Indirect structural health monitoring: estimation and variation of dynamic vehicle-bridge interaction parameters

Abstract

Bridge damage detection by indirect structural health monitoring (SHM) using vehicle-mounted accelerometers relies on vehicle-bridge interaction (VBI) in which the vehicle and bridge form a coupled dynamic system. Changes in bridge frequencies have been proposed as damage indicators. Acceleration amplitudes in frequency sub-bands are correlated with bridge operational mode shapes, changes in which can be indicative of damage location. However, treating the vehicle as combined exciter and sensor carrier introduces confounding, since even the presence of non-moving vehicles changes the effective bridge modal parameters in relation to vehicle-to-bridge mass and natural frequency ratios. Additionally, when natural frequency ratios are close to unity, vehicle movement along the bridge induces non-stationarity for the combined VBI system frequencies, a phenomenon only recently identified and little studied to date. Finally, vehicle kinematics form a filter through which the bridge response is viewed. These confounding effects can reduce or prevent visibility of bridge damage by changing observed frequencies and the associated acceleration amplitudes. The confounding is additional to the challenges inherent to direct SHM: particularly environmental and operational variation (EOV) which can mask the visibility of bridge damage or lead to spurious identification of damage where none exists, and limited availability of baseline data representing the healthy bridge condition. Indirect SHM is often suggested as a way to address the economic and practical challenge of monitoring a large number of bridges simultaneously. This thesis argues that the challenges and confounders mean that it is not yet clear how and where indirect SHM can add value. For indirect SHM to reach a state of operational readiness, greater understanding of the effects of vehicle parameters and vehicle-to-bridge parameter ratios on VBI system dynamics is required.

This study first explores vehicle parameter estimation, leveraging output-only methods to fit a representative reduced-order model reflecting current vehicle condition, configuration, and sensor location. Numerical simulation is used to demonstrate the approach, comparing time- and frequency-domain methods for model fitting in response to synthesised road profiles. Bias due to road profile variation and spectral flatness is considered. Measurements of the wheel hop frequency are presented based on responses to operational driving, alongside estimates of cabin bounce frequency using traditional accelerometers and the application of a smartphone device for the same purpose. Variation in estimated parameters is considered and compared to a second phase of fieldwork featuring a different vehicle suspension condition and configuration. Recommendations are given for future estimation campaigns including vehicle model abstractions, data processing strategies and vehicle speed and road profile roughness class consistency.

Practically, labelled data representing damaged bridge conditions are rarely available and baseline bridge modal parameters may be unknown. In response, this study explores visibility and accuracy of bridge frequencies through prominent peak identification (PPI) assuming limited knowledge of bridge properties. The effects of the (practically immeasurable) vehicle mass ratio, vehicle-to-bridge frequency ratio, vehicle speed and sensor position are explored in a parametric study using finite element simulation. VBI-induced frequency non-stationarity is noted as a confounder and is further explored in the hybrid time-frequency domain, including the first known presentation of such effects for two degree-of-freedom quarter car models. In the presence of vehicle-to-bridge frequency matching, a robust damage identification method would be of great value. The potential for one-class classification (OCC) in latent space after dimensionality reduction of hybrid time-frequency domain inputs is investigated, showing promising results and appearing to facilitate damage identification as well as differentiating between damage locations and intensities.

The study concludes with a laboratory-scale test of the proposed OCC method. Vehicle and bridge system identification by time-domain impulse response curve fitting is complemented by frequency-domain exploration of the vehicle acceleration response in a manner reflecting the earlier field-scale experiments. The potential confounding effects of tyre out-of-roundness or mass asymmetry are discussed. A novel simplified expression is proposed for recovering an estimate of the contact point (CP) response (the acceleration of the bridge deck at the temporally varying vehicle-bridge interface) and the accuracy of such estimates is discussed in relation to vehicle-to-bridge parameter ratios. It is demonstrated that this simplified expression produces comparable outcomes to the existing formulation while reducing the requirement for knowledge of vehicle parameters, while in unfavourable conditions neither approach performs adequately. Investigations of the vehicle response show that despite best efforts, bridge frequencies are not clearly and consistently visible. The initial vehicle-to-bridge acceleration amplitude ratio is found to be unsuited to bridge frequency visibility using indirect SHM, suggesting that prior positive results using this physical model may have managed to suppress vehicle-related acceleration amplitudes in a manner not achieved here. The importance of controlling such behaviour is thus highlighted and recommendations are made for future laboratory-scale experiments relating to VBI phenomena. Despite this, approximate frequency matching between the vehicle axle hop and bridge second bending modes allows one damage condition (simulated by the addition of static mass at bridge quarter-span) to be detected based on changes to VBI-induced time-frequency behaviour of the vehicle. Data derived from traversals of different bridge configurations are compared and the statistical significance of damage detection is confirmed, thus validating the proposed method.