Abstract
Accurate bearing remaining useful life (RUL) can effectively ensure the safe operation of equipment and enhance production efficiency. Despite the widespread application of deep learning-based prediction methods, most rely on supervised learning to directly map input signals to output RUL. However, this often ignores crucial representational properties like smoothness and monotonicity, leading to disorganized and uninterpretable representations that significantly degrade performance. To enhance representation, this paper proposes a novel dual-dimensional contrastive self-supervised learning-based framework named DCSSL for RUL prediction of rolling bearings. It is carried out in two consecutive stages. In the first stage, a strategy combining random cropping and timestamp masking for constructing positive pairs for contrastive learning is proposed. The dual-dimensional contrastive loss function that combines temporal-level and instance-level is devised to enable the model to learn state representations in unlabeled vibration data and mine rolling bearing degradation trends. Then, in the second stage, RUL prediction of labeled vibration data is achieved by fine-tuning the newly constructed prediction head. Experimental validation of DCSSL on a large number of RUL prediction tasks demonstrates its superior performance over other state-of-the-art methods.