Online Milling Chatter Detection Using Deep Residual Network and Transfer Learning

In this paper, a novel approach of the online chatter detection in the milling process is presented based on the deep residual network (Resnet) and transfer learning. Considering the non-stationary characteristic of the chatter signal, the short-time Fourier transform (STFT) is used to transform the vibration acceleration signal into a two-dimensional spectrogram as input to identify the stable, transition and chatter states. A model for variable depth-of-cut conditions at 6000 rpm was trained by fusing the process information with the features down-dimensioned by the deep residual network, and the classification accuracy reached 99.2%. Since the spectrum structure of the signal changes at high speed and the data no longer satisfy the assumption of identically distribution, a model applicable to variable depth-of-cut conditions at 15000 rpm is obtained with 95.6% classification accuracy by using small samples and transfer learning method. The experimental results demonstrate that the method is suitable for early chatter monitoring under variable path and variable parameter conditions in cavity milling.