Apr 09, 2025
Version 2
Analysis of Space Solar Array Arc Images Based on Deep Learning Techniques V.2
- Ahmed Farahat1,
- Afaf M. Abd El-Hameeda2,
- I.M. Selimc2
- 1BeniSeuif University , Egypt.;
- 2National Research Institute of Astronomy and Geophysics (NRIAG), 11421, Helwan, Cairo, Egypt.
Protocol Citation: Ahmed Farahat, Afaf M. Abd El-Hameeda, I.M. Selimc 2025. Analysis of Space Solar Array Arc Images Based on Deep Learning Techniques. protocols.io https://dx.doi.org/10.17504/protocols.io.14egn4z4yv5d/v2Version created by Ahmed Farahat
License: This is an open access protocol distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Protocol status: Working
We use this protocol and it's working
Created: April 08, 2025
Last Modified: April 09, 2025
Protocol Integer ID: 126453
Disclaimer
This protocol is for academic and research purposes only. Results may vary depending on the dataset and system configuration.
Abstract
This protocol presents a comprehensive deep learning-based methodology for detecting defects in solar cells. It covers data preparation, preprocessing, model building (CNN, VGG16, VGG19, and ResNet50), training strategies, evaluation metrics (accuracy, loss, confusion matrix, classification report), and final conclusions. The goal is to enable reliable and automated defect classification using state-of-the-art neural networks.
Guidelines
This protocol describes a deep learning-based pipeline for detecting defects in solar cells using image data. The approach explores different CNN architectures including a custom model, VGG16, and MobileNetV2. We evaluate the effectiveness of each model using standard classification metrics.
Dataset Description
Dataset Description
The dataset includes images of solar cells categorized into two classes: defective and non-defective. Each image is standardized to a 128x128 RGB format for uniformity and compatibility with the neural networks.
Data Preprocessing
Data Preprocessing
Resized images to 128x128 pixels.
Normalized pixel values to a range of [0, 1].
Applied data augmentation: rotation, zoom, flipping.
Split dataset into training (70%), validation (20%), and test (10%) sets.
Model Architectures
Model Architectures
Custom CNN
A simple convolutional neural network was built from scratch consisting of convolutional layers with ReLU activation, followed by max-pooling, dense layers, and dropout for regularization.
VGG16
Transfer learning was applied using the VGG16 architecture pretrained on ImageNet. The top fully connected layers were replaced with custom dense layers. The base layers were frozen during the initial training and fine-tuned later.
MobileNetV2
MobileNetV2, a lightweight and efficient model, was also used for transfer learning. It provided faster training and better inference speed, with comparable or better accuracy than the VGG16 model.
Evaluation Strategy
Evaluation Strategy
All models were evaluated using the following metrics:
Accuracy and loss over epochs.
Confusion matrix.
Precision, recall, and F1-score from classification report.
Results and Discussion
Results and Discussion
The following summarizes model performances:
CNN: ~88% accuracy.
VGG16: ~91% accuracy, better generalization.
MobileNetV2: ~93% accuracy, optimal speed and performance.
MobileNetV2 outperformed the other models in both accuracy and computational efficiency, making it suitable for real-time applications.
Conclusion
Conclusion
The deep learning pipeline presented is highly effective for solar cell defect detection. MobileNetV2 is the recommended model due to its balance of accuracy and inference efficiency.
References
References
Keras Documentation: https://keras.io
TensorFlow Documentation: https://www.tensorflow.org
ImageNet Dataset: https://www.image-net.org