Automated damping identification of long-span bridge using long-term wireless monitoring data with multiple sensor faults

Accurate damping estimation is critical to the serviceability assessment of large and flexible civil infrastructure. However, the frequency of sensor faults found in the long-term monitored data of large-scale structures is a potential cause of errors in the damping estimates. A machine-learning-based fault-data management approach is proposed whereby erroneous data are identified and removed automatically. A support Vector Machine (SVM) is used to automatically detect and recover/isolate multiple types of sensor faults from measured accelerations. The labeled training samples are artificially augmented using digital simulation of a random process. An envelope function is introduced to reflect the time-varying trends of signals. A new feature, the Maximum Correlation Factor, is proposed to measure similarities between the simultaneously measured signals in order to classify faulty and normal data. The performance of the trained SVM classifier was validated via long-term data from the wireless sensor network of a cable-stayed bridge in South Korea. The modal damping ratios were then estimated from the faulty and recovered data. The improved performance of the damping estimation via spike removal and fault isolation was evaluated in terms of the correlation function and stabilization diagram in the output-only modal analysis. The recovered data provided a more robust and consistent damping estimate, and demonstrated the efficacy of the proposed fault-data management strategy that uses a new SVM feature.