Investigating merits of deep self-supervised learning on a fatigue RUL prognostics application

Many critical applications, such as in aerospace, have rare failures, which poses a dilemma for many data-driven prognostics and health management (PHM) frameworks: even though there is a lot of raw sensor data available thanks to the big data revolution, remaining useful life (RUL) prognostics still faces a data scarcity issue, because most data-driven prognostics algorithms need run-to-failure data for training, which is hard to obtain. In this work, we seek to solve this dilemma by exploring a way to exploit the raw sensor data, which, even though does not have a RUL associated with it, typically still contains relevant degradation data but not all the way to failure. To achieve this, we use the self-supervised learning paradigm, which has been widely studied in the machine learning community in recent years. The aim of self-supervised learning is to use large amounts of unlabeled data (i.e., raw sensor data here) but only small amounts of labeled data (i.e., raw sensor data + associated RUL) to accomplish a given task, thus improving training data efficiency. In this work, we apply such a self-supervised learning paradigm to a fatigue RUL estimation problem in aluminum alloy panels, typical of aeronautical applications, based on strain data. We consider the self-supervised learning problem in an autoregressive setting where we have an unsupervised learning phase aimed at predicting the raw strain data at the next time step, followed by a supervised fine-tuning phase, which uses labeled run-to-failure data to fine-tune the RUL predictions. The results show that the investigated approach can significantly improve the accuracy compared to non-self-supervised approaches for the same number of run-to-failure data, or alternatively, requires much less run-to-failure data to achieve comparable accuracy.