Determining extraneous matter and billet length in sugarcane supplies using machine learning

In Australian sugar factories the physical properties of the cane supply are currently not measured prior to processing through the factory (except for some cane consignments at Tully Mill) and thus parameters such as extraneous matter (EM) content and billet lengths are not routinely monitored. This project has used computer-vision and machine-learning methodologies to investigate the feasibility of measuring on-line EM%cane and billet length for each cane consignment entering the factory. A camera and lighting rig were installed above the cane conveyor belt operating between the tippler and shredder at Tully Mill, and computer facilities were installed to allow captured images for each consignment of cane to be analysed using machine-learning models. Tully staff undertook manual sorting and weighing of grabs of cane to provide the mass percentage of stools, tops, billets, and trash. Randomly selected billets removed from the grab samples were photographed on a graduated board to provide estimates of billet length. These data provide the reference data for development of the models. Investigations have shown that the models based on the Tully data provide an acceptable level of accuracy for EM%cane and billet length. The investigations need to be extended to include datasets from other mill districts to determine if universal models can be developed for implementation by the industry.