Fault diagnosis optimization algorithm based on k nearest neighbor

Abstract

In the actual industrial process, it is the key to recognize the fault variables accurately as soon as possible after the fault is detected. Recently, a fault variable recognition method based on k-nearest neighbor reconstruction (FVR-kNN) has been proposed. However, dealing with fault problem caused by multiple variables, the algorithm needs to exhaustive the arrangement of all variables, resulting in high complex computation. And the multivariate estimation in FVR-KNN is not accurate. Thus, this paper proposes a variable recognition optimization algorithm based on FVR-kNN (OFVR-kNN). It optimizes the estimation steps of FVR-kNN in reconstructing multivariate, guaranteeing that the estimations of these potential fault variables have no mutual influence. According to the fault magnitude in corresponding direction, the fault variables are selected in turn. OFVR-kNN does not need to exhaustive all the combinations, greatly reducing the number of reconstructions in fault sample. In this paper, the validity of the optimization method is proved in Tennessee Eastman process.