Anomaly detection and prediction of sensors faults in a refinery using data mining techniques and fuzzy logic

Abstract

Like all manufacturing companies, refineries use many sensors to monitor and control the process of refining, therefore it is very crucial to detect any sensor faults or anomalies as early as possible, and to be able to replace or repair a sensor well in advance of any fault. Objective of this paper is to present a method for detecting anomalies in a sensor data, as well as to predict next occurance of a sensor failure. Data mining techniques to detect anomaly in sensor data and predict the occurrence of next faulty event were introduced. For anomaly detection, this research used MATLAB’s fuzzy logic toolbox tools to find clusters which uses subtractive fuzzy clustering algorithm and generates a model, a Sugeno-type fuzzy inference system. The same toolbox was used to evaluate the model with a promising result. To predict sensor fault, the original time series were used to create a new ‘derived time series’. Two prediction models known as ‘auto regressive integrated moving average’ and ‘autoregressive tree models’ were used against the new time series to predict next occurrence of sensor failure. The results oft hese models were compared. The model developed and introduced in this paper serves as an additional tool, which helps not only engineers and operators of oil refineries, but also other engineers of other disciplines to work more efficiently. Key words: Data mining, derived time series, sensor fault detection and prediction, fuzzy clustring, machine learning, oil refinery.