A novel self-supervised learning framework for gearbox fault diagnosis

This study proposes an innovative self-supervised fault diagnosis framework that combines gated attention mechanisms and multi-head attention mechanisms, consisting of a dual-channel encoder and a cross-domain fusion encoder. In the dual-channel encoder phase, time–frequency domain contrastive loss is proposed to drive the encoder to extract deep fault features. in the cross-domain fusion encoder phase, time–frequency matching loss is utilized to guide the effective integration of different modal features. The advantage of this framework lies in its ability to use unlabeled fault data for pre-training and capture effective fault information representations through the self-supervised learning process. In actual fault diagnosis tasks, only a small number of labeled samples are needed to fine-tune the pre-trained model to significantly improve diagnostic accuracy. We have demonstrated that the proposed method is superior to existing technology in terms of fault diagnosis accuracy, generalization ability, and domain adaptability, fully demonstrating its application potential.