Detection of copper-containing scrap in a post-shredder fraction with machine vision

Conclusion

This work focused on developing a one-stage detector for classifying copper containing particles in a post-shredder fraction. The goal to reduce the copper content to make the feedstock more suitable for use in EAF. During this work a workflow has been developed that can serve as a rough guideline for deploying one-stage detectors in waste management facilities. This work further depicts how the chosen model's performance was enhanced by exporting and converting the model to a target hardware specific format. The best performing model was yolov8n with optimised hyperparameters, exported to the.onnx format. This model achieved a worst-case inference time of 75 ms/image on the limited testing hardware and an mAP50–95 of 0.766. In terms of object-based accuracy, the testing on the independent test data set resulted in 89 % of all copper particles and 86 % of all iron particles being correctly identified.

The trials conducted on the prototype that was constructed to evaluate the created sorting solution in a real-world application showed that a three step sorting setup provides a stable to achieve purities of over 99.5 % from varying input composition and with varying throughput rates. The inference speed and accuracy were noticeably improved by employing hyper parameter tuning prior to training and by converting to.engine from.onnx prior to deployment.

Abstract

Copper contamination in scrap input for electric arc furnaces (EAF) causes quality issues in steel production, as copper cannot be removed during melting in EAFs and leads to surface cracking and brittleness in the final products. This work presents a cost-effective AI-based classification and sorting method using single-stage AI detectors to identify copper particles in post-shredder scrap, offering an alternative to X-ray fluorescence or laser-induced breakdown spectroscopy. After comparing all sizes of YOLOv8 and YOLOv11, the most promising architecture was further subjected to pruning, hyper parameter optimisation and conversion to decrease its inference latency without compromising its prediction accuracy. The thus generated YOLOv8n model achieved a worst-case inference time of 75 ms/image on CPU testing hardware with a mAP50–95 of 77 %. In terms of object-based accuracy, the testing on the independent test data set resulted in 89 % of all copper particles and 86 % of all iron particles being correctly identified. After these offline tests, a prototype consisting of a conveyor belt, low-cost industrial GPU, an industrial camera and an industrial high pressure nozzle bar was built to gauge the model's deployment into an industrial setting by using hyperparameter tuning and conversion to GPU optimised formats. On this, three-stage separation trials with throughputs ranging from 2.5 t/h to 10 t/h with initial copper contents of 10 % and 25 % were conducted. These trials resulted in a purity of the iron fraction of over 99.3 %, calculated by taking the mass of all copper containing particles in the iron fraction.