Kalman Filter for Predictive Maintenance and Anomaly Detection

Abstract

There are various strategies in optimization of anomaly detection problem on sensor data. This paper describes a Gaussian Mixture Model (GMM) and Kalman filter to detect outliers within the sensor data of wire bonding. With limitation on data samples and high dimensional parameters, Principal Component Analysis (PCA) helped to reduce dimensionality without losing important information. The Expectation-Maximization algorithm for estimating Gaussian distribution parameters of GMM provided us a clustering model to fit our data. A weighted distance from the cluster center in the employed GMM model is applied to tune noise variances on measurement errors. The proposed method is validated using real measurements in the context of a manufacturing system.