Deep learning model for enhanced power loss prediction in the frequency domain for magnetic materials

Abstract

AbstractThis paper outlines the methodology for predicting power loss in magnetic materials. It starts by introducing the concept of core loss and the complexity of modelling it. Steinmetz's equation is presented to calculate power loss based on frequency and magnetic flux density, but its limitations are highlighted. As an alternative, a neural network‐based method is introduced. The proposed methodology adopts a long short‐term memory network, expressing the core loss as a function of magnetic flux density, frequency, temperature, and wave classification. Fast Fourier transform was implemented to reduce the data points of the sampled flux density waveform while preserving its characteristics. Analyzing in the frequency domain enabled streamlining the training of the model. The input features were arranged as required, and the network architecture was designed with appropriate layers and optimal activation functions. Through extensive training using the datasets, the model assimilated intricate relationships between input variables and known power loss. Evaluation and validation metrics were subsequently employed to gauge the performance of the trained network. This innovative methodology aims to significantly augment the precision of power loss predictions, providing valuable insights into the nuanced behaviour of magnetic materials.