【作者】 陈远；

【导师】 杨佑发；

【作者基本信息】 重庆大学 ， 防灾减灾工程及防护工程， 2009， 硕士

【摘要】 建筑结构在设计、施工尤其使用过程中,由于遭受人为因素和自然因素而出现老化和破损,亟待损伤识别和维修加固。对建筑结构的损伤识别已经成为土木工程领域一个重要研究课题。结构的损伤会引起相应的动力特性改变,因此,如果建立起结构动力特性变化与结构损伤之间的映射关系就可以实现结构损伤识别。神经网络具有非线性映射能力强、容错性好等优点,非常适合用于结构损伤识别。本文以多层框架为研究对象,从构件(梁、柱)和节点两个层次对框架结构损伤识别进行研究,主要完成以下几个方面的工作:①提出了基于神经网络的框架结构损伤的多重分步识别方法,并在此方法基础上利用APDL语言和MATLAB语言编制程序建立用于框架结构损伤识别的高效神经网络法。②研究了框架结构构件和节点的损伤模型,推导了用于输入神经网络的损伤指标的基本原理,介绍了结构动力特性参数的提取方法。③研究了基于神经网络的框架结构构件损伤识别。根据构件损伤的多重分步识别思路,把构件损伤识别主要分为四步:第一步利用神经网络建立损伤异常过滤器对构件损伤进行预警;第二步以频率构造的组合指标X 1作为神经网络输入向量,对构件损伤进行初步定位;第三步以频率和模态振型构造的组合指标X 2作为神经网络输入向量,对构件损伤进行具体定位;第四步以频率平方变化率RNF作为神经网络输入向量,对构件损伤程度进行识别。④研究了基于神经网络的框架结构节点损伤识别。根据节点损伤的多重分步识别思路,把节点损伤识别主要分为四步:第一步利用神经网络建立损伤异常过滤器对节点损伤进行预警;第二步以频率构造的组合指标X 1作为神经网络输入向量,对节点损伤进行初步定位;第三步以归一化的应变模态差绝对值NSMC作为神经网络输入向量,对节点损伤进行具体定位;第四步以应变模态差绝对值SMC作为神经网络输入向量,对节点损伤程度进行识别。⑤通过理论推导得到了模型参数误差对损伤引起的模态参数改变贡献的表达式,用该式指导神经网络输入参数的选择;在此基础上,又从相对误差的角度进一步研究模型参数误差对神经网络输入向量的影响;最后通过数值算例研究了模型参数误差对框架结构损伤识别的影响。更多还原

【Abstract】 Building structures will inevitably suffer from the influence of artifical factors and natural factors during design and construction ,especially during service, which arises aging and worn. How to identify the damage of building structures has become one of the advancing fronts of civil engineering researches. Structural damages will cause dynamic characteristics to change correspondently. Therefore, if the mapping relationship between structural damage and changes of dynamic characteristics can be established, the damage can be identified using dynamic test of the structures. The neural network technique has great superiority in identifying the damage of structures for its strong non-linear mapping ability, anti-interference capability. In this paper, the multistory frame structure will be used as research object to study damge identification of structural member (beam,column)and joint. In this paper, something has been discussed as the follwing aspects:①The multi-stage damage identification approach based on the neural network for frame structures has been raised in this paper. Under the foundation of this approach, a kind of high efficient neural network methods to identify damage in frame structures has been established by writing APDL and MATLAB programs.②The member damage model and the joint damage model have been researched, fundmental theories of damage indexes used as input vectors of the neural network have been derived , the methods for extracting dynamic characteristics have been put forward.③The multi-stage damage identification approach based on the neural network for frame structures has been used to study identifying damage in structural members. This approach is divided into four steps. Firstly, damage anomalous filter which is set up by BP neural network has been used to alarm the damage in stuctural members. Secondly, the primary location of the member damage is determined by the neural network with inputing the combined damage index X 1. At the third step, the specific location of the member damage is determined by the neural network with inputing the combined damage index X 2. Finally, the damage degree of the member is determined by neural network with inputing the change rate of squared modal frequency.④The multi-stage damage identification approach based on the neural network for frame structures has been used to study identifying damage in structural joints. This approach is divided into four steps. Firstly, damage anomalous filter which is set up by BP neural network has been used to alarm the damage in stuctural joints. Secondly, the primary location of the joint damage is determined by the neural network with inputing the combined damage index X 1. At the third step, the specific location of the member damage is determined by the neural network with inputing the damage index NSMC . Finally, the damage degree of the joint is determined by the neural network with inputing the damage index SMC .⑤An expression accounting for the contribution of model parameter error to the modal parameter change in consequence of damage is derived, which is instructive for selecting the input vectors to the neural network. On this basis, the relative error formulas have been raised to study model parameter error. Numerical simulations have been made to study the influences of model parameter errors in identifying damage in frame strcutures .更多还原