PaperlayerView exact revision on arXiv

Source-linked AI summary

Bayesian System Identification based on Hierarchical Sparse Bayesian Learning and Gibbs Sampling with Application to Structural Damage Assessment

Yong Huang, James L. Beck, Hui Li

arXiv:1701.03550v1stat.APstat.MEstat.ML

TL;DR

Noisy, incomplete modal data complicate the identification of sparse structural stiffness changes. The paper proposes two Gibbs sampling algorithms for Bayesian system identification, and reports better performance than Algorithm 1 and a previous method on the challenging IASC-ASCE Phase II experimental benchmark.

  • Problem

    The paper addresses identifying spatially sparse stiffness changes from noisy, incomplete modal data, where inverse problems may be ill-conditioned or ill-posed.

    Source: §1 · p. 2, §7 · p. 21

  • Method

    The authors present two Gibbs sampling algorithms that sample conditional posteriors for modal and stiffness parameters and differ in their treatment of equation-error precision uncertainty.

    Source: §7 · p. 21

  • Results

    On the challenging IASC-ASCE Phase II experimental benchmark, the reported performance is better than Algorithm 1 and the previous method.

    Source: §7 · p. 21

  • Takeaways & Limitations

    The algorithms provide a Bayesian approach to infer spatial stiffness changes from incomplete modal data.

    Source: §7 · p. 21

    The introduction notes that most existing MCMC methods for structural model updating are efficient only for lower-dimensional problems.

    Source: §1 · p. 3

Abstract

from arXiv · show

The focus in this paper is Bayesian system identification based on noisy incomplete modal data where we can impose spatially-sparse stiffness changes when updating a structural model. To this end, based on a similar hierarchical sparse Bayesian learning model from our previous work, we propose two Gibbs sampling algorithms. The algorithms differ in their strategies to deal with the posterior uncertainty of the equation-error precision parameter, but both sample from the conditional posterior probability density functions (PDFs) for the structural stiffness parameters and system modal parameters. The effective dimension for the Gibbs sampling is low because iterative sampling is done from only three conditional posterior PDFs that correspond to three parameter groups, along with sampling of the equation-error precision parameter from another conditional posterior PDF in one of the algorithms where it is not integrated out as a "nuisance" parameter. A nice feature from a computational perspective is that it is not necessary to solve a nonlinear eigenvalue problem of a structural model. The effectiveness and robustness of the proposed algorithms are illustrated by applying them to the IASE-ASCE Phase II simulated and experimental benchmark studies. The goal is to use incomplete modal data identified before and after possible damage to detect and assess spatially-sparse stiffness reductions induced by any damage. Our past and current focus on meeting challenges arising from Bayesian inference of structural stiffness serve to strengthen the capability of vibration-based structural system identification but our methods also have much broader applicability for inverse problems in science and technology where system matrices are to be inferred from noisy partial information about their eigenquantities.

1. INTRODUCTION

The introduction motivates Bayesian structural identification as a way to represent uncertainty in ill-conditioned inverse problems, and sparse stiffness changes as a useful damage prior. It presents Gibbs sampling to address uncertainty left by earlier approximations and applies the algorithms to IASC-ASCE Phase II benchmark data.

1. INTRODUCTION

  • Noisy, incomplete modal data make inverse solutions vulnerable to ill-conditioning and ill-posedness, motivating methods that account for uncertainty rather than returning only a single estimate.The Bayesian formulation represents the family of plausible parameter values consistent with observations and prior information.Source: §1 · p. 2
  • Damage-related stiffness changes are expected at a limited number of locations, so sparse Bayesian learning can encode this structural prior during model updating.The paper also notes that its methods can infer non-sparse stiffness changes when sufficient data are available.Source: §1 · p. 2
  • Earlier Laplace approximations ignored posterior uncertainty in hyperparameters and system modal parameters, motivating a fuller treatment using MCMC.MCMC generates samples consistent with the posterior distribution of model parameters.Source: §1 · p. 3
  • Two Gibbs sampling algorithms address equation-error precision uncertainty by iteratively sampling parameter groups from conditional posterior PDFs, keeping the sampler's effective dimension low.The algorithms differ in how they handle uncertainty in the equation-error precision parameter.Source: §1 · p. 3
  • The proposed sparse stiffness-loss inversion is illustrated with simulated and experimental IASC-ASCE Phase II benchmark data.The stated goal is to detect and assess spatially sparse stiffness reductions using modal data identified before and after possible damage.Source: §1 · pp. 2–3

2. BAYESIAN SYSTEM IDENTIFICATION BASED ON MODAL DATA

This section formulates structural updating from incomplete modal observations as hierarchical Bayesian inference over stiffness scaling and system modal parameters, accounting for equation-error uncertainty and optional sparsity around calibration. It describes a Gibbs sampling variant that integrates out equation-error precision and samples three parameter groups from conditional posterior PDFs.

