Abstract: In bridge health monitoring system, the data loss caused by sensor failures and external environmental interference seriously affects the reliability of the monitoring system. At present, data recovery methods focus on the problem of recovering part of the missing data in the channel, while paying less attention to the problem of recovering missing data in the entire channel. Therefore, the GRU neural network was improved to recover the missing data of the entire channel and improve the data recovery accuracy. Firstly, an improved GRU model was constructed, which was based on the denoising autoencoding model and used gated recurrent neural network to replace the fully connected layer to learn the spatiotemporal correlation between different data. Meanwhile, attention mechanism and mask mechanism were added to enhance the attention to the missing position and improve the accuracy of data recovery. Secondly, the training data was constructed to train model, and the continuous missing data of each channel was artificially constructed as the input of the model, and the corresponding complete data was used as the output, so as to improve the learning ability of the model on the missing position mechanism. Finally, the missing data of the entire channel was recovered, the effect of data recovery was evaluated by evaluation indexes, and the modal analysis was carried out. The accuracy of the proposed method was verified by numerical simulation and monitoring data of real bridges. The application results of the real bridge show that compared with the existing models, the data recovery accuracy of the proposed method is improved, the mean absolute error is reduced by 21.8%, the root mean square error is reduced by 42.7%, and the model fitting ability is improved by 9.1%.

Key words: bridge engineering; bridge health monitoring; missing data of the channel; gated recurrent neural network; denoising autoencoder model

摘要： 桥梁健康监测系统中，因传感器故障和外界环境干扰导致的数据丢失严重影响监测系统的可靠性。目前数据恢复方法集中于恢复通道部分缺失数据的问题，而对恢复整个通道缺失数据问题关注较少。为此，对GRU神经网络进行改进，用于恢复整个通道缺失数据，提高数据恢复精度。首先，构建改进的GRU模型，其以降噪自编码模型为基础构架，利用门控循环神经网络代替全连接层，学习不同数据间的时空相关性，同时加入注意力机制和掩码机制，加强对缺失位置的关注程度，提高数据恢复的准确性。其次，构造数据训练模型，人为构造每个通道连续缺失数据作为模型输入，相应完整数据作为输出，提高模型对缺失位置机制的学习能力。最后，恢复整个通道缺失数据，用评价指标评价数据恢复效果，并进行模态分析。通过数值模拟和实桥监测数据验证了方法的准确性，实桥中的应用结果表明：与已有模型相比，该方法数据恢复精度提高，平均绝对误差降低了21.8%，均方根误差降低了42.7%，模型拟合能力提高了 9.1%。

关键词: 桥梁工程；桥梁健康监测；通道缺失数据；门控循环神经网络；降噪自编码模型