A Bayesian nonparametric approach for tool condition monitoring

In modern manufacturing systems, the failure of machine tools may cause unexpected system breakdown and bring about tremendous financial losses. With an effective tool condition monitoring (TCM), unnecessary downtime for maintenance can be reduced. Unfortunately, machine tool dynamics are complex, and the accurate relationship between monitoring signals and the tool health states is difficult to describe. In this work, the aim of tool condition monitoring is to estimate and predict the unobserved degree of the tool wear on-line by using the observed raw monitoring sensors. We take a Bayesian nonparametric approach to construct the relationship between raw force signals and the dynamic tool wear accumulation process. Using a Dirichlet process prior over mixture weights, we learn an infinite health state mixture model from training data to describe the continuous wear accumulation process. The nonparametric nature of our model allows control of the model size and self-adaption of the model parameters, and the use of Bayesian method significantly prevents under-fitting and avoids over-fitting. To validate the effectiveness of our model, the proposed approach is applied on the real data from a high-speed CNC milling machine cutters.