Conference Agenda

A method for the magneto-thermal analysis of an induction heating device based on a deep neural network is proposed. The twofold non-linearity of magnetic permeability against temperature and magnetic field – which characterizes the workpiece – is captured by the model, while the use of a Convolutional Neural Network (CNN), trained by a sequence of finite-element analyses, makes it possible to recover the temperature distribution in the workpiece region at a given time. TEAM problem 36 is considered as the case study. The originality of the paper lies in the synthesis of the transient temperature map, by virtue of a coupled magneto-thermal analysis based on a deep learning approach, which is innovative and allows a dramatic reduction in the computational burden.

The new topology optimization method using an explainable deep neural network is proposed. The proposed method uses the Shapley additive explanations (SHAP), which explains the reason for estimating the convolutional neural network. The explanatory results from SHAP are reflected in the crossover method in evolutionary computation to improve the search performance of topology optimization. The proposed method is applied to an interior permanent magnet synchronous motor, and the effectiveness of the method is discussed.

This study compares and examines various deep learning models that accurately predict iron loss characteristics from rotor shapes of

interior permanent magnet synchronous motors under various speed and current conditions. The best iron loss prediction model enabled

an efficiency optimization design to be performed in a short time of 396 ± 14.4 seconds.

The effective modeling of magnetic hysteresis is crucial in developing accurate digital twins for low frequency electromagnetic systems. However, large three-dimensional analysis systems can require the evaluation of the hysteretic performance at hundreds of thousands of points in the components containing magnetic steels. It is essential that any modeling system can evaluate the hysteretic performance in the shortest possible timeframe. The use of Neural Networks raises the possibility of achieving this goal. This paper reviews the various NN architectures which might be considered to address this requirement.

Multi-objective optimization, in which several design objectives are jointly optimized, is frequently employed in electric motor design, where iterative numerical simulations are required to evaluate a large number of design candidates. A trial-and-error design methodology like this is very time-consuming. In this paper, we propose learning-based surrogate models that use deep neural networks (NNs) to accomplish the rapid evaluation of motor designs. A well-trained NN is capable of accurately predicting the responses of given motors almost instantaneously. A motor design candidate can be described with either a list of geometrical parameters of the motor design, or a colored image of the cross-section of the magnetic design. Different deep learning models can be constructed with either parameter-based input or image-based inputs. We compare and show that deep convolutional neural networks (CNNs) with image-based inputs, while taking longer time to train, can achieve higher prediction accuracy for more complicated responses, as compared with models with parameter-based inputs.