A confidence level estimation technique for reliable Data-Driven FRA interpretation in air-core winding faults

Frequency response analysis (FRA) data interpretation is among highly challenging tasks in transformer fault diagnosis. Among different quantitative methods, employing statistical indicators on FRA results for mechanical integrity assessment is well-stablished. However, the boundary values of indices that differentiate normal/healthy, suspicious, and abnormal transformer operating conditions are not yet agreed upon. This will make the FRA data interpretation using SIs subjective and unreliable. In this study, data-driven critical boundaries of the most frequently used SIs are introduced. To address the uncertainty of the decision boundaries, a confidence level estimation technique is developed to quantify the probability of a transformer being in a particular operating condition. Reporting a confidence level along with the classified conditions facilitates objective decision-making that can support utility operators during on-site transformer evaluation. The confidence level estimation technique proposed in this work was inspired by bolstered error estimation method widely applied in pattern recognition. Moreover, this study provides a new comprehensive technique of transformer winding fault modeling using SPICE netlists. The proposed analytical winding model is validated through practical FRA measurements with high precision. It is further used to generate various scenarios of common active part faults, such as disk space variation, radial displacement, and disk tilting. Consequently, the proposed method of confidence level estimation is examined on different winding deformations and demonstrated a high potential for comprehensive transformer winding diagnosis.