Batch Process Monitoring with Attention-Based Two-Dimensional Stacked Graph Convolutional Autoencoder

In batch processes, the efficacy of statistical control is reportedly sensitive to the dynamics and nonlinearity found in the batch data, which can hamper the valid feature extraction for statistical analysis. Normally, a two-way window is applied to aggregate the samples across the batch and time direction to assuage this challenge. However, samples in the window are generally processed on par with each other such that irrelevant features can cause redundancy and make window size critical to be tediously tune during the training phase. In this paper, a linear correlation aided method is proposed to first determine the effective size of a regular-shaped sliding window, and, a Euclidean distance-based measure is used to construct a graph prior to the phase of feature extraction. With these efforts, we presented the Attention-based Two-dimensional Stacked Graph Convolutional Autoencoder (2D-ASGCA) for enhanced learning in the Statistical Process Control (SPC) of batch processes. Moreover, a technique called Deep Reconstruction-based Contribution (DRBC) chart has been applied to perform the variable-wise root cause diagnosis, when a fault is signaled. Finally, two case studies are provided to demonstrate the model validity.