2.1. Linear structural model class and structural stiffness scaling parameters

  • Under low-amplitude linear-dynamics assumptions, stiffness is parameterized by substructure matrices scaled by learnable parameters, optionally targeting damage-prone hot spots.Source: §2.1 · pp. 3–4

2.2. Modal data and system modal parameters

  • With sensors observing only a subset of degrees of freedom, full-system mode shapes and natural frequencies represent the underlying modal parameters.Source: §2.2 · p. 4

2.3 Hierarchical Bayesian modeling

  • Hierarchical Bayesian modeling can learn uncertainty and correlation from data, potentially producing more reliable system-identification results.Source: §2.3 · p. 4

2.3.1 Prior for system modal parameters and structural stiffness scaling parameters

  • Gaussian equation errors allow system modal parameters to deviate from exact model eigenpairs, providing soft constraints that tighten as precision increases.The shared equation-error precision parameter is β; as β → ∞, modal parameters cluster around values implied by the structural model.Source: §2.3.1 · pp. 4–5

2.3.2. Likelihood function for system modal parameters and structural stiffness scaling parameters

  • Identified mode shapes and frequencies have independent Gaussian observation errors, while a Gaussian pseudo-likelihood around calibration encourages sparse stiffness changes.Sparsity is motivated by damage-induced changes localized to a small number of substructures.Source: §2.3.2 · p. 6

2.3.3. Full posterior PDF

  • The posterior’s normalizing integral is intractable, so Gibbs sampling updates four groups: system modes, frequencies, stiffness parameters, and equation-error precision.The algorithm repeatedly samples each group conditional on the other three and the available data.Source: §2.3.3 · p. 7
  • The conditional posterior PDFs required to implement Gibbs sampling can be established analytically.Source: §2.3.3 · p. 7
  • To improve Gibbs sampling efficiency, the algorithm integrates out the equation-error precision parameter β.Source: §2.3.3 · p. 7
  • The algorithm repeatedly draws conditional samples from only three parameter groups.Source: §2.3.3 · p. 7

3.1 Full conditional posterior PDFs without marginalizing over equation-error precision ߚ

The section derives conditional posterior distributions for mode shapes, squared natural frequencies, stiffness scaling parameters, and equation-error precision. Its updates use MAP approximations for selected nuisance parameters, while the stiffness update can impose sparsity through hierarchical priors.

3.1.1 Conditional posterior PDF for ૖

  • The mode-shape posterior integrates over equation-error precision and approximates that integral by evaluating at its MAP value; its mean and covariance are then updated iteratively.The section cycles through the mean, covariance, and MAP update until convergence.Source: §3.1.1 · pp. 7–8

3.1.2 Conditional posterior PDF for ૑ଶ

  • The squared-frequency parameters have a conditional Gaussian posterior, with its mean, covariance, and MAP precision estimate determined by cycling through their update equations.The posterior mean and covariance depend on the current structural stiffness parameters and mode shapes.Source: §3.1.2 · p. 8

3.1.3 Conditional posterior PDF for ી

  • The stiffness-scaling update can impose sparsity through hyperparameters that drive corresponding stiffness changes toward negligible values; setting them to infinity removes the sparsity terms.The section describes this sparse update as analogous to relevance-based feature pruning.Source: §3.1.3 · pp. 8–9

3.1.4 Conditional posterior PDF for ߚ

  • The equation-error precision posterior is approximated by plugging in the MAP value of its hyperprior parameter, whose posterior is optimized under a uniform prior.The resulting conditional posterior for precision is described as a gamma distribution.Source: §3.1.4 · p. 9

3.2 Full conditional posterior PDFs when marginalizing over equation-error precision ߚ

When equation-error precision β is marginalized, Gibbs sampling uses three conditional posterior PDFs for the mode shapes, natural frequencies, and stiffness parameters. The section derives each conditional update and describes how sparsity can be imposed on stiffness changes.

