Anomaly Detection with a Spatio-Temporal Tracking of the Laser Spot

Anomaly detection is an important problem with many applications in industry. This paper introduces a new methodology for detecting anomalies in a real laser heating surface process recorded with a high-speed thermal camera (1000 fps, 32x32 pixels). The system is trained with non-anomalous data only (32 videos with 21500 frames). The proposed method is built upon kernel density estimation and is capable of detecting anomalies in time-series data. The classification should be completed in-process, that is, within the cycle time of the workpiece.