Fault detection and identification using Bayesian recurrent neural networks

摘要: Abstract In the processing and manufacturing industries, there has been a large push to produce higher quality products ensure maximum efficiency of processes, which requires approaches effectively detect resolve disturbances optimal operations. While many types can be compensated by control system, it cannot handle some process disruptions. As such, is important develop monitoring systems identify those faults such that they quickly resolved operators. This article proposes novel probabilistic fault detection identification method adopts newly developed deep learning approach using Bayesian recurrent neural networks (BRNNs) with variational dropout. The BRNN model general complex nonlinear dynamics. Moreover, compared traditional statistic-based data-driven methods, proposed BRNN-based yields uncertainty estimates allow for simultaneous chemical direct identification, propagation analysis. performance demonstrated contrasted (dynamic) principal component analysis, widely applied in industry, benchmark Tennessee Eastman (TEP) real dataset.