An interpretable anti-noise convolutional neural network for online chatter detection in thin-walled parts milling

Abstract

Chatter is a notoriously unstable phenomenon that can adversely affect both surface quality and machining efficiency. To achieve high-performance machining, the development of online chatter detection is of paramount importance. Nevertheless, changes in chatter frequency with cutting position and noise interference during the thin-walled parts milling process present significant challenges to chatter detection. To tackle this issue, an adaptive frequency band attention module (AFBAM) is designed, which is characterized by not relying on prior knowledge (namely modal parameters, frequency spectrum analysis, etc.), and adaptively enhances the frequency band containing abundant chatter information and reduces noise interference by learning time-frequency domain characteristics of signals. After AFBAM highlights the relevant frequency band of input signal, a discriminative feature attention module (DFAM) is constructed to adaptively recalibrate feature responses of each convolutional layer utilizing the global information. DFAM enhances relevant features and suppresses irrelevant features, thus improving the discriminative feature learning and redundant information suppression abilities of network. In addition, both AFBAM and DFAM exhibit clear physical interpretability, which improves the interpretability of network. Based on AFBAM and DFAM, an interpretable anti-noise convolutional neural network for online chatter detection, named AD-CNN, is established. Milling experiments with pocket-shaped thin-walled parts are conducted under different cutting parameters. The results show that the proposed method enables better detection accuracy and anti-noise ability than other state-of-the-art methods. Furthermore, visualization analysis of AFBAM and DFAM brings new insights into the interpretability of convolutional neural network in the field of chatter detection.