MagNet: A Machine Learning Framework for Magnetic Core Loss Modeling
Haoran Li, Seungjae Ryan Lee, Min Luo, Charles R. Sullivan, Yuxin Chen, Minjie Chen
Abstract
This paper presents a two-stage machine learning framework – MagNet – for magnetic core loss modeling. The first stage of MagNet is a waveform transformation network, which generates 2-D images (tensors) and extracts both the frequency and time domain features from the magnetic excitation waveforms; the second stage of MagNet is a convolutional neural network (CNN), which is trained to recognize the patterns in the 2-D images and predict the core loss based on regression. MagNet is supported by a hardware-in-the-loop (HIL) data acquisition system. The system can automatically generate a large amount of data to train the neural network models. MagNet achieved an average relative error of around 5% for single-frequency core loss prediction. In addition to experimental measurements, MagNet can also be trained with data provided on the datasheets of magnetic materials to improve the accuracy.