A shield tunneling posture control method and system based on a transformer and a reverse prediction mechanism

Abstract

The application discloses a shield tunneling posture control method and system based on a Transformer and a reverse prediction mechanism, the method comprising data acquisition, data preprocessing, configuration quantity and target quantity distinction, time sequence sample construction, Transformer reverse prediction model construction and training, posture control execution and adaptive learning; the system comprises a data acquisition unit, a data preprocessing unit, a configuration quantity and target quantity distinction unit, a time sequence sample construction unit, a Transformer reverse prediction model training and reasoning unit, a posture control execution unit and an adaptive learning unit; first, shield tunneling multi-source time sequence data is acquired, and after systematic preprocessing, configuration quantity and target quantity are distinguished, an optimal configuration quantity is inferred in real time by using a reverse prediction model of a Transformer encoder, and is converted into a control instruction to drive a shield machine to execute posture closed-loop adjustment; the problems of complex reverse solving and poor real-time performance of traditional methods are overcome, and the precision, adaptability and engineering practicability of posture control are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and artificial intelligence application technology in underground engineering, specifically to attitude control technology in shield tunneling, and in particular to a shield tunneling attitude control method and system based on Transformer and reverse prediction mechanism. Background Technology

[0002] With the rapid development of urban underground space, the shield tunneling method, with its advantages of high construction efficiency and minimal environmental disturbance, has become the mainstream construction method for projects such as rail transit, municipal utility tunnels, and tunnels crossing rivers and seas. During shield tunneling, attitude control is a key technical aspect to ensure tunnel axis accuracy, prevent segment misalignment, and reduce surface settlement. The attitude parameters of the shield machine include spatial variables such as the direction of advance, pitch angle, yaw angle, and roll angle. These parameters directly determine the tunneling quality, construction safety, and project progress.

[0003] The core issue of tunnel boring machine (TBM) attitude control lies in how to derive and set corresponding configuration parameters, such as total thrust of the propulsion cylinders, propulsion pressure in each direction, cutterhead rotation speed, and penetration depth, based on desired target quantities, such as propulsion speed, jack stroke, or propulsion direction, in order to achieve precise attitude adjustment. This inverse control problem is characterized by high nonlinearity, multivariable coupling, and time dependence. In complex geological environments, such as heterogeneous strata, water-rich sand layers, or interbedded boulders, changes in strata conditions can lead to significant changes in the response modes of tunneling parameters, making it difficult for traditional forward control methods to achieve high-precision attitude adjustment.

[0004] In current engineering practice, attitude control mainly relies on two types of methods. The first type is the traditional control method based on empirical formulas and mechanism derivations. This method uses simplified physical models or empirical relationships to predict the target quantity from the configuration quantity through forward calculation, and then adjusts the configuration quantity through trial and error or iteration to achieve the desired target. However, this method has significant shortcomings: First, it is difficult to accurately characterize complex features such as multivariable coupling and strong nonlinearity, resulting in limited control accuracy under complex geological conditions; second, it lacks adaptability to dynamic changes in working conditions, requiring frequent manual parameter adjustments when the strata change abruptly, which is inefficient; third, the inverse solution process is complex, with significant control lag, making it difficult to respond to attitude deviations in real time.

[0005] The second category comprises data-driven intelligent control methods, such as PID control, fuzzy control, or neural network control. While these methods improve adaptability to some extent, they still suffer from the following drawbacks: First, they lack a direct reverse prediction mechanism. Most methods employ forward modeling followed by optimization algorithms to deduce configuration quantities, resulting in high computational complexity, poor real-time performance, and an inability to meet the low-latency control requirements of engineering sites. Second, they fail to adequately consider historical states. Traditional methods often make decisions based on the current state, neglecting the inertia and hysteresis of the tunnel boring machine (TBM) as a large mechanical device, leading to a lack of dynamic smoothness in the control strategy. Third, they fail to adequately characterize multi-variable coupling relationships. TBM tunneling involves multi-source monitoring data, and the dynamic interaction between configuration quantities and target quantities is complex, making it difficult for traditional models to fully capture these characteristics. First, it limits control precision; second, it has weak long-sequence dependency modeling ability. Traditional recurrent neural networks such as RNN or LSTM suffer from gradient vanishing and memory decay problems, making it unable to effectively capture the long-term evolution of parameters; third, it lacks a clear distinction mechanism between configuration quantities and target quantities. Existing methods do not clearly define the configuration quantities dominated by console commands and the target quantities reflecting the tunneling status, resulting in a high degree of coupling between control signals and state feedback, reducing the interpretability and accuracy of the model; fourth, the model has limited generalization ability. Under complex and changing geological environments and construction conditions, a single static model lacks a continuous learning and dynamic optimization mechanism, making it difficult to adapt to the engineering needs of long-distance, multi-stratum crossings.

[0006] Furthermore, inconsistencies in engineering noise and quality are frequently observed in shield tunneling monitoring data at the construction site, including sensor drift, data gaps, abnormal spikes, and sampling period mismatches. Without a systematic mechanism for anomaly detection, noise reduction, and redundancy suppression, the control model is susceptible to interference from invalid information, leading to deviations. Simultaneously, existing methods often employ fixed data formulas and static parameters, lacking the ability to learn and adapt as tunneling progresses, making it difficult to maintain stable performance in dynamic engineering scenarios.

[0007] In summary, existing technologies suffer from at least the following key problems: a lack of efficient reverse prediction mechanisms, resulting in low control efficiency and poor real-time performance; insufficient utilization of historical state information, leading to a lack of dynamic smoothness in control strategies; incomplete dynamic characterization of multivariate coupling relationships; weak long-sequence dependency modeling capabilities, making it difficult to capture long-term patterns; a lack of mechanisms to distinguish between configuration and target quantities, affecting model accuracy; and insufficient adaptive updating capabilities, failing to adapt to changes in operating conditions. These shortcomings urgently need to be addressed through technological innovation. Summary of the Invention

[0008] This invention aims to overcome several key shortcomings in existing shield tunneling attitude control technologies. Traditional methods lack an efficient reverse prediction mechanism, making it difficult to directly and in real-time deduce control commands based on the desired target quantity. Furthermore, they fail to fully utilize historical time-series information and lack the ability to characterize the dynamic coupling relationships and long-term dependencies among multiple variables, resulting in poor dynamics and limited accuracy of the control strategy. In addition, existing methods do not clearly distinguish between configuration quantities and target quantities, and the models lack adaptive update capabilities, making it difficult to adapt to the complex and ever-changing geological conditions and working conditions during tunneling.

