Abstract
Power transformers are critical assets in modern power grids, where failures can lead to significant operational disruptions and financial losses. Dissolved Gas Analysis (DGA) is a key sensor-based technique widely used for condition monitoring, but traditional diagnostic approaches rely on deterministic thresholds that overlook uncertainty in degradation dynamics. This paper proposes a probabilistic framework for Prognostics and Health Management (PHM) of power transformers, integrating self-adaptive Auto Regressive Integrated Moving Average modeling with a probabilistic reformulation of Duval's graphical methods. The framework enables automated estimation of fault types and failure likelihood directly from DGA sensor data, without requiring labeled datasets or expert-defined rules. Dissolved gas dynamics are forecasted using time-series models with residual-based uncertainty quantification, allowing probabilistic fault inference from predicted gas trends without assuming deterministic persistence of a specific fault type. A sequential pipeline is developed for real-time fault tracking and reliability assessment, aligned with IEC, IEEE, and CIGRE standards. Two case studies validate the method: one involving gas loss in an experimental setup and another examining thermal degradation in a 345 kV transformer. Results show that the framework improves diagnostic reliability, supports early fault detection, and enhances predictive maintenance strategies. By combining probabilistic modeling, time-series forecasting, and sensor-based diagnostic inference, this work contributes a practical and interpretable PHM solution for sensor-enabled monitoring environments in modern power grids.