Machine learning-based identification of cellulose particle pre-bridging and bridging stages in transformer oil
The deterioration of transformer oil quality is influenced by factors including the presence of acids, water, and other contaminates such as cellulose particles and metal dust. The dielectric strength of the oil decreases over time and depending on the service conditions. This study introduces an ef...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Science and Information Organization
2025
|
| Online Access: | http://eprints.utem.edu.my/id/eprint/29354/2/0075804092025153459.pdf http://eprints.utem.edu.my/id/eprint/29354/ https://thesai.org/Downloads/Volume16No3/Paper_37-Machine_Learning_Based_Identification_of_Cellulose_Particle.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The deterioration of transformer oil quality is influenced by factors including the presence of acids, water, and other contaminates such as cellulose particles and metal dust. The dielectric strength of the oil decreases over time and depending on the service conditions. This study introduces an efficient machine learning method to classify the pre-bridging and bridging stages by analyzing the formation of cellulose particle bridges in synthetic ester transformer oil. It is important to note that the prebridging and bridging stages indicate a pre-breakdown condition. The machine learning approach implements the combination of digital image processing (DIP) technique and support vector machine (SVM). The DIP technique, specifically the feature extraction method, captures the feature descriptors from the
cellulose particles bridging images including area,
MajorAxisLength, MinorAxisLength, orientation, contrast,
correlation, homogeneity and energy. These descriptors are used in SVM to assess the pre-bridging and bridging stages in transformer oil without human intervention. Various SVM models were implemented, including linear, quadratic, cubic, fine Gaussian, medium Gaussian, and coarse Gaussian. The results achieved 96.5% accuracy using quadratic and cubic SVM models with the eight feature descriptors. This research has significant implications, allowing early detection of transformer breakdown, prolonging transformer lifespan, ensuring uninterrupted power plant operations, and potentially reducing replacement costs and electricity disruptions due to late breakdown detection. |
|---|