Process monitoring based on Kullback Leibler divergence

Abstract

This article proposes to monitor industrial process faults using Kullback Leibler (KL) divergence. The main idea is to measure the difference between the distributions of normal and faulty data. Sensitivity analysis on the KL divergence under Gaussian distribution assumption is performed, which shows that the sensitivity of KL divergence increases with the number of samples. For non-Gaussian data, a recently proposed kernel method for density ratio estimation is used to estimate the KL divergence. The density ratio estimation method does not involve direct estimation of probability density functions, hence is fast and efficient. For monitoring of non-Gaussian data, the confidence limits are obtained through a window based strategy. Application studies involving a simulation example and an industrial melter process show that the performance of the proposed monitoring strategy is better than the principal component analysis (PCA) based statistical local approach.