基于动态先验特征的包覆药多类型外观缺陷深度检测框架

Dynamic Prior Features-Based Deep Learning Framework for Multidefect Detection of Coated Propellants

包覆药通常被嵌入固体火箭或导弹发动机的动力系统中,其外观质量直接影响该类动力系统的性能表现。针对包覆药外观存在的形状、尺寸和表面缺陷,提出了一种基于动态先验特征的包覆药多类型外观缺陷深度检测框架,包括:1)将基于深度分类器的形状缺陷检测和基于深度分割网络的尺寸缺陷检测模型集成,去除不同任务间的冗余特征,同时将深度分割网络当前迭代形成的过程特征作为动态先验特征,作用于深度分类器参数下一次迭代更新,加快模型收敛速度;2)将深度分割网络产生的过程特征映射至基于卷积自编码器的表面缺陷检测模型中,指导检测模型快速聚焦于包覆药,抑制任务无关特征重复提取。实验结果表明,该方法在模型功耗、检测效率及检测准确率等方面具有较好的表现。

Coated propellants(CPs)are extensively used in the dynamical systems of rockets and missiles.The appearance quality of the CPs has significant impact on the performance of the systems.To this end,a dynamic prior features-based deep learning framework for multidefect detection of CPs,such as shape,size,and surface defects,is put forth in this article:1)An integrated deep model for deep classifier(DC)-based shape defect and deep segmentation network(DSN)-based size defect detection is introduced,which can remove redundant features among different tasks.Particularly,the features generated by the current iteration of the DSN,as dynamic prior features,act on the next iteration of the DC to accelerate the convergence rate,and 2)the dynamic features are also mapped to the convolutional autoencoder-based surface defect detection,which can guide the model to quickly focus on the CPs,while suppressing the repeated feature extraction of task-independent features.Experimental results on an image dataset from a real-world manufacturing line show that the proposed framework has the superiority in terms of the power consumption,detection efficiency,and detection accuracy.