Evaluation of nonlinear scaling and transformation for nonlinear process fault detection

On-line fault detection of nonlinear processes involving dynamic dependencies and similar/ overlapping fault signatures, is a fairly challenging and daunting task. Early detection and unambiguous diagnosis require that the monitoring approaches are able to deal with these daunting features. This paper compares two broad multivariate statistical approaches proposed in the literature for the detection task: (i) nonlinear transformations to generate linear maps and their dynamic variants in high dimensional feature space, as exemplified by kernel principal component analysis and dynamic kernel principal component analysis, and (ii) nonlinear scaling of the data to promote better self aggregation of data classes and hence improved discrimination, as exemplified by correspondence analysis. Using the Tennessee Eastman benchmark problem, we compare the performance of the above methods with respect to the known metrics such as detection delays, false alarm rates (Type I error) and missed detection rates (Type II error). As well, we compare the methods on the basis of computational cost and provide summarizing remarks on the ease of deployment and maintenance of such approaches for plant-wide fault detection of complex chemical processes.