A shield tunneling parameter self-adaptive prediction method and system based on a Transformer architecture

Abstract

This invention discloses an adaptive prediction method and system for shield tunneling parameters based on the Transformer architecture. The method includes data acquisition, data preprocessing, differentiation between active and passive variables, construction of time-series samples, Transformer model construction and training, joint prediction of multiple parameters, adaptive learning, and output and visualization of prediction results. The system includes a data acquisition unit, a data preprocessing unit, a unit for differentiating between active and passive variables, a unit for constructing time-series samples, a Transformer model training and prediction unit, an adaptive learning unit, and a prediction result output unit. By acquiring multi-source operating parameters in real time and performing systematic data preprocessing, decoupling control features and geological feedback features, constructing time-series samples using a sliding window, and inputting them into the Transformer architecture for training and achieving joint prediction of multiple variables, this method solves the problems of weak long-sequence modeling ability, insufficient characterization of multi-variable coupling, and poor real-time performance of traditional methods, significantly improving prediction accuracy and engineering applicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and artificial intelligence application technology in underground engineering, and in particular to an adaptive prediction method and system for shield tunneling parameters based on the Transformer architecture. Background Technology

[0002] With the rapid development of urban underground space, the shield tunneling method has become the mainstream construction method for projects such as rail transit, municipal utility tunnels, and cross-river / sea tunnels due to its advantages such as high construction efficiency and minimal environmental disturbance. During shield tunneling, accurate prediction of key parameters such as cutterhead torque, propulsion speed, jack thrust, and penetration depth is crucial for scientifically formulating tunneling parameters, identifying potential risks such as surface subsidence, cutterhead jamming, or segment damage in advance, and optimizing construction decisions. By predicting the future trends of these key parameters, construction personnel can adjust their operating strategies in a timely manner, thereby improving construction safety and efficiency.

[0003] The parameters of tunnel boring machines (TBMs) are influenced by a combination of factors, including geological conditions, equipment status, and operational behavior, exhibiting highly nonlinear, strong time-series dependence, and abrupt changes. Especially in complex geological environments, such as heterogeneous strata, water-rich sand layers, or interbedded boulders, parameter fluctuations are more drastic and often concealed and delayed. Failure to accurately predict and respond to parameter changes can lead to increased equipment wear, hindered tunneling, and even safety accidents.

[0004] Currently, the prediction methods for tunnel boring machine (TBM) parameters in engineering practice mainly rely on the following two types of technologies: The first type is traditional methods based on empirical formulas and mechanism derivations. These methods are usually based on simplified physical models or empirical relationships, which are difficult to accurately characterize complex features such as multivariate coupling and strong nonlinearity. Their prediction accuracy is limited under complex geological conditions, and they lack adaptability to dynamic changes in working conditions. The second type is prediction methods based on traditional time series models, including statistical methods such as autoregressive models, moving average models, and autoregressive moving average models, as well as deep learning methods such as recurrent neural networks and long short-term memory networks. However, these methods have the following significant shortcomings: First, they have weak long-sequence dependency modeling capabilities. Traditional recurrent neural networks or long short-term memory networks are prone to gradient vanishing or memory decay when processing long-term time-series data, making it difficult to effectively capture the long-term evolution patterns between parameters. Second, they fail to adequately characterize multivariate coupling relationships. Shield tunneling involves multi-source monitoring data, and the parameters are dynamically closely correlated, while traditional models struggle to fully represent these complex interactive features. Third, they have low real-time prediction efficiency. Recursive calculation methods lead to serialization of the training and inference processes, failing to meet the real-time prediction needs of high-frequency, high-dimensional data in shield tunneling. Fourth, they lack a mechanism to distinguish between active and passive variables. They fail to adequately differentiate between active variables dominated by control commands, such as propulsion speed and cutterhead rotation speed, and passive variables reflecting ground feedback, such as cutterhead torque and propulsion cylinder pressure. This results in a high degree of coupling between control signals and ground information, reducing prediction accuracy and interpretability. Fifth, their model generalization and adaptive capabilities are limited. Under varying geological environments and construction conditions, static model parameters struggle to adapt to changes at different stages, lacking continuous learning and dynamic optimization mechanisms.

[0005] Furthermore, tunnel boring machine (TBM) monitoring data often suffers from issues such as sensor drift, missing data, abnormal spikes, and inconsistent sampling periods at the engineering site. Without systematic data preprocessing, such as anomaly detection, denoising, and redundancy suppression, the model is susceptible to noise interference, leading to prediction biases.

[0006] In summary, existing technologies suffer from at least the following key problems: insufficient ability to model long-sequence dependencies; inadequate characterization of multivariate dynamic coupling features; poor real-time prediction performance; lack of a mechanism to distinguish between active and passive variables; and a lack of adaptive update capabilities. These deficiencies limit the accuracy and reliability of shield tunneling parameter prediction, making it difficult to support intelligent decision-making in complex construction environments. Summary of the Invention

[0007] To achieve the above-mentioned objectives, this invention provides an adaptive prediction method and system for shield tunneling parameters based on the Transformer architecture, which enables high-precision, low-latency, and multi-variable joint prediction of key parameters, thereby improving the scientific nature and safety of the construction process.

