This paper addresses the following research question: How can we reliably uncover causal relationships from irregular, noise-contaminated tunnel boring machine (TBM) operating parameters while achieving high-accuracy attitude prediction? To address this question, we develop an enhanced causal discovery approach tailored for TBM attitude prediction. While training a causally driven neural network, we jointly impute missing data points and discover causal graphs to handle irregular, noise-laden TBM operating parameters. We use a case study to validate the feasibility and effectiveness of our proposed approach. Our approach, coupled with a simplified GRU, achieves an R2 of 0.987 for pitch angle prediction, outperforming other models (e.g., long short-term memory and common gated recurrent unit with or without causally driven). Additionally, our research indicates that cutterhead motor torque, cylinder pressure adjustments, and earth pressure are critical determinants of TBM offset and angle deflection. We suggest that our newly developed approach can help site managers better understand how and why the generated outputs from deep learning materialize, which can be used to improve decision-making effectiveness in tunnel construction.
Chen et al. (Tue,) studied this question.