Fault Diagnostics for Multivariate Non-normal Processes

Abstract

There is a great challenge in carrying out multivariate process capability analysis and fault diagnostics on a high dimensional non-normal process, with multiple correlated quality characteristics, in a timely manner. This paper proposes a hybrid capable of performing process capability analysis and fault diagnostics on multivariate non-normal processes. The proposed hybrid first utilizes the Geometric Distance (GD) approach, to reduce dimensionality of the correlated data into fewer number of independent GD variables which can be assessed using univariate process capability indices (PCIs). This is followed by fitting Burr XII distribution to independent GD variables. The independent fitted distributions are used to estimate both yield and multivariate process capability in a time efficient way. Finally, machine learning approach, is deployed to carry out the task of fault diagnostic by identifying and ranking the correlated quality characteristics responsible for the poor performance of individual GD variables. The efficacy of the proposed hybrid is assessed through a real manufacturing example and four simulated scenarios. The results show that the proposed hybrid is robust in estimating both yield and multivariate process capability carrying out fault diagnostics beyond GD variables, and identifying the original characteristic responsible for poor performance.