A contrastive semi-supervised remaining useful life prediction method with incomplete life histories on turbofan

Abstract

With the emergence of deep learning, its technique has been widely used in remaining useful life (RUL) prediction for turbofans. Due to its complex nature, RUL prediction poses significant challenges such as incomplete life data and the labor-intensive process of data labeling. To address the issue, many studies have turned to semi-supervised learning. However, most of these studies have utilized unlabeled data solely from the complete fault history, overlooking the overhang history, which leads to a notable decrease in prediction accuracy. To tackle this problem, this paper proposes a novel methodology that combines contrast learning with variational autoencoders (VAE). Through a symmetric structure, the proposed approach effectively learns the similarity between labeled and unlabeled data, thereby enhancing the prediction accuracy of variational autoencoders. Additionally, the K-nearest neighbor (KNN) regression algorithm is employed to label the unlabeled data, and screening rules are established to eliminate data with poor labeling effects. The effectiveness and stability of the proposed method are rigorously evaluated through numerous comparative experiments.