Research On Intelligent Defect Detection Algorithm Of High Frequency Transformer Based On Deep Learning

The appearance defect detection and classification of high-frequency transformer produced by an electronic company are all done manually at present,which have some problems such as false detection,low efficiency,being influenced by people’s emotion and fatigue,and difficult to guarantee the consistency of detection.The machine vision detection system is developed,but the system is based on the traditional vision detection principle,the feature extraction is too dependent on human experience,and the defect notification is mainly surface feature,so the hidden data distribution information can not be used.The combination of depth learning and vision detection in visual intelligence detection of surface defects is a research hotspot and has a broad application prospect.In this paper,the visual inspection algorithm of high frequency transformer surface defect based on deep learning is studied.Firstly,the characteristics of traditional visual inspection methods and depth learning visual inspection methods for surface defects are analyzed.Aiming at the characteristics of high frequency transformer,such as many kinds of defects and complex detection background,the idea of combining deep learning with visual detection method is conceived,and the requirement analysis of the project and the performance comparison of deep learning detection model are carried out,the YOLO V3 network model is chosen as the reference model of surface defect visual inspection in this paper.Secondly,in order to solve the problem that the initial parameters of the original Yolo V3 network model do not have high accuracy in visual inspection of the appearance defects of the high frequency transformer,the Yolo V3 network model is pre-trained by using the aluminum defect data set provided by alitianchi competition,the obtained weight parameters are used as the initial values of the surface defect detection network of high frequency transformer.The experimental results show that the overall detection accuracy of the model can be improved by 1%.Third,a new convolutional neural network was added to the original YOLO V3 network to solve the problem that the improved YOLO V3 network model could not detect the character mold and the similar defects accurately,the filter of positive and negative samples and the adjustment of the prior box are realized,which provide better initial values for the subsequent border prediction.Experimental results show that the improved network model maintains the excellent feature extraction ability and improves the overall detection accuracy by about 4%.Fourthly,in view of the problem that the selection of Priori boxes of small data sets may affect the distribution effect of fitting data,a new convolution module is added to adjust the width and height of priori boxes and sent to the final prediction layer,provides a better initial value for subsequent border regression.Fifthly,after the collection of high-frequency transformer defect data is enhanced,the new model is tested and compared with the test results of other methods,which shows that the precision of this model is not as good as that of Faster R-CNN,for the real-time and accuracy of the Algorithm,the Algorithm in this paper is better than that of SSD,and compared with the traditional visual detection result of the appearance defect of high frequency transformer,the traditional method has higher detection accuracy but less robustness,while the improved YOLO V3 vision detection algorithm has faster detection speed and stronger robustness than the traditional method.In this paper,"YOLO V3 network-based depth learning visual inspection algorithm for high-frequency transformer appearance defect" provides an approach to intelligent inspection method for high-frequency transformer appearance defect.