Sequential Information Bottleneck Network for RUL Prediction

This work studies the problem of predicting remaining useful life in complex systems. We propose a generic algorithm framework, sequential information bottleneck network(SIBN), which leverages the information criterion to preserve the most relevant information in the learning process. We derive the variational low bound of the information criterion when the data are in the sequential form. By using reparameterization trick, the sequential variational low bound can be optimized using standard stochastic gradient methods, e.g. SGD or Adagrad. The performance of the proposed framework SIBN is investigated on two benchmark datasets: the turbofan dataset and the ball bearing dataset. Experimental results show that our SIBN framework can accelerate network convergence and improve the prediction performance.