基于卷积神经网络的稳定图自动分析方法

Automatic analysis of stabilization diagrams using a convolutional neural network

提出一种基于卷积神经网络(CNN)的稳定图自动分析方法。在获得不同结构的稳定图之后,按照各自的频率识别精度要求,将稳定图均分成若干个频带,得到单一模态稳定图作为CNN训练样本;通过平移、改变稳定点标记等技术手段对样本进行扩充,再将预处理好的训练样本代入CNN,通过跟踪损失函数在训练过程中变化规律,对如学习率等CNN参数进行调优,最终得到可自动判别稳定图中虚假模态的CNN;以3自由度弹簧质量数值模型、7自由度弹簧质量数值模型、以及一座钢筋混凝土框架结构大楼、瑞士Z24桥加速度实测数据验证了所搭建CNN模型的有效性。训练和预测结果表明,搭建的CNN亦可用于其他一般结构的稳定图自动分析,具有一定的通用性。在无需人为提取任何特征参数,也无需设定任何阈值的情况下,即可自动且准确、快速地剔除稳定图上的虚假模态。

A convolutional neural network(CNN)methodology was proposed to automatically interpret stabilization diagrams.Once the stabilization diagrams of different structures had been obtained,they were equally distributed into several frequency bands according to the accuracy requirements of each frequency identification,which were called single mode stabilization diagrams.These frequency bands samples were used as learning samples of the CNN.After that,these learning samples were expanded with some technical methods such as translating and changing the label of stable poles on the stabilization diagram.Then,the preprocessed learning samples were substituted into the constructed CNN.The parameters of CNN,such as learning ratio,were tuned by tracking the changing rule of losing function during the learning process.Finally,a CNN which can automatically eliminate the spurious modes on the stabilization diagram was obtained.The constructed CNN was verified by a 3 degree of freedom(DOF),a 7DOF spring-mass model as well as the accelerometer data of a reinforced concrete frame structure and the Swiss Z24 bridge.The robust learning and prediction results prove that the constructed CNN is effective for analyzing any stabilization diagram of different structures.It can automatically and accurately eliminate the spurious modes on the stabilization diagram immediately without extracting any characteristic parameters or setting any thresholds of them.