24 installed, and a collection frequency of 5 times per second.

[0101] After the collected data is transmitted to the data processing center, data cleaning is first performed. For fiber Bragg grating sensor data, since it is sensitive to strain changes, some noise data caused by slight environmental fluctuations may appear. The wavelet threshold denoising method is used for processing. By setting a suitable threshold, the wavelet coefficients less than the threshold are set to zero, thereby removing the noise component. For other sensor data, if individual data points are found to be beyond the normal range, the 3σ criterion is also used for detection and determination as abnormal values, and then replaced with the average value of adjacent data points. After completing data cleaning, all data are standardized so that various types of data can be analyzed at the same scale.

[0102] Feature extraction and model training: An innovative feature extraction method based on multimodal fusion and deep autoencoder (DAE) is adopted. First, the collected sensor data of different types (such as strain data of fiber Bragg grating sensors, geological structure data of sonar detectors, position and posture data of laser total stations, etc.) are divided into different modes according to their data characteristics. For each mode of data, feature encoding is performed through an independent deep autoencoder. For example, the deep autoencoder corresponding to the strain data of the fiber Bragg grating sensor has 3 hidden layers, and the number of neurons in each layer is 128, 64, and 32 respectively. The data is encoded and decoded through forward propagation and back propagation algorithms to extract the encoding vector that can represent the key features of the strain data.

[0103] The encoding vectors extracted from each modality are fused, and a fusion method based on the attention mechanism is adopted here. By calculating the importance weight of each modality encoding vector in the overall feature representation, the features of different modalities are weighted and fused to obtain a comprehensive feature vector. In this embodiment, a 20-dimensional comprehensive feature vector is obtained after fusion.

[0104] Based on this comprehensive feature vector, a hybrid neural network (HNN) model is trained, which combines the advantages of convolutional neural network (CNN) and recurrent neural network (RNN). The comprehensive feature vector in the historical construction data and the corresponding shield posture data are used as training sets. When training the HNN model, the CNN part is used to extract local spatial features in the comprehensive feature vector. The convolution kernel size is set to 3×3, and the number of convolution layers is 2. The spatial relationship between different sensor data is captured through convolution operations. The RNN part uses a long short-term memory network (LSTM) to process the time series characteristics of the shield posture data. The number of hidden layer neurons is set to 64, and the change law of the shield posture over time is captured through its internal gating mechanism. The entire HNN model is trained by the back propagation algorithm combined with the stochastic gradient descent optimizer, and the parameters of the model are adjusted so that the model can accurately predict the change of the shield posture.

[0105] Attitude control and optimization: During the shield construction of the undersea tunnel, the data acquisition module collects data in real time, and the data processing center uses the trained HNN model to predict the changing trend of the shield attitude. For example, when passing through a complex geological area on the seabed, the HNN model predicts that the shield machine will have a large lateral deviation trend. According to the model prediction results, the decision module finely adjusts the thrust of the jacks on the left and right sides of the shield machine. At the same time, combined with the surrounding geological structure information fed back by the sonar detector, the thrust adjustment strategy is further optimized to enable the shield machine to pass through the area smoothly. Every 8 meters of excavation or every 40 minutes, the newly collected construction data is added to the training data set, the HNN model is retrained, and the model parameters and control strategy are updated. Through continuous optimization, the control accuracy of the shield attitude has reached a very high level in the construction of the undersea tunnel, which effectively ensures the smooth progress of the undersea tunnel project and avoids safety accidents such as seawater leakage and tunnel structure damage caused by the loss of control of the shield attitude.

[0106] Comparative efficiency test: Select multiple groups of actual data of submarine tunnel shield construction, each group of data covers different construction stages and geological conditions. Each group of data is divided into training set, test set and validation set according to the ratio of 80% for training and testing and 20% for final verification. The following three methods are used for comparative tests: Method 1: The traditional modeling method based on principal component analysis (PCA) and support vector machine (SVR) is adopted, that is, only PCA is used for feature extraction (the number of principal components extracted is similar to the dimension of the feature vector after multimodal fusion in this embodiment), and then the extracted features are input into the SVR model for training and prediction. Method 2: Single modal feature extraction based on deep autoencoder (DAE) is adopted (only DAE feature extraction is performed on a certain type of sensor data, such as the position and posture data of the laser total station), and then the extracted features are input into a simple multi-layer perceptron (MLP) model for training and prediction. The modeling method based on multimodal fusion and deep autoencoder (DAE) feature extraction method combined with hybrid neural network (HNN) proposed in this embodiment (hereinafter referred to as the method of the present invention).

