Comparing the Fault Diagnosis Performances of Single Neural Networks and Two Ensemble Neural Networks Based on the Boosting Methods

This work employs potential and distinct features of artificial neural networks, in particular the feed-forward multilayer structure, to achieve fault diagnosis in chemical plants. Artificial neural networks can automatically store information by learning from historical fault data without using any qualitative or quantitative model of the system. However, different limitations are encountered in utilizing a single neural network classifier in fault diagnosis of complex large-scale chemical processes. These limitations are mostly originated from the low reliability and also high generalization error of a single neural network classifier which may be faced in real applications. An attractive remedy for these issues is to use an ensemble of different neural network models instead of relying on a single one. In this work, boosting as a general well-known approach to improve the learning performance of learning algorithms by building ensemble models is investigated. In particular, an efficient boosting method called the Stage-wise Additive Modeling using a Multi-class Exponential Loss Function (SAMMELF) is utilized. This technique can effectively overcome the main limitations of the early boosting methods for fault diagnosis applications. The potential of the SAMMELF algorithm is proved through its application to the challenging diagnosis problem of the Tennessee-Eastman Process (TEP) where the whole sets of the predefined TEP faults are considered in the classification problem.