Prediction of shield machine attitude parameters based on decomposition and multi-head attention mechanism

Abstract

To mitigate the impact of shield attitude prediction errors on operational decision-making, a framework centered on decomposition and deep learning is proposed to predict multiple shield attitudes. The shield time series data is decomposed into trends and fluctuations by integrating detrended fluctuation analysis and variational mode decomposition. A deep learning model augmented by the multi-head attention mechanism is proposed to independently predict trends and fluctuations. The multi-head attention mechanism enhances the prediction accuracy in simultaneously predicting six attitude parameters. Most prediction errors are allocated to fluctuations through decomposition. The precise prediction of trends provides significant insights into shield attitudes and reduces the risk of misleading outcomes. Compared with existing methods, the proposed method achieves greater precision while requiring fewer inference resources to predict all six attitude parameters. The contribution of multi-head attention and the reason behind prediction error allocation are analyzed via experiments and parameter sensitivity analysis.