[0008] An adaptive prediction method for tunnel boring machine (TBM) parameters based on the Transformer architecture includes the following steps: S1. Data Acquisition: Real-time acquisition of multi-source operating parameters during the tunneling process. These multi-source operating parameters include cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. Among these, propulsion cylinder pressure and jack speed cover data acquisition from multiple directions. All parameters are collected synchronously with a unified sampling frequency and timestamp to form multi-dimensional time series data. S2, Data Preprocessing: The multidimensional time series dataset is systematically preprocessed to improve data quality and model robustness; the processed data constitutes a high-quality input feature set; the systematic preprocessing includes the following steps: S201. Outlier Detection and Handling: Outliers in each parameter sequence are identified and removed using a box plot method based on the interquartile range; specifically, the first quartile is used... Q 1 With the third quartile Q 3 Calculate the interquartile range based on the difference Q R Set the outlier threshold as [ Q 1 -1.5 Q R , Q 3 +1.5 Q R Values ​​outside this range are identified as outliers and removed; missing positions after removal are filled using median interpolation or linear interpolation with a sliding window. S202, Signal Denoising: Discrete wavelet transform is used to decompose and reconstruct time series data, suppressing high-frequency noise and retaining the main trend signal reflecting the stratigraphic changes; Specifically, the db10 wavelet basis in the Daubechies wavelet family is used to perform one layer of wavelet decomposition and reconstruction, combined with an adaptive thresholding method to suppress high-frequency noise components. S203, Feature Dimensionality Reduction: Principal Component Analysis (PCA) algorithm is applied to transform the denoised data, extracting principal components whose cumulative contribution rate reaches a predetermined threshold, thereby reducing data dimensionality, eliminating redundant information, and forming a high-quality input feature set; preferably, before performing PCA, feature correlation analysis is performed to calculate the correlation coefficient between each feature. If the absolute value of the correlation coefficient between any two features is higher than a preset threshold, the correlation coefficient between each feature and the engineering geology category label is further calculated, and features with smaller absolute values ​​of correlation coefficients with the engineering geology category label are removed. S204. Data Normalization: All feature data after dimensionality reduction are normalized to a uniform scale. The Z-score normalization method is used to convert each feature data into a standard distribution with a mean of zero and a variance of one, thereby eliminating the interference caused by different physical dimensions and numerical magnitudes on model training and ensuring that each feature has comparability and a balanced contribution during training. S3. Distinguishing between active and passive quantities: Based on the response relationship between various operating parameters and shield control signals, the operating parameters are divided into two categories: active parameters and passive parameters. Active parameters refer to parameters directly controlled by the shield control console commands, used to characterize the control characteristics of the tunneling equipment, including propulsion speed and cutterhead rotation speed. Passive parameters refer to parameters reflecting the dynamic response of the geological environment to the tunneling process, used to characterize the geological feedback characteristics, including cutterhead torque, propulsion cylinder pressure, and penetration depth. Preferably, to achieve accurate determination and effective separation of the above-mentioned feature categories, the following steps are adopted for feature decoupling: First, calculate the covariance matrix between the control signal and each operating parameter to quantify the degree of response of the parameter to the control command; second, analyze the response coefficient of each parameter based on the covariance matrix, and the response coefficient is defined as a measure of the correlation between the parameter and the control signal; finally, determine the feature category according to the preset response coefficient threshold. When the response coefficient is higher than the threshold, it is determined to be an active quantity, and when it is lower than the threshold, it is determined to be a passive quantity, thereby achieving effective separation of control features and formation feedback features.

[0009] S4. Construction of time series samples: Supervised learning samples are constructed based on the sliding window method to predict the future tunneling state using parameter changes at several historical moments, forming a multi-dimensional feature vector. The specific steps include: S401, Sliding window design: Set the window size to... N A time step, in the past N Using the actual parameter values ​​at each time step as input features, predict the single-step parameter value at the next time step or the multi-step parameter values ​​at multiple future time steps; window size N The time steps are dynamically adjusted based on the data sampling frequency and forecasting needs, with a default value of 5 time steps. S402, Sample Construction Strategy: For samples of length... T Time series data is extracted sequentially using a sliding window to generate... T - N training samples; each sample contains _ training samples; N The input features at each time step and their corresponding target output values; for multi-step prediction tasks, the target output value includes future... M The parameter values ​​for each time step. M It can take different values ​​such as 1, 3, or 5; S403. Data Augmentation: Employ sliding window sliding, resampling, and noise injection methods to enhance the diversity and coverage of training samples; enhance the robustness of the model to data perturbations and improve the model's generalization ability by adding small-amplitude noise conforming to a Gaussian distribution. S404, Sample Balancing: The training samples under different working conditions and geological conditions are balanced. A stratified sampling strategy is adopted to ensure that the proportion of samples under various working conditions in the training set is balanced, so as to avoid overfitting to specific working conditions during the model training process. By constructing time series samples using the above method, multi-dimensional feature vectors are formed. The system can effectively extract time series evolution patterns from historical data, establish a mapping relationship from historical states to future states, and provide structured supervised learning samples for the model to learn the dynamic changes of parameters. S5, Transformer Model Construction and Training: The multidimensional feature vectors, after preprocessing and construction using time-series samples, are input into the prediction model based on the Transformer architecture for training. The Transformer model includes an input embedding layer, a positional encoding module, a multi-head self-attention structure, a feedforward network, residual connections and layer normalization, encoder stacking, and an output layer. Input embedding layer: Maps multidimensional feature vectors to the model dimensionality space through a linear transformation, achieving a high-dimensional representation of the features; specifically, it maps 16-dimensional input features to... d model Dimensions, the default value is 128 dimensions, to enhance the expressive power of the model; The positional encoding module employs a combination of sine and cosine functions to add positional information to each time step in the input sequence, enabling the model to recognize the chronological order of time points in the multidimensional feature vector. The positional encoding is added element-wise to the multidimensional feature vector, enhancing feature representation while preserving temporal information. Multi-head self-attention structure: Divides multi-dimensional feature vectors into multiple subspaces, with a default setting of 8 heads; each head independently learns the dependencies between different time steps and different variables; by capturing multiple dependency patterns in parallel, it models the global dependency structure of all time points at once, effectively solving the problem of long sequence dependencies. Feedforward network: Employs a two-layer fully connected network structure, with the first layer having a dimension of 4. d model The second dimension is d model The ReLU activation function is used to enhance the model's non-linear expressive power and learn complex feature transformations. Residual connections and layer normalization: Residual connections and layer normalization operations are added after the self-attention layer and the feedforward network layer, respectively, to stabilize the training process, accelerate model convergence, and prevent gradient vanishing, thereby supporting the effective training of deeper network structures. Encoder stacking: Stack the self-attention mechanism, feedforward network, and corresponding normalization module described above. N The default value is 4 layers, which extract higher-level temporal features layer by layer to form a hierarchical representation from low-level features to high-level semantics. Output layer: The encoder output is mapped to the target dimension space through a fully connected layer, corresponding to the predicted values ​​of multiple key tunneling parameters; the output layer supports two configuration modes: single-step prediction and multi-step prediction. During the model training phase, mean squared error or mean absolute error is used as the loss function, and the Adam algorithm is selected as the optimizer. Through multiple rounds of iterative learning, the dynamic mapping relationship between variables is gradually mastered. During the training process, learning rate decay, early stopping strategy and gradient pruning method are combined to improve training stability and model generalization ability.

