Joint autoencoder-regressor deep neural network for remaining useful life prediction

Abstract

Ubiquitous availability of IoT technologies allows processing of large amounts of data to improve prognostics tasks in industrial applications. One such important task of prognostics is the prediction of remaining useful life of a system from past performance data. In practice, although failure points are pinpointed, the actual start of degradation is not necessarily available, but usually modeled simply with a linear degradation assumption. In this paper we present a data-driven approach to remaining useful life prediction using joint autoencoder-regression network, a deep neural network model incorporating a convolutional neural network autoencoder and a long-short term memory network regressor trained end-to-end. We also present a new fault detection-based approach to modeling remaining useful life degradation. This model allows a better estimate of the start and progress of equipment degradation ending with a failure. We demonstrate the effectiveness of the proposed algorithms on two datasets. The first one is C-MAPSS frequently used as a benchmark among prognostic researchers. The second one is PHME20, a recent prognostic dataset from a prognostics competition. These experiments show that the proposed algorithms are capable of predicting remaining useful life as good as the state of art methods. The results also show that fault detection-based labeling outperforms linear labeling.

Keywords

• Predictive maintenance
• Remaining useful life
• Fault detection
• Deep learning
• C-MAPSS