[0009] Therefore, the technical objective of this invention is to construct an intelligent attitude control system that integrates reverse prediction, time series modeling, and online adaptive capabilities, which can accurately and efficiently deduce the optimal configuration based on the desired target quantity, thereby providing a real-time, accurate, and robust control strategy for tunnel boring.

[0010] To achieve the above technical objectives, the first aspect of this invention provides a shield tunneling attitude control method based on Transformer and reverse prediction mechanism, comprising the following steps: S1. Data Acquisition: Multi-source monitoring parameters are collected in real time during the tunnel boring machine (TBM) excavation process. These parameters include cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. The propulsion cylinder pressure data covers seven directions: top, top right, bottom right, right, top left, bottom left, and left. The jack speed data covers six directions: top right, bottom right, right, top left, bottom left, and left. All parameters are collected synchronously at a uniform sampling frequency, typically set to 1 Hz, and based on a uniform timestamp, forming a multi-dimensional time-series dataset with temporal consistency. The dataset is stored in a structured format, including a timestamp field and numerical fields for each parameter, ensuring data integrity and time-series alignment requirements for subsequent processing. S2, Data Preprocessing: The multidimensional time series dataset is systematically preprocessed to improve data quality and model robustness. This systematic preprocessing includes five steps: outlier detection and handling, data cleaning, missing value handling, signal denoising, and feature standardization, as detailed below: S201. Outlier Detection and Handling: The box plot method based on interquartile range is used to identify and remove outliers in the parameter sequences of multidimensional time series data; for missing segments formed by consecutive outliers, the sliding window midpoint interpolation method is used to fill them; for missing positions formed by isolated outliers, the linear interpolation method of adjacent time points is used to fill them. S202. Data cleaning: Delete all unlabeled or invalidly labeled data rows in the dataset, such as empty rows with the label NA, to ensure the validity of the training data; S203. Missing value handling: First, delete all columns with missing values, and then fill the remaining missing values ​​with the mean of the non-missing values ​​of each feature column to maintain data integrity. S204. Signal Denoising: Discrete wavelet transform is used to decompose and reconstruct the multidimensional time series dataset to suppress high-frequency noise and retain the low-frequency main signal that reflects the trend of stratigraphic change. Specifically, the db10 wavelet basis function in the Daubechies wavelet family is used to perform one-level wavelet decomposition, and the signal is reconstructed after suppressing the high-frequency detail coefficients based on the adaptive threshold method. S205. Feature Standardization: A standardization method is used to transform all features to a standard normal distribution with a mean of 0 and a variance of 1, in order to eliminate the interference of different physical units and numerical ranges on model training. The multidimensional time series dataset processed by the above steps constitutes a high-quality input feature set for subsequent model training; S3. Distinguishing between configuration quantity and target quantity: To construct a clear input for the reverse prediction model, the monitoring parameters are divided into two categories: configuration parameters and target parameters, based on the response relationship between each monitoring parameter and the instructions of the shield tunneling console. The configuration parameters refer to the operating parameters that are directly controlled by the shield tunneling control console and can be actively set. The set includes 11 parameters, such as cutterhead rotation speed, total thrust of propulsion cylinders, penetration depth, propulsion cylinder pressure in seven zones, and cutterhead torque. The target quantity refers to the feedback parameters that mainly reflect the actual state and spatial attitude of the tunneling. The set includes the propulsion jack stroke and average propulsion speed of the upper and lower sections, as well as the propulsion jack speed of the six sections, for a total of 9 parameters. The distinction between configuration variables and target variables is achieved through the following steps: First, the covariance matrix between the timing signals of control commands and the timing signals of each monitoring parameter is calculated to quantify their linkage relationship. Second, the response coefficients of each parameter are calculated based on the covariance matrix. Then, a threshold for the response coefficients is set based on engineering experience; parameters with response coefficients higher than the threshold are identified as configuration variables, and those lower are identified as target variables. Finally, the two types of parameters are feature-separated and normalized separately. This distinction mechanism effectively decouples control variables from state variables, providing clearly structured and physically meaningful input features for subsequent inverse prediction models.

