Remaining Useful Life Prediction on C-MAPSS Dataset via Joint Autoencoder-Regression Architecture

Abstract

The maintenance costs of industrial systems often exceed the initial investment cost. Predictive maintenance, one of the most effective methods in reducing overall maintenance costs, has become an area of interest for data-driven researchers after the increasing automation, monitoring capabilities and development techniques introduced with the new industrial revolution. In this study we introduce joint autoencoder-regression architecture for remaining useful life prediction, and demonstrate it on the NASA Turbofan Engine Degredation Dataset. The architecture incorporates InceptionTime networks for the autoencoder and short-long-term memory for the remaining useful life prediction. In the first stage, the models are trained and optimized using genetic algorithms, and then the models are fine-tuned with noise inducing and network pruning techniques. The results show that InceptionTime network-based joint autocode-regression architecture is competitive with the recent studies on the dataset, and that noise induced models show performance close to the state-of-the-art models.