An accurate and adaptative cutterhead torque prediction method for shield tunneling machines via adaptative residual long-short term memory network. (15th February 2022)

• Record Type: Journal Article
• Authors: Jin, Yanrui; Qin, Chengjin; Tao, Jianfeng; Liu, Chengliang
• Is Part Of: Mechanical systems and signal processing. Volume 165(2022)
• Publication Date: 2022-02-15
• Subjects: Cutterhead torque prediction -- Domain adaptation -- Long short-term memory neural network -- Residual block -- Shield tunneling machine; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.108312
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: It proposes an adaptative residual LSTM network for cutterhead torque prediction. It uses domain adaptation to learn common representation from the labeled dataset. Its accuracy and performance outperform the existing data-driven methods. It improves prediction performance by reducing 0.057 MSE, 0.069 MAE and 5.2% MAPE. Abstract: During the tunnel excavation, accurate cutterhead torque prediction is helpful to adjust shield machine operation parameters for avoiding cutterhead jamming, which greatly improves the safety and efficiency and reduces the cost. However, due to complex and changeable geological environments, it is improper to directly apply the trained model into other domains under changeable geological conditions. To solve the cross-domains issues, this paper presents a novel Adaptive Residual Long Short-Term Memory Network (ARLSTM) to predict cutterhead torque across domains. To begin with, residual blocks are applied into extracting beneficial features from the tunneling parameters automatically. And then, these extracted features are fed into feature regression predictor for obtaining final prediction results. Furthermore, the introduction of domain classifier is used for reducing probability distribution discrepancy from different domains via adversarial procedure. Finally, we used the actual dataset collected from Singapore Metro T225 project for evaluating ARLSTM. The experimental results present that ARLSTM improves the prediction performance by reducing 0.0571 MSE on average, 0.0695 MAE on average and 5.20% MAPE on average using the knowledge of source domain dataset. Meanwhile, compared with other data-driven methods, the comparison results present that the proposed network structure achieved better prediction performance. Consequently, ARLSTM has potential as a promising precise tool for cutterhead torque prediction used in the tunneling process.