Vibration-based gear continuous generating grinding fault classification and interpretation with deep convolutional neural network

Abstract

Continuous generating grinding plays an essential role in modern gear manufacturing, while the increasing need for high-quality gear products pushes the quality control of the grinding process. This paper proposed a vibration-based process monitoring method to classify faults generated during grinding processes. Gear grinding measurements have been conducted during experiments with different parameters of the feed rate, infeed, and gear eccentricity. The grinding vibration signals have been gathered to form a dataset with seven gear quality categories. A novel deep learning model is presented in this paper exploiting the power of Wavelet Packet Decomposition (WPD) and Deep Convolutional Neural Network (DCNN). The WPD is adopted to transform the vibration signal into 2D time-frequency representations, used as inputs to the DCNN model for classification. Experimental results show that, the proposed DCNN can achieve a high classification accuracy of 95.83%, demonstrating the methodology's efficiency. In addition, this paper exploits the Gradient-weighted Class Activation Map (Grad-CAM) technique to interpret the deep neural network's decisions in order to facilitate the model deployment in industrial production. The Grad-CAM visualizations are utilized to show the activation regions on the spectrograms, which further indicate that the attentions of the DCNN have a strong correlation with the manufacturing parameters of different grinding passes.