3.2 Full conditional posterior PDFs when marginalizing over equation-error precision ߚ

  • Marginalizing β leaves three conditional posterior PDFs for Gibbs sampling of the mode shapes, natural frequencies, and stiffness parameters.These are the three parameter groups sampled in this section.Source: §3.2 · p. 10

3.2.1 Conditional posterior PDF for ૖

  • The mode-shape conditional posterior is obtained by integrating over β, with its posterior mean and matrix determined iteratively using the MAP value.The derivation uses the substitution τ = βη.Source: §3.2.1 · p. 10

3.2.2 Conditional posterior PDF for ૑ଶ

  • The natural-frequency conditional posterior is likewise obtained by integrating over β, and its parameter values are found by iteration.The derivation uses the substitution βρ.Source: §3.2.2 · pp. 10–11

3.2.3 Conditional posterior PDF for ી

  • For stiffness parameters θ, the conditional posterior can promote sparsity in Δθ; many optimized γ_j values tend toward zero.Taking all γ_j to infinity removes the sparsity constraint when sparse changes are not desired.Source: §3.2.3 · p. 11

3.3 Pseudo-code for Gibbs sampling algorithms

The two Gibbs samplers repeatedly draw conditional samples for modal and stiffness parameters, with Algorithm 1 sampling β and Algorithm 2 integrating it out. The section also describes chain behavior and a visual procedure for choosing burn-in and checking convergence.

3.3 Pseudo-code for Gibbs sampling algorithms

  • Gibbs samples target the full posterior when the chain is ergodic; a single chain may be ineffective for locally identifiable cases with separated high-posterior regions.The text suggests parallel chains from multiple prior-based starting points for such cases.Source: §3.3 · pp. 11–12
  • Algorithm 1 samples mode shapes, natural frequencies, β, and stiffness parameters, whereas Algorithm 2 integrates out β and samples the other three parameter groups.Both algorithms generate samples by successively drawing from the conditional posterior PDFs.Source: §3.3 · p. 12

3.3.1. Burn-in period determination

  • The authors choose burn-in by visually inspecting sequential chain plots and check convergence by plotting stiffness samples of interest.The number of post-burn-in samples is assessed through the plotted stiffness-sample convergence.Source: §3.3.1 · p. 13
  • The modal data depend nonlinearly on stiffness scaling, but the formulation replaces that inverse problem with coupled linear-in-parameter problems.This structure permits analytical construction of the conditional posterior PDFs used for Gibbs sampling.Source: §3.3.1 · p. 13
  • With an ergodic chain, identified stiffness scaling is unchanged by linear rescaling of measured mode shapes and mass and stiffness matrices.The text states that identification results therefore do not change when the units of these quantities change.Source: §3.3.1 · p. 13

4. GIBBS SAMPLING FOR DAMAGE ASSESSMENT

Damage assessment combines calibration-stage and monitoring-stage modal data to sample stiffness parameters while carrying calibration uncertainty into the monitoring inference. Damage probability is then estimated by comparing sampled stiffness values against a prescribed reduction threshold.

4.1 Generating posterior samples of ી conditional on modal data from both monitoring and calibration stages

  • Sparsity may be imposed on Δθ = θ − θ_u during damage monitoring because damage-induced stiffness loss is typically localized to a small number of substructures.The calibration stage assumes no sparsity constraint on stiffness changes.Source: §4.1 · p. 13
  • The two-stage procedure uses calibration samples of undamaged stiffness in monitoring-stage sampling to account for uncertainty in the baseline stiffness.Calibration samples are paired with monitoring-stage stiffness samples to characterize the posterior for the possibly damaged structure.Source: §4.1 · pp. 13–14

4.2 Evaluation of damage probability

  • For each substructure, damage probability is estimated as the probability that monitoring-stage stiffness falls below a prescribed fraction of calibration-stage stiffness.The estimate uses paired posterior samples from the calibration and monitoring stages.Source: §4.2 · pp. 14–15

5. COMPARISON OF NEW ALGORITHMS WITH THAT IN [15]

The comparison discusses robustness, uncertainty characterization, and computational cost relative to the earlier method. Algorithm 2 is reported to outperform the comparison algorithms, while the Gibbs approach requires more computation for full posterior uncertainty quantification.

5.1 Theoretical benefits for real applications

  • The proposed Gibbs algorithms characterize posterior uncertainty in stiffness and modal parameters even when the model class is not globally identifiable.The discussion highlights this as useful when calibration data do not yield a unique stiffness peak with small uncertainty.Source: §5.1 · p. 15
  • Learning hyperparameters balances model-class data fit against information extracted from the data, penalizing underfit and overfit models.The authors connect this trade-off to the Bayesian Ockham Razor.Source: §5.1 · p. 15
  • Algorithm 2 outperforms the compared algorithms, and its marginalization over β addresses the sensitivity to equation-error precision reported for [15].The proposed algorithm samples Student-t conditional PDFs with heavier tails than Algorithm 1's Gaussian PDFs, which the authors say makes it robust to noise and outliers.Source: §5.1 · p. 16

5.2 Comparison of the computational costs

  • The proposed Gibbs samplers require more computation than the fast method in [15], trading computation time for fuller posterior uncertainty quantification.Algorithm 2 reduces the effective sampling dimensions from four to three and requires fewer iterations to characterize parameter distributions.Source: §5.2 · p. 16

6. ILLUSTRATIVE EXAMPLES

The examples apply the proposed Gibbs-sampling algorithms to simulated unbraced damage cases and simulated and experimental IASC-ASCE Phase II braced-damage cases. They report generally successful damage detection and quantification, alongside greater uncertainty and some false-detection risk with experimental data.

