Signal decomposition enhanced multi-step prediction model of tunnel boring machine based on attention mechanism

To improve short-horizon prediction of critical tunnel boring machine operational parameters, this study develops DRFormer, a multi-step forecasting framework that integrates target-dependent signal decomposition with a hybrid deep-learning architecture. The raw multivariate time-series signals are decomposed into multi-scale components to separate high-frequency fluctuations from low-frequency trends, thereby improving feature representation and reducing the influence of measurement noise and nonstationarity. Following decomposition, a one-dimensional residual network extracts local continuity and multi-scale short-range features, while temporal pattern attention adaptively emphasizes informative periodic, trend-related, and transient patterns. A ProbSparse self-attention encoder–decoder subsequently captures longer-range temporal dependencies with fewer redundant attention operations. In addition, an importance-guided variable selection strategy identifies target-relevant auxiliary parameters from the training data to represent cross-parameter coupling and strengthen model robustness. Evaluations using real-world shield-tunneling data demonstrate that DRFormer provides accurate and stable multi-step forecasts for five key operational parameters, with mean absolute percentage errors generally below 2% and coefficients of determination consistently above 0.99. Comparative and ablation analyses show that the proposed modules make complementary contributions and that DRFormer achieves broad improvements over established forecasting approaches in prediction accuracy, stability, and interpretability. These results highlight its potential for intelligent control, operational optimization, and proactive decision-making in shield tunneling engineering.