On Transfer Learning of Traditional Frequency and Time Domain Features in Turning

TL;DR

This study evaluates traditional frequency and time domain features for chatter detection in turning processes, revealing their limited transferability across different configurations and comparing them with advanced methods like WPT and EEMD.

Contribution

It systematically assesses the transfer learning capabilities of traditional features versus advanced methods in chatter prediction for turning operations.

Findings

Fourier spectrum features achieve up to 96% accuracy within same configurations.

Transferability of traditional features drops significantly across different configurations.

EEMD outperforms traditional features in transfer learning scenarios.

Abstract

There has been an increasing interest in leveraging machine learning tools for chatter prediction and diagnosis in discrete manufacturing processes. Some of the most common features for studying chatter include traditional signal processing tools such as Fast Fourier Transform (FFT), Power Spectral Density (PSD), and the Auto-correlation Function (ACF). In this study, we use these tools in a supervised learning setting to identify chatter in accelerometer signals obtained from a turning experiment. The experiment is performed using four different tool overhang lengths with varying cutting speed and the depth of cut. We then examine the resulting signals and tag them as either chatter or chatter-free. The tagged signals are then used to train a classifier. The classification methods include the most common algorithms: Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Gradient Boost (GB). Our results show that features extracted from the Fourier spectrum are the most informative when training a classifier and testing on data from the same cutting configuration yielding accuracy as high as %96. However, the accuracy drops significantly when training and testing on two different configurations with different structural eigenfrequencies. Thus, we conclude that while these traditional features can be highly tuned to a certain process, their transfer learning ability is limited. We also compare our results against two other methods with rising popularity in the literature: Wavelet Packet Transform (WPT) and Ensemble Empirical Mode Decomposition (EEMD). The latter two methods, especially EEMD, show better transfer learning capabilities for our dataset.