New Method Picks Best System Models from Vibration Data

Monitoring the health of structures like bridges and buildings often relies on vibration data, but accurately estimating system parameters and unknown inputs from limited sensors has been a persistent challenge. A new study introduces a method that automatically selects the most plausible model from multiple guesses, enhancing the reliability of real-time structural health monitoring without requiring extensive prior knowledge.

The key finding is that by using the Kullback-Leibler divergence—a statistical measure of similarity between distributions—researchers can evaluate different initial parameter sets and identify the one that best matches the observed vibration data. This approach was tested on linear and nonlinear systems, such as multi-degree-of-freedom models, and consistently selected the model with the least estimation error, as shown in figures like Fig. 2 and Fig. 3, where parameter errors were minimized.

Methodology involved combining two filters: the unscented Kalman filter (UKF) and the residual-based Kalman filter (RKF). These filters estimate dynamic states, parameters, and unknown inputs from vibration measurements, such as accelerations, which are integrated to derive displacements and velocities. The process runs multiple simulations with different initial parameter guesses, computes the Kullback-Leibler divergence for each, and picks the set with the smallest divergence value, indicating the best fit to the data.

Results analysis from the paper demonstrates the method's effectiveness. For instance, in a 3-degree-of-freedom linear system with a pulse input, the approach correctly identified the initial parameter set that underestimated values by 25% as the most accurate, reducing parameter errors significantly. In nonlinear systems, like the Duffing oscillator, it handled complex behaviors and maintained accuracy even with ambient noise inputs, as illustrated in Fig. 12 and Fig. 13. The method also performed well under varying noise levels, though higher noise (e.g., 20% root-mean-square) led to poorer estimates, as seen in Fig. 5.

Contextually, this advancement matters for real-world applications in civil engineering and infrastructure monitoring. It allows for better assessment of structural integrity during normal operations, such as detecting damage or faults without shutting down systems. This could lead to safer buildings and bridges by providing more dependable, real-time insights into their condition, using readily available vibration data from sensors.

Limitations noted in the paper include the method's reliance on certain assumptions, like Gaussian noise and the need for identifiability conditions where inputs are not zero. In cases with insufficient excitation or high noise, estimates may diverge or be less accurate. Future work could explore extending this to larger-scale systems and validating it with experimental data to address these constraints.