[0107] Evaluation indicators: Prediction accuracy: Calculate the proportion of the number of samples that correctly predict the shield posture change to the total number of samples. The higher the prediction accuracy, the stronger the model's ability to predict the shield posture change. Mean square error (MSE): Calculate the average of the squared difference between the predicted value and the actual value. The smaller the MSE, the higher the prediction accuracy. Control effect evaluation: Observe whether the shield machine's posture can be kept within the normal range after being controlled according to the model prediction results during the actual construction process, and whether safety accidents such as seawater leakage and tunnel structure damage occur. Quantify it into a score. For example, if the shield machine's posture always remains within the normal range and no safety accidents occur, the score is 10 points; if there is a slight posture abnormality but no safety accident occurs, the score is 6 points; if there is a serious posture abnormality and a safety accident occurs, the score is 2 points.

[0108] Test results: In terms of prediction accuracy, the method of the present invention reaches an average of more than 90%, while the average accuracy of method one is about 75%, and the average accuracy of method two is about 80%. This is because the method of the present invention can comprehensively utilize the information of different sensor data through multimodal fusion, and the HNN model can better capture the complex characteristics of shield posture changes, and has higher prediction accuracy than the traditional PCA-SVR method and single-mode DAE-MLP method. In terms of mean square error (MSE), the method of the present invention reaches an average of less than 0.05, the average MSE of method one is about 0.15, and the average MSE of method two is about 0.1. The low MSE of the method of the present invention benefits from its precise feature extraction and powerful model fitting ability, which can more accurately predict the changes in shield posture, making the predicted value closer to the actual value. In terms of control effect evaluation, the average score of the method of the present invention is above 8 points, the average score of method one is about 6 points, and the average score of method two is about 7 points. Through accurate prediction and reasonable control strategy, the method of the present invention can better keep the shield machine posture within the normal range, effectively avoid the occurrence of safety accidents, and thus show obvious advantages in control effect.

[0109] Example 6

[0110] In mountain tunnel construction, due to the complex and changeable geological conditions and the presence of many rock layers and broken zones, the data acquisition module of the shield machine is configured in a targeted manner. A large number of microseismic sensors are installed at key parts of the shield machine, such as the cutter head, shield body, and propulsion cylinder. A total of 30 sensors are installed to monitor microseismic activities during construction, with a measurement accuracy of ±0.01 (Richter magnitude). At the same time, thermal imagers are installed around the shield machine to detect the temperature distribution of various components of the shield machine, with a collection frequency of 1 time per second. Conventional laser total stations, pressure sensors, and earth pressure sensors are also arranged according to the characteristics of mountain tunnel construction. The laser total station has a collection frequency of 2 times per second, and the pressure sensors are installed at 35 jack positions with a collection frequency of 7 times per second. The earth pressure sensors are evenly distributed around the cutter head and shield body, with a total of 20 installed, and a collection frequency of 4 times per second.

[0111] After the collected data is transmitted to the data processing center, data cleaning is first performed. For microseismic sensor data, since it is extremely sensitive to tiny vibrations, a large amount of noise data may appear. An adaptive filtering algorithm is used for processing. The algorithm can automatically adjust the filtering parameters according to the real-time changes of the data and effectively remove the noise components in the microseismic sensor data. For other sensor data, if individual data points are found to be beyond the normal range, the 3σ criterion is also used for detection and determination as outliers, and then replaced with the average value of adjacent data points. After completing data cleaning, all data are standardized so that all types of data can be analyzed at the same scale.

[0112] Feature extraction and model training: A feature extraction method based on local feature preservation and sparse coding (Sparse Coding) is adopted. First, the various types of sensor data collected (such as microseismic activity data of microseismic sensors, temperature distribution data of thermal imagers, position and posture data of laser total stations, etc.) are processed in blocks, and the data of each sensor is divided into several small blocks according to a certain time window or spatial area. For example, for microseismic sensor data, it is divided into blocks according to a time window of every 10 seconds, and sparse coding operations are performed on each small block of data. By defining a suitable dictionary (composed of some basis vectors), the sparse coding algorithm is used to enable each small block of data to find the most matching basis vector combination in the dictionary, so that the data is represented in the form of sparse vectors. In this process, the local features of the data can be retained because the sparse coding of each small block of data is based on its own local characteristics.

[0113] The sparse vectors obtained after sparse coding of all small blocks of data are spliced ​​to obtain a comprehensive feature vector. In this embodiment, a 25-dimensional comprehensive feature vector is obtained after splicing.

