Incorporation of neural network analysis into a technique for automatically sorting lightweight metal scrap generated by ELV shredder facilities

Abstract

In an attempt to improve the technique of automatic sorting of lightweight metal scrap by sensing apparent density and three-dimensional shape, realized by the combination of a three-dimensional (3D) imaging camera and a meter to weigh a moving object on a conveyor belt, neural network analysis was integrated into the scrap identification algorithm, and its effect on the sorting accuracy of this technique was examined using approximately 1750 pieces of scrap sampled at three different end-of-life vehicle (ELV) shredder facilities. As a result, the newly developed algorithm, in which an unknown fragment is identified by passing through two discriminant analyses and one neural network analysis, was demonstrated to greatly decrease the time required for data analysis to prepare the identification algorithm without reducing the sorting accuracy. The average sorting accuracy for a mixture of three types of lightweight metal fragments was found to be 85%, based on the fact that the fist-sized fragments of cast aluminum, wrought aluminum, and magnesium sampled at the three ELV shredder facilities had similar apparent densities and 3D shapes. It was also suggested that still higher sorting performance is possible by repeating the procedure of modifying the database and re-learning of the neural network in the identification algorithm.