A Sequentially-Adaptive Deep Variational Model for Multirate Process Anomaly Detection

Deep learning based process monitoring methods are attracting increasing research attention in recent years, which generally assume that the process variables are uniformly sampled. In practice, however, the process data are generally collected at multiple different rates, resulting in structurally-incomplete training data. Under such circumstances, how to build effective deep models to fully mine the multirate sampled data has become a constraint in achieving better process monitoring performance. In this paper, a sequentially-adaptive deep variational model is designed in which the knowledge that existed in variables with different rates is comprehensively extracted through deep generative neural networks. The multirate samples are first divided into multiple data blocks in which each block is collected at a uniform rate. A deep generative model is then constructed to model the uncertain data distribution and extract probabilistic feature representations considering the slowness principle. To restrain the small data problem in the blocks with slow rates, a sequentially-adaptation strategy is designed to adapt the knowledge from the fast blocks with sufficient training data and enhance the overall modeling performance. The effectiveness is demonstrated through a real-world industrial thermal power plant case.