"Using Domain Adaptation to Counter Distribution Shift Between Training And Application Domain"
Using Domain Adaptation to Counter Distribution Shift Between Training And Application Domain
Sprache des Titels:
Over the last years, the popularity of artificial intelligence methods used in industrial applications has been steadily increasing. Especially neural networks and deep learning achieve promising results. However, this kind of machine learning model requires a large amount of labeled data to train the model successfully. In the context of industrial applications, collecting labeled data can be laborious, costly, or even unfeasible in some scenarios. This thesis studies the application of Domain Adaptation, a possibility to transfer knowledge from a source domain to a related target domain, where no labeled data from the target domain is needed. The task under investigation is automatic error detection in electrical drives. In this context, Domain Adaptation is used to avoid the costly process of manipulating real motors to provoke a motor fault and collect labeled data. Instead, the neural network is trained on easy to produce, simulated fault data and adapted using real-world measurements from a motor running in its nominal state. The specific category of domain adaptation methods used is feature-divergence-based methods, which try to extract common factors between simulations and measurements by minimizing a divergence measure between the internal feature representations inside the neural network. An introductory example, dealing with recognizing different types of handwritten digits, shows the effect of successful domain adaptation. The acquired knowledge is transferred and successfully applied to increase predictive performance on the electrical motor fault detection task. Several domain adaptation algorithms are shown to outperform a baseline model that does not use any domain adaptation mechanism. Furthermore, this thesis conducts a visual and quantitative analysis to answer why domain adaptation can increase fault detection performance.