[0114] Based on this comprehensive feature vector, a model based on the combination of variational autoencoder (VAE) and random forest (RF) is trained. The comprehensive feature vectors in the historical construction data and the corresponding shield posture data are used as training sets. During the training process, the VAE part is used to encode and decode the comprehensive feature vector. By adjusting the parameters of its encoder and decoder, VAE can learn the potential feature representation of the comprehensive feature vector. The RF part takes advantage of the random forest and classifies and predicts the shield posture changes by constructing multiple decision trees. When training the RF part, the number of decision trees is set to 100, and other parameters of RF are adjusted through cross-validation and other methods, so that RF can accurately predict the changes in the shield posture.

[0115] Attitude control and optimization: During the shield construction process in mountainous areas, the data acquisition module collects data in real time, and the data processing center uses the trained VAE-RF combined model to predict the changing trend of the shield attitude. For example, when passing through a rock crushing zone, the VAE-RF model predicts that the shield machine will have a large pitch angle change trend. According to the model prediction results, the decision module finely adjusts the thrust of the front and rear jacks of the shield machine, and combines the microseismic activity information fed back by the microseismic sensor to further optimize the thrust adjustment strategy, so that the shield machine can pass through the crushing zone smoothly.

[0116] Every 6 meters of excavation or every 30 minutes, the newly collected construction data is added to the training data set, the VAE-RF combined model is retrained, and the model parameters and control strategies are updated. Through continuous optimization, the control accuracy of the shield attitude has reached a high level in mountain tunnel construction, effectively ensuring the smooth progress of mountain tunnel projects and avoiding safety accidents such as tunnel collapse and equipment damage caused by the loss of control of the shield attitude.

[0117] Comparative efficiency test: Select multiple sets of actual data of shield construction of mountain tunnels, each set of data covers different construction stages and geological conditions. Divide each set of data into training set, test set and validation set according to the ratio of 80% for training and testing and 20% for final verification. The following three methods are used for comparative tests: Method 1: Use the traditional modeling method based on artificial feature extraction (such as simple calculation of statistical features such as mean and variance) and decision tree (DT), that is, manually calculate some simple statistical features of sensor data, and then input these features into the decision tree model for training and prediction.

[0118] Method 2: Use single feature extraction based on principal component analysis (PCA) (only perform PCA feature extraction on a certain type of sensor data, such as the position and posture data of the laser total station), and then input the extracted features into a support vector machine (SVR) model for training and prediction. This embodiment proposes a modeling method based on local feature retention and sparse coding (SparseCoding) feature extraction method combined with a variational autoencoder (VAE) and a random forest (RF) (hereinafter referred to as the method of the present invention).

[0119] Evaluation indicators: Prediction accuracy: Calculate the proportion of samples with correctly predicted shield posture changes to the total number of samples. The higher the prediction accuracy, the stronger the model's ability to predict shield posture changes. Mean square error (MSE): Calculate the average of the squared difference between the predicted value and the actual value. The smaller the MSE, the higher the prediction accuracy. Control effect evaluation: By observing the actual construction process of the shield machine, after controlling it according to the model prediction results, whether the posture of the shield machine can be maintained without a safety accident, the score is 10 points; if there is a slight posture abnormality but no safety accident occurs, the score is 6 points; if there is a serious posture abnormality and it causes a safety accident, the score is 2 points.

[0120] Experimental results: In terms of prediction accuracy, the method of the present invention reaches an average of more than 88%, while the average accuracy of method one is about 70%, and the average accuracy of method two is about 75%. This is because the method of the present invention can more deeply mine the characteristics of sensor data through local feature retention and sparse coding, and the VAE-RF combined model can better utilize these features for prediction, and has higher prediction accuracy than the traditional artificial feature extraction-decision tree method and PCA-SVR method. In terms of mean square error (MSE), the method of the present invention reaches an average of less than 0.06, the average MSE of method one is about 0.12, and the average MSE of method two is about 0.1. The low MSE of the method of the present invention is due to its precise feature extraction and powerful, and the VAE-RF combined model can better utilize these features for prediction, and has higher prediction accuracy than the traditional artificial feature extraction-decision tree method and PCA-SVR method.

[0121] In terms of control effect evaluation, the average score of the method of the present invention is above 8 points, the average score of method 1 is about 5 points, and the average score of method 2 is about 6 points. Through accurate prediction and reasonable control strategy, the method of the present invention can better keep the posture of the shield machine within the normal range, effectively avoid the occurrence of safety accidents, and thus show obvious advantages in control effect.

