Prototype of AI-powered assistance system for digitalisation of manual waste sorting
Julian Aberger, Somayeh Shami, B. Hacker, José Osmar Medina Pestana, Karim Khodier, Renato Šarc
Abstract
• Artificial Intelligence assistance system for manual waste sorting. • Use case specific requirements for digitalisation for manual sorting. • Evaluation of existing datasets in the waste sector shows improvement potential. • Real data acquisition in near industry scale experiment has been carried out. • Classification accuracy of 81% achieved on an experimental acquired dataset. Global waste generation is projected to reach 3.40 billion tons by 2050, necessitating improved waste sorting for effective recycling and progress toward a circular economy. Achieving this transformation requires higher sorting intensity through intensified processes, increased efficiency, and enhanced yield. While manual sorting remains common, smaller plants often use positive sorting to recover recyclables, and larger plants combine automated systems with manual sorting. Negative sorting is employed to remove impurities and improve material quality. However, innovation in manual sorting has stagnated. Advances in Machine Learning and Artificial Intelligence offer transformative potential for waste management, with digitalisation and improved recyclate quality becoming priorities. Despite these trends, manual sorting is still largely treated as a digital black box. The presented research outlines the design of a novel, human-centric AI-powered assistance system to support sorting workers by enhancing decision-making and real-time assistance during the sorting process, driving the digitalisation of manual sorting. Potential use cases, system requirements, and essential components were explored. High-quality use case-specific data is essential for model training. Therefore, publicly available datasets were evaluated but found inadequate, necessitating use-case-specific data acquisition through near-industry-scale experiments. This data was used to train and develop key system components, such as object recognition, classification, and action recognition models. Results indicate that transfer learning with a balanced dataset is effective for waste-sorting applications. The classification model achieved 81% accuracy on an experimental acquired balanced dataset, outperforming the accuracy of the pre-trained model on its original dataset.