Vib-Mixup: augmentation of vibration signals for fault classification in industrial assets

Recent advancements in vibration sensors and machine learning have facilitated large-scale asset health monitoring in industrial machines. Predicting failures of rolling bearings in rotating machinery can significantly help decrease the risk of unforeseen machine downtime improving productivity. Bearing failure classification can be tackled by learning representations using deep neural networks. Representation learning with limited data in vision and language has been tackled with data augmentation and regularization. However there is lack of consensus on the extensibility of these approaches for vibration signals used for bearing failure prediction. In this work, we propose a mixup of vibration signals (Vib-Mixup) and its variants that uses convex combination of vibration signals or their representations, along with their corresponding failure label combinations for training bearing failure prediction models. Our results using publicly available bearing failure data show that, variants of mixup is a useful strategy with stronger performance over baselines, especially in smaller training data regime.