Predictive maintenance decision using statistical linear regression and kernel methods

In this paper, we develop a predictive maintenance (PdM) method to determine the most effective time to apply maintenance to an equipment and study its application to a real semiconductor etching chamber. More specifically, we first apply linear regression to predict the (output) equipment health condition from the (input) operational parameters. This choice of linear model also allows us to propose an algorithm to reduce the number of operational parameters to be monitored for PdM purposes using t-statistics. Then, we follow a cross-validation based procedure to generate prediction error samples and apply a kernel method to construct the corresponding probability density function of the prediction error. Finally, the PdM decision can be made based on the likelihood of the predicted health condition exceeding a certain maintenance threshold. Our analysis using real data from a semiconductor etching chamber shows that the proposed PdM decision with the reduced dimension linear regression performs comparably to the one using full-scale linear model and can be used for better maintenance planning compared to the existing practice of fixed-schedule maintenance.