Time varying and condition adaptive hidden Markov model for tool wear state estimation and remaining useful life prediction in micro-milling

•Developed an improved hidden Markov model for tool wear monitoring under switching cutting conditions in micro-milling.•Investigated the relationship between cutting condition, cutting time and tool wear state transition.•Considered the effect of cutting condition on the observation in the proposed model.•Verified the effectiveness of the approach with various experimental studies. The tool wear monitoring (TWM) system which can estimate the tool wear state and predict remaining useful life (RUL) of the tool plays an important role in micro-milling because of the high precision requirement for work-pieces and the high tool wear rate. Due to its ability in modelling the non-stationary physical process, hidden Markov model (HMM) has been broadly used in TWM, but almost all of researches have been done under fixed cutting conditions. In order to monitor tool wear under switching cutting conditions, an improved HMM is proposed in this paper. A hazard model is constructed to describe the time varying and condition adaptive state transition probability. Multilayer perceptron (MLP) which is powerful in approximating a nonlinear function is adopted to compute the observation probability. Then, the state transition probability and observation probability are integrated to estimate the tool wear state and predict the RUL online using forward algorithm. Experiments on variant cutting conditions are conducted to verify effectiveness of the proposed model.