MagLearn – Data-driven Machine Learning Framework with Transfer and Few-shot Training for Modeling Magnetic Core Loss

In response to the MagNet Challenge 2023, this paper describes the solution developed by the University of Bristol team, awarded the 3<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rd</sup> Place Outstanding Performance among 24 competing teams worldwide. The core loss of magnetic components has been a challenge for engineers to model due to the lack of full physics models. Classic Steinmetz-Equation-based approaches show significant limitations under power electronics excitations. Data-driven approaches have emerged in recent years as a new solution to this problem as an active research area. Based on the datasets supplied by PowerLab Princeton, this work employs a machine learning framework to predict the core loss of magnetic components from a range of flux density waveforms, e.g. sinusoidal, rectangular, trapezoidal, as the input. The proposed approach builds on an LSTM neural network to extract features from the input B waveforms and predict the power loss value. Designed for the small and imbalanced datasets supplied in the competition, a machine learning pipeline is proposed in this work featuring transfer learning and few-shot training, which is realized through data augmentation and alignment. As a modification to decouple the output from the phase shift of the input waveform, a random shift/flip algorithm is applied in both pre- and post-processing blocks. The performance of the proposed approach is validated through the experimentally measured testing sets, demonstrating a high prediction accuracy.