Rotor Winding Image Detection Method Based on Model-Based Transfer Learning

Abstract

Rotor is a core component of the electric motor. The qualification of rotor winding is one of the core factors for the proper functioning of the rotor, which is still detected by manual operation. Hence, it is important to achieve automatic detection of the rotor windings and enhance the detection accuracy. Recently, convolutional neural network (CNN) has been successfully applied to image recognition., but it requires a large number of labeled samples and there is almost no dataset bias between the target dataset and the source dataset. The challenges of using CNN to recognize rotor winding are that the winding image dataset of different types of rotor exist large dataset bias and the labeled examples are limited. We proposed a new model-based transfer learning method to deal with the challenges. To solve the dataset bias problem., we proposed a new image binarization method to get binary rotor winding images. Using the binary images to train and test model can significantly reduce the interference of dataset bias. Meanwhile., we proposed a method to build model-based transfer learning model which is based on the pre-trained Inception-V3 model trained with the ImageNet dataset., the method is used to solve the problem of limited labeled samples. The comparing experiments show that the model-based transfer learning model trained and tested with binary images significantly outperform existing other models., and can achieve stable and accurate detection of the rotor images.