Multiple faults detection using artificial neural network

Abstract

This thesis investigated issues on the development of efficient fault detection scheme for detection of single and multiple faults due to sensor failure and leakage in the process stream. The proposed scheme consisted of two stage mechanism constructed using artificial neural network (ANN). The first stage was a process estimator that was designed to estimate the normal and unfaulty behaviour of the plant. In order to produce reasonably accurate estimation without including the history data of the output, two types of model have been studied. A group of multi input single output (MISO) Elman network and a multi input multi output (MIMO) Feedforward network have been used, and results revealed that MISO model had better generalisation ability compared to MIMO model. The difference between the actual plant signal and this estimated ‘normal’ plant behaviour, termed as residual was fed to the second stage for fault classification. In the development of fault classifiers, the MISO models had been proven to be better than MIMO model. The effect of adding input with time delayed signals to the network had also been studied. In both cases, successful implementations were obtained. Finally, the proposed fault detection scheme was applied for detection of sensor faults and stream leakage in the Precut column of a fatty acid fractionation plant. The proposed scheme was successful in detecting both single and multiple faults cases imposed to the process. The strategy was also successful in detecting leakage in the process stream even when the percentage of the leakage was as little as 0.1%. The results obtained in this work proved the potential of neural network in detecting multiple faults and leakage in chemical process plant.