Multirate Nonlinear Process Fault Detection Based on Multiscale Hierarchical Variational Autoencoder

In industrial processes, data are frequently gathered at various sampling rates, influenced by factors such as the diverse characteristics of process variables and the use of different sensors. However, the fault detection model for multirate data faces challenges in handling irregular data, complex modeling, and poor generalization capabilities. To effectively use and regularize multirate data and enhance detection models, this article proposes a multiscale hierarchical fault detection (MsHFD) model. This innovative model integrates and reconstructs irregular multirate data into multiscale data, progressively advancing through different scales, and uses variational autoencoder (VAE) for nonlinear feature extraction. The hierarchical model not only preserves data change trends but also incrementally incorporates variable relationships, thereby improving detection efficiency and providing a foundation for faulty variable identification. Finally, comparative experiments with traditional linear models and nonprobabilistic deep models demonstrate that the proposed model in multiphase flow (MPF) process achieves superior fault detection performance.