A Semi-Supervised Remaining Useful Life Prediction Approach Based on Self-Attention Mechanism-Sequential Variational Autoencoder

With the accelerating pace of smart technologies and digital transformation across industrial domains, ensuring the reliability of sophisticated systems has become fundamentally important for manufacturing productivity [1] - [4]. Nevertheless, the complex systems inevitably become degradation due to operational conditions, environmental changes, and material aging. As a logical consequence, accurate prediction of remaining useful life (RUL) in complex systems represents a fundamental component of effective maintenance strategies and asset management [5]. Accurate RUL prediction can help users plan maintenance resource allocation [6]. Furthermore, the unnecessary downtime and repair costs are decreased, thereby promoting production efficiency. The primary motivations for this research emerge from the following factors: • The degradation data of complex systems have strong temporal characteristics. Consequently, how to focus on key parts of the time window, and make full use of the temporal information requires further research [7]. • Conventional deep networks are usually regarded as black box model, the meaning of whose internal latent spaces is difficult to interpret. Furthermore, most of the RUL prediction approaches are point estimation, which requires in-depth work on uncertainty description [8]. • Conventional data-driven approaches demand massive labeled samples for supervised model training. How to effectively utilize unlabeled degradation data from a semi-supervised perspective requires further work [9].