Prediction of Shield Tunneling Attitude Based on WM-CTA Method

[Objective] The attitude of a shield machine is a critical parameter that significantly affects tunnel construction, directly determining construction safety and project quality. To ensure that shield tunneling closely aligns with the designed alignment and to improve engineering construction quality, this study proposes a novel shield attitude prediction model, called WM-CTA, based on deep learning technology. [Methods] The WM-CTA model primarily consists of two frameworks: a data preprocessing module (Wavelet Transform and Maximum Information Coefficient) and a prediction module (Convolutional Neural Network and Attention Mechanism). The preprocessing module, composed of Wavelet Transform (WT) and the Maximum Information Coefficient (MIC) algorithms, was used to perform noise reduction and parameter correlation analysis on the raw data, thereby generating enhanced inputs. The Convolutional Neural Network (CNN) integrated with a channel-wise attention mechanism explored parameter weight differences and extracted local data features. Subsequently, the Temporal Convolutional Network (TCN) was employed to capture temporal dependencies and dynamic variations in the data. Finally, the Attention Mechanism (AM) was applied to extract key temporal node information. The model’s prediction performance was validated using monitoring data from a section of a shield tunnel under construction in Shenyang. Experiments were conducted on data for noise reduction and correlation analysis, followed by analysis of the model’s prediction performance and generalization ability. [Results] Experimental results showed that the monitoring curves processed with wavelet transform had improved smoothness with reduced frequency of abrupt changes between data points. Correlation analysis indicated that shield construction parameters exerted greater influence on shield attitude than soil parameters, enabling dimensionality reduction of input parameters. Compared with four baseline models, the proposed WM-CTA model achieved minimum MAE and RMSE and maximum R2 value. [Conclusion] The experiments verify that the WM-CTA model delivers optimal prediction performance with high computational efficiency. Furthermore, the model exhibits strong generalization ability, providing valuable references for similar future engineering projects.