Unsupervised Order-Enabled Contrastive Learning-Based Dual Autoencoder Fault Detection for Variable Speed Bearing

Bearing fault detection is crucial in ensuring the reliability and safety of rotating machinery, as undetected faults can lead to unexpected equipment failure and costly down-time. Traditional fault detection methods particularly based on constant speed assumptions, face challenges in handling the complexities of variable speed conditions. In this paper, we propose a fault detection method for bearings under variable-speed conditions, combining order information and contrastive learning. Specifically, the impact of speed variations is effectively mitigated through angular resampling. Then, order envelope signals are extracted to amplify weak signals. Subsequently, a contrastive learning-based dual-channel asymmetric autoencoder structure is introduced, where two autoencoders independently process and reconstruct signals. By minimizing the contrastive loss between hidden layers of normal samples, the model learns the inherent consistency of normal data representations. This consistency allows the model to detect abnormalities by identifying deviations in contrastive loss. Experiments validate the performance of the proposed method in detecting bearing faults under variable speed conditions.