Vision Measurement for Levitation Gap of Magnetic Levitation Ball Based on Transfer Learning

Abstract

In recent years, with the rapid development of deep learning, convolutional neural network has been widely used in pattern recognition, target classification and other fields. This paper takes maglev system as the research object and proposes a new method for levitation clearance measurement, aiming at the phenomenon of under-fitting in the trained convolutional neural network when the training data is relatively small. This paper introduces the principle of migration learning, using the VGG16 network trained in the ImageNet competition which retains the weight and network architecture of the convolution part of VGG16 and reconstructing the fully connected network part by conducting training with the training set. The experimental results show that the suspension gap measurement achieved by transfer learning has high measuring precision and good measuring effect. In the training set, the maximum error of VGG16 network using transfer learning was 0.198mm and the average error was 0.066mm.In the test set, the maximum error was 0.198mm and the average error was 0.073mm.On the full data, the maximum clearance error is 0.198mm, and the average clearance error is 0.069mm.