A new approach for developing EPB-TBM disc cutter wear prediction equations in granite stratum using backpropagation neural network. (October 2022)

• Record Type: Journal Article
• Authors: Ding, Xiaobin; Xie, Yuxuan; Xue, Haowen; Chen, Rui
• Is Part Of: Tunnelling and underground space technology. Volume 128(2022)
• Publication Date: 2022-10
• Subjects: Disc Cutter Wear Prediction -- Backpropagation Neural Networks -- Tunnel Boring Machine -- Operational Parameters; Tunneling -- Periodicals; Underground construction -- Periodicals; Tunnels -- Periodicals; Underground areas -- Periodicals; 624.193
• DOI: 10.1016/j.tust.2022.104654
• Languages: English
• ISSNs: 0886-7798

Abstract

Graphical abstract: Highlights: An adaptable workflow is proposed for quantitative evaluation of cutter wear. Involve 12 parameters in backpropagation neural network analysis. Propose several empirical equations for cutter wear prediction. Sensitivity analysis from multivariate analysis reveals affecting parameters. Abstract: Accurate prediction on Tunnel Boring Machine (TBM) cutter wear can lower the risk of cutter replacement and provide a reference for project management. A workflow is proposed to develop a new empirical disc cutter wear prediction model with Backpropagation Neural Network (BPNN) by incorporating parameters from TBM operation, geological conditions, and cutter layout. The procedure is examined by a case study to develop an empirical equation quantifying cutter wear by the reduction in cutter radius instead of the service life. A dataset with 12 different parameters is established from a 595-ring tunnel section of Guangzhou Metro Line 18 project constructed in granite and migmatitic granite strata. Cumulative Error Rate (CER), defined as the ratio of the error between the predicted and actual value of cutter wear to the actual value, is used for model performance evaluation and error tracing. The top 3 rated BPNN models get the best average CER with 9.10% Bias and 11.34% Variance. Sensitivity analysis are performed on these outstanding models providing references for the development of prediction equations. The log–log equation outperforms quadratic and level-level equations and reaches an average CER with 18.77% Bias and 8.77% Variance, which is a great improvement compared with other published empirical equations. Additionally, error tracing and equation explanation are performed to provide optimization options for future studies.