Generalizable fault detection and diagnosis of bearings via time-frequency embedded pre-training

Abstract Deep learning has demonstrated significant potential in data-driven fault detection and diagnosis (FDD). However, its effectiveness often depends on large amounts of labeled data, which are costly and impractical to obtain in real-world industrial applications such as elevator systems. To address this limitation, this paper proposes a novel time-frequency embedded pre-training network (TFEPTN) that employs self-supervised learning to extract meaningful representations from unlabeled vibration signals. TFEPTN features a dual-branch Transformer-based encoder for time-domain and frequency-domain feature extraction, guided by a time-frequency consistency constraint to enhance the learning of domain-representative features. The model is first pre-trained in an unsupervised manner and then fine-tuned for downstream FDD tasks. Experimental results on datasets of rolling bearings demonstrate that TFEPTN achieves strong generalization and outperforms several state-of-the-art methods under limited supervision and cross-domain conditions.