| High-frequency transformers are an important component of electronic products.Appearance defects in the production process will affect the quality and service life of the products.They must pass the appearance inspection before they can enter the market.At present,the appearance defect detection of high-frequency transformers mostly adopts manual detection methods,but this method is subjectively affected by personnel,and it is prone to missed detections and false detections.When the personnel are in a state of fatigue,the detection effect will drop sharply.There is an urgent need for automatic and intelligent detection methods to improve detection results and liberate manpower.This paper introduces machine vision technology and deep learning methods into the appearance defect detection of high-frequency transformers.Aiming at the appearance defects of different types of high-frequency transformers,the visual intelligent detection methods for the appearance defects of high-frequency transformers are studied.Firstly,the main types of appearance defects of high-frequency transformers are introduced,and their causes and characteristics are analyzed.According to the characteristics of defects and the difficulty of image analysis,they can be divided into three categories: tin staining/missing inner rubber,broken inner rubber and broken skeleton.A visual inspection system was constructed,hardware and related software such as industrial cameras,light sources,lenses,etc.were selected,and an appropriate lighting arrangement was selected to collect images.Secondly,use machine vision-based methods to analyze high-frequency transformer surface defects,use HSV color model channel separation methods to highlight defect areas,and then analyze and compare image denoising,image enhancement,and morphology commonly used in machine vision defect detection methods And other methods,select the best method to remove noise and further highlight the defect area.After image segmentation and area feature screening,the identification and location of the tin-staining/inner glue missing defect can be realized.The experimental results show that the accuracy of this method for detecting tin staining/inner glue missing can reach 99%.Third,for the defects of the inner rubber of high-frequency transformers that are difficult to detect by typical machine vision methods,an algorithm for identifying the defects of inner rubber of high-frequency transformers based on frequency domain analysis is designed.Use spatial domain image preprocessing to remove noise,design frequency domain filters to enhance the defect components,and then return to the spatial domain for image segmentation,and then extract several features with large differences between the defect area and the interference area to form a feature vector.The vector is input to the GMM(Gaussian Mixture Model)classifier to realize the identification of the inner rubber damage defect.Experimental results show that the algorithm has strong practicability,and the detection accuracy rate is95.6%.Fourthly,based on the image analysis method,it is difficult to detect the skeleton damage defect with very low contrast between the defect area and the background area and the image background is complex.The deep learning-based high-frequency transformer skeleton damage recognition algorithm is studied.A high-frequency transformer skeleton image database is constructed,the I Inception V4 network with the fully connected layer removed is used as the feature extractor,and the redesigned classifier forms a new training model.In the training,the transfer learning method is adopted to reduce the time of model training,and the network model is fine-tuned using the high-frequency transformer skeleton image database.The experimental results show that the method has better adaptability.Compared with the image analysis method,the recognition effect is good,and the accuracy of the recognition of the skeleton damage defect is 85.8%. |
PDF Full Text Request