[0122] The present invention has been described in detail above in conjunction with the embodiments. However, those skilled in the art will appreciate that, without departing from the spirit of the present invention, the specific parameters in the above embodiments may be changed to form a plurality of specific embodiments, which are all within the common variation range of the present invention and will not be described in detail herein.

Claims

1. A real-time intelligent control method for shield machine attitude based on machine learning, characterized in that: The following steps are involved: Data collection and preprocessing steps: the data collection module collects shield construction data according to the set frequency and method and transmits it to the data processing center, which performs integrity check, data cleaning, outlier detection and standardization processing; In the feature extraction and model training steps, a feature extraction method combining PCA and DBN is used. First, PCA is used to reduce the dimension of high-dimensional sensor data, and then the principal components are input into DBN to learn deep-level features. Based on the extracted feature vectors, the SVR model is trained for shield posture prediction and the decision tree model is used to generate control strategies. Attitude control and optimization steps: During the shield construction process, the data acquisition module collects data in real time, the data processing center uses the SVR model to predict the trend of shield attitude changes, the decision tree model generates the control strategy, and the control execution module adjusts the jack thrust to achieve real-time control. The newly collected data is regularly added to the training data set, the model is retrained and the control strategy is updated, and the model and strategy are corrected according to the feedback information.

2. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: When the construction distance or time interval in the posture control and optimization step reaches a set value, the model and control strategy are optimized.

3. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: The feature vectors in the historical construction data are taken as input, and the corresponding shield posture data are taken as output. By adjusting the kernel function parameters and penalty factors of SVR, the model parameters are optimized using the cross-validation method, so that the model can accurately predict the changes in the shield posture. At the same time, the decision tree model is trained to generate the shield posture control strategy. According to different posture prediction results and construction conditions, the decision tree model determines the thrust adjustment plan of each jack.

4. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: Regularization terms are added to prevent the model from overfitting, early stopping is used to stop training early based on the performance of the validation set, and ensemble learning is used to improve the predictive ability of the model.

5. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: The rate of change of attitude data is calculated through differential operation, the correlation analysis method is used to determine the correlation characteristics between construction parameters, and the cluster analysis method is used to explore the intrinsic relationship between geological data and attitude data. The process of feature construction and selection is described in detail to improve data quality and model performance.

6. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: Calculate the rate of change of posture data, the relationship characteristics between construction parameters, and the correlation characteristics between geological data and posture data, and enhance the expressiveness of data through feature engineering.

7. The real-time intelligent control method of shield machine attitude based on machine learning according to claim 1 is characterized in that: Establish a feedback mechanism to monitor the changes in the shield machine's posture in real time, and make feedback adjustments to the control strategy based on the actual adjustment effect to ensure that the shield machine's posture can be stably controlled within the target range.

8. The real-time intelligent control method of shield machine attitude based on machine learning according to any one of claims 1 to 7, characterized in that: A real-time intelligent control system for shield machine attitude based on machine learning is used, including: The data acquisition module includes using a high-precision laser total station to measure the absolute position coordinates and attitude angle of the shield machine, using a pressure sensor to measure the thrust of the jack, using an earth pressure sensor to monitor the pressure of the soil on the shield machine, and using a geological radar to detect the geological conditions in front of the shield machine. The collected data is transmitted to the data processing center through a high-speed data transmission line, and the data is encrypted and verified during the transmission process; The data processing center is used to clean the collected data, detect and remove outliers, and perform data standardization operations. It also uses a combination of principal component analysis and deep belief networks to extract features. Based on the extracted feature vectors, a support vector machine regression model is trained for shield posture prediction and a decision tree model is used for shield posture control strategy generation. The control execution module is used to control the jack action of the shield machine to adjust the shield posture according to the model prediction results and control strategies of the data processing center. It has real-time monitoring and feedback functions, can feed back the control execution results to the data processing center, and monitor the working status of the jack and the pressure of the hydraulic system during the control process, and activate the emergency plan when abnormal conditions are found; The data acquisition module also includes a feature extraction method based on local feature retention and sparse coding to process the various types of sensor data collected in blocks, and uses a sparse coding algorithm to enable each small block of data to find the most matching basis vector combination in the dictionary, thereby representing the data in the form of sparse vectors.

9. The real-time intelligent control system for shield machine attitude based on machine learning according to claim 8 is characterized in that: It also includes taking the comprehensive feature vectors in the historical construction data and the corresponding shield posture data as training sets. During the training process, VAE is used to encode and decode the comprehensive feature vectors. By adjusting the parameters of its encoder and decoder, VAE can learn the potential feature representation of the comprehensive feature vector; the RF part takes advantage of the random forest and classifies and predicts the shield posture changes by constructing a decision tree.

Images