Machine Vision-based Methods For Appearance Defect Detection Of Solid Propellant

Solid propellant(SPs),as a typical high-energy material,are extensively used in rockets and missiles.The appearance quality of the SPs has a significant impact on the performance of the systems.Owing to the poor manufacturing technology,the appearance of SPs may show various types of defects,such as size defects,shape defects,and surface defects,and it is,therefore,crucial to conduct the appearance defect detection of SPs.However,most of the existing visual-based defect detection methods mainly focus on a single defect pattern.It is,therefore,of great necessity to develop a new and holistic framework to facilitate the multidefect detection of SPs to simultaneously accommodate the shape,size,and surface defect detections.Furthermore,due to scarcity of surface defective samples,developing small samples-based surface defect detection technology has become an urgent need.In order to overcome these challenges,this dissertation devotes to carrying out an appearance defect detection method of SPs based on machine vision.The primary research contributions and innovative outcomes are summarized as follows:(1)Development of a deep semantic segmentation-based size defect detection.A deep semantic segmentation network,as the pre-processing method,is utilized to generate binary segmentation masks of SPs.The traditional edge extraction algorithm is,then,applied to capture the statistical characteristics of SPs.Furthermore,the real size of SPs can be calculated by spatial coordinate transformation rules.If the measured size is outside the appropriate range,it indicates that the SP belongs to the type of size defects.Experiments reveal that the proposed method has a great size defect detection accuracy of 95.90%.(2)Development of an integrated deep framework-based shape defect detection.A multitask learning-based integrated deep framework,which can simultaneously detect the shape and size defects of SPs is proposed by integrating a deep classifier and a deep segmentation network.Note that the gradient flow is forbidden form the deep classifier to shared feature extractor,aiming at preventing the unstable features from corrupting the learning.Experiments show that the shape defect detection accuracy of the proposed method is 99.43%,reaching the desired accuracy.(3)Development of a cascade attention mechanism-based surface defect detection.As the first stage,a deep semantic segmentation network is used to locate and segment the region of interest,and simultaneously generates a visual attention mechanism for surface defect detection in the second stage.The attention mechanism ensures that the surface defect detection method can focus on the surface of SPs,while totally suppress the interference of background information.Experiments reveal that the proposed model can reaches a surface defect detection accuracy of 97.87%.(4)Development of an intelligent detection system for appearance defects of SPs.The above-mentioned defect detection algorithms are integrated,and an intelligent quality inspection system is developed.The system cooperating with hardware equipment and control equipment can complete the defect detection tasks of SPs.The detecting speed of the developed system is 15.9 frames per second,fulfilling the actual production needs.