Concept Discovery in Power Electronics Training Using Non-Negative Matrix Factorization
Abstract
Power electronics training involves mastering complex concepts that are crucial for designing and controlling electronic power systems. Traditional teaching methods often struggle to identify and address students' conceptual weaknesses efficiently. This paper proposes a novel approach called Concept Discovery via Non-Negative Matrix Factorization (CD-NMF) to analyze learners' responses and uncover underlying concept patterns. The primary problem is the challenge of extracting interpretable and actionable insights from large-scale training data, which is often sparse and nonnegative by nature. CD-NMF leverages the mathematical properties of NMF to decompose student response matrices into meaningful latent components corresponding to core concepts in power electronics. These components serve to reveal concept mastery levels and identify misconceptions. The interpretability of NMF facilitates educators' understanding of students' learning status without requiring labeled data or complex feature engineering. Experiments on a dataset collected from power electronics training sessions demonstrate that CD-NMF outperforms baseline matrix factorization and clustering methods in terms of concept detection accuracy and educational relevance. The approach successfully identifies distinct concept clusters aligning with curriculum design, providing a practical tool for personalized intervention and curriculum adjustment. In conclusion, CD-NMF is an effective and interpretable datadriven method for concept discovery in power electronics education. It enhances training efficacy by pinpointing knowledge gaps and supporting tailored feedback, potentially improving learning outcomes significantly. Future work will explore integrating CD-NMF with adaptive learning systems to automate personalized curricula dynamically.