The Research On Qualification Detection Algorithms Of Rotor Winding Parts Based On Dcgan And Attention Mechanism

The motor rotor has a huge demand in modern industrial,as to the problem caused by mainly using manual inspection for the qualification test of winding shape formed by commutator and enameled wire,which is low efficiency and proned to be disturbed,this paper is devoted to the study of a new fast and accurate detection algorithm for rotor winding qualification.The main research contents are as follows:(1)An on-line qualification detection process for rotor winding is designed.Through the morphological analysis of the part to be tested,and with the combanation of the precise registration of the control system and the transmission system,the rotor winding image is positioned and segmented by the hook template to obtaine the rotor winding part with less influence on the deflection angle,and then the qualification test of whole rotor can be finished in one rotation cycle by the detection algorithm based on deep learning.(2)A visual attention mechanism is studied to improve the processing of image weak features.In view of the fact that it is difficult to extract those features when the part image of rotor winding is similar to the background area,and the attention mechanism of the mixed spatial information and channels information is used to enhance the extraction of rotor winding shape information,which can suppress some irrelevant details at the same time.Experiments show that the method can strengthen the detection of the qualification of weak reflective parts.(3)A generative adversarial network is studied to deal with the problem of sample imbalance.In order to reduce the influence of under-fitting caused by insufficent training negative samples on the detection model,a modified generative adversarial networks is constructed by dynamic game between recognizer and generator to expand the unqualified shape samples,such as missing and broken parts.Experiments show that this method effectively increases the diversity of negative samples,and improve the performance of the detection model.(4)A residual network classification model based on sample expansion and visual attention mechanism is proposed.Based on the framework of residual network,an online detection and classification model is designed by using DCGAN and visual attention mechanism,which can accurately and effectively identify the qualification of winding shape formed by commutator and enameled wire.Experiments show that the final detection accuracy can reach up to about 98%,which meets the requirements of real-time online detection.