Mamba-attention: A self-supervised framework for efficient remaining useful life prediction

Accurate remaining useful life (RUL) prediction provides significant economic, technical, and safety benefits by forecasting the lifespans of industrial equipment. Current state-of-the-art (SOTA) RUL prediction models predominantly rely on recurrent neural networks (RNNs) and attention-based models. While RNNs are adept at capturing temporal dependencies, they struggle with issues such as vanishing gradients and exhibit inefficiency in processing long sequences. Meanwhile, attention-based models improve the capture of global context but are limited by their high computational complexity. The proposed Mamba-Attention model addresses these limitations by integrating a selective state space model with attention mechanisms, achieving SOTA performance with faster training times on C-MAPSS dataset. Industrial equipment is frequently replaced prematurely, and obtaining labeled data is challenging due to cost and time constraints. To overcome these difficulties, this research introduces a novel framework that leverages unlabeled data via self-supervised learning approaches, including temporal loss, N-tuplet loss, and pseudo-label loss. By pre-training with self-supervised losses and fine-tuning on minimal labeled data in few-shot scenarios, the proposed approach remains effective in producing accurate RUL predictions, even under limited supervision, as demonstrated on the C-MAPSS and CALCE datasets. This reduces the reliance on costly labeled datasets and offers practical solutions for industrial applications.