Adaptive spatiotemporal neighborhood feature learning based on manifold regularization autoencoder for industrial process monitoring

Background:

Process monitoring systems are crucial for process safety in modern industries. However, existing methods typically extract spatial features from the data without considering the dynamic spatiotemporal information inherent in industrial process data, which affects the accuracy of monitoring.

Method:

This paper proposes an Adaptive Spatiotemporal Manifold Regularization Autoencoder (ASMRAE) for process monitoring which aims to extract features while preserving the spatiotemporal neighborhood structure. Specifically, an adaptive temporal neighborhood weight calculation method is designed to adjust the spatial structure within the temporal neighborhood, adapting to changing temporal information. The spatiotemporal neighborhood topology is described probabilistically to accommodate the uncertainty inherent in industrial process data. By minimizing the probability distribution of the spatiotemporal neighborhood structure between the raw data and the extracted features, the resulting features capture spatiotemporal information. Finally, $T^2$ and SPE statistics are constructed for monitoring purposes.

Significant Findings:

The experimental results show that capturing accurate spatiotemporal structures can enhance monitoring performance. The effectiveness of the proposed method is validated in the Vinyl Acetate Monomer process and Aluminum electrolysis process.