Classification by Neural Net of a Particle Stream in an Eddy Current Drum Separator

Abstract

At present, eddy current separation is primarily used for the recovery of non-ferrous metals from industrial waste (e. g. car scrap), rather then the recovery of similar metals from domestic waste. This is due to the complexity of domestic waste and to the difficulties in its characterisation. This work approached the problem in order to study the possibility of developing and implementing the control system of an eddy current drum separator, based on image analysis of the fed stream. The investigated control sets settled the working parameters of the device (i. e. belt and drum speed and splitter position) according to the data supplied by a control model based on a neural net. The goal was to establish the critical parameters that can lead the neural net to a proper classification of the input stream, starting from selected OWF (organic wet fraction of the waste) and shredded particles. After experimental tests, carried out under different conditions (belt and drum speed) and on particles of different materials, morphological (area, diameter, minor and major axis, aspect), optical (metal particles surface response to the lighting system) and trajectory (second-order curve) parameters are extracted by image analysis of the scene. In order to achieve a proper classification, the most important parameters were area and trajectory parameters. Better classification results were obtained taking into account a special feature, ECD (eddy current dimension), extracted after image analysis.