A Study on Dual-task VAE with Weibull expansion for RUL estimation and application to Bearing data

Remaining Useful Life (RUL) is one of the most fundamental assessment measures in reliability engineering. Although several widely-used public datasets have information on failure times, they do not have information on RUL. As one of the simple methodologies, we can consider a linear degradation concerning the difference between the future failure time and the current time. However, since it is not always the case that such a linear degradation occurs, then we need to generate complex teacher labels incorporating the detection of changing points in the degradation states. In one of the previous studies, one proposed a failure prediction model that took into account the degradation states and achieved some improvement in prediction accuracy. However, this model sometimes does not perform well; for example, it cannot be adapted to specific tagging methods. Therefore we propose Dual-task VAE incorporating Weibull distribution as a prediction model which does not depend on the tagging methods. A numerical example with an actual data set is also presented to show the performance of our model