Hybrid AI-Subject Matter Expert Solution for Evaluating the Health Index of Oil Distribution Transformers

Abstract

Reliability of oil distribution transformers is paramount, ensuring a stable flow of electricity and shielding from potential fire hazards. The internal insulation system of these transformers utilizes a combination of oil and paper. As the oil circulates through the active part of the system, it collects gaseous and physical traces of existing or past defects or degradations, providing a holistic view of the transformer's health, and allowing for early detection of problems and predictive maintenance. While various and mainly data-driven methods have been developed to calculate a transformer health index from oil samples, they lack accuracy due to limited data. This paper proposes a novel hybrid approach that leverages both Artificial Intelligence and Subject Matter Expertise to enhance the health estimation of oil distribution transformers. Our methodology utilizes a substantial dataset exceeding 65,600 analyzed oil samples, coupled with the valuable knowledge of domain experts. This combined approach achieves an accuracy exceeding 95%, suitable for real-world industrial applications. Furthermore, we introduce a risk management feature that strengthens the ability to identify transformers at high risk of failure. Notably, the health index estimation is implemented as a semi-automatic process, retaining the "expert in the loop" principle for managing critical and ambiguous cases.