[0011] S4. Construction of time series samples: Based on the sliding window method, time series samples for supervised learning are constructed. The core of this method is to predict the required allocation amount based on historical states and the current desired target amount. The specific construction process includes the following steps: S401, Sliding Window Design: Set the window length to... NEach time step, in the past continuous N The actual values ​​of all parameters within each time step are used as input features, and the configuration amount at the current time step is used as the prediction target; window size N It can be configured according to the data sampling frequency and the real-time requirements of prediction; S402. Sample Construction Strategy: For time series data of length T, a sliding window is used to generate samples. T - N There are 10 training samples. The input features for each sample include: history. N The target quantity sequence and history at each time step N The input sequence contains the configuration quantity sequence for each time step and the target quantity for the current time step; the configuration quantity for the current time step is padded with zero values; the complete input sequence includes... N +1 time steps, each time step has 20 feature dimensions, including 9 target dimensions and 11 configuration dimensions; the sample output is the configuration dimension of the current time step, with 11 dimensions; S403. Data normalization processing: Min-Max normalization method is applied to the configuration quantity and target quantity features respectively to linearly transform the data to the [0,1] interval, eliminating the influence of different physical dimensions on model training; S404. Dataset partitioning: Divide the sample data into training set, validation set and test set according to time sequence, with the training set accounting for 70%, the validation set accounting for 15% and the test set accounting for 15%, to maintain the continuity of the time series. Using the above time-series sample construction method, the system establishes a reverse prediction model to achieve an accurate mapping from historical state information and current target quantity to current configuration quantity. S5, Transformer Backward Prediction Model Construction and Training: The multidimensional features, after preprocessing and construction using time-series samples, are input into the inverse prediction model based on the Transformer encoder architecture for training. The Transformer inverse prediction model includes the following key modules: Input projection layer: Used to map the 20-dimensional input feature vector to the model's hidden dimension space through a linear transformation. d model Setting it to 128 enables high-dimensional representation of features; Position encoding module: The position encoding matrix generated by sine and cosine functions adds absolute position information to each time step of the input sequence, enabling the model to recognize the temporal order; Multi-head self-attention mechanism: Set the number of attention heads to 8, with each head having a dimension of 16, and map the input sequence to multiple subspaces to compute self-attention weights in parallel in order to capture the long-range dependencies between different time steps and different variables; Feedforward network: It adopts a two-layer fully connected network structure, with the first layer having a dimension of 512 and the second layer having a dimension of 128. The ReLU activation function is used to enhance the non-linear expressive power of the model. Residual connections and layer normalization: Residual connections are introduced after the multi-head self-attention layer and the feedforward network layer, followed by layer normalization. This design helps stabilize the training process, accelerate convergence, and alleviate the gradient vanishing problem. Encoder stacking: A complete encoder layer consisting of a multi-head self-attention mechanism, feedforward network, residual connections, and layer normalization is stacked. The number of stacked layers is... N It can be set according to the model capacity requirements, thereby refining higher-level temporal feature representations layer by layer; Output layer: Take the feature vector of the last time step output by the encoder, and map it to an 11-dimensional output space through a fully connected layer, corresponding to the configuration quantity prediction value at the current time step; During the model training phase, mean squared error was used as the loss function, and the Adam optimization algorithm was used to update the parameters. The initial learning rate was set to 0.0001, the batch size was set to 64, and the number of training epochs was set to 50. To improve the training effect, dynamic adjustment of the learning rate, early stopping strategy, and gradient pruning technique were adopted simultaneously to enhance the stability of model training and the final generalization performance. S6, Attitude Control Execution: During the actual tunneling process of the tunnel boring machine (TBM), the system, based on the desired attitude target, calls a fully trained Transformer backpropagation model to infer in real time the optimal configuration required to achieve the target, thereby driving the TBM to complete precise closed-loop attitude control. The specific execution process is as follows: S601, Real-time data acquisition: The system collects target quantity and configuration quantity data in real time at the current moment and at N historical time steps to form a complete time series sample that meets the model input requirements; S602. Data Preprocessing: Perform the same preprocessing procedures as the training phase on the collected real-time data, including outlier detection and processing, data cleaning, missing value handling, signal denoising and feature standardization, to ensure that the quality of the input data is consistent with that of the model training phase. S603, Data Normalization: Use the Min-Max normalizer saved during the training phase to normalize the configuration quantity and the target quantity respectively; S604, Model Inference: Input the preprocessed time series data into the trained Transformer inverse prediction model and infer the predicted value of the configuration quantity at the current time. S605. Result Inverse Normalization: Using the normalization parameters saved during the training phase, the normalized configuration quantity output by the model is inversely normalized to the original physical dimensions to obtain actual parameter values ​​that can be directly used for control. S606, Control command generation: Convert the denormalized configuration parameters into control commands that the tunnel boring machine control system can recognize, including the total thrust setting value of the propulsion cylinder, the propulsion pressure setting value in each direction, the cutterhead speed setting value, etc. S607, Attitude Adjustment Execution: After receiving control commands, the tunnel boring machine (TBM) control system drives the corresponding actuators to make adjustments, achieving precise control over the TBM's propulsion direction, speed, and attitude. For example, when it is desired to increase the propulsion speed, the system predicts and sets a higher total thrust of the propulsion cylinders; when it is desired to adjust the propulsion direction, the system predicts and sets the propulsion pressure difference in different directions.

[0012] S7, Adaptive Learning: To enable the model to continuously adapt to dynamically changing geological conditions and construction processes during tunneling, an online adaptive learning mechanism was established. This mechanism continuously monitors the model's predictive performance on newly acquired data and triggers incremental learning based on performance degradation, thereby achieving dynamic fine-tuning and continuous optimization of network weights and ensuring the model's predictive accuracy and robustness throughout the entire tunneling cycle. Its specific operation process is as follows: First, the system calculates the predicted configuration quantity corresponding to the target quantity in the new batch of measured data in real time, compares it with the actual configuration quantity of the batch, and calculates performance evaluation metrics, such as mean squared error. Second, the system sets a performance threshold; when the evaluation metric continuously exceeds the set threshold, it automatically triggers the incremental update process of the model. Then, the system uses newly collected data batches to perform small-scale incremental training on the already trained Transformer model at a learning rate much lower than the initial training rate, optimizing the network weights. During this process, strategies such as weight decay can be used to balance adaptation to new operating conditions and retention of historical knowledge, avoiding catastrophic forgetting. Finally, the system manages and evaluates the performance of the model versions before and after the update, completing the iterative upgrade of the model while ensuring stability.

[0013] Through the above mechanism, the system can continuously learn and adapt to new working conditions during the tunneling process, and always maintain a high-precision and stable prediction capability for configuration quantities.

[0014] The second aspect of this invention provides a shield tunneling attitude adaptive control system for implementing the above-mentioned method. The system adopts a modular architecture design, including a data acquisition unit, a data preprocessing unit, a configuration quantity and target quantity differentiation unit, a time series sample construction unit, a Transformer inverse prediction model training and inference unit, an attitude control execution unit, and an adaptive learning unit. Each unit is connected in sequence and realizes data interaction and command transmission through a system bus, together forming a complete intelligent control closed loop. The data acquisition unit is used to collect multi-source operating parameters in real time during the shield tunneling process, including cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke and jack speed, and record them synchronously according to a unified sampling frequency and timestamp. The data preprocessing unit is used to systematically preprocess the collected raw data, sequentially performing outlier detection and processing, data cleaning, missing value processing, signal denoising and feature standardization operations to improve data quality and model input stability. The configuration quantity and target quantity differentiation unit is used to differentiate multi-source monitoring parameters into configuration quantities and target quantities based on the response relationship between parameters and control commands; The temporal sample construction unit is used to construct supervised learning samples based on the sliding window method. It takes the target quantity and configuration quantity sequence of the past N steps and the current target quantity as input and the current configuration quantity as output to generate a sample set that meets the training requirements of the Transformer model. The Transformer inverse prediction model training and inference unit is used to build and train an inverse prediction model based on the Transformer encoder architecture, realize high-precision mapping from target quantity to configuration quantity, and support online inference and real-time prediction. The attitude control execution unit is used to convert the configuration quantity prediction value output by the model into actual control commands and drive the shield machine execution mechanism to adjust the attitude, thereby realizing closed-loop control. The adaptive learning unit is used to dynamically trigger the incremental update mechanism of the model based on the difference between the newly collected data and the prediction results, and to achieve continuous adaptation of the model to changes in working conditions through weight fine-tuning.

