Simultaneous Data Reconciliation and Gross Error Detection Based on Denoising Autoencoder

Hong Yang, Yuan Yuan, Zhisheng Chen, Xiaodong Xu, Biao Luo

Abstract: Industrial process data are usually affected by random and gross errors leading to deviation from the true value and violation of process constraints. Traditional data reconciliation methods rely on an accurate process model, which can lead to incorrect reconciliation results if the model is not accurate enough. This article proposes a deep learning approach based on denoising autoencoder (DAE) for dealing with simultaneous data reconciliation and gross error detection (SDR & GED) in steady-state systems without directly using the process models/constraints. Reconciled values that satisfy the process constraints can be gradually obtained by iterative training. In addition, a pretraining method is proposed to accelerate the training procedure. The proposed DAE framework is equipped with a 1-D convolutional neural network, which can reduce the model complexity. Furthermore, DAE is capable of handling a variety of noises so that data with different types of noise distributions can still be reconciled effectively. Finally, loop analysis and process constraints’ construction for the hot section of an olefin plant are performed, and the effectiveness of the proposed method is verified by the study of this system.