Article Details

Pipeline infrastructure is vital to the global energy sector, yet leakages remain a persistent threat with severe environmental and economic consequences. This study compares an enhanced lightweight deep learning model for IoT-based oil pipeline leak detection with three established architectures: InceptionV3, MobileNetV2, and ResNet-50. The proposed model integrates Convolutional Neural Networks (CNN), Knowledge Distillation (KD), and an Autoencoder (AE) to achieve high detection accuracy while maintaining a compact model size suitable for edge device deployment. A locally collected thermal image dataset of 1,506 images, augmented to 9,036 samples, was used for training and evaluation. Performance was assessed at 10 and 20 training epochs using accuracy, precision, recall, and F1-score. The proposed CNN + KD + AE model consistently outperformed all baseline models at both epoch milestones, achieving an accuracy, precision, recall, and F1-score of 0.98 with a compact model size of 4.3 MB, compared to 8.49 MB for MobileNetV2 and 10.63 MB for ResNet-50. These results confirm the suitability of the proposed framework for real-time, resource-constrained pipeline monitoring applications.