An Enhanced Approach for Remaining Useful Life Prediction of Bearings Using Incomplete Lifecycle Data

Accurately predicting the remaining useful life (RUL) of rolling bearings is crucial for ensuring optimal equipment operation and timely maintenance. This study addresses key limitations of traditional deep learning approaches, including limited lifecycle data and a significant reliance on supervised learning. To address these challenges, we propose the Incomplete Lifecycle Data Enhanced Network (ILDENet), a self-supervised learning network incorporating contrastive learning to improve feature discrimination. The network operates in three distinct phases: first, feature parameters indicative of bearing degradation trends are extracted across time, frequency, and time-frequency domains. Second, pre-training is performed in a mask self-supervised learning mode with contrastive learning on extensive incomplete lifecycle data, effectively distinguishing similar and dissimilar degradation patterns. Finally, the model undergoes fine-tuning in a supervised phase using complete lifecycle data. To further improve prediction accuracy, we employ a channel-independent transformer network, while RUL estimation is accomplished by integrating multi-source features through a multi-layer perceptron (MLP) layer. The proposed method is rigorously evaluated on both the IEEE PHM 2012 and XJTU-SY datasets and benchmarked against state-of-the-art approaches. Experimental results demonstrate that the proposed algorithm effectively captures the complex dynamic behavior of rolling bearings, achieving high accuracy and generalization in RUL prediction.