Comparative study of boosting and bagging based methods for fault detection in a chemical process

This paper presents a comparative study of boosting and bagging based methods for fault detection and isolation in a chemical process. Boosting trains a number of weak classifiers sequentially on differently weighted versions of training samples, whereas, bagging trains a number of classifiers independently on bootstrap samples. The joint prediction of the ensemble of classifiers has relatively high accuracy on the training set. In the case of boosting, two of the well-known algorithms i.e., AdaBoost and gradient boosting were used, whereas, the random forest was used in the case of bagging. The decision tree was used as a base classifier in all three cases. The effectiveness of these techniques and their comparisons with the conventional technique was demonstrated by applying them to a benchmark Tennessee Eastman process.