Intelligent fault diagnosis scheme for rotating machinery based on momentum contrastive bi-tuning framework

Existing fine-tuning methods mainly leverage the discriminative knowledge and discard the intrinsic structure of data. In this paper, we propose a novel framework Momentum Contrastive Bi-Tuning (MCBiT) for intelligent diagnosis of rotating machinery, which can fully exploit both the discriminative knowledge of labels and the in-trinsic structure of target data in a boosting fine-tuning way. One-dimensional vibration signals are transformed by Gramian Angular Difference Field (GADF) and fed into MCBiT, which enhances the conventional fine-tuning by integrating two branches on the ImageNet-pretrained backbone: a classifier with an instance-contrastive cross-entropy loss to better exploit label knowledge; and a projector with a categorical contrastive learning loss to mining the intrinsic structure of data. Our proposed approach outperforms state-of-the-art methods on six publicly available rotating machinery fault diagnosis datasets and our experimental-collected dataset at different data scales. The promising performance of our proposed MCBiT contributes toward more practical data-driven approaches that can realize timely deployment under challenging real-world environments.