Self-supervised Learning for Efficient Remaining Useful Life Prediction

Canonical deep learning-based Remaining Useful Life (RUL) prediction relies on supervised learning methods which in turn requires large data sets of run-to-failure to ensure model performance. In a class cases, is difficult collect practice as it may be expensive and unsafe operate assets until failure. As such, there need leverage that are not but still contain some measurable, thus learnable, degradation signal. this paper, we propose utilizing self-supervised pretraining step learn representations the will enable efficient training downstream task RUL prediction. The chosen time series sequence ordering, involves constructing tuples each consisting $n$ sequences sampled from reordered with probability $p$. Subsequently, classifier trained resulting binary classification task; distinguishing between correctly ordered shuffled tuples. classifier's weights then transferred RUL-model fine-tuned using data. We show proposed scheme can retain performance when fraction full set. addition, indications enhance even To conduct our experiments, use set simulated turbofan jet engines.