Conference Agenda

This paper proposes a novel approach to predict the performance of an interior permanent magnet motor by employing a convolutional neural network (CNN) for different magnet positions and core material distribution. The approach improves the prediction accuracy by using the magnetic flux density distribution vector created by topology optimization as an input to the CNN. It is shown that the proposed method can predict the average torque and torque amplitude with a higher accuracy as compared to the conventional methods.

This work presents a unified framework, within which an iterative finite element solver is directly integrated into the training loop of a neural network on the basis of differentiable programming. We apply this framework to a transient eddy current problem on a coaxial cable with the goal of quantifying the error of a coarse solver, given high resolution trajectory data. To that end, we extend a previously proposed method in [1] from structured, Cartesian grids to unstructured triangular grids.

Learning from examples is the golden rule in the construction of behavioral models using neural networks (NN). When NN are trained to simulate physical equations, the tight enforcement of such laws is not guaranteed by training process. In addition, there can be situations when providing enough examples for a reliable training can result difficult if not impossible. To alleviate these drawbacks of NN, recently a class of NN incorporating physical behavior has been proposed. Such NN are called “Physical-Informed Networks” (PIN). In this contribution, their application to both direct and inverse ElectroMagnetic (EM) problems will be presented. A throughout comparison of different neuron activation functions, to take properly into account local and global field EM characteristics represents the innovation content of this work.

The deep learning has been rapidly progressed. Especially, some deep neural networks for solving partial differential equations are aggressively developed. Such deep learning techniques are called the physics-informed neural networks (PINNs). Although PINNs have great potentials to drastically change the current numerical analysis processes, their current drawbacks are heavy computational cost due to the backpropagation learning. On the other hand, there is a neural network approach called the extreme learning machine (ELM), whose computational cost is relatively light because it does not use the backpropagation-based learning. Thus, it is an interesting approach to combine ELM and PINNs for the fast construction of the neural networks for solving partial differential equations. In this paper, a new physics-informed extreme learning machine (PIELM) for solving magnetostatic problems is presented. The proposed approach is applied to a simple test model. It is shown that the magnetic flux density distributions can be obtained without solving finite element equations and using backpropagation learning.

Traditional electrostatic simulation are meshed-based methods which convert partial differential equations into an algebraic system of equations and their solutions are approximated through numerical methods. These methods are time consuming and any changes in their initial or boundary conditions will require solving the numerical problem again. It is possible to use an autoencoder to capture a finite number of electrostatic field solutions with different boundary conditions. The autoencoder can reconstruct an infinite number of electrostatic field corresponding to respective parameter changes via a regression model in the latent space. However, performing inverse prediction on the regression model is an ill-posed problem. In this work, we propose an end-to-end deep learning approach to model parameter changes to the boundary conditions. The proposed method is demonstrated on the test problem of a long air-filled capacitor structure. The proposed approach is compared to other computational methods such as the physics informed neural net (PINN) and finite difference method (FDM). It is shown that our method can easily outperform both FDM and PINN under dynamic boundary condition as well as retaining its full capability as a forward model

Global efforts to reduce the consumption of Rare-Earth elements (REE) are being made by researchers, companies and governmental entities. In the context of internal permanent magnet synchronous machines (IPMSM), this reduction is achieved either by the use of REE free permanent magnets (PM), such as Ferrite PMs or by reducing the volume of REE PM. In both cases, the reduction of PM torque can be counterbalanced with the generation of reluctance torque. This paper proposes a new design approach to include both PM torque and reluctance torque in an objective function. Additionally, the capacity of a 1-dimension convolutional neural network (1DCNN) is also explored for the first time as an approximation model for Spoke Type Machines (Spoke). The results show that 1DCNN can calculate the torque of a given geometry with an average error of 4.16%. For the electrical machine in analysis, the inclusion of both torque components on the objective function increases the torque per unit of volume of PM.

Using Deep Galerkin method (DGM), one of the deep learning methods for solving partial differential equations, static magnetic field simulation of a model in which a magnetic sphere is placed in a uniform magnetic field was performed. Comparing the simulation results with the analytical solutions shows good agreement. This method has the advantage that it is not necessary to generate a mesh in the computational domain. Moreover, by training deep neural network with physical quantities such as relative magnetic permeability as variables, it was shown that it is possible to recalculate immediately after changing the value of physical quantities.

In the paper, an innovative approach for the fast estimation of the mutual inductance between transmitting and receiving coils for Dynamic Wireless Power Transfer Systems (WPTS) is implemented. To this end, a Convolutional Neural Network (CNN) is used; an image representing the geometry of two coils, which are partially misaligned, is the input of the CNN, while the output is the corresponding inductance value. An analytical approach to the computation of the inductance value allows for a fast training of a first CNN. Subsequently, a transfer learning technique will be used for refining the CNN training, using Finite Element Analyses of the magnetic field. This way, thanks to the fast and accurate mutual inductance estimate, it is possible to appropriately tune the load resistance of the WPTS, acting e.g. on the voltage of a DC bus, in order to maximize the power transfer efficiency.

This paper presents feature extraction and visualization of synchronous reluctance motors (SynRMs) using convolutional neural networks (CNNs) as knowledge to support non-expert designers. We generate SynRM images via topology optimization, to which the classes indicating the magnitudes of the average torque and the torque ripple values are assigned. Using pairs of images and their class information, we train a multi-task CNN that simultaneously predicts the two types of classes in addition to individual single-task CNNs. A class activation mapping method visualizes the features focused by each CNN. The experiments provide qualitative and quantitative comparisons between different CNNs.