Multi-Point Deflection Framework Accurately Identifies Stochastic Moving Loads on Bridges

Structural health monitoring of civil infrastructure relies on inverse dynamic analysis to accurately identify moving traffic loads, providing the empirical foundation for structural integrity evaluations, safety assessments, and fatigue life predictions. In a study published in Advances in Bridge Engineering, researchers Jiaqiang Zhang, Yuefei Liu, and Xueping Fan introduced a specialized AR(1)-Tikhonov-SVD computational framework driven by multi-point dynamic deflection responses to reconstruct stochastic moving forces acting on beam structures. Conventional moving load identification methodologies have predominantly addressed deterministic loading profiles or relied on idealized independent and identically distributed random assumptions that treat dynamic variations as uncorrelated white noise. However, operational vehicle-bridge interactions involve complex stochastic fluctuations caused by road surface roughness, vehicle braking maneuvers, and suspension dynamics that exhibit pronounced temporal continuity over time. To overcome these long-standing analytical simplifications, the research team formulated an integrated probabilistic modeling and regularized inversion strategy capable of capturing continuous temporal load correlations while resolving the severe ill-posedness inherent to inverse structural dynamic problems.

The foundational theoretical derivation developed by the investigators models an isotropic, elastic simply supported bridge beam characterized by computational span length, constant flexural rigidity, and uniform mass per unit length, supported by rigid boundaries at both ends. When a concentrated moving force traverses the beam structure at a constant forward velocity, the transient dynamic flexural behavior of the system is governed by the fourth-order Euler-Bernoulli partial differential equation. By applying the classical separation of variables technique across spatial coordinates and temporal variables, the dynamic deflection response at any given point along the span is expressed analytically through infinite modal series expansions representing structural vibration frequencies and representative generalized disturbance frequencies. Discretizing this closed-form analytical relationship across multiple predefined cross-sectional measurement positions and uniform sampling time intervals produces a compact linear system of equations relating structural dynamic deflections to applied discrete load histories through a global system transfer matrix. This unified analytical formulation provides the dynamic foundation necessary for reconstructing load histories directly from recorded multi-point displacement data.

To realistically represent time-varying load amplitudes without imposing prohibitive computational demands, the research team implemented a discrete first-order autoregressive stochastic model, designated AR(1), to generate temporally correlated load factor sequences following a normal distribution. In this stochastic formulation, an autoregressive coefficient governs the correlation between consecutive temporal increments, with the authors selecting an autoregressive coefficient value of 0.95 to reflect the high inertial persistence and filtering characteristics imparted by vehicular suspension systems during bridge crossings. The stochastic load factor sequence is subsequently coupled through element-wise multiplication with a nominal deterministic load pattern, establishing a continuous stochastic moving load profile that incorporates stationary variance while accurately mirroring short-term dynamic dependencies. The authors emphasized that selecting a parsimonious single-parameter autoregressive model avoids the parameter estimation instability, overfitting tendencies, and solution divergence that frequently compromise higher-order autoregressive moving average representations during ill-posed inverse reconstructions, thereby establishing an efficient bridge between oversimplified white noise assumptions and realistic physical load dynamics.

A central mathematical impediment in reconstructing moving loads from dynamic structural responses lies in the severe ill-posed nature of the governing system equations, wherein the dynamic transfer matrix is inherently ill-conditioned. Because tiny measurement perturbations or numerical discretization errors can trigger severe amplification and catastrophic divergence in the reconstructed load profile, the authors implemented classical Tikhonov regularization combined with Singular Value Decomposition to stabilize the solution space. Tikhonov regularization transforms the unstable inverse formulation into a well-behaved constrained optimization problem by augmenting the standard least-squares residual norm with an additional penalty term weighted by a positive regularization parameter. Applying Singular Value Decomposition decomposes the rectangular transfer matrix into orthogonal coordinate matrices and a descending diagonal matrix of singular values. This algebraic decomposition establishes a rigorous spectral filtering perspective, damping the destabilizing high-frequency solution oscillations driven by near-zero singular values and guaranteeing an analytically unique, stable regularized load time history across the discretized observation period.

Although Tikhonov regularization provides a sound mathematical framework for solution stabilization, the ultimate quality and accuracy of the identified load profile depend almost entirely on selecting an appropriate regularization parameter. Traditional inverse problem methodologies frequently rely on empirical parameter selection schemes such as the L-curve technique, which requires subjective human judgment, or the discrepancy principle, which demands precise prior knowledge of measurement noise levels that is rarely available in practical bridge monitoring installations. To eliminate these operational constraints, the authors embedded the Generalized Cross-Validation criterion directly into the regularized decomposition workflow to adaptively optimize the regularization parameter in a completely data-driven manner. Grounded in statistical predictive theory, the Generalized Cross-Validation algorithm identifies the optimal balance between minimizing response fitting residuals and suppressing unphysical load oscillations by locating the global minimum of a continuous cross-validation objective function. This automated optimization enables robust, objective parameter calculation without requiring prior noise statistics or manual tuning.

