Transfer Learning with 2D Vibration Images for Fault Diagnosis of Bearings Under Variable Speed

Abstract

One of the most critical assignments in fault diagnosis is to decide the finest set of features by evaluating the statistical parameters of the time-domain signals. However, these parameters are vulnerable under variable speed conditions, i.e., different loads, and speeds to capture the dynamic attributes of various health types. Therefore, this paper proposes a vibration imagining-based diagnosis approach for bearing under variable speed conditions. First, a Discrete Cosine Stockwell Transformation (DCST) coefficient-based preprocessing step is proposed to create an identical health pattern for variable speed conditions. Then, from that 2D coefficient matrix, a vibration image is created to capture those health patterns into grayscale. Finally, a Transfer Learning embedded Convolutional Neural Network (TL-CNN) is proposed to inspect the comprehensive structure of the 2D vibration images for final classification. The experimental results show that the proposed method achieved 100% classification accuracy on a publicly available dataset.