TLDR: This study introduces KS-Net, a deep learning model, to classify the rotor shapes (2D, V, Nabla) of Interior Permanent Magnet Synchronous Motors (IPMSMs) using electromagnetic parameters (Ld, Lq, φa). Compared to traditional methods like Finite Element Method (FEM), this machine learning approach offers a faster and more cost-effective alternative. KS-Net achieved 99.98% accuracy, demonstrating competitive performance against other machine learning algorithms like Cubic SVM (100% accuracy) and Quadratic SVM (100% accuracy), proving the viability of data-driven methods for motor design and fault diagnosis.
Electric motors are at the heart of countless modern systems, from electric vehicles to industrial machinery. Among these, Interior Permanent Magnet Synchronous Motors (IPMSMs) are highly valued for their efficiency and precise control capabilities. A crucial factor determining an IPMSM’s performance is the geometric structure of its rotor. Traditionally, analyzing these rotor shapes has involved complex and computationally expensive methods like the Finite Element Method (FEM).
However, a new study introduces an innovative, data-driven approach to classify IPMSM rotor shapes, offering a faster and more cost-effective alternative. Researchers Kıvanç DOÄžAN and Ahmet ORHAN developed a custom deep learning model called KS-Net, which can determine the rotor type (specifically 2D type, V type, or Nabla type) based on the motor’s electromagnetic parameters: d-axis inductance (Ld), q-axis inductance (Lq), and static magnetic flux (φa).
The study highlights that rotor shape directly impacts key motor characteristics such as torque production, efficiency, and magnetic saturation. Therefore, accurately predicting this shape from motor parameters is vital for both optimizing new designs and monitoring existing systems for potential faults. Machine learning algorithms have increasingly found applications in electric machines for tasks like parameter estimation and fault diagnosis, and this research extends their utility to rotor shape prediction.
To evaluate KS-Net, the researchers used a balanced dataset of 9,000 samples, with 3,000 samples for each of the three rotor topologies. These samples were generated through systematic simulations and characterized by the three electromagnetic parameters. The KS-Net model, a deep feed-forward neural network with three hidden layers (128, 64, and 32 neurons, using ReLU activation), was compared against several established machine learning algorithms: Cubic SVM, Quadratic SVM, Fine KNN, Cosine KNN, and Fine Tree. All models were trained and tested using 10-fold cross-validation in the MATLAB Classifier Learner interface.
The results were highly impressive. Cubic SVM and Quadratic SVM achieved perfect classification, identifying all samples flawlessly with 100% accuracy. The proposed KS-Net model demonstrated remarkable competitiveness, achieving 99.98% accuracy with only two misclassifications. Fine KNN also performed exceptionally well with 99.99% accuracy, making just a single error. Cosine KNN achieved 99.93% accuracy, while the Fine Tree model, though still high, showed the lowest performance at 99.06%.
These findings underscore the significant potential of data-driven approaches in electric motor design. By accurately predicting rotor shape from electromagnetic parameters, engineers can accelerate design processes, develop automated rotor identification systems, and lay the groundwork for data-driven fault diagnosis. The KS-Net model, in particular, offers a robust and efficient alternative to traditional, resource-intensive FEM-based analyses.
Also Read:
- Advanced Motor Diagnostics: A Hypergraph Contrastive Approach to Multimodal Fault Detection
- AI-Driven Floorplanning for Better Analog Circuit Routing
The researchers plan to further enhance KS-Net’s generalization capability by training it on larger and more diverse datasets, optimizing its hyperparameters, and incorporating additional parametric inputs related to rotor geometry. Future work also aims to move beyond classification to optimize rotor design using regression models and develop fully automated design systems. For more details, you can refer to the full research paper here.
