Metal Recovery and Recycling using Machine Learning and Eddy Current Inspection

Abstract

The need to recover and recycle material towards building a circular economy is becoming a global imperative. Non-ferrous metals, in particular, are highly recyclable and can be extracted using processes such as eddy current separation. However, their further separation into recyclable groups based on metal or alloy continues to pose a challenge. This thesis explores the efficacy of magnetic induction spectroscopy (MIS) to discriminate between non-ferrous metals obtained from multiple industrial-sourced waste streams. Different machine learning algorithms are described, which were trained and tested on the different measurements to determine the best algorithm. The thesis also explores how adding extra features from a camera system can improve classification. The MIS system was installed onto an industrial system to separate wrought aluminium from cast aluminium. The industrial system's recovery and purity were 25.77 - 37.99 and 34.96 - 58.06%pt lower than those obtained in the laboratory, respectively. The waste stream and target metal are important factors that determine the sensor combination used. Removing copper from aluminium within 'Fridge metals' was sorted with a 99.04% recovery rate and 100% purity rate when just colour was used. The developed vision system was susceptible to surface contamination and lighting conditions. Removing wrought aluminium from cast aluminium from 'Twitch' was best sorted with the combined induction and colour measurements. When induction and colour were used independently, their F1 score was not as high as when they were combined. Wrought was removed with a recovery rate of 71.21% - 92.56% and a purity rate of 83.92% - 88.05% when using induction and colour, depending on the dataset and ratio of wrought and cast. The 'Twitch' results improved on those obtained with the industrial system. 'Twitch' showed that different sensors may have conflicting predictions for the same piece that passes over multiple sensors, leading to more ejections into the target bin. Removing stainless steel from 'Zurika' was best sorted with induction only, which obtained a recovery rate of 90.76% and a purity rate of 87.03%. A sample within 'Zurika', which contained stainless steel and copper, showed how a piece's position, orientation and metal content can change what metal is detected with the current measurement chosen for classification. The sample showed how the time series data has the potential to detect multiple metals present. Finally, the sample of multiple metals highlighted the need for a method to label similar samples as it could contaminate the final product.

Date of Award	31 May 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Anthony Peyton (Main Supervisor) & Michael O'Toole (Co Supervisor)