Bu çalışma, elektrik panolarındaki yangın güvenliği konusuna odaklanmaktadır. Nesnelerin İnterneti (Internet of Things - IoT) ve Makine Öğrenimi (Machine Learning - ML) teknolojilerinin entegrasyonu yoluyla elektrik panolarına yönelik yangın güvenliği önlemlerinin nasıl geliştirilebileceği üzerine çalışılmıştır. Elektrik panolarındaki yangın güvenliği önlemlerinin etkinliğini artırmak amacıyla gerçekleştirilen çalışmalar kapsamlı bir şekilde irdelenmiştir. Araştırma bulguları, yangınların büyük bir kısmının elektrik arızalarından kaynaklandığını ortaya koymuştur. Özellikle gevşek bağlantılar, aşırı yüklenme, izolasyon hataları ve diğer teknik arızalar, yangınların başlıca çıkış noktaları arasında yer almaktadır. Yangın algılama sistemlerinin gelişimi, yangına dayanıklı malzemelerin kullanımı, düzenli bakım ve denetim prosedürleri, kanunlar, yönetmelikler ve standartlar tarafından kesin sonuçlara bağlanmıştır. Fakat günümüzde hala elektrik panolarında çıkan yangınlar sebepli can kayıplı ve maddi hasarlı birçok vaka yaşanmaktadır. Yeni nesil teknolojiler ve kestirimci bakım yöntemleri, yangın riskini azaltmada öne çıkan stratejilerdir. Kızılötesi tarama ve termal izleme teknikleri, özellikle erken yangın tespiti ve arızalanma risklerini minimize etmede etkili olmaktadır. Bu bağlamda, elektrik panolarında yangın güvenliği üzerine potansiyel iyileştirme alanları tespit edilmiştir. Elektrik panoları yangın güvenliğini artırmak için nesnelerin interneti (IoT) ve makine öğrenmesi (ML) modelleri ile yeni bir yaklaşım üzerinde durulmuş, yangın meydana gelmeden koşulların izlenmesine imkân veren program geliştirilmiştir. Ayrıca, bu program, yapılan deneylerin anlık olarak izlenmesini ve sonuçların raporlanmasını sağlamıştır. Deneylerde Arduino Mega ve Raspberry Pi cihazları verileri toplamıştır. Toplanan veriler pano ortam sıcaklığı, pano ortam nemi, karbonmonoksit, karbondioksit, hidrokarbon gazlarını, voltaj, akım, parametrelerini ve kablo odak sıcaklığı, dış ortam sıcaklığını, dış ortam nemini içermektedir. Çalışmada toplam 21 deney tamamlamış, 7 deneyde yanma ve yangın meydana gelmiştir. Ayrıca deneylerden 3521 satır x 11 sütun gerçek veri elde edilmiş ve bunlardan 3478 satır ve 9 sütunu ML algoritmalarında kullanılmıştır. Kullanılan ML algoritmaları denetimli öğrenme yönteminden olan, karar ağacı (KA), ikili lojistik regresyon, Gaussian Naive Bayes (GNB), Gaussian Destek Vektör Makinesi (SVM) ve K-En yakın komşu (KNN) yöntemidir. Her bir algoritma 5-katlı çapraz doğrulama yöntemiyle değerlendirilmiştir. Çalışmadan elde edilen, farklı modellerin test performansları şu şekilde ölçülmüştür. Karar ağacı (ince ağaç) modeli %98.7 doğruluk oranı ile yüksek bir performans sergilemiştir. İkili lojistik regresyon modeli %95.3, GNB modeli %91.2 doğruluk oranı ile değerlendirilmiştir. Öte yandan, Gaussian SVM ve KNN yöntemleri her ikisi de %99.3 doğruluk oranı ile en yüksek performansı gösteren modeller olmuştur. Sonuç olarak, bu tez IoT ve ML teknolojilerinin, elektrik panolarında yangın güvenliğini artırmak ve yangın risklerini azaltmak için yapılmıştır. Geliştirilen sistemler, endüstriyel tesislerde yangın güvenliği yönetimini iyileştiren yenilikçi çözümler sunmakta ve yangın öncesi risk değerlendirmeleri yaparak erken müdahale olanaklarını artırmaktadır. Bu bütünleşik yaklaşım, yangın algılama ve müdahale süreçlerinde önemli gelişmelere yol açarak, can ve mal kaybını minimize etmede büyük rol üstlenebilir.
