Edge Training of Lightweight Bearing Fault Diagnosis Model via Dataset Distillation

Rolling bearings are critical components in industrial machinery, and effective fault diagnosis is essential for ensuring safe equipment operation. However, training high-performance fault diagnosis models typically requires large-scale and diverse fault data, which makes it nearly impossible to train directly on edge AI devices. To address this challenge, this paper proposes a rapid training method based on dataset distillation, aimed at significantly accelerating the training process of rolling bearing fault diagnosis models. Inspired by dataset distillation techniques in the image domain, we synthesize a small number of highly representative data points that encapsulate the knowledge from the original large datasets, such as the CWRU dataset. These synthetic data points serve as an efficient "refined"training set, enabling fault diagnosis models to achieve performance comparable to those trained on the full dataset, but with remarkably fewer gradient descent steps. This study details the application of dataset distillation to time-series vibration signals for fault diagnosis tasks, eliminating the intermediate step of converting them into two-dimensional images. Experimental results validate that the proposed method substantially reduces model training time while maintaining high diagnostic accuracy, offering a novel approach for efficient and lightweight industrial fault diagnosis.