Unsupervised deep learning for gear health monitoring

Abstract

Deep learning has revolutionized many fields in recent yearsby replacing expert-designed, handcrafted features withlearned representations. Gear health monitoring is a fieldwhere expert-designed features are heavily used for predictivemodeling. This paper investigates how unsuperviseddeep learning can be applied to gear health monitoring tomake predictions on low frequency scales using high frequencydata given small, sparsely labeled data sets. Deepconvolutional autoencoders are trained and used to generatelearned features. The learned features are compared withrelevant handcrafted features via their performance in trainingmachine learning models to predict discrete gear fatiguestates. The learned features performed poorly against thehandcrafted features, however models trained on featuresets tended to outperform those exclusively trained on handcraftedfeatures. The top performing model was a multi-layerperceptron trained on both feature sets that leveraged theability of the condition indicators to represent healthy andfailure states and the ability of the learned features to representthe intermediate worn state. This work demonstrates thatunsupervised deep learning techniques can be used to bolsterthe performance of handcrafted features in small, sparselylabeled data sets in gear health monitoring.