To address the challenge of efficiently calculating the air gap magnetic flux and optimizing motor geometry, we experimented with several different neural network architectures. The solution revolved around training a neural network to learn the relationship between motor geometry parameters and the resulting flux distribution, thus eliminating the need for complex simulations and reducing computation time.

Review of Related Work

• CNNs for Steady Flow Approximation: Ideas from this research highlighted how convolutional neural networks could model physical phenomena efficiently, influencing the choice of network architectures. [2]
• Physics-Informed Neural Networks (PINNs): The use of PINNs in other applications demonstrated how physics-based constraints could enhance model accuracy by incorporating governing equations directly into the loss function. [3][4]
• Dynamic System Simulation: Techniques from prior work in dynamic systems helped refine the strategy for auto-adjusting geometry parameters using neural networks.[5]

1. Data Collection and Preprocessing

We began by gathering a dataset that included various motor geometries and their corresponding magnetic flux distributions. This data was derived from Finite Element Method Magnetics (FEMM) [1] simulations, which is a common approach for calculating flux distributions. The input parameters for the model included motor geometry features like rotor and stator dimensions, material properties, and other design specifications. These features, along with the corresponding magnetic flux distribution for each geometry, formed the basis of our training data.

2. Neural Network Architecture Design

Given the varying nature of motor geometries, we designed distinct neural network architectures for different types of geometry. We experimented with two major types of geometries:

Flat-type Geometry Neural Architecture: This architecture was designed to model geometries where the rotor and stator were aligned in a flat configuration.

V-type Geometry Neural Architecture: For geometries with a more angled rotor and stator, this architecture was used.

Each network was tailored to capture the specific relationships between the input parameters and the magnetic flux distribution of that geometry type. The architecture for both networks consisted of a series of fully connected layers, with the final layer producing the flux distribution across the air gap.

3. Control Flow for Auto-adjusting Geometry Parameters

To go beyond just predicting flux, we developed a second neural network, NN1, that could automatically adjust the geometry parameters to achieve a desired flux distribution. The input to NN1 was a set of key flux features, such as peak flux values, the angle of peak flux, periodicity, and points where the flux reached specific fractions of the peak value. The output of NN1 was a set of geometry parameters that, when fed into NN2 (the trained neural network), would result in the desired flux distribution.

We synthetically generated training data for NN1 by sampling a wide range of flux distributions and extracting key features. This allowed NN1 to learn the mapping from desired flux features to the corresponding motor geometry parameters.

Once trained, NN1 could take a set of desired flux features as input and output the necessary geometry parameters. These parameters could then be passed to NN2, which would compute the final flux distribution based on the optimized geometry. This approach not only automated the flux prediction but also provided a mechanism for optimizing motor geometry in real-time to achieve specific performance outcomes.