[0010] S6. Multi-parameter joint prediction: Based on the trained Transformer model, multiple key tunneling variables are jointly predicted, outputting predicted values ​​for 16 key parameters, including cutterhead rotational speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. This step includes the following technical features: S601, Multivariate Simultaneous Prediction: The model outputs the prediction results of multiple related parameters simultaneously in parallel, fully modeling the dynamic coupling relationship between parameters, overcoming the defect of univariate prediction methods that ignore the mutual influence between variables, and improving the overall consistency and engineering applicability of the prediction results. S602. Prediction accuracy evaluation: For each prediction parameter, calculate its prediction error index, including mean square error, mean absolute error and mean absolute percentage error. The system evaluates the prediction quality of each parameter and provides a quantitative basis for model optimization. S603. Quantification of prediction uncertainty: Using ensemble learning or Monte Carlo sampling methods, the uncertainty of the predicted value is estimated, and the corresponding confidence interval is generated to provide a risk quantification reference for construction decisions. S604, Multi-step prediction mechanism: The model supports both single-step and multi-step prediction modes. Single-step prediction outputs the parameter value at the next time step, while multi-step prediction outputs the parameter sequence for multiple consecutive time steps in the future, in order to meet the different needs of short-term and medium-term prediction in engineering.

[0011] S7, Adaptive Learning: To adapt to changing working conditions during tunneling, the system dynamically adjusts network weights based on the difference between predicted and measured data after new data is acquired. This incremental fine-tuning and continuous optimization maintains the model's stable predictive ability for different stages of the tunneling process. Specifically, this includes the following mechanisms: Online monitoring mechanism: Real-time monitoring of the model's predictive performance on new data batches, calculating prediction error metrics such as mean squared error and mean absolute error to evaluate the model's predictive accuracy under current operating conditions.

[0012] Adaptive triggering mechanism: When the prediction error exceeds a set threshold, such as when the loss function value is greater than 0.1, or when the prediction error continues to rise in multiple consecutive batches, the model update process is automatically triggered to avoid model performance degradation.

[0013] Incremental learning strategy: Use a small learning rate, such as 0.0001, to incrementally train the model, updating only some network layers or using a weight decay strategy, so as to maintain the memory of historical data while adapting to new working conditions and avoid catastrophic forgetting problems.

[0014] Model version management: Saves model versions at different stages, supports model rollback and performance comparison, and ensures system stability and traceability.

[0015] Continuous optimization mechanism: The model is comprehensively evaluated regularly. When incremental learning fails to effectively improve prediction performance, the model is retrained or its structure is adjusted to achieve continuous evolution and optimization.

[0016] Through the above mechanism, the system can dynamically adapt to changes in the tunneling environment, improving the robustness and practicality of the prediction model.

[0017] S8. Prediction Result Output and Visualization: The model's parameter predictions will be output in real time as structured data and visualizations. The core information output in this step includes predicted values ​​of key tunneling parameters, future trends, and prediction confidence intervals. The output supports both single-step and multi-step prediction modes, providing predictions for multiple future time steps simultaneously, thus meeting the needs for assessing short-term and medium-term parameter evolution trends.

[0018] The second aspect of the present invention provides an adaptive prediction system for shield tunneling parameters based on the Transformer architecture. The system includes a data acquisition unit, a data preprocessing unit, an active and passive quantity differentiation unit, a time series sample construction unit, a Transformer model training and prediction unit, an adaptive learning unit, and a prediction result output unit. The data acquisition unit is used to collect multi-source operating parameters of the tunnel boring machine; The data preprocessing unit is used to perform anomaly detection, noise reduction, dimensionality reduction and normalization. The active and passive quantity differentiation unit is used to distinguish between control characteristics and formation feedback characteristics; The time-series sample construction unit is used to construct training samples based on the sliding window method; The Transformer model training and prediction unit is used to predict parameters using the Transformer architecture. The adaptive learning unit is used to dynamically update the model parameters based on real-time data; The prediction result output unit is used to output the prediction results in a visual form and provide them to the shield tunneling monitoring system through a data interface, enabling construction personnel to view the prediction information and make decisions.

[0019] In summary, compared with the prior art, the beneficial effects of the present invention are: 1) Outstanding ability to model long sequence dependencies: The Transformer architecture is applied to the task of predicting shield tunneling parameters. Its global self-attention mechanism effectively captures long sequence dependencies, effectively solving the gradient vanishing and memory decay problems of traditional recurrent neural networks or long short-term memory models. It significantly enhances the ability to model the long-term evolution of parameters and realizes a deeper and more comprehensive model of the dynamic evolution of parameters. 2) Significant advantages of multivariate joint prediction: By uniformly representing and jointly predicting the tunneling parameters of multiple dimensions and physical quantities, the model can simultaneously output the predicted values ​​of up to 16 key parameters. With the help of the multi-head attention mechanism, the dynamic coupling relationship between parameters is explicitly modeled, realizing multivariate collaborative prediction and improving the consistency and reliability of the prediction results in engineering applications. 3) High prediction accuracy and strong real-time performance: Based on the parallel computing mechanism of Transformer, recursive operations are eliminated, and the training and inference efficiency is greatly improved. The average absolute percentage error of key parameters can be controlled within 5%, and the predicted value matches the true value well. At the same time, the inference latency is less than 10 milliseconds, which meets the real-time monitoring needs of high-frequency data streams in engineering sites. 4) Decoupling mechanism between active and passive variables: By proposing a parameter classification method based on response coefficients, the monitoring parameters are divided into active variables that reflect control commands and passive variables that reflect formation feedback. This effectively decouples control signals and formation information at the feature level, significantly improving the interpretability and prediction accuracy of the model and solving the prediction confusion problem caused by the high coupling between control and formation features in previous methods. 5) End-to-end data governance and feature enhancement process: A systematic preprocessing process from anomaly detection, wavelet denoising, feature selection to standardization was designed. Combined with time-series sliding window and data augmentation strategies, high-quality training samples were constructed, which effectively improved the robustness and generalization ability of the model in real engineering noise environment. 6) Online adaptive and continuous learning mechanism: The model has the ability to learn incrementally and optimize dynamically. It can automatically trigger weight fine-tuning or structural updates based on real-time monitoring data, so as to achieve rapid adaptation to different geological conditions and construction stages. This overcomes the problem of the decline in prediction performance of traditional static models under new working conditions and reflects the self-evolution characteristics of intelligent systems in dynamic environments. 7) High system integration and intuitive visualization: The prediction results are output to the shield tunneling monitoring system in real time through a standardized interface, supporting the visualization of single-step prediction, multi-step prediction, confidence interval and risk warning. This solves the technical problems of low system integration and unintuitive prediction result display in existing methods, making it easier for construction personnel to make decisions and improving the level of intelligent construction. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an adaptive prediction method for tunnel boring machine parameters based on the Transformer architecture, as shown in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a shield tunneling parameter adaptive prediction system based on the Transformer architecture, as shown in an embodiment of the present invention. Figure 3 This is a schematic diagram of the Transformer architecture shown in an embodiment of the present invention; Figure 4 This is a comparison chart of Transformer parameter prediction results shown in the embodiments of the present invention; Figure descriptions: 1-Data acquisition unit, 2-Data preprocessing unit, 3-Active and passive quantity differentiation unit, 4-Time series sample construction unit, 5-Transformer model training and prediction unit, 6-Adaptive learning unit, 7-Prediction result output unit. Detailed Implementation

