Transfer learning-based method for automated e-waste recycling in smart cities

Research Motivation

As urbanization increases, so does the volume of electrical and electronic waste (e-waste). Traditional manual sorting is inefficient and poses health risks due to hazardous materials. This paper explores how Artificial Intelligence (AI) can support the Circular Economy by automating the classification of e-waste, specifically focusing on smartphones as a case study. The goal is to reduce human intervention and improve the accuracy of waste segregation in smart cities.

To overcome the challenge of training deep neural networks from scratch—which typically requires massive datasets and significant computational resources—the authors employed transfer learning. They utilized the pre-trained AlexNet model, which was originally trained on the large ImageNet dataset. By fine-tuning the output layers of AlexNet, the researchers adapted the model to classify a smaller, custom dataset containing 650 images across 12 smartphone classes from 6 different brands. The authors optimized the model by tuning the learning rate, selecting the Stochastic Gradient Descent with Momentum (SGDM) optimizer, and applying data augmentation to prevent overfitting.

Findings and Significance

Using an SGDM optimizer with a learning rate of 3e-4, the model achieved approximately 98% accuracy. This result demonstrates that transfer learning is a highly effective, resource-efficient strategy for automated waste sorting, even when the available training data is limited. By enabling machines to accurately identify and categorize electronic devices, this approach supports the development of circular smart cities, where automated systems facilitate the recovery of valuable materials and reduce the environmental impact of e-waste.