Dynamical Variational Autoencoders for Estimating the Remaining Useful Life of Machinery

The purpose of this work is to show the effectiveness of using generative models, specifically Dynamical Variational Autoencoders (DVAEs) for Remaining Useful Life (RUL) estimation. Many deep learning methods simply output point estimates of the RUL, generative models have the benefit of being optimized to learn an underlying probability distribution, this allows for uncertainty quantification. This work showcases how to construct a conditional DVAE to learn how to sample from p(y1:T |x1:T ), where y1:T is a sequence of RUL estimates and x1:T are a sequence of sensor signals. It is shown why noncausal sensors are important when constructing this conditional model and how one can achieve state of the art results using DVAEs. Because the DVAE is a generative model and learns to sample from p(y1:T |x1:T ), one can also quantify the uncertainty of the RUL estimates directly with this model. This is tested on NASA’s CMAPSS turbofan engine dataset and an open dataset from a dust filter experimental setup and it is demonstrated it is able to achieve state of the art results.