Parzen window distribution as new membership function for ANFIS algorithm- Application to a distillation column faults prediction

The distillation column is one of the most important unit operations used in the chemical engineering. The continuous distillation process is largely used in many applications such as petrochemical production, natural gas processing, and petroleum refineries, and many others. Corrective maintenance of the chemical reactors represents a consequential problem because it is very costly and it disrupts production for long periods of time. In addition, most of the time, this may lead to harmful effects and disastrous results. The most common solution has been to rely on preventive maintenance. Unfortunately, this has been both expensive and inadequate. Therefore, the optimal solution is to resort to predictive maintenance that involves the design of a pre-crash control system and a higher ex-ante understanding of the future path of the reactor. This research paper aims to propose the Adaptive Neuro Fuzzy Inference System (ANFIS) as a superior technique that can forecast the future path of the distillation column system. In addition, this paper will propose Parzen windows distribution as a new membership function in order to improve ANFIS performance either by reducing consumption time and making processing closer to real-time application, or by minimizing the root means square error (RMSE) between real and predictive data. This methodology was tested on real experimental data obtained from a distillation column with the aim of predicting failures that may possibly occur during the automated continuous distillation process. A comparative study was necessary in order to properly select the superior membership function that can be used for the ANFIS algorithm when ANFIS is applied to the distillation column data. Results demonstrated the importance of the proposed technique since it proved to be highly successful in terms of reducing time consumed. Additionally, Parzen windows had the smallest RMSE for many signals in both normal and degraded modes.