Deep Learning System for E-Waste Management ^†

Abstract

The deep learning system for e-waste management presented in this proposal is a transformative solution designed to address the escalating challenges of garbage collection and management in urban environments. Rapid urbanization has resulted in increased waste generation, necessitating a more intelligent and efficient approach to e-waste collection and disposal. This system integrates cutting-edge technologies, primarily Artificial Intelligence (AI), to improve e-waste management processes, enhance resource utilization, and contribute to the creation of cleaner and more sustainable urban spaces. Urban areas are experiencing unprecedented growth, leading to a surge in the volume of waste generated daily; as such, traditional waste management systems struggle to cope with this influx, resulting in environmental pollution, compromised public health, and inefficient resource utilization. The proposed deep learning model with accuracy of 83% seeks to revolutionize existing practices by leveraging the capabilities of AI. The aim of this research is to develop a sequential neural network using a Keras and TensorFlow image analysis: a deep learning convolutional neural network (CNN) for e-waste management. The Python programming tool will be used to develop the deep learning model as well as a GUI that will facilitate human–computer interactions. The system will be tested and the result evaluated to assess the functionality and adequacy of the system.

1. Introduction

Deep learning system for e-waste management is an emerging technology that combines technological and environmental sustainability in the digital age; it is a vital issue which cannot be managed by relying on individuals with tight schedules noticing each and every phenomenon.. Today, automatic systems are preferred over manual alternatives to make life simpler and easier in all aspects [1]. Although is generally agreed that waste management services are essential services that must be provided in every society, Ref [2] very little is known about what constitutes e-waste. The consequences and harm caused by poor waste management are extremely dangerous to the world and to human health [3] and have served as a wakeup call and demonstrate need to investigate the situation before they cause even more harm to the world at large. This has prompted the need to look for the best practice to employ to properly manage waste by separating the waste components into classes using the latest technological trends. The concept as presented is a transformative solution designed to address the escalating challenges of e-waste management [4]. The concept integrates cutting-edge technologies, primarily artificial intelligence, with optimized waste management processes [5].

Due to the rapid increase in the use of electrical and electronic equipment (EEE) worldwide, E-waste has become a critical environmental issue around the world. Several studies have pointed out that failure to adopt appropriate recycling practices for e-waste may cause environmental disasters and health challenges to humans due to the presence of hazardous materials [6]. This research focuses on the development of an intelligent automated sorting system for segregating various e-waste components by using a deep learning algorithm (sequential neural network) with TensorFlow and Keras and mobile applications; this will prevent the exposure to and damage of the soil by hazardous substances as well as the pollution of the environment, and traditional sorting methods will be a thing of the past as this innovation will make sorting fast and safe [7]. This technologically driven method seeks to transform e-waste management by delivering a cost-effective and long-term solution and is expected to increase the usage of e-waste recycle bins, hence supporting global initiatives and creating a greener, cleaner, and safer environment by smartly monitoring and controlling the collection, segregation, and disposal of e-waste [8] through the concept of deep learning (DL).

2. Research Methodology

The research methodology is the process that demonstrates how a project is carried out with a clear outline of the methods to be used. It shows how the research goal is realized through the systematic execution of the objectives [9]. To accomplish the objectives of the research, a sequential neural network model was developed, trained, and evaluated using python programming language, the TensorFlow and Keras libraries, and visual studio code for the mobile application. The sequential neural network technique is considered the most outstanding in the processing of sequential data of e-waste [10].

3. Data Collection

The dataset used for this research was collected from the Kaggle online dataset repository [11]. The obtained training data were mainly used in the classification of the recyclability status of e-waste, which was classified as batteries, computers, keyboards, mice, printers, washing machines, PCBs, player devices, microwaves, mobiles, televisions, and speakers. There were about 3600 pictures encompassing 300 images of batteries, 300 images of computers, 300 images of keyboards, 300 images of microwaves, 300 images of mobiles, 300 images of mice, 300 image of PCBs, 300 images of player devices, 300 images of printers, 300 images of speakers, 300 images of washing machines, and 300 images of televisions were processed, trained, tested, and evaluated for performance. These divisions were made depending on the pictures contained in the separate respective folders. Figure 1 below provides images of some of the dataset that was used for this research, classified as batteries, computers, keyboards, mice, printers, washing machines, PCBs, player devices, microwaves, mobiles, televisions, and speakers. There were about 3859 images, and of these, 3139 were used for training the model, 360 images were used for the validation and testing of the model with 360 images of batteries, 310 images of computers, 330 images of keyboards, 320 images of microwaves, 330 images of mobiles, 322 images of mice, 330 images of PCBs, 321 images of player devices, 330 images of printers, 306 images of speakers, 324 images of washing machines, and 326 images of televisions.

