Physics-Inspired Multimodal Feature Fusion Cascaded Networks for Data-Driven Magnetic Core Loss Modeling

This article proposes a physics-inspired multimodal feature fusion cascaded network (PI-MFF-CN) for data-driven magnetic core loss modeling based on MagNet database. The proposed methodology consists of two cascaded sub-models: the physics-inspired network model and the multimodal feature fusion network model. Firstly, a network model inspired by physics and related micromagnetism, is developed based on the Landau-Lifshitz-Gilbert (LLG) equation. It provides new sequence information (HLLG (t)) for the next cascaded core loss prediction model. This addresses the limitation where H(t) waveforms are unable to participate in the actual prediction process. With embedded physical micromagenetic parameters (A, K, Ms) in the gradient learning process of the neural network, the trained physics-inspired network can be regarded as the inverse model (B(t)→HLLG(t)) of LLG Equation having physical interpretability. Then, in order to address a series of challenges in multimodal information learning, a multimodal feature fusion-based network model is proposed. This approach combines the advantages of convolutional neural network (CNN) and fully connected neural network (FCNN) to learn hybrid sequence-scale data. Specifically, it employs parallel CNN branches for sequence feature mappings, followed by concatenating these mappings with other scalar data into an FCNN for global learning. To validate the effectiveness of the proposed method, this article trains and optimizes the proposed models based on MagNet database, and then a series of experiments including extensive material validation (Ferroxcube-3C90, 3C94 & TDK-N27, N30, N49, N87, etc.) were carried out. A series of experimental outcomes demonstrate that the proposed PI-MFF-CN-based method is generalized and robust in accurately predicting magnetic core losses.