A Computationally-efficient Semi-supervised Learning Model for the Estimation of State Degradation of a Milling Tool

This paper proposes a computation-efficient semi-supervised machine learning model for the near-to-real-time state of health estimation for a Milling tool. To secure high reliability of the health state estimation, we have adopted the two-stage classification Beta Variational Auto-Encoder (β-VAE) architecture. The computation efficiency is achieved through the correlation between the latent space and input space classification processes of β-VAE for healthy states, where the input space decoder classification is triggered only when the latent space classifies the state to be unhealthy. This can be seen as a double-check process due to the high cost related to processing the false-positive maintenance decisions. We have implemented the proposed model and tested it on the UC Berkeley milling data set. The experiment results show that our state of health estimation model can reduce the computation cost by up to 15% while achieving high classification accuracy.