Fault identification using fast k-nearest neighbor reconstruction

Abstract

Data with characteristics like nonlinear and non-Gaussian are common in industrial processes. As a non-parametric method, k-nearest neighbor (kNN) rule has shown its superiority in handling the data set with these complex characteristics. Once a fault is detected, to further identify the faulty variables is useful for finding the root cause and important for the process recovery. Without prior fault information, due to the increasing number of process variables, the existing kNN reconstruction-based identification methods need to exhaust all the combinations of variables, which is extremely time-consuming. Our previous work finds that the variable contribution by kNN (VCkNN), which defined in original variable space, can significantly reduce the ratio of false diagnosis. This reliable ranking of the variable contribution can be used to guide the variable selection in the identification procedure. In this paper, we propose a fast kNN reconstruction method by virtue of the ranking of VCkNN for multiple faulty variables identification. The proposed method significantly reduces the computation complexity of identification procedure while improves the missing reconstruction ratio. Experiments on a numerical case and Tennessee Eastman problem are used to demonstrate the performance of the proposed method.