Time-frequency synchronisation contrastive learning-driven multi-sensor remaining useful life prediction

Deep learning techniques play a crucial role in predicting the remaining useful life (RUL) of mechanical equipment. Nevertheless, obtaining a substantial number of labeled samples is a challenge, and the prediction accuracy tends to decline when the labeled samples are insufficient. Moreover, existing RUL prediction methods usually extract the degradation characteristics only from one single domain, which is insufficient for a high-accuracy prediction. To address these challenges, a time-frequency synchronization contrastive learning-driven (TFSCL) multi-sensor remaining useful life prediction model is proposed. The proposed TFSCL utilizes a large amount of unlabeled data for model pre-training and key feature extraction, and it introduces a novel time-frequency fusion contrastive loss function to optimize the pre-training process. It employs a dual-channel structure at the sensor and timestamp levels, incorporating an attention mechanism that adaptively adjusts sensor feature weights, enabling more accurate extraction of critical information while effectively mitigating interference from irrelevant data. To validate the effectiveness of the proposed TFSCL, two case studies are conducted, with different labeling ratios being used. The experimental results demonstrate that even with lower labeling ratios, the proposed TFSCL model still achieves a satisfactory prediction effect and outperforms other advanced methods.