A wavelet-based data-driven modelling for tool wear assessment of difficult to machine materials

In this work, wavelet packet decomposition along with principle component analysis are utilised for feature extraction using two low cost sensing methods: vibration and power sensors, in end-milling of gamma prime-strengthened alloy. The high wear rate of this material induces a rapid transition from a sharp state to a dull state of the tool, and hence limits the number of available data for model establishment. To address this challenge, a neural network with Bayesian regularisation is designed and its performance is compared with regression and time-series models. A maximum of 4% estimation error for Bayesian regularisation neural network, compared to 33% and 17% estimation error of the latter models, shows the good potential of this technique when a limited dataset is available. In addition, the use of low cost measuring sensors in this paper aligned well with the industrial applications to detect and avoid unscheduled downtime in machining situations.