基于振动信号深度表征学习的高层建筑结构振动识别

VIBRATION RECOGNITION OF HIGH-RISE BUILDING STRUCTURES BASED ON DEEP REPRESENTATION LEARNING ON VIBRATIONAL SIGNALS

振动识别对于实现结构振动控制、性态监测评估具有重要作用。“智慧城市”背景下的高层建筑结构对其振动响应识别提出了更高的要求,传统基于时频域转换与图神经网络的振动识别方法在计算成本和工程适用性上的局限性愈发显著。该文提出了一种基于振动信号深度表征学习的振动识别方法。在该方法中通过搭建自编码网络对一维频域信号进行自监督学习用以获取振动表征信息,然后基于原始信号和重构信号进行振动敏感参数定义和计算,其结果作为特征工程输入到线性分类器模型中进行训练,进而实现对振动的快速识别。该文以某超高层建筑在不同环境激励下的振动响应识别为例,验证了所提出方法的可靠性。同时为证明提出方法的优越性,引入了目前普遍认为最优的MFCC+CNN振动识别方法进行对比,结果表明,该文提出的方法在计算时间上节省了近20倍,并且整体识别精度从0.74提高到0.95以上。

Vibration recognition plays an important role in realizing structural vibration control and health monitoring and evaluation.Under the background of smart city,high-rise building structures have appealed for higher requirements in their vibrational response recognition.Traditional vibration recognition methods based on time-frequency domain transformation and graph neural network reveal more significant limitations in terms of computational cost and engineering applicability.In this study,a novel vibration recognition method based on deep representation learning of vibration signals is proposed.In present method,self-supervised learning is performed on one-dimensional frequency-domain signals by establishing an autoencoder network to obtain vibration representation information.Then,the vibration sensitive parameters are defined and calculated based on original signals and the reconstructed ones,and the results of which are utilized as feature engineering and fed into the downstream linear classifiers for training,thereby realizing the rapid recognition of vibration.To verify the reliability of the proposed method,the vibration response identification of a super high-rise building under different ambient excitations is conducted for validation.At the same time,in order to prove the superiority of proposed method,the MFCC+CNN vibration recognition method,which is generally considered to be the best at present,is introduced for comparison.The results show that the proposed method is able to save nearly 20 times of computing time,and succeed to improve the overall recognition accuracy from 0.74 to above 0.95.