In order to combine machine learning and additional domain knowledge, Paul Wurth developed an addition to the existing TPDS system – the machine vision pipeline. This processing pipeline applies semantic segmentation to the camera video stream, and a machine learning algorithm to classify images on the pixel level into the different areas of interest: raceway, tuyere, lance, and injection. Additional transformation steps make use of the semantic understanding of the image, and the information contained in the raw images is converted into a structured format that can be shown as time-series data.
Advanced domain knowledge has been digitized in rule-based logic using RulesXpert, the rule logic engine from the DataXpert platform. The logic fuses sensor data collected from the plant and time-series data generated by the machine learning algorithm, leading to significantly enhanced anomaly detection performance.
These improvements could be achieved with no changes to the underlying camera equipment in the field, allowing for easy upgrades of existing TPDS installations by upgrading the server hardware (adding computational resources) and TPDS software only.
Contact: Marc Weydert, Pierre Van Dorpe, Steve Czarnuch, Aissatou Ndiaye Ba, Madita Bayer