13 August 2024 in IEEE Transactions on Industrial Informatics
Many machines generally operate under time-varying speed conditions. The significant and frequent changes of speeds often lead to missed defect detections and false alarms, which hinders the further applications of health monitoring technology. In this article, a novel self-supervised learning motivated network that is composed of extractor, projector, spyder and detector, i.e., EPSDNet, is proposed for self-adaptive health monitoring of machines under time-varying speed conditions. First, contrastive learning is used in EPSDNet to implement self-extraction of trend manifold over speeds, which adjusts the representational features and reduces their entropy. Second, a nonlinear regression network, i.e., Spyder, is proposed to identify the relationship between vibration signals and corresponding rotational speeds, which facilitates the adaptive feature alignment for each transient speed condition. Finally, a refined skip connection-based feature fusion is proposed to balance the calculation cost and signal reconstruction among encoder and decoder of detector. The real-time health indicator can be obtained based on the reconstruction error of detector for machinery. The effectiveness of EPSDNet is validated on two bearing test rigs with time-varying rotational speeds. The results show that EPSDNet outperforms other state-of-the-art methods for machinery health monitoring, which provides a novel approach to address the challenging issue of time-varying speed conditions.
25 June 2025 in Advanced Engineering Informatics
13 June 2024 in IEEE Transactions on Power Electronics
Valve-regulated lead-acid (VRLA) batteries are of paramount importance in ensuring the safety and stability of large intelligent data centers (IDCs). The health management of VRLA battery suffers from the limitation of sparse charge-discharge data and nonideal samples in the high-voltage direct current system of IDC. In this article, a novel semisupervised few-shot model, i.e., densely-connected contrastive observer (DCober), is proposed for feature self-extraction of multisource signals and few-shot diagnosis of VRLA batteries. First, a view-augmented method is proposed to implement image augmentation for multichannel temporal images, which improves the feature self-extraction performance. Second, self-supervised contrastive learning is used in DCober to implement feature representation learning on massive nonlabeled data in real industrial scenarios. Finally, a similarity-based few-shot finetuning method is proposed for finetuning enhancement on the extension of limited labeled data, which improves the fault recognition performance of DCober in few-shot scenarios. The effectiveness of DCober is validated through a real industrial application of VRLA battery in Tencent T-block IDC. The testing results show that DCober outperforms other state-of-the-art methods and obtains a recognition accuracy of 95.63% in 1-shot data scenario.
21 January 2026 in IEEE Transactions on Systems, Man, and Cybernetics: Systems
Machinery fault diagnosis with limited fault data under time-varying conditions remains a persistent challenge. In practical industrial settings, healthy samples are generally abundant during daily monitoring, while fault data is sparse. This hinders further development of dynamic fault recognition technology. In this article, a novel self-supervised vibration representation and speed-joint network, i.e., VRSNet, is proposed to conduct few-shot fault diagnosis under time-varying speed conditions. First, contrastive learning (CL) is employed to achieve feature self-representation using only healthy vibration signals. This enhances the discriminability of extractor and facilitates the self-extraction of feature manifolds across varying speeds. Second, a dimension-lifting network named approximator (APP) is proposed to identify the potential relationship between rotational speeds and corresponding vibration representations, which facilitates condition alignment for each transient speed and cross-modal feature fusion. Finally, a vibration-speed-joint fine-tuning method is proposed to implement few-shot training for fault recognition, where the norm-freezing strategy effectively preserves the pretraining knowledge of both vibration and speed branches for generalization enhancement. VRSNet is verified through two cases under time-varying speed conditions. The results demonstrate that it outperforms those state-of-the-art (SOTA) methods and achieves a recognition accuracy of 84.75% in a 1-shot fault data scenario with varying speeds. Even when confronted with faults at previously unseen speeds, VRSNet can still recognize them accurately.