Conference Agenda

To precisely calculate iron losses of permanent magnet synchronous motors (PMSMs) under the influence of multiple physical fields, an advanced deep learning based estimation strategy is proposed, which combines convolutional neural network (CNN) with network-based deep transfer learning (NDTL). The advantage of the proposed model lies in its ability to accurately address iron loss calculation challenges across various operating conditions, even with a limited amount of sample data. Specifically, simulation data encompassing various loads and influencing factors are first generated using an analytical iron loss model, followed by designing a pre-trained CNN-based iron loss model to efficiently extract pertinent features from the data. Subsequently, an NDTL algorithm is employed to fine-tune the key parameters of the CNN model enabling it to adapt to novel feature distributions and thereby substantially enhancing model performance. Finally, the effectiveness of the proposed NDTL-CNN approach is tested using data from various states on an actual PMSM applied in electric vehicles. Experimental results indicate that the model using the proposed algorithm achieves high estimation accuracy and good practicality, solving the problem of poor generalization ability of traditional deep neural network approaches.

Parametric surrogate models of electric machines are essential for efficient design optimization and operational monitoring. Addressing geometry variations, spline-based computer-aided design representations play a pivotal role. In this study, we propose a novel approach that integrates isogeometric analysis with deep learning to enable rapid and physically consistent predictions by directly learning spline basis coefficients. The effectiveness of this method is demonstrated using a parametric nonlinear magnetostatic model of a permanent magnet synchronous machine.

Partial discharges pulses are phenomena of localized ionization which occurs in the presence of high electric fields in limited portions of an insulating system, without leading to a complete electrical failure. These events can manifest themselves in different forms, depending on the location and configuration of the defect within the insulating material. The techniques traditionally used for interpreting the data collected include the use of time and frequency domain analysis, and are distinguished according to the nature of the voltage applied (AC or DC). Recognition and classification of discharges in DC still remains an open problem. In the present paper, Artificial learning is used to classify three different types of experimental partial discharges. Machine learning algorithms, including Decision Trees and Discriminat Analyses, and a neural network of the LSTM type have been implemented. The optimization of the hyperparameters of the various algorithms provided high levels of accuracy while maintaining low computational costs.

A physics-informed neural network (PINN) is developed to find the bases of the unknown source in the linear inverse sources problems (ISPs). The singular value decomposition (SVD) is applied to the Green’s matrix of the ISP system. The PINN solves a set of bases of the unknown source by using the projection of the measurement data onto the left singular vectors from the SVD as the label data, and the PINN training results show that the right singular vectors from the SVD can serve as a set of bases of the unknown source in the linear ISPs. The conclusions benefit future explorations on the SVD-based solution approaches for linear ISPs.

Fault diagnosis in wiring networks is crucial for reliable industrial system operation. This paper introduces a novel method combining Time Domain Reflectometry (TDR) with the Inception-ResNet V2 network for fault diagnosis. By converting TDR signals into images, resized images are fed into the Inception-ResNet V2 network using transfer learning. The benefits of this method include efficient convolutions for real-time detection, and no need for predefined transformations or manual feature extraction. Its feasibility is proven with application to a common wiring network configuration.

The state-of-the-art design processes for electrical machines (eMachine) are template-based methods and design from scratch approaches. Over the past four decades, significant advancements have been made to make these methods effective in terms of electromagnetic performance. However, these techniques cannot provide information on the topology and the components of an eMachine from a given CAD drawing. This information is imperative for the automatic selection of an appropriate simulation profile to meet the required KPIs, enhancing the numerical computation of electromagnetic fields, and improving the user experience through an improved user interface. This paper uses object detection to identify the components of an eMachine and uses that information to classify the topological category.

This paper explores the integration of datasets with various fidelities and transfer learning in designing electric machines (E-machines). By reusing knowledge from prior tasks, transfer learning reduces the cost and time of creating datasets and training mod-els. We test the approach on various E-machine types and fidelity data. The results show that transfer learning enhances efficiency, generalizes well across tasks, and works with smaller and low-fidelity datasets. This provides a scalable and cost-effective solution for the inverse problems of E-machine design.

This paper introduces a novel approach to automatic unstructured mesh generation, leveraging deep neural network to predict an optimal finite element mesh distribution. A posteriori error estimation method is employed to generate training samples. Compared to traditional mesh adaptation techniques, this approach generates an optimal mesh in a fraction of the time. We demonstrate its performance using standard test problems, such as Poisson’s equation. The results show that, in comparison to conventional adaptive mesh generation methods, the neural network approach is highly competitive, striking an excellent balance between accuracy and computational cost while effectively predicting the mesh density distribution across the solution domain.

