Failure Prediction Using Gated Recurrent Unit and Autoencoder in Complex Manufacturing Process

Abstract

Big loss is caused by the failures in complex manufacturing process or in a production line. The design of the efficient and effective failure detection and prediction algorithms is the key for reducing the loss, and more and more algorithms rely on advanced machine learning technologies. The design of failure detection and prediction algorithms is however particularly challenging due to the high dimensionality, extremely imbalanced classes and the non-stationary distribution of the multivariate time series. For multivariate time series in real complex manufacturing process, it's really hard to decide whether the variable is dependent or independent because there is always variation along the production line. In this study, a novel failure prediction approach which combines gated recurrent unit and autoencoder is designed to improve the performance of imbalanced learning. The failure prediction algorithm is applied in a real pulp and paper mill to detect and predict the sheet break during the production. The results show that the proposed method can perform better than other related work.