[0015] In summary, compared with the prior art, the beneficial effects of the present invention are: 1) An efficient and direct reverse prediction mechanism was established: The Transformer architecture was innovatively used to realize the end-to-end direct mapping from the expected target quantity to the required configuration quantity, which abandoned the complex process of "forward modeling and iterative reverse calculation" in the traditional method, and fundamentally improved the efficiency and real-time performance of control decision-making. 2) Deeply integrated time-series dynamic modeling was achieved: By introducing historical multi-step state information as model input, the inertia and hysteresis characteristics of the tunnel boring machine system were fully characterized, thereby generating a smoother and more stable control strategy and enhancing the coherence and reliability of dynamic control. 3) It has a strong ability to capture long-sequence dependencies: By utilizing the multi-head self-attention mechanism of the Transformer architecture, it can capture the global dependencies of all time points at once, effectively solving the gradient vanishing and memory decay problems of traditional recurrent neural network models when processing long-sequence data, and significantly improving the model's ability to capture the long-term evolution law between parameters. 4) Improved the overall modeling accuracy of multivariable coupling relationships: The model can jointly predict all configuration parameters in an integrated manner, and explicitly model the complex interaction and coupling relationships between configuration and target quantities and within each configuration quantity through the attention mechanism, which significantly improves the overall accuracy of multivariable system control. 5) Enhanced model interpretability and physical consistency: Through a strict mechanism for distinguishing between configuration quantities and target quantities, the control input and state feedback of the system are clearly defined, making the model structure fit the actual control logic and greatly improving the transparency and reliability of the model decision-making process. 6) Ensures high precision and low latency in engineering applications: Thanks to the parallel computing characteristics of Transformer, the model is highly efficient in both training and inference stages. The prediction error of key configuration quantities in actual tests is significantly lower than that of traditional methods, and the latency of a single inference is extremely low, with a single prediction time of less than 10 milliseconds, which fully meets the stringent requirements of real-time control on site. 7) A systematic data governance process was constructed: a multi-level preprocessing pipeline was integrated from anomaly detection, cleaning, repair to noise reduction and standardization, which effectively improved the quality and consistency of the original engineering data and laid a solid foundation for the robust performance of the model; 8) It endows the model with the ability to continuously evolve and adapt: ​​an online monitoring and incremental learning mechanism is designed, which enables the model to dynamically adjust and optimize based on new tunneling data, continuously adapt to changes in different strata and working conditions, and overcome the limitation of insufficient generalization ability of static models. 9) Achieved high-level closed-loop integration with the control system: The system design can be directly embedded into the existing control system of the tunnel boring machine to form a complete closed loop of state perception, intelligent decision-making and precise execution, which greatly improves the automation level and overall safety of the construction process. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of a shield tunneling attitude control method based on Transformer and reverse prediction mechanism, as shown in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a shield tunneling attitude adaptive control system according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of the configuration quantity and target quantity differentiation mechanism shown in an embodiment of the present invention; Figure 4This is a schematic diagram of the temporal sample construction and Transformer inverse prediction model architecture shown in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the attitude control execution process according to an embodiment of the present invention; Figure 6 This is a graph showing the model training curve results as illustrated in an embodiment of the present invention; Figure 7 This is a comparison chart of the configuration quantity output by the model shown in the embodiment of the present invention and the actual configuration quantity. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] Example 1

[0019] This embodiment provides specific implementation steps for a shield tunneling attitude control method based on Transformer and reverse prediction mechanisms. This method integrates data acquisition, preprocessing, feature differentiation, temporal modeling, model training, control execution, and adaptive learning to achieve efficient and direct mapping from desired target quantities to configured quantities, thereby improving the accuracy, real-time performance, and adaptability of shield tunneling attitude control.

[0020] S1. Data Acquisition: Multi-source operating parameters during the tunneling process are acquired in real-time through the tunnel boring machine (TBM) monitoring system. Acquired parameters include cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. Specifically, propulsion cylinder pressure data covers seven directions: up, upper right, lower right, right, upper left, lower left, and left. Jack stroke data covers two directions: up and down. Jack speed data covers six directions: upper right, lower right, right, upper left, lower left, and left. All parameters are sampled at a uniform frequency of 1 Hz and collected synchronously based on a unified timestamp, forming a multi-dimensional time-series dataset. The data is stored in a structured format, such as CSV, containing a timestamp field and numerical fields for each parameter to ensure temporal consistency and integrity.

[0021] S2. Data Preprocessing: Systematically preprocess the collected multidimensional time series data to improve data quality and model robustness. The preprocessing process includes the following sub-steps: S201. Outlier Detection and Handling: Outliers in each parameter sequence are identified using a boxplot method based on interquartile range. First, the first quartile of each parameter sequence is calculated. Q 1 With the third quartile Q 3 This leads to the interquartile range.Q R Set the outlier detection threshold to [ ]. Q 1 -1.5 Q R , Q 3 +1.5 Q R Values ​​outside this range are identified as outliers and removed. For missing data resulting from removal, outliers are filled in based on their continuity: consecutive outliers are filled using a sliding window median interpolation method; isolated outliers are filled using linear interpolation between preceding and following time points. The location and number of outliers are recorded for subsequent analysis.

[0022] S202. Data Cleaning: Delete all rows of data with invalid or unlabeled data to ensure that all data used for training are valid records. For example, if the original dataset contains 51,313 valid records, all valid data should be retained after data cleaning.

[0023] S203. Missing value handling: First, delete all columns with missing values ​​in the data. Then, fill the missing values ​​in the remaining feature columns with the arithmetic mean of the non-missing data in that column to maintain the integrity of the dataset.

[0024] S204. Signal Denoising: Discrete wavelet transform is used to decompose and reconstruct time series data to suppress high-frequency noise and retain the main trend signal reflecting the changes in the strata; specifically, the db10 wavelet basis in the Daubechies wavelet family is used to perform one layer of wavelet decomposition, extract approximation coefficients and reconstruct the signal; the threshold selection adopts an adaptive method, which is dynamically adjusted according to the noise level of the signal to achieve a balance between effective denoising and preservation of detailed information. S205. Feature Standardization: The Z-score standardization method is used to transform all feature data into a standard normal distribution with a mean of 0 and a variance of 1, in order to eliminate the interference of different physical units and numerical ranges on model training. The parameters required for standardization are calculated from the training set and simultaneously applied to the test set and the online prediction stage to ensure the consistency of data distribution.

[0025] S3. Distinguishing between configuration quantity and target quantity: To construct a clear input for the reverse prediction model, the monitoring parameters are categorized into two types based on the response relationship between each monitoring parameter and the shield tunneling control console commands: configuration parameters and target parameters. Configuration parameters refer to actively settable operational parameters directly controlled by the shield tunneling control console commands, including cutterhead rotation speed, total thrust of the propulsion cylinders, penetration depth, propulsion cylinder pressure in seven zones, and cutterhead torque, totaling 11 parameters. Target parameters refer to feedback parameters primarily reflecting the actual tunneling state and spatial attitude, including the propulsion jack stroke in the upper and lower zones, average propulsion speed, and propulsion jack speed in six zones, totaling 9 parameters. Specifically, propulsion cylinder pressure includes seven zones: upper, upper right, lower right, right, upper left, lower left, and left; average propulsion speed includes six zones: upper right, lower right, right, upper left, lower left, and left.

