This paper proposes a condition monitoring (CM) framework to detect the health-to-fault operational behavior of wind turbine (WT) converters based on SCADA data from early-stage operation. A novel method for characterizing operating data distribution is proposed to exploit the limited and unbalanced SCADA data. By providing weighting factors for cost functions, this method significantly improves the robustness and generalization of the CM method. The CM framework is also enabled for the full life cycle including early-stage and long-term operation, by employing an online learning algorithm. Validation with both healthy and faulty WTs demonstrates the potential to detect abnormality a few days before actual converter faults occur.