A comparison of PCA- and energy-based fault detection and isolation in a physical heated twotank process

Abstract

The demands for petrochemical (PC) plant productivity and safety are increasing and with it the need for effective fault detection and isolation (FDI) methods. One of the more well-known and widely tested methods is principal component analysis (PCA), which uses process history as its basis for FDI. Another viable and more recently developed technique, namely energy graph-based visualisation (EGBV), has shown promising FDI results in complex processes. Comparison studies of FDI methods are a frequent occurrence in literature and with good reason - they provide valuable insight into the strengths and weaknesses of each. For this reason, a valid research question may be asked concerning EGBV and PCA: How does EGBV compare to the more established PCA FDI method? This work aims to answer this question, using a modelled and practical heated two-tank system as the benchmark process. To address this research question and conduct a fair comparison, the following objectives were followed. First, a literature survey was conducted to clarify the relevant FDI terminology and theory. Subsequently, an existing Simulink® model of the process was modified to match the practical plant’s performance, so that it may act as a reference system for the comparison. This system was then used to verify the application of EGBV and PCA, followed by their validation using the practical process data. Finally, an objective comparison was made between EGBV and PCA using the results from the model and the practical system. The literature survey revealed three key FDI performance metrics that allowed for an

objective comparison between the methods, namely robustness, sensitivity and promptness. The Simulink® model of the heated two-tank process was successfully adapted to match the practical plant specifications and to include twenty process faults varying in type, location and magnitude. The model modifications were verified and subsequently validated using the practical process’s results. The model closely resembled the practical plant’s performance. The theory of EGBV and PCA was explained and applied to the model to verify its implementation. Both methods were subsequently validated using practical results. EGBV successfully detected fifteen out of the twenty faults in the practical process, while PCA detected only nine. PCA was found to be the more robust method, exhibiting fewer false alarms than EGBV. The EGBV method, however, responds quicker to faults and is a more sensitive method, depending on the chosen EGBV detection philosophy. EGBV is also found to have superior fault isolation ability compared to PCA. Additionally, EGBV exhibits more ideal FDI method attributes than PCA. Ultimately, the preferred FDI method will depend on the priorities of the plant practitioner.