[0026] The method for distinguishing between configuration quantity and target quantity includes the following steps: S301, Covariance Analysis: Calculate the covariance matrix between the control command timing signal and the timing signals of each monitoring parameter to quantify the degree of response of the parameters to the control signal; S302. Response coefficient calculation: Based on the covariance matrix, calculate the response coefficient of each parameter. The response coefficient is defined as the correlation coefficient between the parameter and the control signal, and its value range is [-1, 1]. S303. Threshold Determination: When the response coefficient is greater than the set threshold (e.g., 0.7), it is determined to be a configuration quantity; when the response coefficient is less than the set threshold, it is determined to be a target quantity. The threshold can be adjusted according to the actual engineering situation.

[0027] S304. Feature Separation: Normalize the configuration quantity and the target quantity respectively, and use the Min-Max normalization method to achieve effective separation of control features and state feedback features, thereby improving the interpretability and accuracy of the reverse prediction model.

[0028] This mechanism effectively decouples control variables from state variables, providing input features with clear structure and explicit physical meaning for the inverse prediction model.

[0029] S4. Construction of time series samples: Based on the sliding window method, time series samples for supervised learning are constructed. The core of this method is to predict the required allocation amount based on historical states and the current desired target amount. The specific construction process includes the following steps: S401, Sliding Window Design: Set the window length to... N Each time step, in the past continuous N The actual values ​​of all parameters within each time step are used as input features, and the configuration amount at the current time step is used as the prediction target; window size NThe window size can be configured according to the data sampling frequency and prediction real-time requirements. The default value is 5 time steps. For higher frequency data, such as 10Hz, the window size can be increased to 10 to 15 time steps. S402, Sample Construction Strategy: For samples of length... T Time series data, generated through a sliding window T - N There are 10 training samples. The input features for each sample include: history. N The target quantity sequence and history at each time step N The input sequence contains the configuration quantity sequence for each time step and the target quantity for the current time step; the configuration quantity for the current time step is padded with zero values; the complete input sequence includes... N +1 time steps, each with 20 feature dimensions, including 9 target dimensions and 11 configuration dimensions; the sample output is the configuration dimension of the current time step, with 11 dimensions; for multi-step prediction tasks, the target output value includes future... M The parameter value for each time step, M can take different values ​​such as 1, 3 or 5; S403. Data normalization processing: Min-Max normalization is applied to the configuration quantity and target quantity features respectively to linearly transform the data to the [0,1] interval, eliminating the influence of different physical units on model training; and applied to the validation set, test set and online prediction. S404. Dataset partitioning: Divide the sample data into training set, validation set and test set according to time sequence, with the training set accounting for 70%, the validation set accounting for 15% and the test set accounting for 15%, to maintain the continuity of the time series. Using the time-series sample construction method described above, the system establishes a reverse prediction model to achieve an accurate mapping from historical state information and current target quantity to current configuration quantity. For example, for 167,690 data records, the training set contains 117,379 samples, the validation set contains 25,152 samples, and the test set contains 25,154 samples.

[0030] Under this design, the model can systematically learn the mapping relationship from historical states and current target quantities to current configuration quantities, and establish a reverse prediction model.

[0031] S5, Transformer Backward Prediction Model Construction and Training: The multidimensional features, after preprocessing and construction using time-series samples, are input into the inverse prediction model based on the Transformer encoder architecture for training. The Transformer inverse prediction model includes the following key modules: Input projection layer: Used to map the 20-dimensional input feature vector to the model's hidden dimension space through a linear transformation. dmodel Setting it to 128 enables high-dimensional representation of features; The position encoding module uses a position encoding matrix generated by sine and cosine functions to add absolute position information to each time step of the input sequence, enabling the model to recognize the temporal order. The position encoding is added to the input embedding, which preserves the temporal information while enhancing the feature representation.

[0032] Multi-head self-attention mechanism: A multi-head attention structure is adopted, with 8 attention heads and each head having a dimension of [missing information]. d model / 8=16, mapping the input sequence to multiple subspaces to compute self-attention weights in parallel, in order to capture long-range dependencies between different time steps and different variables; avoiding the gradient vanishing problem of traditional RNN / LSTM.

[0033] Feedforward network: Employs a two-layer fully connected network structure, with the first layer having a dimension of 4. d model =512, the second dimension is d model =128, using the ReLU activation function to enhance the nonlinear expressive power of the model; Residual connections and layer normalization: Residual connections are introduced after the multi-head self-attention layer and the feedforward network layer, followed by layer normalization. This design helps stabilize the training process, accelerate convergence, and alleviate the gradient vanishing problem. Encoder stacking: A complete encoder layer consisting of a multi-head self-attention mechanism, feedforward network, residual connections, and layer normalization is stacked. The number of stacked layers is... N It can be set according to the model capacity requirements, such as adopting N =3, thereby refining higher-level temporal feature representations layer by layer; Output layer: Take the feature vector of the last time step output by the encoder, and map it to an 11-dimensional output space through a fully connected layer, corresponding to the configuration quantity prediction value at the current time step; During model training, mean squared error (MSE) is used as the loss function, and the Adam optimization algorithm is used for parameter updates. The initial learning rate is set to 0.0001, the batch size to 64, and the number of training epochs to 50. A learning rate scheduling strategy is employed during training: the learning rate is reduced when the validation set loss no longer decreases, resulting in a decay factor of 0.5 and a patience value of 5. An early stopping strategy is also used: training stops when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting. The maximum gradient norm of gradient clipping is set to 1.0 to stabilize the training process and prevent gradient explosion.

[0034] Training results show that the model can converge quickly on both the training and validation sets, and the final validation loss can be reduced to about 0.0006, indicating that the model has good learning ability and generalization performance.

