Distribution independent threshold setting based on one-class support vector machine

In this paper, threshold setting issues for data-driven fault detection are addressed. It is state of the art that multivariate analysis based threshold setting schemes are widely applied, which generally require detailed knowledge about the distribution of the process data. The often used Hotelling's T-2, SPE threshold setting is based on the assumption of Gaussian distributed process data. In industrial applications, the distribution of data sets is often unknown or non-Gaussian. Alternatively, the fault detection is formulated as classification or outlier detection (one-class) problem which can be solved e.g. by means of machine learning algorithms. The classifier parameter choice is normally done by expert knowledge or using iterative approaches like cross validation. Such a procedure has considerable influence on the fault detection performance. The availability of training and evaluation data collected under faulty conditions is mostly very limited or time and cost consuming and thus often problematic. This paper presents an iterative threshold setting algorithm, which only uses fault-free data for parameter optimization. For this purpose, a one-class support vector machine which is restricted to convex data sets (including non-Gaussian) is used. The effectiveness of the proposed threshold setting scheme is assessed based on false alarm rate, fault detection rate and randomized algorithm evaluation. Additionally, random uniform distributed uncertainties (scaling and rotation) and offset faults (inside an ellipse) are taken into account. Finally, a comparison study with principal component analysis and Hotelling's T-2, SPE threshold setting schemes is demonstrated. Copyright (C) 2020 The Authors.