A missing manufacturing process data imputation framework for nonlinear dynamic soft sensor modeling and its application

Abstract

Data-driven soft sensors have been widely used in manufacturing processes for product quality prediction. However, in engineering practice, traditional linear soft sensors are considered to be insufficient when the manufacturing processes present high dimensionality as well as strong nonlinearity and dynamics. Moreover, due to the economic and technical limitations, missing data problem is widespread, which affects the generalization and accuracy of soft sensors. To overcome these issues, a missing manufacturing process data imputation framework is presented for nonlinear dynamic soft sensor modeling with the purpose of quality prediction. Specifically, the generate adversarial imputation network is introduced for data imputation, and the loss functions of discriminator and generator are reasonably designed. Subsequently, a mix gated unit combining bidirectional gated recurrent unit with bidirectional minimal gated unit is proposed for nonlinear dynamic soft sensor modeling aiming at improving the performance of quality prediction. Finally, the efficiency and flexibility of the proposed framework are demonstrated through a representative manufacturing process, the hot rolling process, in which the percentages of missing data are simulated from 20% to 80%. The ideal performance indicators, root mean square error and mean absolute error, have been obtained. It can be shown that the proposed framework can provide better prediction accuracy than the competitive methods in each scenario.