[0035] S6, Attitude Control Execution: During the actual tunneling process of the tunnel boring machine (TBM), the system, based on the desired attitude targets, such as the desired propulsion speed and jack stroke, calls a fully trained Transformer backpropagation model to infer in real time the optimal configuration required to achieve the target, such as the total thrust of the propulsion cylinders, pressure in each direction, and cutterhead rotation speed, thereby driving the TBM to complete precise closed-loop attitude control. The specific execution process is as follows: S601, Real-time Data Acquisition: The system collects current and historical data in real time. N = Five time steps of target quantity and configuration quantity data to form a complete time series sample that meets the model input requirements; S602. Data Preprocessing: Perform the same preprocessing procedures as the training phase on the collected real-time data, including outlier detection and processing, data cleaning, missing value handling, signal denoising and feature standardization, to ensure that the quality of the input data is consistent with that of the model training phase. S603, Data Normalization: Use the Min-Max normalizer saved during the training phase to normalize the configuration quantity and the target quantity respectively; S604, Model Inference: Input the preprocessed time series data into the trained Transformer inverse prediction model and infer the predicted value of the configuration quantity at the current time. S605. Result Inverse Normalization: Using the normalization parameters saved during the training phase, the normalized configuration quantity output by the model is inversely normalized to the original physical dimensions to obtain actual parameter values ​​that can be directly used for control. S606, Control command generation: Convert the denormalized configuration parameters into control commands that the tunnel boring machine control system can recognize, including the total thrust setting value of the propulsion cylinder, the propulsion pressure setting value in each direction, the cutterhead speed setting value, etc. S607, Attitude Adjustment Execution: After receiving control commands, the tunnel boring machine (TBM) control system drives the corresponding actuators to make adjustments, achieving precise control over the TBM's propulsion direction, speed, and attitude. For example, when it is desired to increase the propulsion speed, the system predicts and sets a higher total thrust of the propulsion cylinders; when it is desired to adjust the propulsion direction, the system predicts and sets the propulsion pressure difference in different directions.

[0036] S7, Adaptive Learning: To enable the model to continuously adapt to dynamically changing geological conditions and construction processes during tunneling, an online adaptive learning mechanism was established. This mechanism continuously monitors the model's predictive performance on newly acquired data and triggers incremental learning based on performance degradation, thereby achieving dynamic fine-tuning and continuous optimization of network weights and ensuring the model's predictive accuracy and robustness throughout the entire tunneling cycle. Its specific operation process is as follows: First, the system calculates the predicted configuration quantity corresponding to the target quantity in the new batch of measured data in real time, compares it with the actual configuration quantity of the batch, and calculates performance evaluation metrics, such as mean squared error. Second, the system sets a performance threshold; when the evaluation metric continuously exceeds the set threshold by 0.1, it automatically triggers the incremental update process of the model. Subsequently, the system uses newly collected data batches to perform small-scale incremental training on the already trained Transformer model at a learning rate much lower than the initial training rate, for example, 0.0001, optimizing the network weights. During this process, strategies such as weight decay can be used to balance adaptation to new working conditions and retention of historical knowledge, avoiding catastrophic forgetting. Finally, the system manages and evaluates the performance of the model versions before and after the update, completing the iterative upgrade of the model while ensuring stability. Through these mechanisms, the system can dynamically adapt to changes in the tunneling environment, improving the robustness and practicality of the prediction model.

[0037] Example 2

[0038] This embodiment provides an adaptive prediction system for shield tunneling parameters to implement the method described in Embodiment 1. The system adopts a modular architecture design, including a data acquisition unit, a data preprocessing unit, a configuration quantity and target quantity differentiation unit, a time series sample construction unit, a Transformer inverse prediction model training and inference unit, an attitude control execution unit, and an adaptive learning unit. Each unit is connected in sequence and realizes data interaction and command transmission through a system bus, together forming a complete intelligent control closed loop. The data acquisition unit is responsible for direct communication with the tunnel boring machine's control system to acquire multi-source raw monitoring data in real time during the tunneling process. Its hardware includes multiple analog / digital acquisition interfaces, a synchronization clock module, and a data buffer memory. This unit connects to the tunnel control network via an industrial standard bus (such as CAN bus or industrial Ethernet) and synchronously acquires the following parameters at a configurable sampling frequency (typically set to 1 Hz): cutterhead rotation speed, cutterhead torque, total thrust, propulsion cylinder pressure in seven directions (upper, upper right, lower right, right, upper left, lower left, left), average propulsion speed, penetration depth, jack stroke in the upper and lower directions, and jack speed in six directions (upper right, lower right, right, upper left, lower left, left). A timestamp synchronization module ensures that all heterogeneous data streams are based on a unified time reference, eliminating timing misalignments caused by sensor or communication delays. After initial packaging, the acquired raw data is temporarily stored in the data buffer module, forming a continuous, time-stamped multi-dimensional time-series data stream for subsequent units to access as needed.

[0039] The data preprocessing unit performs standardization and pipelined purification and enhancement processing on the raw time-series data stream output by the data acquisition unit to improve data quality and provide robust input for subsequent modeling. This unit adopts a pipelined processing approach, sequentially including an outlier detection and processing module, a data cleaning module, a missing value processing module, a signal denoising module, and a feature standardization module. The outlier detection module identifies and removes outliers from each parameter sequence based on the interquartile range boxplot method. The data cleaning module deletes invalid or unlabeled data rows. The missing value processing module deletes columns with completely missing values ​​and fills the remaining missing values ​​with the column mean. The signal denoising module uses the discrete wavelet transform algorithm, employing the db10 wavelet basis from the Daubechies wavelet family for one-level wavelet decomposition and reconstruction to suppress high-frequency noise. The feature standardization module uses the Z-score method to convert all parameters into a standard normal distribution with a mean of zero and a variance of one.

[0040] The configuration and target quantity differentiation unit distinguishes multi-source monitoring parameters into configuration quantities and target quantities through covariance analysis and response coefficient determination mechanisms. This unit includes a covariance calculation module, a response coefficient analysis module, and a feature classification module. The covariance calculation module quantifies the linkage relationship between the control command timing signal and the timing signals of each monitoring parameter. The response coefficient analysis module calculates the response coefficient of each parameter, and the feature classification module classifies the parameters into configuration quantities and target quantities based on preset thresholds. Configuration quantities include 11 parameters: cutterhead rotation speed, total thrust of the propulsion cylinder, penetration depth, propulsion cylinder pressure in seven zones, and cutterhead torque. Target quantities include 9 parameters: propulsion jack stroke in the upper and lower zones, average propulsion speed, and propulsion jack speed in six zones. This mechanism effectively decouples control variables from state variables.

[0041] The time-series sample construction unit constructs supervised learning samples based on the sliding window method, supporting the training and inference of the reverse prediction model. This unit includes a sliding window module, a sample generation module, a data normalization module, and a data partitioning module. The sliding window module sets the window length to N time steps, with a default of 5 steps, continuously extracting historical data segments from the time series. The sample generation module uses the target quantity sequence of the past N steps, the configuration quantity sequence of the past N steps, and the current target quantity as input features, and the current configuration quantity as output to generate complete sample pairs. The data normalization module applies the Min-Max normalization method to the configuration quantity and the target quantity, linearly transforming the data to the zero-to-one interval. The data partitioning module divides the samples into training, validation, and test sets in chronological order, with proportions of 70%, 15%, and 15%, respectively, maintaining the continuity of the time series.

