A Blind Source Separation-based Approach To Multi-excitation Monitoring Data Analysis Of Long-Span Bridge

Buffeting is one of the wind-induced bridge vibrations, which is a relatively small, but frequent response. These charactors make buffeting a threat to bridge’s fatigue performance and driving safety.SHM system uses accelerometer to measure bridge vaibration. Accelerometer’s signal is complicated due to multiple loads’ effect. Further more, wind load and vehicle load are the two main control loads. To purely research the bridge wind load response, we need to weaken, even remove vehicle load efect. Therefore, this thesis mainly focuses on investigating how to separate wind-induced vibration and vehicle-induced vibration in a mixed signal.Main contents are included as follows:The signal separation method of wind-induced vibration and vehicle-induced vibration based on Independent Component Analysis(ICA) is proposed. Firstly, the ICA theory is introduced. Then considering bridge vibration characteristics, the single channel signal is filtered by a band-pass filter bank, the frequency-independent pseudo multichannel signals are used to do ICA, so the independent compenents are obtained. Lastly, independent components are clustered into wind-induced vibration subspace and vehicleinduced vibration subspace, two signals are reconstructed.The signal separation’s performance evaluation method is presented. Firstly, autoencoder theory and deep neural networks theory are introduced. Secondly, a deep neural networks classifier based on stacked autoencoder is constructed, and the classifier is trained using Sutong Yangtze River Bridge’s vibration data. Lastly, the evaluation method based on sample classification is proposed.Finally, combine ICA and DNN, construct a bridge vibration signal separation algorithm flow, acceleration data collected from Sutong Yangtze River Bridge’s SHM system is used to separate wind-induced vibration and vehicle-induced vibration, separation ability of the algorithm flow is illustrated.