Towards Geometry-Preserving Domain Adaptation for Fault Identification

In most industries, the working conditions of equipment vary significantly from one site to another, time a year and so on. This variation poses severe challenge for data-driven fault identification methods: it introduces change in data distribution. contradicts underlying assumption machine learning methods, namely that training test samples follow same Domain Adaptation (DA) methods aim address this problem by minimizing distribution distance between (source) (target) samples. However, area predictive maintenance, idea is complicated fact different classes – categories also across domains. Most state-of-the-art DA assume target domain complete, i.e., we have access examples all possible or faulty during adaptation. reality, often very difficult guarantee. Therefore, there need adaptation method able align source domains even cases having an incomplete set data. paper presents our work progress as propose approach such setting based on maintaining geometry information way, model can capture relationships preserve them constructed domain-invariant feature space, situations where some are entirely missing. examines using artificial sets demonstrate effectiveness geometry-preserving transformation. We started investigations real-world maintenance datasets, CWRU.