A Transductive Learning Approach to Process Fault Identification

Abstract

The problem of fault identification is considered using recent developments in machine learning that allow the use of unlabeled data to optimally define decision boundaries that separate data belonging to different categories. Traditionally, fault identification uses either hardware redundancy or software redundancy for trouble-shooting the source of faults in a system. Unfortunately, this imposes a data redundancy cost on such systems. Instead of performing model inversion that requires an accurate model, in transduction estimates of the values of a function at specified points are required, instead of learning a general rule on the entire input domain. Transductive learning is motivated from similar arguments underlying state-of-the-art classification and regression methods such as support vector machines. However, transduction is more fundamental as it is a step used in proving learning error bounds in classical statistical learning theory. Use of transduction allows a flexible ordering of the classes of functions from which a model is selected and, therefore, the error bounds are provably tight. The potential of the proposed framework is assessed using data from metallurgical process systems. It is shown that for higher dimensional and large multiclass systems, the proposed framework gives betterperformances with respect to classification error minimization.