The Proposed Model

Thies study used sequential neural network (SNN) architecture in a deep learning algorithm to form four (4) mother classes, and each mother class contained three (3) categories of e-waste dependents out of batteries, computers, keyboards, mice, printers, washing machines, PCBs, player devices, microwaves, mobiles, televisions, and speakers. Twelve convolutional neural networks were divided into four blocks (A, B, C, and D), each of which contained three of the classes of the e-waste components—one for each mother class—which also prevented the recognition of a class that did not exist. For instance, if the mother class is computer, we know that the child class cannot be a battery. At first, this model uses the image as the input; at that point, the SNN creates the bounding box and utilizes the mother class as an output. With these data, the real image is cropped, allowing the CNN’s input to identify the mother class. After that, the output generated is the child class, which is converged with the mother class and creates the final class. Finally, before the detection of the model, the output is predicted using the bounding box. The mechanisms of this model can be found in Figure 2 below.

This diagram depicts a hierarchical convolutional neural network (CNN) architecture designed for object detection and classification, specifically applied to e-waste.

4. Result and Discussion

Python 3, Jupyter 4.0.11 with Anaconda 3, TensorFlow 2.16 for deep learning, and Microsoft Visual C++ 2015–2019 redistribution (×64) 14.29.30153.0 are the software requirements for this research.

4.1. Model Architecture

The model architecture displayed is a convolutional neural network (CNN) implemented using the Keras Sequential API. Once the training pipeline is successfully built up and configured, TensorFlow started initializing the model training [5]. A lot of computational power is required for training an enormous network in SNN. To prepare our neural network training, we used a HP laptop outfitted with an Intel core i5 processor. Along with Python version 3.9, open-source software was used for high-performance mathematical calculations. Its adaptable architecture allows the easy deployment of calculations over a variety of stages [12]. (The input shape expects input images of different shapes (224, 224, and 3), and the model includes three convolutional layers (conv2d_3, conv2d_4, and conv2d_5) with increasing filter sizes (16, 32, and 64) to capture different levels of feature complexity. In the pooling layers: maximum pooling layers are interspersed between convolutional layers to reduce the spatial dimensions and computational load while retaining important features [1]. The flatten_1 layer converts the 3D feature maps into a 1D feature vector to prepare it for the fully connected layers, and the model has two dense layers (‘dense_2’ with 128 neurons and ‘dense_3’ with 12 neurons) to perform the final classification task. The final layer suggests the model output probabilities for 12 different classes. This architecture is typical for image classification tasks, leveraging convolutional layers for feature extraction and dense layers for classification [13]. The input shapes are images of 224, 224, and 3, and the model includes three convolutional layers with filter sizes of 16, 32, and 64, followed by maximum pooling layers. There are also two dense layers with 128 and 12, and the final dense layer has 12 neurons, indicating classification into 12 different classes, as depicted in

Table 1. Model architecture.

Layers (Types)	Outputs Shapes
Rescaling_5 (Rescaling)	(None, 224, 224, 3)
Conv2d_15 (conv2D	(None, 222, 222, 16)
Max_pooling2d_15 (MaxPooling2D)	(None, 11, 111, 16)
Conv2d_16 (Conv2D)	(None, 109, 109, 32)
Max_pooling2d_16 (MaxPooling2D)	(None, 54, 54, 32)
Conv2d_17 (Conv2D)	(None, 52, 52, 64)
Max_pooling2d (MaxPooling2D)	(None, 26, 26, 64)
Flatten_5 (FLATTEN)	(None, 43264)
Dense_10 (DNESE)	(None, 128)
Dense_11 (DNESE)	(None, 128)

Total params: 5,563,052 (21.22 MB); Trainable params: 5,563,052 (21.22 MB); Non-trainable params: 0 (0.00 B); Model Training.

below.

The training process in Figure 3 below shows the first five epochs focuses on the learning dynamics, convergence behavior, potential issues, and interpretation of the results. In the initial training epochs, the model starts with a low training accuracy of 0.4861 and a high training loss of 1.477, indicating random initialization and significant prediction errors. The validation loss of 2.769 further reflects poor generalization. By Epoch 2, significant improvements are observed with training accuracy rising to 0.6398 and training loss decreasing to 1.209, accompanied by a notable drop in validation loss to 1.2647, showing better performance on unseen data. Over Epochs 3 to 5, training accuracy steadily increases from 0.6639 to 0.8126, and training loss decreases from 1.0463 to 0.5668. The validation loss follows a generally downward trend, and validation accuracy surpasses 0.8 by Epoch 5, indicating the model’s effective learning and improved generalization.

