Research On Deep Learning-based Defect Detection Technology Of CRH’s PCB

In recent years,with the rapid development of our country’s high-speed EMU,the total mileage and the number of EMU climbed rapidly,each type of EMU also gradually entered the advanced maintenance stage,all kinds of maintenance technology has gradually become the hot research direction of the factory,and all kinds of EMU controller board detection has not yet appeared high-tech maintenance technology.At present,the manual visual examination is still the main method,but the manual visual examination is easy to cause visual fatigue,resulting in false detection and false detection.Not only the detection efficiency is low,but also the maintenance quality is not high,which does not meet the actual production needs.Aiming at the difficult problems of feature extraction such as unclear board defect types,small feature targets and fuzzy existing data,this paper first collected and labeled the defect database according to the actual situation,and produced a board defect data set PCB-data with nearly a thousand defects.Then,according to the characteristics of small defect targets and uneven distribution in the data set,the database was enhanced.The training effect of the data set is improved.At the same time,the Wiener filter is introduced to repair the jitter and blur phenomenon caused by handheld shooting in the data set,which increases the number of available defect images.Aiming at the problem that the accuracy of YOLOv5 s algorithm is not high when detecting board defects and the ability to extract multi-scale information is not strong,this paper proposes a C3 improved module based on Res2 Net network structure,and introduces the atrous space convolutional pooling pyramid module to verify the improved network,and the actual increase is 3.6%.The feature expression of the detection model is enriched.Aiming at the problem of insufficient accuracy of the algorithm model,the attention mechanism is introduced to allocate information processing resources efficiently,and the location of the attention mechanism is analyzed.The experiment shows that compared with the improved model YOLOv5s-BA proposed in this paper,the m AP value can be improved by 1.15% by adding SE attention to the appropriate location.The final accuracy reaches94.52%,which meets the actual production requirements.In view of the current situation of low automation and intelligence in the field of board defect detection,this paper developed an offline EMU PCB board defect detection system,and deployed the model algorithm proposed in this paper,and realized the multi-form defect detection of the board maintenance production line.