Gear Power Honing process monitoring for small training-sample size based on linear subspace learning methods

Abstract

Gear finishing process plays a key role in the manufacturing flow of today's automotive transmissions, where tight quality standard and Noise, Vibration, and Harshness (NVH) requirements must be satisfied. In this context, process monitoring systems represent a robust and flexible tool to ensure the continuous achievement of such demanding targets. Thanks to computing technology advancements and large amounts of data availability, predictive Machine Learning (ML) models are increasingly employed in industrial application. ML models require large amounts of data to achieve good performance levels and generalization. In the production field, the lack of extensive datasets limits the application of these approaches for process monitoring. In this work, we propose a Subspace Learning strategy prior to classification, to effectively extract the most relevant information from a limited number of available training samples. The original data are projected onto a lower-dimensional subspace, defined by Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), that retains the most discriminative characteristics, such that the Support Vector Machine (SVM) classifier can operate more efficiently and achieve improved prediction performance even on small training sample. The approach is validated on time-frequency vibration signals acquired on a Gear Power Honing (GPH) machine. Results show that the proposed approach achieves high accuracy even with few samples per class, thus representing a valid solution for process monitoring.