Domain Adaptation via Simulation Parameter and Data Perturbation for Predictive Maintenance

Conventional data-driven predictive maintenance (PdM) solutions learn from samples of run-to-failures (R2F) to estimate the remaining useful life of an asset. In practice, such samples are scarce or completely missing. Simulation models can be oftentimes used to generate R2F samples as a replacement. However, due to the complexity of the assets, creating realistic simulation models is tedious, or even impossible. Thus generated R2F data cannot be used to create reliable PdM models as they are highly sensitive to noises in the sensors or small deviations in system working condition. To address this, we present a new concept of simulation data generation based on supervised domain adaptation for a regression problem where the remaining useful life (RUL) or the health index (HI) of the system is predicted. Apart from input and output domain shift, given the changes in the dominant failing component and its degradation process, the function mapping sensor readings to RUL and/or HI is also prone to changes and thus is a random process itself. Therefore, we aim to generate R2F training data from different working conditions and possible failure types using parameter randomization in the simulation model. By sampling from various configurations within simulation model’s parameter space, we ensure that the trained data-driven PdM model’s performance is not impacted by the initial conditions and/or the changes in the degradation of the system’s condition indicators. Our results indicate that the model is robust to signal reading manipulation and showcases a more spread-out feature importance across a wider range of sensor readings for making predictions. We also demonstrate its applicability on the real-world factory physical system whilst our models were mainly trained using generated data.