Markovian and Non-Markovian sensitivity enhancing transformations for process monitoring

- Process diagnosis requires causal information to be successful completed.
- Current process monitoring methods are based on acausal models.
- We propose a plug-in approach that integrates causal information into standard process monitoring.
- Smearing-out effect is reduced without compromising fault detection and diagnosis.
- Results obtained recommend the use of the static Non-Markovian SET as pre-processing for the Hotelling's T2 methodology.

Process monitoring is a key activity in modern industrial processes. Even though abnormality detection can be rather effectively done with resort to acausal correlation models of the variables normal operating conditions associations, fault diagnosis and troubleshooting do require causal information. In this article, we propose a new plug-in approach that brings the causal network structure into a classical monitoring scheme based on the Hotelling's T2 methodology. The modular plug-in nature associated to a well-known monitoring scheme aims at facilitating the access to the benefits of using more information about the system structure in fault analysis and diagnosis. The pre-processing module consists of a Sensitivity Enhancing Transformation (SET) that incorporates the network structure inferred from normal operation data, which has recently conducted to significant improvements for monitoring the correlation structure of industrial processes. Additionally, we consider both Markovian and Non-Markovian network structures in the development of the SET. The proposed methodology was tested with two simulated case studies (a CSTR and the Tennessee Eastman benchmark) and compared with several alternative approaches. The results obtained recommend the use of the static Non-Markovian SET as pre-processing for the Hotelling's T2 methodology.