ABSTRACT
Unplanned worn disc cutters maintenance can cause casualty and financial loss. Good knowledge of cutter wear will facilitate the design of the excavation plan and elongate cutter life. Cutter wear is subject to multiple influential factors, which can be regarded as a nonlinear multivariate question. Back propagation neural network (BPNN), a robust machine learning method in this field, can shed light on it. A shield tunneling section from Metro Line 18 in Guangzhou, China, only encounters hard rock strata. There are 49 manually measured cutter wear. The tunnel boring machine records over 250 types of parameters per second with a real-time logging system. According to 28 types of influential parameters from previous studies, 14 input parameters are selected to reflect the effect of machine, geology, and operation on the output, cutter wear, which is quantified as the average radial reduction of cutter ring. By extrapolation and interpolation, a dataset with 1434 samples is established from the Pan-nan section. Cutter wear is distributed to each ring within the inspection section with a published model. To overcome the inherent weakness of BPNN, we apply SMBO (Sequential Model-based Optimization) and GA (Genetic Algorithm) and compare their effectiveness. SMBO and GA returns model with R2 of 0.968 and 0.971. Error tracing reveals GA model tends to overestimate records with slight wear.