The authors conducted extensive numerical simulations using a finite element model of a simply supported structural steel beam in ANSYS to rigorously evaluate the performance of their AR(1)-Tikhonov-SVD identification framework. Modal extraction using the Block Lanczos method verified that the simulated natural frequencies of the first ten structural vibration modes aligned precisely with analytical theoretical predictions. Using dynamic deflection responses extracted at nine equally spaced cross sections under an applied stochastic load that produced a peak mid-span deflection of 4.578 millimeters, the proposed inversion method achieved a Relative Percentage Error of 8.41 percent under noise-free baseline conditions, verifying the fundamental accuracy of the integrated formulation.

Because field-deployed structural health monitoring sensors are inevitably exposed to environmental electromagnetic interference and measurement hardware noise, the researchers evaluated the noise robustness of their identification framework by superimposing synthetic Gaussian white noise onto the simulated deflection responses. Evaluating the framework under noise levels of 0.5 percent, 1.0 percent, and 2.0 percent demonstrated that the Relative Percentage Error increased systematically from the baseline value to 15.10 percent under two percent noise contamination. Visual analysis of the Generalized Cross-Validation functional landscape for the two percent noise case revealed an unequivocal global minimum point corresponding to an optimal regularization parameter of 3.3887 times ten to the negative eighteenth power. The authors highlighted that the existence of this pronounced mathematical minimum demonstrates how the Generalized Cross-Validation mechanism reliably balances data fidelity against noise suppression, preventing wild under-regularized oscillations while avoiding over-smoothing that would eliminate genuine stochastic load fluctuations.

Spatial sensor layout constitutes another critical design parameter influencing inverse identification fidelity, prompting the investigators to evaluate how varying the number of measurement stations impacts reconstruction accuracy. Testing symmetrical sensor configurations arranged around the beam mid-span indicated that increasing the number of deflection measurement points consistently reduced identification errors across all evaluated noise levels. Under two percent Gaussian white noise, moving from a single mid-span measurement point to a distributed nine-point sensor array reduced the Relative Percentage Error from 19.11 percent down to 15.10 percent, demonstrating the clear benefits of multi-point spatial data synergy. However, the simulation data simultaneously revealed distinct diminishing marginal returns: expanding from one to three measurement points reduced the reconstruction error by 2.07 percent, whereas expanding instrumentation from seven to nine sensors yielded an improvement of only 0.14 percent. Consequently, the authors emphasized that practical bridge monitoring programs must strategically balance spatial sensor density against instrumentation costs and data acquisition overhead.

In practical engineering structures, structural damping dissipates mechanical vibrational energy and alters dynamic response characteristics, which could potentially impair inverse identification accuracy if not properly understood. To evaluate this operational factor, the researchers incorporated Rayleigh damping with a two percent damping ratio into the simply supported beam model and compared the resulting load reconstructions against undamped baseline simulations. While the added structural complexity induced by damping caused a marginal increase in identification error under idealized noise-free conditions, the presence of structural damping actually enhanced identification accuracy across all noise-contaminated scenarios. The authors explained that structural damping acts as an intrinsic physical low-pass filter, naturally attenuating extraneous high-frequency vibrational components and smoothing noise-corrupted response signals. By working in synergy with Tikhonov mathematical regularization, mechanical damping suppresses high-frequency noise amplification, confirming that the proposed AR(1)-Tikhonov-SVD framework demonstrates robust performance and practical engineering adaptability when applied to damped structural systems.

Robustness verifications under non-Gaussian impulsive noise, spatially correlated sensor interference, and field-measured composite noise confirmed stable load reconstructions, while velocity tests from 10 to 25 meters per second restricted error growth below fifteen percent. Compared to computationally intensive deep learning models requiring thousands of seconds of training or iterative schemes prone to convergence instability, the closed-form SVD and adaptive GCV pipeline ensures predictable, training-free efficiency. While currently formulated for a single moving force on an elastic beam, extending this probabilistic framework to multi-axle vehicles, continuous girders, and physical experiments will advance real-time bridge safety assessment.

Subject of Research: Stochastic moving load identification on bridge structures using multi-point deflection responses

Article Title: Stochastic moving load identification driven by multi-point deflection responses: an AR(1)-Tikhonov-SVD framework and robustness verification

Article References: Zhang, J., Liu, Y., & Fan, X. (2026). Stochastic moving load identification driven by multi-point deflection responses: an AR(1)-Tikhonov-SVD framework and robustness verification. Advances in Bridge Engineering, 7(1), Article 9. https://doi.org/10.1186/s43251-025-00196-y

Image Credits: AI Generated

DOI: 10.1186/s43251-025-00196-y

Keywords: structural health monitoring, moving load identification, bridge engineering, Tikhonov regularization, singular value decomposition, generalized cross-validation, autoregressive modeling, finite element analysis, Stochastic, moving, load, identification