Non-linear Analysis Dictionary Learning with Orthogonal Constraint for Fault Detection

Deep learning techniques use non-linear modifications to learn an abstract representation of the original data. It is possible to determine exceedingly complex functions using such non-linear transformation modules. Neural network-based representation learning has recently been adopted vastly for chemical process monitoring, fault detection and diagnosis problems. But these systems and methods lack consideration of orthogonal constraints which is crucial to maximize the detection performance. Moreover, independent transforms are required to be learnt because different faults need different parameters. To deal with these challenges, this research suggests an innovative analysis dictionary learning with orthogonal constraints-based method where the analysis dictionary and features are learned in such a way that analysis dictionary analyses the data to produce coefficients/features. We also present its kernelized version to account for non-linearity in the data. The experimental results demonstrate the effectiveness of proposed approach on the Tennessee Eastman (TE) benchmark process based upon the missed detection rate (MDR) and false alarm rate (FAR) using hotelling T² and the squared prediction error (SPE) statistics with confidence limit approximated by kernel density estimation (KDE). The proposed methods outperform SOTA methods by considerable margin for most fault types.