Self-supervised domain adaptation for machinery remaining useful life prediction

Remaining useful life (RUL) prediction presents one of the most crucial tasks in modern machinery prognostics and health management systems. As a powerful data-driven solution, deep learning has shown its promising potential in accurately predicting the RUL based on historical condition monitoring data. However, deep learning-based methods typically require the training and test data to be drawn from the same distribution or domain, which is usually not the case in real-world application scenarios. Unsupervised domain adaptation (UDA) methods have been proposed to address this domain shift problem, but most of them focus only on learning domain-invariant feature representations while forcing the prediction error to be low on the source labeled data. Empirical observations have shown that this kind of domain adaptation is insufficient to guarantee good generalization in the target domain. To overcome this limitation, we propose a novel self-supervised domain adaptation (SSDA) framework that additionally incorporates the intrinsic information of the target domain data into the domain adaptation process without the need for its RUL labels. We developed a dual Siamese network-based training pipeline that enables the optimization for the self-supervised task in both the source and target domains to be realized jointly in conjunction with the base UDA objectives. Evaluation results from extensive experiments on the benchmark C-MAPSS dataset of aircraft turbofan engines show the superiority of our proposed framework over other state-of-the-art methods. On average, we achieve an improvement of 20.1% and 51.2% on two different performance metrics.