4.2. Plotting Performance

Figure 4 below provides the performance graphs of the accuracies and losses of the trained model, illustrating the training loss and accuracy and validation loss and accuracy of 20 epochs for a deep learning model.

Analyzing training and validation loss and accuracy graphs is essential for evaluating model performance in deep learning. Validation loss and accuracy quantify the model’s ability to generalize to new data, whereas training loss and accuracy show how effectively the model is learning from the training set. Effective learning is ideally indicated by a gradual decrease in training loss and an improvement in training accuracy. While validation accuracy should rise and stabilize, validation loss should fall and then stabilize. An overfitted model is one that memorizes training data rather than allowing the model to generalize [14]; this is shown through an increasing validation loss or decreasing validation accuracy. Regularization, dropout, and early halting are examples of mitigation strategies that can help increase the model’s capacity for generalization. In Figure 4a which is the training loss decreases steadily, while validation loss decreases initially but starts increasing after a certain number of epochs. The model is likely overfitting. Initially both training and validation loss decreases. However, the increase in validation loss after a certain point suggests the model is starting to memorize the training data rather than generalizing. Training loss decreases consistently, indicating the model is learning the training data. Validation loss decreases initially but then increases, indicating overfitting after a certain epoch, while in Figure 4b the training accuracy increases steadily, reaching close to 100%, while validation accuracy increases initially but then stabilizes or decreases. The model is learning the training data well but is not generalizing effectively to unseen data. This again points to overfitting. Training accuracy increases and reaches close to 100%, indicating the model fits the training data well. Validation accuracy increases initially but then stabilizes or decreases, again suggesting overfitting. Training and validation graphs are essential tools in diagnosing the training behavior of deep learning models. Proper interpretation of these graphs can provide insights into model performance, indicating whether the model is overfitting, underfitting, or training effectively. By employing techniques such as regularization, dropout, and early stopping, we can mitigate common issues and enhance the model’s generalization capability.

4.3. Evaluate Model Performance

Accuracy, precision, recall, F1-score, and confusion matrix are the most common and effective metrics used to evaluate deep learning models [15], and they were used to evaluate the effectiveness and performance of the model. The model summaries are depicted in

Table 2. Evaluation report of various metrics.

Categories	Precision	Recall	F1_Scores	Support
Battery	0.91	1.00	0.95	30
Keyboard	0.86	0.83	0.85	30
Microwave	0.80	0.80	0.80	30
Mobile	0.79	0.83	0.83	30
Mouse	0.84	0.70	0.76	30
PCB	1.00	0.90	0.95	30
Player	0.84	0.87	0.85	30
Printer	0.76	0.73	0.75	30
Television	0.84	0.87	0.85	30
Washing Machine	1.00	0.83	0.91	30
Computer	0.94	1.00	0.97	30
Speaker	0.83	0.97	0.89	30
Accuracy			0.86	360
Macro Average	0.87	0.86	0.86	360
Weighted Average	0.87	0.86	0.86	360

as follows: precision = 83%; recall = 81%; F1 score = 81%; and accuracy = 83%.

The precision score was calculated to be 83%, indicating a high percentage of positively classified events that were correctly identified. The model’s capacity to precisely identify real positive instances was demonstrated by the recall score of 81%. The F1-score was 81%, demonstrating a high degree of correlation between the predicted and true labels. The confusion matrix in Figure 5 provides specific information about how well the model classified each class. The confusion matrix is a fundamental tool for evaluating the performance of a classification model. It provides a detailed summary of the classification outcomes for each class by showing the number of correct and incorrect predictions [16]. This is particularly useful for understanding the model’s behavior across different categories, identifying strengths and weaknesses, and guiding further improvements [17]. The provided confusion matrix displays the classification results for a multi-class deep learning model across 12 categories. Each cell in the matrix represents the number of instances where the predicted class matches the true class. The diagonal elements represent correctly classified instances, while the off-diagonal elements represent misclassifications.

Testing Performance

The model was tested with various e-waste components for accuracy and prediction, such as in Figure 6 and Figure 7, and the model correctly predicted the classes with high accuracy.

5. Conclusions

The study presents a novel deep learning-based system for e-waste management that leverages sequential neural networks (SNNs) and convolutional neural networks (CNNs) to effectively classify and sort e-waste components. The model demonstrates an accuracy of 83%, showing significant potential in the ability to address the growing environmental and health challenges associated with e-waste. By automating the sorting process and reducing human exposure to hazardous materials, this system contributes to a safer and more efficient waste management solution. Additionally, the integration of artificial intelligence in waste sorting offers a cost-effective, scalable, and sustainable approach to recycling, supporting global efforts toward environmental protection and resource conservation. Future work could enhance the system’s generalizability and precision by incorporating larger and more diverse datasets, as well as integrating real-time applications for smart cities and industries.