This study examines how fire safety measures for electrical panels can be enhanced through the integration of Internet of Things (IoT) and Machine Learning (ML) technologies. Electrical panels are known as critical structures in energy transmission, with a high risk of fire due to a variety of electrical faults such as loose connections, overloads, insulation errors, lack of maintenance, and various technical faults. Despite technological advancements in fire detection systems and fire-resistant materials, these panels still pose serious risks in many industrial facilities, occasionally causing material damage and loss of life. A fundamental reason why electrical panels are so risky is that they serve as central nodes of energy systems and are critical points of electrical distribution. Any malfunction in these panels can lead to widespread outages and even major fires. Therefore, reducing the fire risk in these panels, making electrical panels safer, and being able to respond quickly in the event of a fire is of great importance. In this context, the integration of IoT and ML technologies offers a new approach to fire safety by enabling continuous monitoring of electrical panels and conducting risk assessments before a fire occurs. IoT devices can continuously collect data in the panels and transfer this data to a central system. The collected data can be processed by ML algorithms, which can generate alerts in the event of any risk and provide information for preventive maintenance. This allows potential faults to be detected and rectified before they lead to more serious problems. This advanced technology-based approach enhances the management of electrical panels and connected systems and raises fire safety standards. Such a system not only prevents fires but also reduces maintenance costs and enhances operational efficiency. Therefore, this study thoroughly examines how IoT and ML technologies can be integrated into fire safety solutions for electrical panels and how this integration can minimize fire risks. The study aims to reduce fire risks and advance fire safety measures in electrical panels using IoT and ML technologies. The developed system aims to detect potential fire hazards in advance and minimize them, providing real-time monitoring and predictive analysis to make electrical panels safer. For this purpose, the design and implementation of an IoT-based sensor network integrated with advanced machine learning algorithms have been adopted. This methodology includes various sensors equipped to monitor critical environmental and electrical parameters using platforms like Arduino Mega and Raspberry Pi. These platforms continuously record parameters directly related to fire, such as temperature, humidity, carbon monoxide, carbon dioxide, and hydrocarbon gases, as well as electrical variables like current, voltage, and frequency changes. Additionally, a thermal camera has been integrated into the system to continuously measure and record the temperature of the cable inside the panel during the experiment. This integration significantly enhances the sensitivity and response time of fire detection and prevention systems. During the experiments, the collected data were classified based on a heat release graph under four different stages of fire development (S1, S2, S3, S4) and normal (N) conditions. This classification is used to closely monitor and analyze the various stages of a fire from its inception: S1 (non-flaming, pre-ignition conditions), S2 (ignition, flash phase and well flame development), S3 (fully developed fire stage), and S4 (post-fire and decay) conditions. These stages are critical for managing fire risks and developing intervention strategies. For experimental purposes, an electrical panel was constructed with its front cover cut to specific dimensions and sealed with flexiglass. This allowed the experiment conditions to be monitored and recorded externally via camera. The experiments involved a cable between a contactor and a capacitor, with various cable types used: original cable, cut cable, crushed cable, weakened cable, thin cable, etc. Additionally, the capacitors in the panel were diversified, including 5 kvar and 2.5 kvar capacitors. In summary, experiments were conducted with both 5 kvar and 2.5 kvar capacitors. A total of 21 experiments were conducted, with fires occurring in 7 of them, and real data comprising 3521 rows × 11 columns (total 38731 data) were collected. Of these, 3478 rows and 9 columns (total 31302) were included in the analysis methods base cluster. The collected data and thermal images are of great importance for the development of fire detection systems and the optimization of fire prevention strategies. Future studies will aim to enhance the accuracy of fire detection algorithms and improve fire response processes using this data. The integrated system in this study reshapes fire safety management in industrial facilities, providing more reliable and effective solutions that minimize loss of life and property. To process the data obtained from experiments conducted on electrical panels and minimize fire incidents, the academic version of the Matlab program was used. Five different supervised machine learning classification algorithms were selected in Matlab: Decision Trees, Logistic Regression, Gaussian Naive Bayes, Support Vector Machines, and K-Nearest Neighbors. These algorithms were trained on data primarily based on critical features such as carbon monoxide, carbon dioxide, cable temperature, panel interior temperature, and panel humidity, identified through a statistical method called analysis of variance (ANOVA) used for feature selection. ANOVA was used to determine which features provided the most information for fire detection and ultimately led to the decision to use these five features. Each machine learning model was thoroughly tested using a five k-fold cross-validation method, which involves training and testing the model on different parts of the data set to provide a robust evaluation of the model's generalization capability and stability. In five k-fold cross-validation, 80% of the data set was used for training, and the remaining 20% was set aside for testing. Cross-validation measures the performance of each algorithm more objectively and demonstrates the consistency of the algorithm across different data sets. The results demonstrate that the integration of IoT and ML technologies significantly enhances the capacity to detect and predict potential fire events in electrical panels. Notably, the Support Vector Machines and K-Nearest Neighbor algorithms stood out with extraordinarily high accuracy rates exceeding 99%, underscoring their reliability and effectiveness in fire detection. The decision tree model also exhibited impressive results with an accuracy rate of 98%, demonstrating its ability to model complex data structures. While the logistic regression model showed strong performance with a 95% accuracy rate, the Gaussian Naive Bayes model performed less well, with an accuracy rate of 91%. This indicates that certain algorithms may better suit specific data structures or feature distributions, and different models may need to be preferred depending on the characteristics of the data set. The joint integration of IoT and ML technologies in this study offers a new approach to fire safety, not only detecting fire events but also predicting them before they occur. This proactive approach significantly enhances fire safety conditions, especially in high-risk environments such as industrial facilities. The integration of these technologies with traditional fire safety measures transforms fire safety management into a more dynamic and interactive affair, shifting from reactive to proactive and predictive approaches. The use of next-generation technologies allows fire safety systems to play an active role not only during a fire but from the onset of fire risk, enabling real-time collection and analysis of environmental and operational data to identify potentially hazardous abnormal conditions early on. This has great potential to reduce possible damages and especially to protect human life, as well as to identify maintenance needs of panels, which is of great importance not only for safety but also for operational continuity in industrial facilities. Future studies might also explore the integration of automatic extinguishing mechanisms within the IoT and ML framework, which would enable seamless collaboration between fire detection systems and automatic extinguishing systems, providing a prompt and effective response as soon as a fire occurs. This type of integration would strengthen the capabilities of fire risk management systems further.
Except where otherwise noted, this item's license is described as http://creativecommons.org/licenses/by/4.0/