[0042] The Transformer backpropagation model training and inference unit is the core computing unit, responsible for building and training the backpropagation model based on the Transformer encoder architecture and performing online inference. This unit is deployed on a GPU computing platform or a high-performance embedded device and includes a model training module, a model inference engine, and a result output module. The model training module uses the Adam optimization algorithm with mean squared error as the loss function, combined with dynamic learning rate adjustment, early stopping strategy, and gradient pruning techniques to improve training stability. The model inference engine implements the forward propagation of the Transformer network. The network structure includes an input projection layer, a position encoding module, a multi-head self-attention mechanism, a feedforward network, residual connections and layer normalization, encoder stacking, and an output layer. The input projection layer maps 20-dimensional features to a 128-dimensional hidden space. The position encoding module uses sine and cosine functions to add absolute position information. The multi-head self-attention mechanism uses eight heads to capture long-range dependencies. The feedforward network uses a two-layer fully connected structure with dimensions of 512 and 128, respectively, and uses the ReLU activation function. The encoder stack has three layers. The output layer maps the features from the last time step to 11-dimensional configuration predictions. The results output module provides the predictions, confidence intervals, and error estimates.

[0043] The attitude control execution unit converts the predicted configuration values ​​output by the model into actual control commands and drives the tunnel boring machine's actuators. This unit includes a result denormalization module, a control command generation module, and an actuator control module. The result denormalization module uses the normalization parameters saved during the training phase to restore the predicted values ​​to their original physical dimensions. The control command generation module converts the parameters into commands such as the total thrust setpoint of the propulsion cylinder, the propulsion pressure setpoint in each direction, and the cutterhead speed setpoint. The actuator control module adjusts the actuators through the tunnel boring machine's control system to achieve precise closed-loop control of the propulsion direction, speed, and attitude.

[0044] The adaptive learning unit enables continuous online optimization of the model, adapting to dynamic changes in operating conditions. This unit includes a performance monitoring module, an update judgment module, and an incremental training module. The performance monitoring module calculates prediction error metrics such as mean squared error in real time for new data batches. The update judgment module triggers the update process when the error continuously exceeds a set threshold. The incremental training module uses a small learning rate to incrementally train the model, combined with a weight decay strategy to avoid catastrophic forgetting, ensuring the model's robustness throughout the entire mining cycle.

[0045] Through modular design, parallel computing optimization, and real-time data processing mechanisms, this system achieves high-precision, low-latency intelligent control, significantly improving the automation level of shield tunneling attitude control and construction safety.

[0046] Example 3

[0047] This embodiment further illustrates the overall architectural features and technical advantages of the shield tunneling attitude adaptive control system. The system adopts a modular, parallel, and high-availability design philosophy, achieving efficient, stable, and scalable intelligent control services. Its core architectural features include modular design, parallel computing optimization, real-time data processing, high-availability design, scalability, and data security mechanisms. Its core architectural features are as follows: Modular Design: The system adopts a modular architecture, dividing the complete prediction process into seven functional units: data acquisition, preprocessing, feature differentiation, sample construction, model training and prediction, adaptive learning, and result output. Each unit has a clearly defined function and a single responsibility, interacting through standardized data interfaces and communication protocols. This loosely coupled design allows each module to be developed, tested, deployed, and upgraded independently, greatly improving the system's maintainability, scalability, and code reusability. For example, algorithm upgrades in the data preprocessing unit or architectural replacements in the model prediction unit can be completed without affecting the normal operation of other modules.

[0048] Parallel computing optimization: The core prediction model of the system is built on the Transformer architecture, fully utilizing its global self-attention mechanism and parallel computing characteristics. Compared with traditional recurrent neural network structures, this system completely abandons sequential recursive computation, supporting full parallel processing in both model training and inference phases. Combined with GPU hardware acceleration and batch processing technology, the system can process a large number of time-series data samples simultaneously, significantly improving computational efficiency. This optimization enables the model to achieve real-time prediction with millisecond-level latency when processing high-frequency, high-dimensional data streams generated by tunnel boring machines, fully meeting the stringent timeliness requirements of engineering sites.

[0049] Real-time data processing: To ensure the real-time nature of predictions, the system employs efficient data stream processing mechanisms such as RabbitMQ and Kafka. Each functional unit communicates with the message middleware via a high-speed data bus, ensuring low-latency, high-throughput data flow between processing stages. The system supports real-time processing and online inference of continuously input data streams, eliminating the need to wait for batch data accumulation, thus achieving an end-to-end real-time pipeline from data acquisition to prediction result output. This mechanism ensures that construction personnel can obtain the latest parameter predictions and risk warning information in a timely manner.

[0050] High availability design: The system supports distributed deployment and features automatic fault recovery, load balancing, and fault tolerance mechanisms. When a module fails, the system can automatically switch to a backup module or degrade the service to ensure the continuity of predictive services.

[0051] Scalability: The system supports deployment on GPUs or high-performance embedded devices, and computing resources can be flexibly configured according to actual needs. Horizontal scaling is supported, increasing the system's processing power by adding computing nodes.

[0052] Data security: The system employs security mechanisms such as data encryption, access control, and audit logs to ensure data security and privacy protection. Prediction results are transmitted through a secure data interface to prevent data leakage and unauthorized access.

[0053] Through the above architecture design, the system achieves high-precision, low-latency real-time prediction. The predicted values ​​match the actual values ​​well, the inference latency is low, and the single prediction time is <10ms, making it suitable for real-time monitoring needs in engineering sites.

[0054] Figure 1 The flowchart of the present invention shows the complete process from shield tunneling monitoring data acquisition, data preprocessing, distinction between configuration quantity and target quantity, construction of time series samples, training and inference of Transformer inverse prediction model, execution of attitude control to adaptive learning, as well as the data flow between each step.

[0055] Figure 2 The diagram shows the system structure of the present invention, which includes a data acquisition unit, a data preprocessing unit, a configuration quantity and target quantity differentiation unit, a time series sample construction unit, a Transformer inverse prediction model training and inference unit, an attitude control execution unit, and an adaptive learning unit. The units interact with each other through a data bus, and the main functional modules of each unit are labeled.

[0056] Figure 3The diagram illustrates the principle of the mechanism for distinguishing between configuration quantities and target quantities, showing the process of distinguishing and extracting configuration quantities (dominant parameters controlled by the control signal) and target quantities (state feedback parameters), including covariance matrix calculation, response coefficient analysis, and feature classification judgment steps, as well as feature comparison before and after the distinction.

