Minimum Data Requirements for Data-Driven Reliability-Oriented Design by Active Learning

To achieve data-driven power electronic design, a critical question exists: how much data is required? or questions about the data quality are not sufficiently addressed yet. To address this issue, this paper proposes an active learning framework that identifies optimal training samples with minimal data while achieving the desired accuracy. A case study on the reliability-oriented design (ROD) of a traction converter is presented to validate the proposed approach. Unlike conventional ROD methods that depend on extensive simulations and experimental campaigns, the proposed method integrates surrogate modeling with active learning to substantially reduce the volume of required data without compromising predictive accuracy. Experimental results confirm the effectiveness of the proposed framework.