【作者】 徐阳；

【导师】 李惠； 李顺龙；

【作者基本信息】 哈尔滨工业大学 ， 工程力学， 2019， 博士

【摘要】 桥梁结构是一个国家和地区的重要经济命脉。桥梁结构在全寿命服役周期中,不可避免地遭受到环境侵蚀、往复荷载及突发灾害(如地震)等复杂因素的耦合作用,产生结构渐变损伤的萌生、发展和累积,导致服役性能不断劣化。目前,国内外学者一般采用基于动力反演的模态方法进行结构损伤识别、模型修正和安全评估。然而,此类方法往往只处理有限测点的不完备加速度监测信息,并且依赖对早期微小损伤不敏感的频率这一结构整体属性。另外,结构损伤往往伴随着复杂的非均质背景干扰,常规识别方法在实际场景中的普适性较差。实际工程中的目视巡检结果严重依赖于主观意识、量化不准确,成本昂贵。针对以上难题,本文研究基于计算机视觉的不同类型桥梁结构局部损伤自主智能识别方法,包括研究局部像素信息阈值处理、统计特征无监督高斯聚类建模、基于深度受限玻尔兹曼机高层次特征提取、基于深度有向无环图卷积神经网络多层次特征融合以及基于区域推荐机制的目标检测算法,发展拉索腐蚀疲劳退化评估、钢箱梁微小疲劳裂纹识别、钢筋混凝土桥墩结构多类型地震损伤识别定位等方法。主要研究内容如下:提出基于图像的在役拉索腐蚀状态识别及疲劳寿命评估方法。研究基于腐蚀过程和表观图像统计特征的概率建模方法,建立腐蚀特征空间与疲劳寿命控制参数统计学映射关系;基于上述方法对某桥服役18年的腐蚀钢丝疲劳寿命进行预测,结果表明各类应力幅下疲劳寿命预测误差均小于16%。提出基于计算机视觉的强干扰背景下钢箱梁微小疲劳裂纹自主智能识别方法。构建深度堆栈受限玻尔兹曼机和有向无环图卷积神经网络,提取并融合图像初级细部特征至高级抽象特征的多层次特征;将实际桥梁钢箱梁内部采集的350幅图像用于训练和验证网络模型,预测结果表明,构建的深度神经网络对多座跨海大桥各类样本的迁移识别准确率均大于93%。针对复杂非均质背景下多分类地震损伤的识别定位难题,提出基于计算机视觉的钢筋混凝土桥墩多类型地震损伤分类识别和区域定位方法。构建基于区域推荐注意力机制的深度卷积神经网络,采用具有一定概率保证率的矩形识别框实现多类型地震损伤分类定位,对混凝土开裂和剥落、钢筋暴露和屈曲的平均识别准确率大于80%,平均覆盖率大于88%。更多还原

【Abstract】 Bridges are of significance to countries and regions as economic lifelines.Bridges inevitably suffer from environmental erosion,dynamic loads and sudden disasters(e.g.earthquakes)in the whole service period,resulting in the gradual initiation,development and accumulation of structural damages,causing the continuous performance deterioration and even catastrophic accidents.At present,structural damage detection,model correction and safety asessment are commonly performed by dynamic inversion methods based on modal identification,which only process incomplete acceleration monitoring information on limited measuring positions.In addition,these methods rely on the overall property of structural frequency and are not sensitive to slight damages,especially when coupled with complicated non-homogeneous interferences.Conventional identification methods are always not universal in the practical applications of real scenes.In application,manual inspections are frequently adopted to visually inspect structures and evaluate damage degrees according to previous experiences,which heavily depend on subjective consciousness and are often descriptive,inaccurate and unreliable.Meanwhile,they always consume expensive labor,time and financial costs.To solve these issues,this study focuses on structural damage detection by optical images,which are obtained by ordinary consumer cameras from real-world structures and contain complex disturbances.In consideration with the unique characteristic of the individual task for different structural components,this study proposes autonomous monitoring and smart detection frameworks t hrough local pixel threshold processing,unsupervised gaussian clustering,establishing stacked Restricted Boltzmann Machine,fusion directed acyclic graph convolutional neural networks and region proposal networks to accomplish goals of primary image processing,image statistical feature modeling,multi-level feature fusion and attention-based object detection.Specifically,this study investigates the corrosion fatigue degradation of high strength steel wires in stay-cables,tiny fatigue crack identification in steel box girders and multi-type seismic damage classification and localization of reinforced concrete pier columns.The main contents and results are as follows:A novel non-destructive assessment method for in-service cable corrosion status identification and fatigue life degradation is proposed based on the corrosion random process of high-strength steel wires and image statistical characteristics,breaking the destructive limits of traditional cable replacement and inspection engineering.By time-dependent statistical modeling of corrosion process and unsupervised clustering of Gaussian mxiture model,the mapping relation model from corrosion feature space to key parameters of fatigue life is established and successfully applied on the corrosion degradation assessment of cable bridges in China.Prediction errors of corrosion fatigue life keep within 1 6% under a variety of stress amplitude codnitions.A refined identification method is proposed for tiny fatigue cracks in real-world steel box girders accompanied with complicated background disturbances,which solves the problems that conventional modal identification methods based on dynamic inversion are lack of sensitivity on early damages and can only identify late severe damages.By constructing deep Restricted Bolzmann Machines and fu sion directed acyclic graph convolutional neural networks,multi-level feature extraction and fusion of primary detailed and advanced abstract features are achieved.Transfer tests on several large-span bridges with steel box girder in China show that identification accuracies on all kinds of test samples exceed 93%,which varifies its robustness and portability.A two-stage identification and localization framework for multi-type seismic damages of reinforced concrete structures is established based on the regional proposal and attention-wise mechanism,conquering the bottleneck of multi-scale damage detection coupled with inhomogeneous background.Multi-scale candidate regions are generated with different sizes and aspect ratio.A multi-task loss function is constructed based on the combination of cross entropy classification and smooth L1 regression of rectangular damage regions’ coordinates.Rectangular boxes are regressed for localization as well as the corresponding category labels and guarantee probabilities.The average precision of identifications for concrete cracking,spalling,steel bar exposure and buckling exceeds 80%,and the average coverage of damaged regions exceeds 88%.This study investigates structurl damage detection methods for civil infrastructures based on computer vision techniques,overcomes the deficiencies of traditional methods and greatly improves the results’ accuracy and stability of structural condition assessment and highly enhances the automation and intelligence of state evaluation for civil engineering.It makes great contributions as a foundation for the popularization of intelligent detection,automatic damage identification and state assessment of civil infrastructures.更多还原