桥梁拉索动力响应监测数据中存在大量的低质量数据，现有的监测数据检测研究集中于时域波形异常的明显异常数据，然而监测数据中还存在时域波形正常，但频域特征混乱的数据，这类数据无法准确地获取桥梁拉索动力特性。针对该问题，将现有的异常数据检测拓展为数据质量评价，同时对明显异常数据和频域混乱数据进行检测。采用卷积神经网络(CNN)和数据频域特征建立了桥梁拉索动力响应监测数据质量评价方法，实施流程包括：采用快速傅立叶变换(FFT)将时域数据序列转化为功率谱密度函数(PSDF)，利用格拉姆角场(GAF)方法对PSDF序列进行可视化，进而搭建CNN模型对监测数据质量进行自动化评价。以某斜拉桥的拉索加速度监测数据为例开展了应用研究，研究结果表明，与时域序列检测方法相比，PSDF序列检测方法能够更好地区分正常与频域混乱数据，具有更高的评价准确率；利用两个传感器监测数据建立的CNN模型对所有26个传感器监测数据质量评价准确率均在94%以上；此外，该方法建立的评价模型应用到另一座类似桥梁的监测数据质量评价中，准确率也达到了95%。

There was a large amount of low-quality data in the monitoring data of bridge cable dynamic response. The existing data detection research focused on obviously abnormal data with abnormal time-domain waveforms. However, there was chaotic data in frequency-domain characteristics with normal time-domain waveforms in the monitoring data, which can’t accurately obtain the dynamic characteristics of bridge cables. Aiming at this problem, the existing abnormal data detection was extended to data quality evaluation, and the obvious abnormal data and frequency-domain chaotic data were detected at the same time. The data quality evaluation method of bridge cable dynamic response monitoring was established by using a convolutional neural network (CNN) and data frequency-domain features. The implementation process included: the time-domain data sequence was transformed into a power spectral density function (PSDF) by fast Fourier transform (FFT), the Gramm angular field (GAF) method was used to visualize the PSDF sequence, and a CNN model was designed and built to evaluate the data quality automatically. Taking the cable acceleration monitoring data of a cable-stayed bridge as an example, the applied research was carried out. The results show that compared with the time-domain sequence detection method, the PSDF sequence detection method can better distinguish normal and pseudo-normal data, and has a higher evaluation accuracy rate; The accuracy rate of the CNN model established by using the monitoring data of two sensors to evaluate the quality of all 26 sensor monitoring data is above 94%; In addition, the evaluation model established by this method is applied to the monitoring data quality evaluation of another similar bridge with an accuracy of 95%.

国家自然科学基金资助项目(51878027)、北京市教委青年拔尖人才培育计划项目(CIT TCD201904060)、北京建筑大学基本科研业务费项目(X20174, X21073)、山东省交通运输厅科技计划(2021B66)