Abstract
This work presents a novel algorithm for failure prediction in manufacturing processes using online unsupervised learning based on Kullback-Leibler divergence (KLD). The proposed method continuously monitors sensor data by comparing the probability distributions of a test window against those of a reference window to detect deviations that signal potential system degradation. These distributions are modeled as multivariate Gaussians to capture interdependencies between sensor signals. The algorithm is applied to real-world data from an electric arc furnace in the steel industry, demonstrating its ability to predict failures without prior offline training. Experimental results reveal that multivariate KLD analysis offers a more favorable balance between early fault detection and false alarm rates than univariate approaches. The method provides a lightweight, data-efficient, and practical solution for predictive maintenance in industrial settings where labeled failure data is limited or unavailable.