[0021] 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.

[0022] Example 1

[0023] This embodiment provides an adaptive prediction method for tunnel boring machine (TBM) parameters based on the Transformer architecture. The specific implementation steps are as follows: S1. Data Acquisition: Multi-source operating parameters during the tunneling process are acquired in real-time through the shield tunneling monitoring system. Acquired parameters include cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. Propulsion cylinder pressure data covers four directions: right, top, left, and bottom. Jack stroke data covers two directions: top and bottom. Jack speed data covers six directions: right, upper right, lower right, left, upper left, and lower left. All parameters are sampled at a frequency of 1 Hz and collected synchronously according to a unified timestamp, forming multi-dimensional time-series data.

[0024] S2. Data Preprocessing: Systematically preprocess the collected multidimensional time series data to improve data quality and model robustness; the preprocessing 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.

[0025] S202, 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. S203. Feature Reduction: Key physical quantities highly correlated with the tunneling process were selected from the raw monitoring data. A combination of correlation analysis and principal component analysis was used to ultimately retain 16 key parameters: cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. These parameters comprehensively cover the main physical quantities of the tunneling process.

[0026] S204. Data Normalization: All feature data after dimensionality reduction are standardized to a uniform scale. The Z-score normalization method is used to convert each feature data into a standard normal distribution with a mean of 0 and a variance of 1, eliminating the interference of different physical units and numerical magnitudes on model training. The mean and standard deviation parameters required for standardization are calculated from the training set and simultaneously applied to the test set and online prediction stage to ensure the consistency of data distribution.

[0027] S3. Distinguishing between active and passive quantities: Based on the response relationship between various operating parameters and shield control signals, the operating parameters are divided into two categories: active parameters and passive parameters. Active parameters refer to parameters directly controlled by the shield control console commands, used to characterize the control characteristics of the tunneling equipment, including but not limited to propulsion speed and cutterhead rotation speed. Passive parameters refer to parameters reflecting the dynamic response of the geological environment to the tunneling process, used to characterize the geological feedback characteristics, including but not limited to cutterhead torque, propulsion cylinder pressure, and penetration depth. In specific implementation, to achieve accurate determination and effective separation of the above-mentioned feature categories, the following steps are adopted for feature decoupling: First, calculate the covariance matrix between the control signal and each operating parameter to quantify the degree of response of the parameter to the control command; second, analyze the response coefficient of each parameter based on the covariance matrix. The response coefficient is defined as a measure of the correlation between the parameter and the control signal, and its value range is [-1, 1]; finally, determine the feature category according to the preset response coefficient threshold. When the response coefficient is higher than the threshold, it is determined to be an active quantity, and when it is lower than the threshold, it is determined to be a passive quantity, thereby achieving effective separation of control features and formation feedback features; in specific implementation, the preset response coefficient threshold is 0.7.

[0028] S4. Construction of Time-Series Samples: Supervised learning samples are constructed based on the sliding window method to predict the mining status at future times using parameter changes at several historical moments, forming a multi-dimensional feature vector. This includes the following steps: S401, Sliding window design: Set the window size to... N A time step, in the past NThe actual parameter values ​​at each time step are used as input features to predict the single-step parameter value at the next moment or the multi-step parameter values ​​at multiple future moments. The window size N is dynamically adjusted according to the data sampling frequency and prediction requirements, with a default value of 5 time steps. For higher frequency data (such as 10Hz), the window size can be appropriately increased to 10 to 15 time steps. S402. Sample Construction Strategy: For time series data of length T, consecutive time segments are sequentially extracted using a sliding window to generate... T - 5 training samples; each sample contains _ training samples; 5 The input features at each time step and their corresponding target output values ​​are combined, resulting in a total input dimension of 16 dimensions × 5 time steps = 80 dimensions. For multi-step prediction tasks, the target output value includes future... M The parameter values ​​for each time step. M It can take different values ​​such as 1, 3, or 5; S403. Data Augmentation: Employ sliding window sliding, resampling, and noise injection methods to enhance the diversity and coverage of training samples; enhance the model's robustness to data perturbations and improve its generalization ability by adding small-amplitude noise conforming to a Gaussian distribution; specifically, add Gaussian noise (standard deviation of 0.01) to the input features to enhance the model's robustness to data perturbations.

[0029] S404, Sample Balancing: The training samples under different working conditions and geological conditions are balanced. A stratified sampling strategy is adopted to ensure that the proportion of samples under various working conditions in the training set is balanced, so as to avoid overfitting to specific working conditions during model training.

[0030] By constructing time-series samples using the aforementioned method, multi-dimensional feature vectors are formed. The system can effectively extract time-series evolution patterns from historical data, establish a mapping relationship from historical states to future states, and provide structured supervised learning samples for the model to learn the dynamic changes in parameters.

