A Novel Transfer Learning Framework for Chatter Detection Using Convolutional Neural Networks

Abstract

Detection and avoidance of regenerative chatter play a crucial role in ensuring the high quality and efficiency of machining operations. Predominant analytical approaches provide stability lobe diagrams for machining processes. Deep learning is a general term given to the most recent and successful group of machine learning methods that proved great promise in many areas of human life. This study purposes a novel transfer learning framework that combines analytical solutions and convolution neural network (CNN) under a novel transfer learning framework. Stability lobes and numerical time-domain solutions of analytical methods are used to train and label, arguably one of the most successful CNN architectures, AlexNet. This approach eliminates the need for a time-consuming and costly experimental data collection phase for training. Furthermore, an ensemble empirical mode decomposition based signal pre-processing method is developed. An IMF-based multi-band ensemble approach is proposed where only intrinsic mode functions relevant to each modal frequency of the system are selected based on their entropy increase and used in training multiple AlexNet instances. The measured data were collected during shoulder milling from a CNC-vertical milling machine. The results revealed considerable success in several scenarios ranging from 82 to 100%, without using any experimentally measured data in training.