Abstract:

Objective Glass fiber is widely used in transportation, shipbuilding, aerospace, construction, wind power, national defense equipment, leisure sports and other fields, and is stored in the form of glass fiber yarn clusters in the production and transportation process, so the quality monitoring of glass fiber yarn clusters is particularly important. However, conventional manual detection has disadvantages such as low efficiency, high missing detection, and long lag time. Therefore, a machine vision-based appearance defect detection system for glass fiber yarn was proposed and designed to meet the demand of the production line.
Method Based on the combination of conventional image algorithm and deep learning algorithm, firstly, the conventional method is adopted to preprocess the image, and the RGB and HSV color space channels are adopted to identify the model label of the glass fiber yarn clusters, and secondly, the suspected defective glass fiber yarn clusters image is transmitted to the improved MobileNetV2 deep learning model for defect judgment. Finally, a complete set of software and hardware system for the detection of appearance defects of glass fiber yarn clusters was designed, and the system used Siemens S7-200 PLC as the hardware controller to complete the automatic transmission and sorting of glass fiber yarn clusters in the process of detection.
Results The image data collected in the industrial field is used as a dataset sample, and the resolution of the upper end face image is 1 942 pixel×1 942 pixel, and the bottom end image resolution is 3 500 pixel×3 500 pixel. 3 000 normal samples, 25 847 defective samples, 1 000 images of red, yellow and blue varieties were selected for verification by the variety classification algorithm. In order to ensure that no detections are missed in the preclassification task, the parameter threshold is set as low as possible during the algorithm design, so that all images of suspected defects continue to be input to the subsequent deep learning model for further review. In the variety classification task, the recognition accuracy of each species reached 99%. In the follow-up defect review task, ResNet, AlexNet, VGG16 and MobileNetV2 models were experimented, respectively, and the overall classification accuracy reached about 95%, of which the classification accuracy of ResNet50 reached 96.12%, which was 0.62% higher than that of MobileNetV2, but its training time and model parameters were much larger than those of MobileNetV2. MobileNetV2 achieves a classification accuracy of 97.37%, which is 1.87% higher than the original model, proving the effectiveness of the improvement. On the whole, the proposed detection algorithm meets the requirements of the appearance defect detection system of glass fiber yarn clusters.
Conclusion A software and hardware system for the detection of appearance defects of glass fiber yarn clusters based on machine vision is proposed and designed. The conveyor mechanism controlled by the Siemens S7-200 PLC can be freely and conveniently embedded in the actual production line, and the multi-angle camera is adopted to collect the appearance image of the glass fiber yarn clusters during the conveying process. According to the requirements of production inspection, effective pretreatment, variety classification and defect detection algorithms were designed. The influence of the reflection effect of the plastic packaging on the surface of the glass fiber yarn clusters on the image feature extraction is solved, the image preprocessing method is designed to effectively reduce the computation amount of the system, and the proposed SA-MobileNetV2 lightweight deep learning model can meet the needs of the actual production line to detect and sort one glass fiber yarn clusters per 5 s on average. The system is simple to operate and has complete data, which helps enterprises to improve production processes and monitor product quality, and has broad engineering application value.

Key words: machine vision, glass fiber yarn, defect detection, image processing, elastic algorithm service framework