[0031] S5. Transformer Model Construction and Training: Input the multi-dimensional feature vectors constructed from preprocessed and time-series samples into the prediction model based on the Transformer architecture for training; The Transformer model includes an input embedding layer, a positional encoding module, a multi-head self-attention structure, a feedforward network, residual connections and layer normalization, encoder stacking, and an output layer. Input embedding layer: Maps multidimensional feature vectors to the model dimensionality space through a linear transformation, achieving a high-dimensional representation of the features; specifically, it maps 16-dimensional input features to... d model Dimensions, default value dmodel The value is 128 dimensions to enhance the model's expressive power; The positional encoding module employs a combination of sine and cosine functions to add positional information to each time step in the input sequence, enabling the model to recognize the chronological order of time points in the multidimensional feature vector. The positional encoding is added element-wise to the multidimensional feature vector, enhancing feature representation while preserving temporal information. Multi-head self-attention structure: Divides multi-dimensional feature vectors into multiple subspaces, with a default setting of 8 heads; each head independently learns the dependencies between different time steps and different variables; by capturing multiple dependency patterns in parallel, it models the global dependency structure of all time points at once, effectively solving the problem of long sequence dependencies. Feedforward network: Employs a two-layer fully connected network structure, with the first layer having a dimension of 4. d model The second dimension is d model The ReLU activation function is used to enhance the model's non-linear expressive power and learn complex feature transformations. Residual connections and layer normalization: Residual connections and layer normalization operations are added after the self-attention layer and the feedforward network layer, respectively, to stabilize the training process, accelerate model convergence, and prevent gradient vanishing, thereby supporting the effective training of deeper network structures. Encoder stacking: Stack the self-attention mechanism, feedforward network, and corresponding normalization module described above. N The default value is 4 layers, which extract higher-level temporal features layer by layer to form a hierarchical representation from low-level features to high-level semantics. Output layer: The encoder output is mapped to a 16-dimensional output space through a fully connected layer, corresponding to the predicted values ​​of 16 key tunneling parameters; the output layer supports two configuration modes: single-step prediction and multi-step prediction. During the model training phase, mean squared error or mean absolute error is used as the loss function, the Adam algorithm is selected as the optimizer, the initial learning rate is set to 0.001, the batch size is 64, and the number of training rounds is 100. Through multiple rounds of iterative learning, the dynamic mapping relationship between variables is gradually mastered. During the training process, a learning rate decay strategy is adopted. When the validation set loss no longer decreases, the learning rate is reduced to improve training stability and model generalization ability.

[0032] S6. Multi-parameter Joint Prediction: Based on the trained Transformer model, multiple key tunneling variables are jointly predicted, outputting predicted values ​​for 16 key parameters, including cutterhead torque, cutterhead rotation speed, average advance speed, total thrust, and penetration depth. This step includes the following technical features: S601, Multivariate Simultaneous Prediction: The model outputs the prediction results of multiple related parameters simultaneously in parallel, fully modeling the dynamic coupling relationship between parameters, overcoming the defect of univariate prediction methods that ignore the mutual influence between variables, and improving the overall consistency and engineering applicability of the prediction results. S602. Prediction accuracy evaluation: For each prediction parameter, calculate its prediction error index, including mean square error, mean absolute error and mean absolute percentage error. The system evaluates the prediction quality of each parameter and provides a quantitative basis for model optimization. S603. Quantification of prediction uncertainty: Using ensemble learning or Monte Carlo sampling methods, the uncertainty of the predicted value is estimated, and the corresponding confidence interval is generated to provide a risk quantification reference for construction decisions. S604, Multi-step prediction mechanism: The model supports both single-step and multi-step prediction modes. Single-step prediction outputs the parameter value at the next time step, while multi-step prediction outputs the parameter sequence for multiple consecutive time steps in the future, in order to meet the different needs of short-term and medium-term prediction in engineering.

[0033] S7, Adaptive Learning: To adapt to changing working conditions during tunneling, the system dynamically adjusts network weights based on the difference between predicted and measured data after new data is acquired. This incremental fine-tuning and continuous optimization maintains the model's stable predictive ability for different stages of the tunneling process. Specifically, this includes the following mechanisms: Online monitoring mechanism: Real-time monitoring of the model's predictive performance on new data batches, calculating prediction error metrics such as mean squared error and mean absolute error to evaluate the model's predictive accuracy under current operating conditions.

[0034] Adaptive triggering mechanism: When the prediction error exceeds a set threshold, such as when the loss function value is greater than 0.1, or when the prediction error continues to rise in multiple consecutive batches, the model update process is automatically triggered to avoid model performance degradation.

[0035] Incremental learning strategy: Use a small learning rate, such as 0.0001, to incrementally train the model, updating only some network layers or using a weight decay strategy, so as to maintain the memory of historical data while adapting to new working conditions and avoid catastrophic forgetting problems.

[0036] Model version management: Saves model versions at different stages, supports model rollback and performance comparison, and ensures system stability and traceability.

[0037] Continuous optimization mechanism: The model is comprehensively evaluated regularly. When incremental learning fails to effectively improve prediction performance, the model is retrained or its structure is adjusted to achieve continuous evolution and optimization.

[0038] Through the above mechanism, the system can dynamically adapt to changes in the tunneling environment, improving the robustness and practicality of the prediction model.

[0039] S8. Prediction Result Output and Visualization: The parameter prediction results generated by the model are structured and visualized, including: Single-step prediction output: Outputs the next time step. t Predicted values ​​for 16 key tunneling parameters with +1, including predicted values, confidence intervals, and prediction error estimates; Multi-step prediction output: Supports outputting predicted values ​​for multiple future time steps, such as... t +1 to t At +5, a parameter change trend curve is generated to help construction personnel understand the short-term evolution trend of the parameters; Visualization of prediction results: The prediction results are displayed in the form of charts, including time series curves and prediction error distribution maps, to intuitively present the prediction effect; Risk warning: When it is predicted that key parameters such as cutter head torque or propulsion cylinder pressure will fluctuate abnormally or exceed the safety threshold, the system will automatically generate and issue a warning message. All forecast results, visualization data, and early warning information are transmitted in real time to the tunnel boring machine (TBM) monitoring system via a standardized data interface, providing decision support for construction personnel. For example, when a forecast indicates that the cutterhead torque is about to increase significantly, the system can issue an early warning so that operators can adjust tunneling parameters in a timely manner and prevent the risk of equipment overload.

