Answers are grounded in this paper and cite their sources.
Force estimates obtained from vibration measurements can absorb errors caused by sensor contamination and an incomplete structural model. This study examines both effects using an ℓ 1 acceleration misfit and a piecewise-smooth structural model. A generalized Gauss–Newton method computes force updates through local convex subproblems solved by the alternating direction method of multipliers. The continuous-time adjoint motivates this discrete formulation but does not establish global optimality. Three synthetic benchmarks consider unilateral contact, breathing-stiffness switching, and simultaneous beam loads. With fixed parameter settings, nonlinear ℓ 1 reconstruction gives relative force errors of 0.280, 0.054, and 0.126. Further beam comparisons select regularization from withheld measurements and include innovation-clipped Kalman smoothing, a residual-based physics-informed neural network, and a neural force model that enforces the discrete dynamics exactly. The latter gives lower force errors on two of three noise realizations. The batch method requires less time for refitting, but not when parameter selection is included. Refining the internal time step reduces the known-force response error from 1.69 × 10 −3 to 4.34 × 10 −6 against an independent continuous-model solution, with unchanged observation and force grids. These results show the value of combining outlier-resistant fitting with the relevant structural physics, while also showing the limits of response-based parameter selection and finite-budget optimization. The evidence is numerical and applies to continuous piecewise-linear restoring forces rather than impulsive contact or arbitrary discontinuities.
No takes yet. Share an insight, caveat, or question.
Mohamed Hamdaoui (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: