Abstract: The safe and stable operation of power transformers in hydropower stations is crucial for ensuring overall grid reliability. Conventional monitoring techniques, which depend on periodic inspections and single-parameter indicators, often fail to provide timely or accurate insights into transformer health. This study presents an in-depth investigation of a machine learning-driven approach to transformer condition monitoring. We introduce an online framework for multi-source data fusion, develop a feature-extraction module tailored to transformer state data, and implement a dual-model fusion strategy to compute a comprehensive health index. A pilot validation on the main transformer of a selected hydropower station demonstrates that the proposed method achieves high accuracy, robustness, and real-time responsiveness in operational settings.