Dynamic Bayesian Network-Based Anomaly Detection for In-Process Visual Inspection of Laser Surface Heat Treatment

Abstract

We present the application of a cyber-physical system for inprocess quality control based on the visual inspection of a laser surface heat treatment process. To do this, we propose a classification framework that detects anomalies in recorded video sequences that have been preprocessed using a clustering-based method for feature subset selection. One peculiarity of the classification task is that there are no examples with errors, since major irregularities seldom occur in efficient industrial processes. Additionally, the parts to be processed are expensive so the sample size is small. The proposed framework uses anomaly detection, cross-validation and sampling techniques to deal with these issues. Regarding anomaly detection, dynamic Bayesian networks (DBNs) are used to represent the temporal characteristics of the normal process. Experiments are conducted with two different types of DBN structure learning algorithms, and classification performance is assessed on both anomaly-free examples and sequences with anomalies simulated by experts.