Simulation-assisted Training of Neural Networks for Condition Monitoring of Electrical Drives: Enhanced Domain Adaptation Methods

Abstract:
Over the past years, Deep Learning (DL) has become a topic of increasing popularity in industrial applications, including automatic fault detection. One major downside of DL is its requirement for a large amount of labeled data to solve a given task. Particularly in industrial fault detection problems, this data acquisition process can be very costly or even infeasible. An appealing alternative is to simulate real-world processes and obtain labeled data by running a simulation. However, a simulation often cannot cover all aspects of the real world, introducing artifacts and imprecision. Domain Adaptation, as a subcategory of Transfer Learning, provides a collection of algorithms that allow identification of discriminative, common signal patterns between simulations and the real world. This paper conducts a thorough investigation on the industrial application of electric motor fault detection using electrical signals. Specifically, the task is to detect errors in a block-commutated 280W motor given monitored phase currents. Given a DL classifier trained on simulated signals, the aim is to adapt it to real-world measurements given only measured signals from the motor running in its nominal (healthy) operating condition. Firstly, we review two techniques of preprocessing electrical signals to achieve a suitable input format for a Deep Neural Network. Secondly, we provide a collection of Domain Adaptation algorithm candidates alongside explanations. Finally, in our experiments, we show that Deep Domain Adaptation is indeed suited for this type of problem and demonstrate its effectiveness.