Fit Desidn Of A Multi-camera Defect Detection System For Hairiness On Yarn Package

At present,the quality control of the yarn packages defects before it leaves the factory relies mainly on manual visual inspection.This method is not only complex,but also causes damage to the yarn packages during transit,which greatly affects the production efficiency.Hairiness defects are difficult to observe directly under natural lighting conditions due to their small size and require research into visual light source systems.Existing inspection algorithms are mainly focused on the more obvious problems such as oil and deformation,but the accuracy of the inspection algorithm needs to be further improved for minor defects such as "hairiness" and "broken threads".Generic intelligent inspection equipment does not provide efficient and adapted solutions for the structural characteristics of the yarn packages,which can easily lead to double-checking,missed inspection and wrong inspection.(1)This study starts from the intelligent manufacturing needs and takes the problem of the adaptability of the inspection system for the appearance defects of yarn packages as the starting point,focusing on the problems existing in the practical application of the defect detection system.The adaptation problem in the inspection process of yarn packages is extracted,each functional module is designed,an adaptation framework is built,and the corresponding adaptation design strategy is proposed by integrating knowledge from different fields such as optics,mechanical structure,algorithms and existing production data of the factory.(2)Highlighting defect characteristics by adaptive design of machine vision light sources.A microscopic vision approach is used to build a pilot experimental platform for the adaptation design of the target detection algorithm for hairiness defects,to capture images of defects on the surface of the yarn packages,to build a microscopic image dataset of defects,to test the effects of light and dark field illumination methods,light color,brightness and angle on the surface imaging,and to derive the best light source environment for the adaptation of the silk cake defect features,so that small defects such as hair feathers can be easily detected in contrast to the background.The machine vision illumination solution is transferred to the industrial inspection environment,and a front-to-back fused multi-source illumination method is designed.(3)Improve the accuracy and recall of the silk cake detection algorithm through algorithm adaptation design.Comparing the detection effects of four target detection algorithms,Center Net,YOLOx,YOLOv4 and Faster r-cnn,it is concluded that the Center Net algorithm without anchor frame works best.The SE-Center Net model was further adapted,optimised and designed to incorporate feature fusion for the problem of small feature loss during convolution.The discrete convolution layer DCNv2 is added to address the problem of relatively discrete distribution of hairiness,and the channel information weights are adjusted by fusing SENet's channel attention mechanism to address the loss of channel features.In experiments with the algorithm adapted to microscopic datasets,the SE-Center Net target detection algorithm achieved AP values of 93.88% and 92.86% for hairiness and broken ends,respectively,verifying the effectiveness of the algorithm.Migrating the algorithm to real production,the MAP for dark-field illumination detection was 93.37% and 94.83% for back-illumination.Model simplification using the TRT inference engine improved the detection speed to 3.6ms while ensuring detection stability and accuracy.(4)Full coverage of the silk cake surface is achieved by adapting the design of the rotating vision inspection mechanism.A rotatable image acquisition mechanism is designed for the hollow cylindrical shape of the bobbin,with optional connection structures and accessories.The camera type,viewfinder range,resolution,object distance and lens were adapted according to the 40?m detection accuracy of the hairiness,and actual shooting tests were carried out to confirm the angular speed of the camera movement.Adaptation of motors and gearboxes according to structure and morphological quality data to complete the adaptation design of the multi-camera rotating acquisition mechanism for the combination of line and surface cameras.