[0057] Figure 4 A schematic diagram illustrating the construction of a Transformer inverse prediction model for time-series samples demonstrates the process of constructing training samples using the sliding window method (including historical data). N The sequence of target quantity and configuration quantity for each step, the target quantity for the current step, and the predicted value of the configuration quantity for the current step output, as well as the key modules of the Transformer model (input projection layer, position encoding module, multi-head self-attention mechanism, feedforward network, residual connection and layer normalization, etc.), illustrate the dimensional relationship between input and output and the model architecture.

[0058] Figure 5 The flowchart for attitude control execution shows the complete closed-loop control process from input of desired target quantity, model inference, configuration quantity prediction, control command generation to attitude adjustment execution.

[0059] Figure 6 The training curves of the model, including the trends of training loss and validation loss with the number of training epochs (used to illustrate the convergence and stability of model training), are used to illustrate the prediction accuracy and reliability of the method of this invention.

[0060] Figure 7 The graph comparing the configuration values ​​output by the model with the actual configuration values ​​is used to illustrate the accuracy and reliability of the method of the present invention.

[0061] Through the implementation of the above embodiments, the present invention achieves high-precision, low-delay prediction of multiple key tunneling parameters under complex geological conditions, effectively improving the safety, efficiency and intelligence level of shield tunneling construction.

[0062] The above describes one or more embodiments of the present invention in a relatively specific and detailed manner, but it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A shield tunneling attitude control method based on Transformer and reverse prediction mechanism, characterized in that, Includes the following steps: S1. Data Acquisition: Real-time acquisition of multi-source monitoring parameters during the shield tunneling process to form multi-dimensional time series data with temporal consistency; S2. Data preprocessing: The multidimensional time series data is systematically preprocessed, including outlier detection and processing, data cleaning, missing value processing, signal denoising and feature standardization, to generate a high-quality input feature set; S3. Differentiation between configuration quantity and target quantity: Based on the response relationship between each monitoring parameter and the shield control command, the monitoring parameters are divided into configuration quantity and target quantity; S4. Construction of time series samples: Based on the sliding window method, supervised learning samples are constructed by taking the parameter sequence of several historical time steps and the current target quantity as input and the current configuration quantity as output. S5. Transformer Model Construction and Training: Construct an inverse prediction model based on the Transformer encoder architecture, input the preprocessed features into the model for training, and realize the mapping from target quantity to configuration quantity; S6. Attitude control execution: Based on the expected target quantity, the trained model is called to infer the optimal configuration quantity and converted into control commands to drive the tunnel boring machine to perform attitude adjustment; S7. Adaptive Learning: Dynamically triggers incremental learning of the model based on the prediction error of newly collected data, continuously optimizing the model weights.

2. The shield tunneling attitude control method based on Transformer and reverse prediction mechanism according to claim 1, characterized in that, In step S2, the outlier detection and processing uses a box plot method based on interquartile range to identify and remove outliers, and fills in missing values ​​using median interpolation or linear interpolation; the signal denoising uses the db10 wavelet basis in the Daubechies wavelet family for one-level wavelet decomposition and reconstruction; the feature standardization uses the Z-score method to convert the data into a standard normal distribution with a mean of 0 and a variance of 1.

3. The shield tunneling attitude control method based on Transformer and reverse prediction mechanism according to claim 1, characterized in that, Step S3, the step of distinguishing between the configuration quantity and the target quantity, includes: S301, Calculate the covariance matrix between the control command timing signal and the timing signals of each monitoring parameter; S302. Calculate the response coefficients of each parameter based on the covariance matrix; S303. Based on the set response coefficient threshold, the parameters are divided into configuration quantities and target quantities; The configuration parameters include cutterhead rotation speed, total thrust of the propulsion cylinder, penetration depth, propulsion cylinder pressure in seven zones, and cutterhead torque; the target parameters include the stroke of the propulsion jacks in the upper and lower zones, the average propulsion speed, and the propulsion jack speed in the six zones.

4. The shield tunneling attitude control method based on Transformer and reverse prediction mechanism according to claim 1, characterized in that, Step S4, the steps for constructing time-series samples include: S401. Set the length of the sliding window to N time steps; S402. Take the target quantity and configuration quantity sequence of N historical time steps and the current target quantity as input features, and the current configuration quantity as output; S403. Perform Min-Max normalization on the configuration quantity and the target quantity respectively; S404. Divide the samples into training set, validation set and test set in chronological order.

5. The shield tunneling attitude control method based on Transformer and reverse prediction mechanism according to claim 1, characterized in that, In step S5, the Transformer inverse prediction model includes: The input projection layer is used to map the input features to the hidden dimension space; The position encoding module uses sine and cosine functions to generate position codes; A multi-head self-attention mechanism is used to set up multiple attention heads to capture long-range dependencies. The feedforward network employs a two-layer fully connected structure and uses the ReLU activation function; Residual connections and layer normalization are used to stabilize the training process; Encoder stacking involves stacking encoder layers multiple times to extract higher-level features. The output layer maps the features from the last time step to the configuration quantity prediction value.

6. The shield tunneling attitude control method based on Transformer and reverse prediction mechanism according to claim 1, characterized in that, In step S7, adaptive learning includes: Real-time monitoring of model prediction errors on new data batches; When the prediction error continuously exceeds the set threshold, an incremental model update is triggered. Incremental training is performed at a learning rate lower than the initial training rate, and a weight decay strategy is used to avoid catastrophic forgetting. Save model versions and support performance evaluation and rollback.

7. A shield tunneling attitude adaptive control system, used to implement the shield tunneling attitude control method based on Transformer and reverse prediction mechanism as described in any one of claims 1 to 6, characterized in that, include: The data acquisition unit is used to collect multi-source operating parameters in real time during the tunnel boring process; The data preprocessing unit is used to perform outlier detection and processing, data cleaning, missing value handling, signal denoising, and feature standardization on time-series data. The configuration quantity and target quantity differentiation unit is used to distinguish the monitoring parameters into configuration quantities and target quantities; Temporal sample construction unit, used to construct supervised learning samples based on the sliding window method; Transformer Backward Prediction Model Training and Inference Unit, used to build, train, and perform backward predictions; The attitude control execution unit is used to convert the predicted configuration quantities into control commands and drive the shield machine's execution mechanism; Adaptive learning units are used to dynamically update model parameters based on new data.

8. The shield tunneling attitude adaptive control system according to claim 7, characterized in that, The data acquisition unit communicates directly with the data bus of the tunnel boring machine control console, with a sampling frequency of 1Hz to 10Hz; the Transformer inverse prediction model training and inference unit is deployed on a GPU computing platform or a high-performance embedded device, supports batch inference, has low inference latency, and a single prediction time of <10ms.

Images