Enhanced stochastic recurrent hybrid model for RUL Predictions via Semi-supervised learning

•A semi-supervised model is proposed to address the challenge of scarcity of labeled data.•Both the commonalities and individualities of degradation behaviors are characterized.•An extra future sensor data prediction task is proposed to conduct HIs meaningfully.•The effectiveness of the proposed method is verified by comparative experiments. Deep learning (DL) methods based on semi-supervised learning (SSL) have risen in popularity to achieve accurate remaining useful life (RUL) predictions when the volume of labeled sensor data is limited. The key to the method performance is the extraction of meaningful latent variables which can be served as health indicators (HIs). In this work, an enhanced stochastic recurrent hybrid model (ESRHM) is proposed through multi-task learning of RUL prediction, sensor data generation, and future sensor data prediction tasks. The extracted time-dependent HIs can contain both deterministic and stochastic information to characterize both the commonalities and individualities of different degradation behaviors via latent variables sharing. The proposed ESRHM is evaluated on the C-MAPSS and the lithium-ion batteries datasets to demonstrate its effectiveness in HIs construction and RUL prediction when coping with limited volume of labeled data.