[0040] Example 2

[0041] This embodiment provides an adaptive prediction system for shield tunneling parameters for implementing the method described in Embodiment 1. The system adopts a modular architecture design, including a data acquisition unit 1, a data preprocessing unit 2, an active quantity and passive quantity differentiation unit 3, a time series sample construction unit 4, a Transformer model training and prediction unit 5, an adaptive learning unit 6, and a prediction result output unit 7. Data acquisition unit 1 communicates directly with the data bus of the tunnel boring machine (TBM) control console, for example, via a CAN bus or Ethernet interface. The sampling frequency can be flexibly configured according to engineering requirements, typically ranging from 1 Hz to 10 Hz. This unit includes a multi-channel data acquisition interface, a timestamp synchronization module, and a data buffer module. The multi-channel data acquisition interface acquires multi-source operating parameters of the TBM during tunneling in real time, including cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. The timestamp synchronization module ensures that all parameters are acquired synchronously according to a unified time reference, eliminating timing deviations. The data buffer module is responsible for temporarily storing the acquired raw data and organizing it into a continuous multi-dimensional time-series data stream for subsequent processing by subsequent units.

[0042] Data preprocessing unit 2 performs systematic preprocessing on the collected raw time-series data sequentially. Internally, it employs a pipelined processing approach, including an anomaly detection module, a denoising module, a dimensionality reduction module, and a normalization module. The anomaly detection module uses a box plot method based on interquartile range to identify outliers in each parameter sequence in real time and removes them according to preset thresholds. Subsequently, a sliding window median interpolation method is used to fill in missing data locations, ensuring data continuity. The denoising module utilizes a discrete wavelet transform algorithm, preferentially selecting the db10 wavelet basis from the Daubechies wavelet family, to perform a one-level decomposition and reconstruction of the time-series data, effectively suppressing high-frequency noise while retaining useful low-frequency signals reflecting formation change trends. The dimensionality reduction module uses feature correlation analysis and principal component analysis to screen key physical quantities highly correlated with the tunneling status from the raw monitoring parameters. After screening, sixteen key parameters are ultimately retained, including cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke, and jack speed. The normalization module performs Z-score standardization on all feature data, converting each parameter into a standard normal distribution with a mean of zero and a variance of one, in order to eliminate the interference of different physical dimensions and numerical scales on model training.

[0043] Active and passive quantity differentiation unit 3 is used to classify the pre-processed monitoring parameters into characteristic categories, achieving effective decoupling between control characteristics and formation feedback characteristics. The unit includes a covariance calculation module, a response coefficient analysis module, and a feature classification module. The covariance calculation module is responsible for calculating the covariance matrix between the shield control signal and each monitoring parameter, quantifying the degree of parameter response to control commands. The response coefficient analysis module further calculates the response coefficient of each parameter based on the covariance matrix; this coefficient characterizes the strength of the correlation between the parameter and the control signal. The feature classification module classifies parameters according to a preset response coefficient threshold: parameters with response coefficients higher than the threshold are classified as active quantities, mainly including propulsion speed and cutterhead rotation speed; these parameters directly reflect the characteristics of control commands. Parameters with response coefficients lower than the threshold are classified as passive quantities, mainly including cutterhead torque, propulsion cylinder pressure, and penetration depth; these parameters mainly reflect the feedback characteristics of the formation environment.

[0044] The Time Series Sample Construction Unit 4, based on the sliding window method, transforms continuous time series data into supervised learning samples suitable for model training. This unit includes a sliding window module, a sample generation module, a data augmentation module, and a sample balancing module. The sliding window module uses a configurable sliding window with a default window length of five time steps, continuously extracting historical data segments from the time series as input features. The sample generation module generates corresponding target output values ​​according to the prediction task requirements, supporting both single-step and multi-step prediction modes to form complete input-output sample pairs. The data augmentation module increases the diversity of training samples and improves the model's robustness and generalization ability to data perturbations through methods such as sliding window sliding, resampling, and adding Gaussian-distributed micro-noise. The sample balancing module employs a stratified sampling strategy to balance samples under different engineering conditions and geological formations, ensuring a balanced proportion of various types of samples in the training set and preventing overfitting of the model to specific conditions.

[0045] The Transformer model training and prediction unit 5 is the core computing unit, responsible for building and training prediction models based on the Transformer architecture and performing online inference. This unit can be deployed on GPU computing platforms or high-performance embedded devices, and mainly includes a model training module, a model inference engine, and a result output module. The model training module supports offline full training and online incremental fine-tuning of the model. Offline training uses a complete historical dataset, employing the Adam optimizer and combining techniques such as learning rate decay, early stopping strategies, and gradient pruning to improve training stability. Online incremental training, when a decline in model performance is detected, fine-tunes the network weights with a smaller learning rate to adapt to new working data. The model inference engine implements the forward propagation process of the Transformer network. The network structure specifically includes: an input embedding layer that maps 16-dimensional features to a 128-dimensional model space; a position encoding module that adds positional information to the sequence using sine and cosine functions; a multi-head self-attention mechanism that sets up eight attention heads to capture the dependencies between different variables and time steps in parallel; a feedforward network that uses a two-layer fully connected structure and introduces the ReLU activation function to enhance nonlinear expressive power; residual connections and layer normalization operations for stabilizing the training process; an encoder stacking structure that stacks four layers by default to extract high-level temporal features layer by layer; and an output layer that maps the encoder output to a 16-dimensional target space through fully connected layers, corresponding to the predicted values ​​of 16 key mining parameters. The inference engine supports batch processing computation, significantly improving prediction efficiency. The result output module is responsible for processing the model inference results, outputting the predicted values ​​of the sixteen key parameters, calculating and providing prediction confidence intervals and error estimates, and supporting both single-step and multi-step prediction output modes.

