Conference Agenda

In designs of electric machines, the number of slots and other integer design variables must appropriately be set in addition to size variables. In short, the whole design problems of electric machines are mixed-integer ones, which are generally difficult to solve. To automatically obtain whole designs of electric machines, in this paper, a multi-objective automatic design method for electric machines is presented based on the tree search approach. This method enables to seek for Pareto solutions through the reinforcement learning-based tree search using various design data. In this work, the present method is applied to a design problem of an induction motor. It is shown that the whole designs of the induction motor can be obtained considering the trade-off relation of the efficiency and the total mass, including the appropriate selections of the number of slots in the stator and rotor. In addition, the validity of the resultant design is evaluated by the finite element method, and the effectiveness of the present method is discussed.

Optimization of electrical machines often requires the evaluation of performance by means of finite element analysis (FEA) whose computational cost is demanding. We show how certain data-centric techniques can be employed in the development of high-accuracy global approximators. Unlike online surrogate models, which partially rely on FEA for increased precision, our method is purely statistical. In order to showcase the process, we analyze the optimization of a motor based on those propelling Tesla passenger cars.

After the creation of a simulation training database, which maps geometric design variables to the objectives of interest, the samples undergo several pre-processing transformations. We show how statistical reasoning and augmented data can be effective in creating robust, surrogate-assisted optimizers. Performance of the model is analyzed and illustrative results are discussed.

In paper presents a novel visualization using deep neural network to interpret the result of topology optimization (TO). A convolutional neural network (CNN) is trained to predict the characteristics of a machine from its image. Then, the trained CNN predicts the variation of a machine characteristic for the change in a local structure. This variation directly represents the degree of influence of the local structure on the machine characteristics. This method is applied to a permanent magnet motor resulted from TO. The proposed method is shown to successfully provide significant regions for the motor characteristic.

This paper presents a novel approach to building vector basis functions used to model current in planar electromagnetic solvers. The approach utilises Physics Informed Neural Network framework to solve partial differential equation problem which yields a solution for building vector basis functions on a generalised polygonal domain. This work demonstrates the capability of PINNs to generate concave and non simply connected vector basis functions and, focusing on quadrilaterals, the capability of a PINN model to store the information about a shape of vector basis functions on a set of different domains.

This paper presents a novel approach for monitoring and diagnosis of electrical faults in induction motors. It is based on a hybrid system consisting of the multiple coupled circuit model and a neural network. This combination enables a machine learning-based classification of different electrical fault cases without the need for measurements of real fault conditions. For this purpose, the multiple coupled circuit model is specifically adapted to the respective motor by identifying the model parameters using the differential evolution algorithm by matching the simulation results of the modeling to real measured data of the induction motor.

In this paper, a new approach to locate lightning striking points is proposed that makes use of the voltage waveform registered by sensors placed on transmission lines. The method takes advantage of the use of Neural Networks (NN); the time domain voltage waveforms are taken in the frequency domain with a Discrete Fourier Transform. Then, to reduce the dimensionality of the problem, a principal component analysis is performed that passes only a feature vector with 10 components to the NN. Preliminary results show that the Location Accuracy (LA) of this approach is far better than traditional systems.

This study presents a novel alternative flux model generation method based on Bayesian algorithms and neural networks. In the proposed method, an alternative flux model of hybrid-field motors is developed using single-layer neural networks. In addition, learning data for neural networks are generated based on the Bayesian approach. Using the proposed method, machine learning models with high estimation accuracy can be developed using a small number of magnetic analyses. To validate the effectiveness of the proposed method, it is applied to a verification model of a hybrid-field motor.

We propose a streamer discharge model based on physics-informed neural networks (PINN) to improve the computational efficiency with regard to the classical approach. The Poisson equation and the convection-diffusion equation are trained to generate sufficient data and construct a deep operator neural network (DeepONet). The performance of the PINN, in terms of accuracy, is analyzed by applying it to different datasets (electron density and potential distribution) and comparing to a reference solution (spatial evolution of electrons).