open access

https://doi.org/10.14943/doctoral.k16021

https://hdl.handle.net/2115/94189

Research on advanced motor design by integrating topology optimization and artificial intelligence techniques

日本語

deep learning

2024-03-25

Hokkaido University

ⅹ, 85p.

In recent years, environmental issues have led to an ever-increasing demand for higher performance and efficiency in electrical equipments. Among them, motors such as permanent magnet motors are said to account for more than half of the electric energy consumption for industory in Japan, making higher performance and efficiency an important social issue. Various shape optimization techniques have been studied to meet this demand. Among them, topology optimization has been attracted much attention in recent years because it does not require the design of dimensional parameters that were previously necessary, and it has the potential to produce novel shapes that exceed human knowledge. However, shape optimization, including topology optimization, has several challenges. I. Inability to consider basic parameters of the device: Shape optimization targets only the geometry of the device components, but in reality, the performance of the device also depends on more basic parameters (such as the size of the device and the number of poles). Simultaneous optimization of these basic parameters and shape of components can lead to more innovative devices. Hereafter, we refer to this type of optimization as integrated optimization. II. Increased computational cost: Stochastic optimization methods such as genetic algorithms, which are often used for shape optimization including topology optimization, have the advantage of not requiring gradient computation and having high global search ability. On the other hand, they require numerical analysis of a large number of shapes. The finite element method (FEM) is commonly used as a numerical analysis method. However, repeated application of this method increases the computational cost and lengthens the design process. Although surrogate model methods that replace FEM with other fast models such as machine learning models are already known as remedy for such problems, the limits of their applicability are still unknown, and further research is essential for more advanced design, including topology optimization and integrated optimization. On the other hand, Artificial Intelligence (AI) technology has been rapidly advancing in recent years, and has already demonstrated superior performance over humans in several fields, such as image recognition, natural language processing, and playing games. This technology can be applied to the optimal design of motors to solve the above issues. Based on the above, this study proposes an advanced method of motor design that combines shape optimization, especially topology optimization, with AI technology. It is shown that integrated optimization of motors can lead to the discovery of innovative structures. In addition, we aim to improve and accelerate the motor design by using surrogate model methods based on deep learning. To solve the problem described in I., a tree search technique which is known as game AI technique is applied. In AI for games such as Go, each board situation of the game corresponds to a node in a tree structure, and the best move is searched and determined by examining the nodes in the manner of tree search. By drawing an analogy between game play and electrical equipment design, by mapping the basic parameters of the equipment to the nodes of a tree structure, it is possible to select the optimal basic parameters. Topology optimization is performed based on the selected parameters, and the results are recorded in the tree structure. By repeating this process, the optimal combination of basic parameters and shape of components can be obtained. This method is applied to the multi-objective optimization of a permanent magnet motor. To solve the problem described in II, the surrogate model methods are applied. There are two types of learning methods in the surrogate model method: offline learning and online (adaptive) learning. The offline learning method speeds up the main optimization stage by building a powerful performance predictor using the dataset preliminary collected. While this method has a high speedup ratio, its performance is highly dependent on the collected data. On the other hand, the online learning method uses data obtained during the optimization process to successively build predictors and apply them to the optimization. The speedup ratio for this method is not as high as the offline learning method, but it does not require data collection through pre-optimization. A comparison of the results by these two methods is presented. On the other hand, topology optimization is a high-dimensional problem, and even if the surrogate model method is applied, the computational cost and difficulty of the problem due to the high dimensionality cannot be solved. On the other hand, the variational autoencoder, one of deep generative model, can be applied to represent a wide variety of motor geometries in a low-dimensional latent space to achieve low-cost topology optimization. This method is applied to various motor geometries.

2024-03-25

甲第16021号

博士(情報科学)

(主査) 教授 五十嵐 一, 教授 野口 聡, 教授 北 裕幸, 教授 近野 敦

情報科学院(情報科学専攻)

博士論文

Version of Record