Intelligent fault diagnosis scheme for Plant-wide processes

Abstract

For fault detection and diagnosis in complex systems, model based methods are not very practical. Instead, data-driven techniques are widely used especially machine learning methods such as classification and k-mean clustering. However to our knowledge, hierarchical clustering technique based on KLD is not explored in this field. This study suggests a new approach to build an unsupervised model for fault diagnosis using hierarchical clustering with the statistical multivariate technique Kullback-Leibler divergence as an index to compute the dissimilarity degree between the different data distributions. These datasets are preprocessed through the principal component analysis (PCA) which allows to reduce the dimensionality and generates a principal and a residual subspace that can both be used to train our model which can be visualized in a dendrogram. In order to produce the optimal model, we set various CPVs to obtain different numbers of retained components, hence different models. The proposed method was applied to plant-wide Tennessee Eastman process to test the accuracy and the elapsed time of our algorithm. The results show the effectiveness of our optimal model with an accuracy of 86.36% of correct predictions. Keywords: Fault Diagnosis (FD), Machine Learning (ML), Hierarchical Clustering, Kullback- Leibler Divergence (KLD), Principal Component Analysis (PCA), Dissimilarity Index, Tennessee Eastman Process.