Image-based remaining useful life prediction through adaptation from simulation to experimental domain

Degradation profoundly affects the performance of industrial systems, necessitating operational safety prognostics. However, the availability of run-to-failure data is often limited, and labels in real-world scenarios are scarce. To address the challenge, this work utilizes the simulation domain to extract degradation knowledge and then adaptively transfers this knowledge to the experimental domain, aiming at estimating the remaining useful life (RUL). The relative RUL in the simulation domain is adopted, focusing on the degradation trend and avoiding the determination of absolute RUL. The feature disentanglement technique captures degradation-relevant features. To improve model performance, Bayesian optimization is introduced to search for optimal hyperparameters, and a two-task learning approach is designed to achieve the objectives of both domains. A few labeled experimental samples are used to adjust the predictor to appropriate scale. The case study on infrared degradation image streams validates the effectiveness of this domain adaptation scheme. Further analysis and discussions demonstrate the superiority of the model and the associated optimization strategy.