Fault detection in three-phase power transformers using machine learning techniques

Alenezi, Mohammed 2025. Fault detection in three-phase power transformers using machine learning techniques. PhD Thesis, Cardiff University. Item availability restricted.

	PDF - Accepted Post-Print Version Restricted to Repository staff only until 18 May 2027 due to copyright restrictions. Download (7MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (14MB)

Abstract

Power transformers play a vital role in electrical power systems, and their failure can result in costly outages, equipment damage, and compromised grid reliability. Accurate, early fault detection is therefore essential to ensure system stability. However, traditional protection methods—such as differential protection schemes, dissolved gas analysis (DGA), and frequency response analysis (FRA)—struggle to reliably distinguish between internal faults, and external disturbances, especially under dynamic or asymmetric operating conditions. This research presents a hybrid machine learning-based fault detection framework for power transformers that combines empirical experimentation, advanced signal processing, and metaheuristic optimization. A custom three-phase transformer prototype was built and tested at Cardiff University, where a wide range of internal and external fault scenarios were systematically simulated, including inter-turn, turn-to-ground, line-to-line, and phase to ground faults. The experimental setup featured voltage and current acquisition systems under controlled laboratory conditions, enabling the collection of high-fidelity real-world signals. The study employs Discrete Wavelet Transform (DWT) and Matching Pursuit (MP) to extract robust time–frequency features from voltage and current signals. These features form the foundation for classification using multiple supervised learning algorithms, including Decision Trees (DT), Support Vector Machines (SVM), Random Forests (RF), and Logistic Regression (LR). To optimize both feature selection and classifier performance, two bio inspired optimization techniques—Particle Swarm Optimization (PSO) and Dwarf Mongoose Optimization (DMO)—were applied. The PSO-RF model achieved a classification accuracy of 97.71%, while the DMO-RF model surpassed this with 98.33% accuracy, 98.80% precision, and a 99.04% F1-score, demonstrating superior generalization. Furthermore, a voltage-signal-based ML framework was introduced in the final phase of this research (Chapter 6), showcasing a lightweight, interpretable system suitable for real-time deployment. Using raw phase voltage data, the framework was tested on 6000 samples and demonstrated exceptional diagnostic performance: 99.9% cross-validation accuracy and 95% test accuracy, with a false alarm rate as low as 0.47%. The results confirm the feasibility of using voltage-only signals for scalable and cost-effective transformer fault monitoring. The novelty of this work lies in the integration of real-world transformer data, advanced signal decomposition, hybrid optimization, and interpretable Machine learning (ML) models. The proposed framework not only improves classification accuracy but also reduces computational overhead, making it ideal for integration into Supervisory Control and Data Acquisition (SCADA) systems or digital relays. Overall, this study provides a practical and high-performance solution to transformer protection, with significant implications for real time monitoring, grid reliability, and operational safety in modern power systems.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. Power Transformers 2. Faults Detection 3. Inrush Current 4. Machine Learning 5. Classification 6. Optimization Algorithms
Date of First Compliant Deposit:	18 May 2026
Last Modified:	20 May 2026 11:34
URI:	https://orca.cardiff.ac.uk/id/eprint/187083

Actions (repository staff only)

Edit Item