[0046] The Adaptive Learning Unit 6 is used to achieve continuous online optimization of the model, ensuring its ability to adapt to dynamically changing construction conditions. This unit includes a performance monitoring module, an update judgment module, and an incremental training module. The performance monitoring module calculates the model's prediction error metrics, such as mean squared error and mean absolute error, in real time on newly collected data batches, continuously evaluating the model's prediction accuracy under current conditions. The update judgment module makes judgments based on preset performance thresholds. When the prediction error exceeds a specific threshold, or when the error in multiple consecutive batches shows a continuous upward trend, this module automatically triggers the model update process. After the update is triggered, the incremental training module uses a small learning rate to incrementally train the model. The training process can employ strategies such as updating only some network layers or introducing weight decay, aiming to adapt to new data characteristics while effectively retaining the memory of historical conditions and avoiding catastrophic forgetting problems.

[0047] The prediction result output unit 7 is responsible for processing the model's prediction results and outputting them in various formats to support human-computer interaction and construction decision-making. This unit integrates a result generation module, a visualization module, an early warning module, and a data transmission module. The result generation module structures the raw prediction values ​​output by the model, generating complete prediction information including single-step prediction values, multi-step prediction sequences, prediction confidence intervals, and error estimates. The visualization module intuitively displays the prediction results and historical measured data in chart form, such as time series prediction curves, comparison charts of predicted and measured values, and prediction error distribution charts, supporting real-time updates and historical data playback. The early warning module continuously monitors the prediction results. When it detects that the predicted values ​​of key tunneling parameters, such as cutterhead torque or propulsion cylinder pressure, are about to experience abnormal fluctuations or exceed preset safety thresholds, it automatically generates and issues multi-level early warning information, including early warning parameters, early warning time, and recommended measures. The data transmission module transmits the structured prediction results, visualized chart data, and early warning information to the shield tunneling centralized monitoring system in real time and reliably through standardized data interfaces, such as RESTful API or WebSocket, and supports simultaneous access and data subscription by multiple monitoring clients. Through the collaborative work of the above seven functional units, this system realizes a complete closed loop from data acquisition, preprocessing, feature engineering, model training and prediction, online adaptive learning to result output and visualization, providing intelligent parameter prediction support with high precision, low latency and continuous evolution capability for shield tunneling construction.

[0048] Example 3

[0049] This embodiment further illustrates the overall architectural features and technical advantages of the adaptive prediction system for tunnel boring machine (TBM) parameters. The system adopts a modular, parallel, and highly available design philosophy, achieving efficient, stable, and scalable intelligent prediction services. 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.

[0050] 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.

[0051] 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.

[0052] 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.

[0053] 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.

[0054] 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.

[0055] 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.

[0056] 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.

[0057] Figure 1 This is a flowchart illustrating an adaptive prediction method for shield tunneling parameters based on the Transformer architecture, as shown in an embodiment of the present invention. It displays the complete process from shield tunneling monitoring data acquisition, data preprocessing, differentiation between active and passive variables, construction of time-series samples, Transformer model training and prediction, adaptive learning to prediction result output and visualization, as well as the data flow between each step.

[0058] Figure 2 This is a schematic diagram of the structure of an adaptive prediction system for shield tunneling parameters based on the Transformer architecture, as shown in an embodiment of the present invention. It includes a data acquisition unit, a data preprocessing unit, an active and passive quantity differentiation unit, a time series sample construction unit, a Transformer model training and prediction unit, an adaptive learning unit, and a prediction result output unit. The units interact with each other through a data bus, and the main functional modules of each unit are labeled.

[0059] Figure 3 This is a schematic diagram of the Transformer architecture shown in an embodiment of the present invention. It includes key modules such as a position encoding module, a multi-head self-attention mechanism, a feedforward network, residual connections, and layer normalization. It demonstrates the layer-by-layer extraction and prediction output mechanism of parameter features in the network and annotates the parameter configuration of each module.

[0060] Figure 4 The above is a comparison chart of the Transformer parameter prediction results shown in the embodiments of the present invention. It shows the prediction effect of the model on the test set for key parameters such as cutterhead torque, cutterhead rotation speed, average feed speed, total thrust and penetration. It includes a comparison between the predicted curve and the actual curve with time on the horizontal axis and parameter values ​​on the vertical axis, which is used to illustrate the prediction accuracy and reliability of the method of the present invention.

[0061] It should be noted that this invention is not limited to the selection of specific network structure parameters, filtering methods or dimensionality reduction algorithms. Those skilled in the art can make reasonable adjustments to the number of network layers, convolutional kernel size, learning rate, sampling frequency or model update cycle according to actual application needs. As long as it does not deviate from the basic idea of ​​this invention, that is, to achieve adaptive classification and dynamic perception of the strata in shield tunneling through feature decoupling mechanism and lightweight convolutional neural network, it should fall within the protection scope of this invention.

Claims

1. An adaptive prediction method for tunnel boring machine (TBM) parameters based on the Transformer architecture, characterized in that, Includes the following steps: S1. Data Acquisition: Multi-source operating parameters during the tunneling process are collected in real time through the shield monitoring system to form multi-dimensional time series data. The multi-source operating parameters include cutterhead rotation speed, cutterhead torque, propulsion cylinder pressure, propulsion speed, penetration depth, jack stroke and jack speed. S2. Data Preprocessing: The multidimensional time series data is systematically preprocessed, including outlier detection and processing, signal denoising, feature dimensionality reduction, and data normalization, to generate a high-quality input feature set. Outlier detection uses a box plot method based on interquartile range to remove outliers; denoising uses discrete wavelet transform decomposition and reconstruction to preserve the low-frequency main trend; dimensionality reduction filters key physical quantities highly correlated with the tunneling status from the original monitoring data; and normalization standardizes all parameter data. S3. Distinguishing between active and passive quantities: Based on the response relationship between each operating parameter and the shield control signal, the operating parameters are distinguished into active and passive quantities. The active quantity represents the control characteristics of the tunneling equipment, and the passive quantity represents the ground feedback characteristics. S4. Construction of time series samples: Based on the sliding window method, the parameter values ​​of several historical time steps are used as input features, and the parameter values ​​of at least one future time step are used as target output to construct supervised learning samples. S5. Transformer Model Construction and Training: The multidimensional features that have been preprocessed and constructed from samples are input into the prediction model based on the Transformer architecture for training. The model includes an input embedding layer, a position encoding module, a multi-head self-attention structure, a feedforward network, residual connections and layer normalization, encoder stacking, and an output layer. S6. Multi-parameter joint prediction: The model performs joint prediction on multiple key tunneling variables and outputs the predicted values ​​of multiple parameters; S7. Adaptive Learning: Based on the difference between the actual parameter values ​​of newly collected data and the model output, the network weights are dynamically adjusted to achieve online incremental learning and continuous optimization of the model. S8. Prediction Result Output: The parameter prediction results are output in real time in a visual form, including single-step prediction, multi-step prediction, prediction confidence interval and risk warning information. These are provided to the shield tunneling monitoring system through a data interface, allowing construction personnel to view the prediction information and make decisions.

