Abstract: Aiming at the problems that were difficult to capture the damage feature information and the identification results were inaccurate when there were different types of damages in truss rod elements, a damage identification method was proposed using JMRAE. Firstly, JMRAE was applied to intercept the signals according to different scale numbers, and the Sigmoid function and ReLU function were combined to extract the features. ZCA was introduced to reduce the features dimension to retain important information and reduce data redundancy. Then, SoftMax classifier was applied to solve the local features of different segments in the hidden layers, and feature fusion was performed to determine the structural states. Finally, the numerical three-dimensional truss structure model and the laboratory-built truss were used for validation and comparative study with the refined composite multiscale dispersion entropy(RCMDE), kurtosis, and back-propagation(BP) neural network methods. The results show that the proposed method has higher damage identification accuracy.

Key words: joint multiple reconstructions autoencoder(JMRAE), zero-phase component analysis(ZCA), SoftMax classifier, feature fusion, damage identification

摘要： 针对桁架杆单元存在不同损伤类型时损伤特征信息难以捕捉且识别结果不准确的问题，提出了利用联合多重重建自编码器（JMRAE）进行损伤识别的方法。首先，运用JMRAE按照不同尺度数分段截取信号，将Sigmoid函数和ReLU函数进行组合以提取特征量，引入零相位成分分析（ZCA）降低特征量维度，以保留重要信息并减少数据冗余。然后，运用SoftMax分类器求解隐含层中不同片段的局部特征量，并进行特征量融合以判断结构状态。最后，运用三维桁架结构数值模型和实验室搭建桁架进行验证，并与精细复合多尺度散布熵（RCMDE）、峰度和反向传播（BP）神经网络方法进行对比研究，结果表明所提方法具有更高的损伤识别准确性。

关键词: 联合多重重建自编码器, 零相位成分分析, SoftMax分类器, 特征量融合, 损伤识别