An unscented universal filter for joint input-state-parameter estimation of a partially measured system: an application to drive-by structural health monitoring

Abstract

Drive-by structural health monitoring infers the condition of a bridge from the response of an instrumented vehicle, whose measurements are dominated by the road roughness and the vehicle dynamics. The bridge responses, contact forces and bridge parameters depend on one another through the vehicle-bridge interaction (VBI), so estimating them jointly is desirable. This paper proposes a novel filter in the Universal Filter family, the Unscented Universal Filter (UUF), for joint input-state-parameter estimation of nonlinear systems with unknown inputs. The unknown input is estimated by weighted least squares from the innovation, without any evolution model or statistics, and the system inversion imposes no rank condition on the feedforward matrix. The nonlinear propagation is linearised by the scaled unscented transform, and the linearisation residuals are propagated into the error covariances. For the acceleration-only configuration, auxiliary integrators form pseudo-measurements from the acceleration measurements themselves. The UUF is applied to the coupled VBI system, with the contact forces as the unknown input and the local stiffness of the bridge as the tracked parameter, from the responses of the vehicle body and at most a few bridge responses. In numerical case studies on a simply supported beam and the Old Ada Bridge, a steel truss bridge in Japan, the UUF is compared with classical and state-of-the-art joint estimation filters, all tuned by the same practical strategy. Across the damage severities, locations, road classes and noise levels examined, the UUF yields the lowest total error in most scenarios, and its error varies little with damage severity, location and road class. On the Old Ada Bridge, it identifies the stiffness of a damaged member from acceleration-only measurements, without the suspension parameters of the vehicle, and with a rank-deficient feedforward matrix.