Data quality evaluation for bridge structural health monitoring based on deep learning and frequency-domain information

Abnormal data recognition is of great importance in structural health monitoring. Most of the existing studies focused on detecting obvious abnormal data, which has obvious abnormal time-domain waveform. Pseudo normal data, which is normally looking in time domain but chaotic in frequency domain, did not receive enough attention and were likely to be misclassified as normal data. As a result, structural performance may be incorrectly evaluated because pseudo normal data are not recognized and eliminated from the monitoring data. This study developed a novel quality evaluation framework for monitoring data of bridge dynamic response. The main novelty of the proposed framework is that the frequency-domain information is used to characterize the quality of the monitoring data so that the normal data, obvious abnormal data, and pseudo normal data can be accurately distinguished. In the framework, the frequency-domain information was obtained by fast Fourier transform (FFT), and the Gramian angular field images converted from FFT results were used to train a designed convolutional neural network (CNN). The cable acceleration data of the Waitan cable-stayed bridge were taken as an example to verify the accuracy of the proposed framework. Compared with the CNN models based on time-domain images and time-frequency stacked images, this framework can better recognize pseudo normal data from the monitoring data. In large-scale testing, the classification accuracy of all channels was more than 96%. Finally, two cable acceleration sensors of another cable-stayed bridge were used to demonstrate the feasibility of the framework in cross-object application. The results show that the framework has good accuracy and robustness in large-scale monitoring data quality evaluation and cross-object application.