2. The adaptive prediction method for shield tunneling parameters based on Transformer architecture according to claim 1, characterized in that, In step S2, data preprocessing includes the following steps in sequence: S201. Outlier Detection and Handling: Outliers in each parameter sequence are identified and removed using a box plot method based on the interquartile range; specifically, the first quartile is used... Q 1 With the third quartile Q 3 Calculate the interquartile range based on the difference Q R Set the outlier threshold as [ Q 1 -1.5 Q R , Q 3 +1.5 Q R Values ​​outside this range are identified as outliers and removed; missing positions after removal are filled using median interpolation or linear interpolation with a sliding window. S202, Signal Denoising: One-level wavelet decomposition and reconstruction is performed using the db10 wavelet basis in the Daubechies wavelet family, and an adaptive thresholding method is used to suppress high-frequency noise; S203, Feature Dimensionality Reduction: Apply principal component analysis algorithm to transform the denoised data and extract principal components whose cumulative contribution rate reaches or exceeds a predetermined threshold; S204. Data normalization: The Z-score normalization method is used to make each parameter have a statistical distribution with a mean of zero and a variance of one.

3. The adaptive prediction method for tunnel boring machine parameters based on Transformer architecture according to claim 1, characterized in that, In step S3, the step of distinguishing between active and passive quantities includes: calculating the covariance matrix between the control signal and each monitoring parameter, and calculating the response coefficient of each parameter; when the response coefficient is greater than a set threshold, it is determined to be an active quantity; when the response coefficient is less than the set threshold, it is determined to be a passive quantity; the active quantity includes the feed speed and the cutter head rotation speed, and the passive quantity includes the cutter head torque, the feed cylinder pressure, and the penetration depth.

4. The adaptive prediction method for tunnel boring machine parameters based on Transformer architecture according to claim 1, characterized in that, In step S4, the time series sample construction steps include: S401, Sliding window design: Set the window size to... N A time step, in the past N Using the actual parameter values ​​at each time step as input features, predict the single-step parameter value at the next moment or the multi-step parameter value at multiple future moments; S402, Sample Construction Strategy: For samples of length... T Time series data is used to sequentially extract continuous time segments using a sliding window to generate... T - N training samples; each sample contains _ training samples; N The input features at each time step and their corresponding target output values; for multi-step prediction tasks, the target output value includes future... M Parameter values ​​for each time step; S403, Data Augmentation: Employ sliding window sliding, resampling, and noise injection methods to improve the diversity and coverage of training samples; add small-amplitude noise conforming to a Gaussian distribution; S404. Sample Balancing: The training samples under different working conditions and geological conditions are balanced by adopting a stratified sampling strategy to ensure that the proportion of samples under various working conditions in the training set is balanced.

5. The adaptive prediction method for tunnel boring machine parameters based on Transformer architecture according to claim 1, characterized in that, The Transformer model includes an input embedding layer, a positional encoding module, a multi-head self-attention structure, a feedforward network, residual connections and layer normalization, encoder stacking, and an output layer. The input embedding layer maps multidimensional feature vectors to the model dimension space through a linear transformation; The position encoding module uses a position encoding method that combines sine and cosine functions; The multi-head self-attention structure divides the multi-dimensional feature vector into multiple subspaces, with a default setting of 8 heads; The feedforward network employs a two-layer fully connected network structure and uses the ReLU activation function; Residual connections and layer normalization are added after each sub-layer; Encoder stacking involves stacking the aforementioned self-attention mechanism, feedforward network, and corresponding normalization module. N layer; The output layer maps the encoder output to the target dimension space through a fully connected layer.

6. The adaptive prediction method for tunnel boring machine parameters based on Transformer architecture according to claim 1, characterized in that, The adaptive learning steps include: real-time monitoring of the model's prediction performance on new data batches and calculation of prediction error metrics; automatic triggering of the model update process when the prediction error exceeds a set threshold or when the prediction error continues to rise in multiple consecutive batches; incremental training of the model using a small learning rate, updating only some network layers or using a weight decay strategy; saving model versions at different stages, supporting model rollback and performance comparison.

7. A shield tunneling parameter adaptive prediction system based on Transformer architecture, applied to the shield tunneling parameter adaptive prediction method based on Transformer architecture as described in claim 1, 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 sequentially perform outlier detection and processing, signal denoising, feature dimensionality reduction, and data normalization on the collected time-series data. Feature decoupling unit is used to distinguish operating parameters into active and passive variables; Temporal sample construction unit, used to construct supervised learning samples based on the sliding window method; The Transformer model training and prediction unit is used to build and train prediction models based on the Transformer architecture and perform parameter prediction. Adaptive learning units are used to dynamically update model parameters based on real-time data; The prediction result output unit is used to output the prediction results in the form of structured data and visualization, and provide them to the tunnel boring machine monitoring system.

8. The adaptive prediction system for tunnel boring machine parameters based on Transformer architecture according to claim 7, characterized in that, The prediction result output unit includes: The results generation module is used to generate single-step predicted values, multi-step predicted sequences, prediction confidence intervals, and error estimates. The visualization module is used to display the prediction results and historical measured data in the form of charts; The early warning module is used to generate early warning information when key parameters are predicted to be abnormal or exceed thresholds. The data transmission module is used to transmit prediction results, visualization data, and early warning information to the tunnel boring machine monitoring system in real time through a standardized data interface.

Images