Using Deep Learning for Efficient Labeling of Bearing Images - Overcoming the Challenge of Limited Labeled Data Availability

Abstract

The development of complex deep learning models requires extensive labeled datasets, which are scarce in specialized fields due to the labor-intensive and time consuming nature of data labeling. This project aims to enhance the labeling process by using initial labels generated by deep learning models, later refined by human experts, a method commonly known as "Model Assisted Labeling". Collaborating with SKF, the project focuses on assessing damaged bearing images for remanufacturing suitability, particularly identifying specific failure modes. Given the abundance of raw, unlabeled images and the limited high-quality labeled data, the proposed system processes raw image data, providing preliminary labels for expert review. This system also filters out irrelevant and duplicate images, optimizing the preprocessing phase and saving both time and resources. Using techniques like transfer learning and data augmentation, our system improves deep learning model performance and generalization. The findings indicate that these techniques can effectively develop high-performing models to aid in the labeling workflow.