基于卷积神经网络的结构损伤识别研究进展

为系统梳理基于卷积神经网络的工程结构损伤识别方法的发展脉络和研究现状,分别从结构损伤的识别目的和在不同类型结构中的应用两方面进行了归类、分析和评价。介绍了卷积神经网络的基本结构和评价指标,回顾了卷积神经网络的研究和应用历程。在损伤的识别目的方面,主要针对混凝土结构损伤的分类、定位和分割,详细介绍了基于不同类型卷积神经网络的结构损伤识别方法,即基于分类的方法、基于回归的方法和像素级的图像分割算法; 分析了各类方法所使用的卷积神经网络模型的结构特点、计算流程、训练方法和损伤识别性能。在不同类型结构的损伤识别方面,分析了卷积神经网络在砌体结构、钢结构桥梁和古建筑木结构裂缝识别中的应用。最后,基于对卷积神经网络优缺点的思考,提出了发展建议和展望。结果表明:训练样本中结构损伤的多样性对模型的损伤识别效果影响较大; 现有基于卷积神经网络的损伤分割方法模型参数较多,计算量大; 采用数据增广和迁移学习方法可有效防止模型过拟合,提高模型训练效率; 针对微小损伤和不同类型结构损伤的识别,此类方法的性能有待提高。

Research Progress on Structural Damage Detection Based on Convolutional Neural Networks

In order to systematically sort out the development and research status of the damage detection of engineering structures based on the convolutional neural networks(CNN), the method for structural damage detection were categorized, analyzed, and evaluated in the aspects of detection objects and applications in different types of structures. The basic architecture and assessing metrics of CNN were introduced, and the researches and applications of the CNN were reviewed. In terms of detection objects, for the classification, the location, and the segmentation of concrete structural damages, the methods based on different types of CNN for structural damage detection, namely the classification based, the regression based, and the pixel level image segmentation based methods, were described in details. The features of network architectures, the analysis procedures, the training methods, and the performance of damage detection were analyzed. In terms of the damage detection of different types of structures, the applications of CNN based methods in the damage detection for masonry structures, steel bridges, and wooden architectural heritages were also analyzed. Finally, the suggestions and expectations for CNN methods were proposed based on considering the advantages and disadvantages. The results show that the variety of training samples significantly influence the effects of the model for detection. The existing damage segmentation methods require numerous parameters and high computational cost. Using the data augmentation and transfer learning method can effectively avoid the over-fitting problems and improve the training effectiveness. The performance of such methods needs to be improved for the detection of small damages and different types of damages.