A cascade deep-learning approach is proposed for optimizing the design and control of a dual-frequency induction heating system used in semiconductor manufacturing. The system is composed of two independent power inductors, fed at different frequencies, to achieve a homogeneous temperature profile along a graphite susceptor surface, crucial for enhancing layer quality and integrity. The optimization process considers both electrical (current magnitudes and frequencies) and geometrical parameters of the coils, which influence the power penetration and subsequent temperature distribution within the graphite disk. A two-step procedure based on Deep Neural Networks (DNNs) is employed. The first step, namely optimal design, identifies the optimal operating frequencies and geometrical parameters of the two coils. The second step, namely optimal control, determines the optimal current magnitudes. The DNNs are trained using a database generated through Finite Element (FE) analysis.

In the design and optimization of motors, accurately predicting key performance indicators (KPIs) such as torque, temperature, and stress is crucial for ensuring efficiency, reliability, and longevity. In this study, the application of hybrid data driven methods to predict these the torque, stress, and temperature for motors with varying geometries is explored. Specifically, a combination of Gradient Boosting and Feedforward Neural Networks (FNN) has been employed to leverage the strengths of both models in capturing complex, non-linear relationships within the data. Gradient Boosting excels in handling tabular, structured data and identifying important features through boosting, while FNNs provide powerful modeling capabilities for learning intricate patterns. To identify the most relevant features for each KPI, a correlation matrix was utilized to analyze dependencies among input variables, ensuring that only the most influential features were used for each prediction task. This hybrid approach effectively combines the advantages of both models, leading to accurate predictions of torque, temperature, and stress. The correlation-based feature selection further enhances model performance by reducing noise and improving interpretability.

Utilizing Neural Networks to identify the regions for topology optimization (computed by numerical analysis) associated with certain performance parameters using the Grad Cam method is on the bleeding edge of research these days. This method seems to work very well when the topology to be optimized is generally well defined and the small changes do not give rise to a large variation of the topology. When that happens the identification of the optimization target regions becomes difficult as the activation map of the neural network by Grad Cam becomes highly different in each design and difficult to interpret. In this paper we propose a method to handle the large topology variation and provide straightforward interpretation of the Grad Cam heat maps which would otherwise be difficult to manage for a large set of very different geometries. The principal idea of the proposed method is the averaging of the activation heat maps for all designs not only for a single picture as it is typical for Grad Cam. Another approach is also proposed based on providing an animation of the Grad Cam heat plots which will be explained in more details in the full paper.

Transcutaneous Energy Transmitters (TETs) are innovative devices designed to wirelessly transfer energy from an external source to an artificial organ (AO) implanted within the body, utilizing inductive coupling between external transmitter and internal receiver coils. A critical objective in TET design is to minimize device size while ensuring reliable and efficient power delivery to the AO across varying operational conditions, such as changes in coil alignment and distance. On the other hand, broad learning system (BLS) models, grounded in pseudo-inverse theory and compressed sensing, offer a robust alternative for regression tasks, supported by mathematical proofs of their universal approximation property. By employing a horizontal architecture comprising feature nodes and enhancement nodes, BLS models deliver a powerful framework for capturing intricate patterns, presenting a compelling option for deep learning applications. To address the limitations of single BLS in terms of generalization ability and robustness to noise, this study proposes an optimized bagged BLS (BagBLS) model combined with Tree-structured Parzen Estimator hyperparameter optimization based on Tree-structured Parzen estimator to TETs modeling. The consistently high coefficient of determination scores obtained from BagBLS studies indicate the degree to which our methodology captures the underlying patterns in the TETs data.

Manta Ray Foraging Optimization (MRFO) is a nature-inspired metaheuristic optimization method that mimics manta ray behavior while foraging in the marine environment proposed originally to single-objective optimization. It aims to solve complex optimization problems by replicating how manta rays find and grab their prey. In this paper, the numerical results are reported to validate a novel multi-objective optimization algorithm inspired by MRFO (MOMRFO) to solve the Testing Electromagnetic Analysis Method (TEAM) workshop benchmark problem 25. Simulation results demonstrate the validity of the proposed MOMRFO to find nondominated solutions that represent good trade-offs among the objectives in the evaluated problem. Furthermore, the results demonstrate that MOMRFO offers superior performance in terms of convergence speed, solution diversity, and overall optimization quality, making it a promising tool for solving complex electromagnetic problems.