Chatter Detection Approach Based on Wavelet Synchrosqueezing and t-Distributed Stochastic Neighbor Embedding for a Turning Process

Chatter is a type of unstable self-induced vibration and a major obstacle in machining processes. Such vibration might reduce workpiece surface quality and production efficiency and can damage cutters and machine tables; therefore, chatter prediction and identification are crucial aspects of machining, and thus, have been widely investigated by experts and scholars in relevant fields. In the present study, an accelerometer was installed on the main spindle casing used in lathe milling, and vibration data were collected for analyses to detect chatter. The collected vibration data were then converted through synchrosqueezing, and instantaneous amplitude- and frequency-domain features were extracted using time–frequency-domain representations. The processed data had more distinct features than the original data and less noise. These data were used to train various neural network models, and their performance levels were compared through cross validation. The results revealed that the VGGNet model achieved the highest prediction accuracy (93.3%). The t-distributed stochastic neighbor embedding (t-SNE) method was also employed for dimensionality reduction while preserving local data structures; t-SNE was noted to be effective for reducing the dimensionality of high-dimensional data to reduce computational loads after synchrosqueezing. With the data in the low-dimensional space, the proposed approach could effectively indicate the occurrence of chatter. The data features and time–frequency-domain signals obtained after dimensionality reduction were imported into the constructed model for retraining, improving the prediction accuracy to 96.7%.