Research On Silicon Rod Defect Detection System Based On Deep Learning

Currently,the detection of defects in square silicon rods heavily relies on manual inspection,which is inefficient and labor-intensive.This study focuses on addressing these issues by employing machine vision techniques for visual inspection of eight common defects found on the surface of square silicon rods,including pits,chipped edges,crystal cracks,bright spots,watermarks,unground to,stains,and line marks.The research is conducted with a focus on industrial applications.The main contributions of this study can be summarized as follows:Firstly,an independent data acquisition device is designed specifically for detecting defects on square silicon rods.To overcome the challenges posed by strong reflection and diffraction on the rod surface,a structured light fill-in method is proposed.The feasibility of this method is verified using Tracepro simulation software.Additionally,a multi-camera inspection method is introduced to meet the requirements of large-area and multi-angle inspection.The collection device is designed based on actual field processes,and information interaction with Siemens 1511-PLC is achieved using OPC UA communication technology,enabling intelligent collection of defect data based on manual inspection results.Secondly,a dataset of square silicon rod surface defects is created.The internal control standard for silicon rod defects is used as a basis for collecting and classifying defective silicon rod data.Images containing defects are collected,classified,and labeled according to the defect classification standard.To address the issue of a small number of samples and imbalances in the number of different defects,data enhancement techniques are applied,resulting in a comprehensive silicon rod defect dataset.Thirdly,a deep learning-based defect detection method for silicon rods is developed.Two models,namely yolov5s and yolov5m from the YOLOv5 neural network framework,are selected for defect feature extraction.Optimal weights for both models are obtained,resulting in detection accuracies of 0.82 and 0.83,and sensitivities of 0.92 and 0.91,respectively.Considering real-time detection requirements in industrial applications,the yolov5s model is selected for deployment.Finally,the industrial deployment,analytical validation,and optimization of the silicon rod defect detection system are carried out The model is deployed in the inspection system using openVINO,and the development of system modules and industrial application deployment are completed.Three evaluation indexes,namely detection rate,multi-detection rate,and over-detection rate,are proposed based on engineering application requirements.To address the low detection rate of certain defects,two methods are proposed:target area restriction for detection to remove background interference and a multiconfidence defect classification method.Experimental verification shows a 2.3%increase in detection rate with the target area restriction method and increases of 6.9%,4%,2%,and 3%in detection rate for pits,bright spots,watermarks,and stains,respectively,using the multi-confidence defect classification method.Simulation experiments are designed to analyze the impact of curtain angle and light source length on imaging quality,thereby identifying the factors influencing the system’s performance.The developed detection system achieves detection rates of 85.2%,94.5%,91%,87.1%,91.5%,95%,88.1%,and 95%for pits,chipped edges,crystal cracks,bright spots,watermarks,unground to,stains,and line marks,respectively.The multi-detection rates are 10.2%,2.3%,1.5%,1.1%,3.6%,1.3%,8.4%,and 4%,while the missed detection rates are 0.2%1.1%,2.3%,0.1%,1.23%,0.2%,1.5%,and 0.1%for pits,chipped edges,crystal cracks,bright spots,watermarks,unground to,stains,and line marks,respectively.