Improved you only look once 8n algorithm-based rotor attitude detection for a spherical motor.

Abstract

Rotor attitude detection (RAD) is one of the key technologies to control permanent magnet spherical motors (PMSpM). This paper proposes an improved you only look once v8n (YOLOv8n) based RAD method for a PMSpM. The visual image datasets collection and annotation method are described, and three different visual feature objects are set for the RAD. To increase the real-time performance of the YOLOv8n algorithm, the backbone and head of the YOLOv8n network are improved by introducing the FasterNet and c2fFaster modules. PyTorch 2.0.1 and CUDA 11.8 with Python 3.11.4 are used to conduct the verification experiments. The models are trained on both the custom PMSpM datasets and the publicly available datasets COCO 2017. The verification shows that the inference time of the improved YOLOv8n algorithm can be as short as 4.31 ms for the PMSpM custom datasets on the basis of ensuring visual object detection accuracy. Thus, the performance of the improved YOLOv8n algorithm is significantly better than that of the YOLOv8n and the multiobject Kalman kernel correlation filter (MKKCF) algorithm. The PMSpM rotor attitude can be calculated using the positional relationships of three visual objects captured by a single monocular industrial camera. To verify the precision of the proposed RAD method, the contact RAD method is taken as the benchmark for a comparison between the MKKCF-based RAD method and the improved YOLOv8n-based RAD method. Typical PMSpM spatial motion experiments demonstrate that the proposed improved YOLOv8n-based RAD method achieves higher precision than the MKKCF-based visual RAD method.