Visual Surveillance Berbasis YOLO Dengan Rule Based : Analisis Komparatif Model YOLO Untuk Identifikasi Kebakaran dan Kebocoran Gas Pada Kapal di Pelabuhan Gresik

Khusnul, Muchlisin (2025) Visual Surveillance Berbasis YOLO Dengan Rule Based : Analisis Komparatif Model YOLO Untuk Identifikasi Kebakaran dan Kebocoran Gas Pada Kapal di Pelabuhan Gresik. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022231044-Master_Thesis.pdf Restricted to Repository staff only Download (73MB) | Request a copy

Abstract

Penelitian ini bertujuan untuk mengevaluasi pengaruh faktor-faktor optimasi terhadap performa model deteksi kebakaran dan kebocoran gas berbasis YOLO. Faktor-faktor yang dianalisis meliputi pemilihan optimizer, pengaturan hyperparameter, teknik augmentasi data, serta variasi dataset. Hasil eksperimen menunjukkan bahwa penggunaan optimizer AdamW dengan learning rate 0.001429 secara signifikan meningkatkan akurasi deteksi, kecepatan konvergensi pelatihan, serta metrik evaluasi seperti mAP50, Precision, Recall, dan F1-score dibandingkan optimizer konvensional seperti SGD. Teknik augmentasi data berupa konversi grayscale, rotasi, dan penambahan noise terbukti efektif meningkatkan ketahanan model terhadap variasi pencahayaan dan gangguan visual. Selain itu, keberagaman dataset memberikan kontribusi signifikan dalam meningkatkan kemampuan generalisasi model dan mengurangi risiko overfitting. Meskipun terdapat kompromi antara kecepatan inferensi dan akurasi, konfigurasi optimasi yang tepat mampu menghasilkan keseimbangan optimal untuk aplikasi real-time di lingkungan maritim. Temuan pada penelitian ini merekomendasikan penerapan kombinasi strategi optimasi guna meningkatkan efektivitas dan keandalan sistem deteksi kebakaran dan kebocoran gas berbasis YOLO.
==================================================================================================================================
This study aims to evaluate the impact of optimization factors on the performance of YOLO-based fire and gas leak detection models. The analyzed factors include the choice of optimizer, hyperparameter tuning, data augmentation techniques, and dataset diversity. Experimental results demonstrate that using the AdamW optimizer with a learning rate of 0.001429 significantly improves detection accuracy, training convergence speed, and evaluation metrics such as mAP50, Precision, Recall, and F1-score compared to conventional optimizers like SGD. Data augmentation methods, including grayscale conversion, rotation, and noise addition, effectively enhance the model’s robustness against variations in lighting conditions and visual disturbances. Furthermore, diverse datasets contribute substantially to improving model generalization capacity and reducing overfitting risk. Although there is a trade-off between inference speed and accuracy, appropriately configured optimization achieves an optimal balance suitable for real-time applications in maritime environments. The findings of this study recommend the implementation of combined optimization strategies to enhance the effectiveness and reliability of YOLO-based fire and gas leak detection systems.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	deteksi kebakaran, kebocoran gas, optimasi model, augmentasi data, optimizer, AdamW, SGD, hyperparameter, real-time inference, fire detection, gas leak detection, YOLO, model optimization, data augmentation, hyperparameters, real-time inference
Subjects:	V Naval Science > V Naval Science (General) > V220 Naval ports, bases, reservations, docks, etc.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	KHUSNUL MUCHLISIN
Date Deposited:	05 Jun 2025 04:52
Last Modified:	05 Jun 2025 04:52
URI:	http://repository.its.ac.id/id/eprint/119112

Actions (login required)

View Item