Railway wheels are safety critical components and are a major cost driver for maintenance. Due to their criticality, they are tightly monitored by different condition monitoring devices. This opens the opportunity to develop algorithms that compose an end-to-end solution to a purely data-driven digital twin of railway wheels. However, challenges arise when dealing with real condition monitoring data. For example, changing in-service factors and operating conditions can cause novel fluctuations in the data that are not related to the asset's health. These can harm the performance of any data-driven model and ultimately, the digital twin.  Furthermore, faults occur rarely in operating systems. A training dataset typically does not cover any or only a subset of all possible fault types. Therefore, the implemented models need to be sensitive to novel variations relating to changing health conditions in order to detect newly emerging fault types. Simultaneously, invariance to variations to non-informative causal factors is required. We propose contrastive learning to learn a feature representation that satisfies both objectives. The conducted experiments on a benchmark dataset show that the learned feature space is  (1) invariant to changing operating conditions while also being (2) suited for the detection of novel fault types and hence, presents a perfect basis for stable fault detection, diagnostics and prognostics.

Download the slides for this talk.Download ( PDF, 3231.31 MB)