TrajecNets: Online Failure Evolution Analysis in 2D Space

Abstract

We propose a novel Recurrent Neural Network (RNN) based autoencoder for embedding the run-to-failure time series sensor data in a 2D feature space. The embedding, extracted from the network, is in the form of a smooth trajectory, which represents the temporal evolution of data from healthy to failure states, hence the name TrajecNets. The visualizable 2D trajectory can be used directly for highly intuitive and interpretable health monitoring, which can in turn be used for Remaining Useful Life (RUL) estimation task, without compromising the performance. We also propose a novel unsupervised failure prediction methodology which uses the 2D trajectories and health curve of the time series to compute evolving failure mode probabilities. Together, the visualizable 2D trajectories and the interpretable failure mode probabilities, health curve and RUL are envisaged to provide system and maintenance engineers, insight into failure dynamics. Experiments on NASA CMAPSS Turbofan benchmark dataset showpromising results on degradation tracking, health monitoring, failure prediction and RUL estimation tasks.