Intelligent Interpretation of Dissolved Gases in Transformer Oil With Electronic Nose and Machine Learning

• Published in: IEEE transactions on industrial informatics Vol. 21; no. 4; pp. 2839 - 2848
• Main Authors: Govindarajan, Suganya, Devarajan, Harimurugan, Ardila-Rey, Jorge Alfredo, Cerda-Luna, Matias Patricio, Araya, Sergi Leandro Torres, Diaz, Cristhian Camilo Delgado
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.04.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Acetylene; Classification; Condition monitoring; Dielectric oil; Dissolved gas analysis; Dissolved gases; electronic nose; Electronic noses; Ethane; Gas analysis; Gases; Genetic algorithms; Machine learning; Metal oxide semiconductors; Mineral oils; Minerals; Monitoring; Oil insulation; Oils; Pollution measurement; power transformer; Sensor arrays; Support vector machines; Transformers
• ISSN: 1551-3203; 1941-0050
• DOI: 10.1109/TII.2024.3507943

Dissolved gas analysis (DGA) is crucial for identifying incipient failures in transformers by analyzing gas concentrations due to degradation. However, its high cost and time-consuming nature limit practical use. To address this, a metal-oxide semiconductor based electronic nose (E-nose) is utilized in this study to detect gases in transformer oil, including hydrogen (H 2 ), methane (CH 4 ), ethane (C 2 H 6 ), ethylene (C 2 H 4 ), and acetylene (C 2 H 2 ). Machine learning techniques are integrated with the E-nose system to enhance classification performance. Experimental results using artificially contaminated mineral oil samples demonstrate promising accuracy in gas classification. Initially, without feature reduction, the F1 score was 0.2972. Feature ranking increased the F1 score to 0.7956, and after implementing dimensionality reduction, it further improved to 0.9313. Subsequently, the combination of support vector machine and genetic algorithm was employed for sensor selection, achieving an F1 score of 0.9869. Among the combinations of 2, 3, and 4 sensors, MQ 8 and TGS 2612 consistently showed the best F1 scores, with TGS 813 and TGS 2611 also contributing significantly. This innovative approach suggests a potential solution for transformer oil condition monitoring, offering a rapid, simple, and cost-effective alternative to traditional DGA analyses. By combining E-nose technology with machine learning, this method holds promise for facilitating routine measurements and ensuring the reliability and efficiency of transformer operations.