AI-DRIVEN PREDICTIVE FAULT DIAGNOSIS AND REMAINING USEFUL LIFE ESTIMATION OF POWER TRANSFORMERS USING DEEP LEARNING TECHNIQUES

Abstract

Power transformers are among the most critical and expensive assets in electrical power systems. Their uninterrupted operation is essential for maintaining the reliability, stability, and efficiency of modern power grids. Unexpected transformer failures often result in substantial economic losses, prolonged power outages, expensive maintenance operations, and serious safety hazards. Conventional maintenance practices, such as scheduled inspections and periodic offline testing, are insufficient for accurately predicting transformer failures because they fail to continuously monitor transformer health and detect early-stage degradation. Consequently, electric utilities worldwide are increasingly adopting Artificial Intelligence (AI) and Internet of Things (IoT)-based condition monitoring systems to enable predictive maintenance.

This research proposes an intelligent deep learning framework for predictive fault diagnosis and Remaining Useful Life (RUL) estimation of power transformers using IoT-enabled monitoring and multiple deep learning architectures. Real-time sensor measurements including transformer oil temperature, winding temperature, dissolved gas analysis (DGA), load current, moisture level, vibration, and partial discharge are continuously collected through IoT devices. The acquired data undergo preprocessing, feature engineering, normalization, and feature selection before being supplied to deep learning models including Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Transformer Networks.

The developed framework performs two major functions. First, it identifies transformer fault categories with high classification accuracy. Second, it estimates the remaining useful operational life of transformers to facilitate predictive maintenance scheduling. The performance of different deep learning models is evaluated using Accuracy, Precision, Recall, F1-score, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² Score. Experimental results demonstrate that Transformer-based deep learning models outperform conventional machine learning algorithms in both fault diagnosis and RUL prediction. The proposed framework significantly improves transformer reliability, reduces maintenance costs, minimizes unexpected failures, and supports intelligent asset management in future smart grid environments.