Comparison of Statistical and Machine Learning Approaches for Fault Detection in a Continuous Catalytic Regeneration Unit

Catalytic reforming process is one of the most important processes in oil refineries that produce high octane number gasoline. However, its operation often leads to high coke formation on the catalyst particles, and subsequent degradation in reactivity. Continuous Catalytic Regeneration (CCR) system regenerates the catalyst continuously and helps maintain acceptable productivity goals. CCR accomplishes this by enabling the reformer to perform on lower coke ratio which is crucial for higher reformate yield. Any upset in the process operation directly affects the product formation and has adverse effects on economic yields. Hence, process monitoring using data-driven fault detection techniques is deemed to be crucial for ensuring reliable operating conditions.

Data-driven fault detection and diagnosis (FDD) methods, which can be divided into two categories, statistical and machine learning (ML) approaches, have become popular in both research and applications. In this study, we will present our key findings of process monitoring using both statistical and ML techniques. For the statistical approach, the commonly used method based on principal component analysis (PCA) will be employed and will be compared with dynamic PCA extension in which the serial dependence in measured variables is considered. For further comparison, on the same data, autoencoders, which is a popular unsupervised ML technique, will be employed for fault detection on real industrial operation data from the CCR process. This comparative study is expected to result in an effective fault detection approach for an operating plant that will also substantially contribute to process sustainability and reliability.