Fault monitoring for chemical processes using neighborhood embedding discriminative analysis

The importance of chemical process safety and the availability of abundant samples keep popularizing the wider application of data-driven fault monitoring techniques. With a goal of efficiently discovering the inconsistency between the online monitored sample and the normal samples, a novel fault monitoring algorithm called neighborhood embedding discriminative analysis (NEDA) is proposed, which can adaptively provide different latent feature generating mechanisms for different monitored samples so that the inherited inconsistency could be uncovered in a timely manner. Instead of extracting representative features from a dataset only given from the normal operating condition, the objective function designed for the NEDA algorithm additionally takes the online monitoring sample into account, and then timely generates but only one projecting vector to point out the specific inconsistency for the corresponding monitored sample. The NEDA algorithm aims to figure out a discriminative projection so that the neighborhood embedding error (NEE) corresponding to the online monitored sample could be maximized, while the NEE associated with the normal samples is minimized. Furthermore, the corresponding NEE for the monitored sample of current interest is employed as the indicator for fault monitoring purposes. As demonstrated through comparisons, the salient performance achieved by the proposed NEDA-based fault monitoring method in monitoring static as well as dynamic processes can be always guaranteed.

Automated summary

The NEDA algorithm aims to discover inconsistencies in monitored samples by providing different latent feature generating mechanisms, with the goal of maximizing the NEE for the monitored sample and minimizing the NEE for normal samples. The corresponding NEE is used as an indicator for fault monitoring purposes. This algorithm demonstrates significant performance in monitoring both static and dynamic processes.