Multi-Class Heteroscedastic Linear Dimensionality Reduction Scheme for Diagnosing Process Faults

Dimensionality reduction is an important factor in fault diagnosis, when dealing with a high-dimensional feature space, since it decreases the computational burden and the model complexity. This paper focuses on the development and comparison of several state-of-the-art linear dimensionality reduction techniques to provide discriminant features for the process fault diagnosis. These techniques, including heteroscedastic discriminant analysis, Fisher's discriminant analysis, Chernoff discriminant analysis and principal component analysis, can handle multi-class feature sets. The attained results show that the heteroscedastic variant outperforms other techniques both in terms of performance measures and speed.