Researches On Detection Of Electrode Contact Surface Defects Based On Deep Learning

In the vacuum interrupter,the vacuum contact is an important component used to control the current flow and extinguish the arc,so as power transmission is constantly developed towards high voltages and high currents,the electrical contact performance and service life of the vacuum contact is receiving more and more attention from researchers.In the manufacturing process of vacuum contacts,there is a need to detect the surface defects of contacts in time to avoid the degradation of the working performance of vacuum contacts and the resulting safety hazards.With the widespread use of deep learning,the automated,efficient,and accurate inspection and analysis of the surface condition of vacuum contacts through machine vision technology is a key research topic in the field of vacuum contacts.A vacuum switch contact surface defect detection system based on Lab VIEW virtual instrument development platform is designed in this paper.In this system,a set of contact image acquisition gantry and an automated image training platform are included.Firstly,the composition and design method of the vacuum contact surface image acquisition bench is introduced,and the key components of the bench are selected and designed in sequence.The acquired images were then used to construct a dataset of contact defects and the dataset was expanded using Albumentation.Subsequently,Anaconda was used to build separate runtime environments for Tensorflow and Pytorch to train different deep learning network models.Finally,the effectiveness of the data set expansion and the optimal detection algorithm for contact surface defects are investigated based on experimental comparisons.Experimental data show that,the expansion of the vacuum contact defect dataset using Albumentation resulted in a 2.4% and 5.5% improvement in completeness and 2.5% and7.6% improvement in accuracy of the algorithms for SSD and Faster＿RCNN,respectively.Experimentation with the expanded dataset,where the SSD and SSDLITE algorithm models under the Tensorflow framework were slow to converge,with accuracy rates still below 50%after 10,000 iterations of training;the optimal accuracy of the Faster＿RCNN series algorithms has reached 90.1%,but the average detection time per image was 1596 ms.The accuracy of the YOLOv5 algorithms in the Pytorch framework were all above 80%,with the best detection accuracy being yolov5 l,which reached 97.5%,saving 1558 ms compared to the best algorithm in the Faster＿RCNN series.A vacuum contact surface defect detection system has been developed and designed to make the overall training and detection process more intuitive,interactive,and practical in this paper.Through relevant comparative experiments,the best detection algorithm for contact surface damage is explored,which improves the efficiency and accuracy of detecting surface defects in vacuum contacts and provides technical support for the subsequent production and design of vacuum contacts,further improving the reliability of contacts working in vacuum interrupters.