Transfer learning-based method for automated ewaste recycling in smart cities

Technical Summary: Transfer Learning-Based Method for Automated E-Waste Recycling in Smart Cities

Problem Statement
The paper addresses the critical challenge of managing the rapidly growing stream of electronic waste (e-waste) within the context of smart cities and the Circular Economy (CE). Traditional waste sorting methods are insufficient for the volume and complexity of modern e-waste, which contains both valuable materials (e.g., gold, copper) and hazardous substances (e.g., mercury, cadmium). Manual sorting is labor-intensive, prone to human error, and inefficient. The authors identify a specific need for automated sorting systems that can accurately classify devices to facilitate efficient recycling, reduce environmental toxicity, and recover resources, thereby supporting the transition toward sustainable "circular smart cities."

Methodology
The study proposes a computer vision-based approach utilizing Transfer Learning to automate the classification of smartphones, used here as a representative case study for e-waste. The technical workflow involves the following steps:

• Dataset Construction: A custom dataset was manually curated from the internet, consisting of 650 RGB images representing 12 distinct smartphone models from 6 brands (Acer, HTC, Huawei, Apple, LG, Samsung). To ensure distinguishability, images focused on the backside of devices where unique features like logos and camera lenses are located. The data was split into 80% for training and 20% for validation.
• Model Architecture: The authors employed AlexNet, a pre-trained Convolutional Neural Network (CNN) originally trained on the ImageNet dataset (1 million images, 1,000 labels). Instead of training a network from scratch, which requires massive datasets and computational resources, the study utilized transfer learning.
  • The pre-trained weights of the first five layers (convolutional and pooling layers) were frozen to retain learned feature representations.
  • The final three layers (fully connected layers) were replaced and re-initialized to output 12 classes corresponding to the specific smartphone models.
• Data Augmentation: To address the small size of the dataset and prevent overfitting, extensive data augmentation was applied. This included random X/Y reflections, translations, scales (range 0.9–1.1), and shears (range -50 to 50 pixels). This process expanded the dataset from 650 to 5,850 images (a 9x multiplier).
• Training Configuration: The implementation was conducted in MATLAB R2020b using a GPU (ASUS Nvidia GeForce RTX 2070S). The training parameters were optimized by testing:
  • Optimizers: Stochastic Gradient Descent with Momentum (SGDM), Adam, and RMSProp.
  • Learning Rates: Various rates were tested to find the optimal convergence speed.
  • Batch Size: Set to 64.
  • Epochs: Maximum of 30 with early stopping capabilities.

Key Results
The study evaluated the model's performance by tuning the optimizer and learning rate while maintaining the data augmentation baseline. The results indicated:

• Optimizer Performance: Among the tested optimizers, SGDM yielded the highest accuracy. Adam and RMSProp showed lower performance, with SGDM achieving a mean accuracy of 98.35% (± 2.0155), compared to 95.26% for Adam and 93.20% for RMSProp.
• Learning Rate: A learning rate of 3e-4 combined with the SGDM optimizer provided the best generalization and stability.
• Final Accuracy: The optimized model achieved an accuracy of approximately 98% on the validation set, demonstrating effective generalization despite the small initial dataset size.

Significance and Claims
The paper claims that its primary contribution lies in demonstrating the efficacy of transfer learning as a viable solution for automated e-waste sorting, particularly when large datasets are unavailable. The authors argue that:

1. Efficiency: Transfer learning significantly reduces the computational burden and time required to design a Neural Network compared to building one from scratch.
2. Accuracy: The proposed method reduces the error rate of e-waste sorting compared to human-based processes, achieving high accuracy (approx. 98%) with a relatively small dataset.
3. Scalability for Smart Cities: The approach supports the broader goals of Circular Smart Cities by enabling digitalization in waste management, which can lead to better resource recovery and reduced environmental impact.

The authors maintain a modest stance, noting that while their method improves sorting efficiency, it is not a standalone solution for the transition to a sustainable economy. They emphasize that automated classification should be viewed as a component of a larger system, potentially integrated with robotics, IoT, and data mining, rather than a "rebound-effect" technology that replaces the need for systemic changes in business models or consumer behavior. The work is presented as a foundational step toward a semi-automated e-waste management system being developed in their Circular Digital Economy Lab.