6.1 Example I: Unbraced damage cases in simulated Phase II benchmark problem

  • Algorithm 2 reached its stationary state after roughly 2,000 burn-in samples, versus 6,000 for Algorithm 1; its higher efficiency is consistent with Gibbs samplers’ effective dimensions decreasing from four to three through integration.Across different sample counts, damage-probability estimates were almost the same for Algorithm 2 but differed for Algorithm 1, indicating that Algorithm 2’s GS Markov chain samples better represented the target PDF.Source: §6.1 · pp. 17–18
  • Both algorithms correctly localized actual damage at θ1,y and θ2,y in DP1U and θ1,y in DP2U; Algorithm 1 also falsely detected DP1U damage at θ1,x.For Algorithm 2, posterior median estimated losses were about 50% at θ1,y and θ2,y in DP1U and 30% at θ1,y in DP2U, close to their actual values.Source: §6.1 · p. 18
  • Compared with the method in, the new Gibbs-sampling methods had a smaller spread in post-burn-in samples and identified undamaged substructures with higher confidence.The paper attributes these improvements to its hierarchical Bayesian model, which learns the equation-error precision prior rate parameter.Source: §6.1 · p. 18

6.2 Example II: Braced damage case in simulated and experimental Phase II benchmark problems

  • The braced-frame example examines four simulated damage patterns with RB as the undamaged calibration case, and five experimental brace-removal configurations with Config. 1 as calibration.Only the partial-sensor simulated-data scenario is reported; experimental acceleration measurements are obtained using sledgehammer impacts.Source: §6.2 · pp. 18–19
  • Algorithm 2 required roughly 200 Markov Chain samples to reach the stationary state, while Algorithm 1 took roughly 4000 samples.For the experimental undamaged case, 16 parameter components produced samples spread over an unacceptably wide area; the three identified modal-data sets poorly supported reliable inference of stiffness-scaling parameters.Source: §6.2 · p. 19
  • About 12% loss in θ₁,₋ᵧ and 5% in θ₃,₋ᵧ were estimated for simulated DP3B; experimental Config. 5 estimates were also identified and quantified, but had large variance.The Config. 5 variance was attributed to larger modeling errors in the real-data case.Source: §6.2 · p. 20
  • Experimental damage probabilities were large for actual damaged substructures in Configs. 2–6, except Config. 6; undamaged substructures more often showed apparent stiffness increases than in simulation.For Config. 6, the estimated damage in θ₂,₊ₓ was underestimated by approximately 20% by all three methods, versus the real stiffness reduction ratio of 45.1%.Source: §6.2 · p. 20
  • Comparisons suggested false damage detections were less likely for Algorithm 2 than for Algorithm 1 and method, owing to smaller model-parameter uncertainty; in Config. 5, its possible false detection had much lower damage probability than two actually damaged substructures.Source: §6.2 · p. 21

7. CONCLUSION AND FUTURE WORK

The conclusion presents two Gibbs samplers for sparse Bayesian structural system identification and reports Algorithm 2’s stronger performance on the challenging experimental benchmark. It also highlights the samplers’ data-driven, sparse, and scale-invariant stiffness identification features.

7. CONCLUSION AND FUTURE WORK

  • For incomplete modal data, stiffness changes are represented through soft prior constraints, while hyperparameters are learned from data without parameter tuning.Sparse Bayesian learning produces sparse inferred stiffness losses to regularize ill-posed identification and damage assessment.Source: §7 · p. 21
  • On the experimental benchmark, Algorithm 2 outperformed Algorithm 1 and the prior sparse Bayesian learning method; it also reached stationarity faster in simulation and produced more concentrated experimental samples.The conclusion describes the experimental case as having larger modeling errors and the simulated case as having smaller modeling errors.Source: §7 · p. 21
  • The algorithms identify system natural frequencies alongside mode shapes and use structural eigenvalue equations without solving a nonlinear eigenvalue problem.Frequencies are introduced as parameters to represent the structural system’s actual natural frequencies.Source: §7 · p. 21
  • Algorithm 2 analytically marginalizes equation-error precision, reducing the sampler’s effective dimension and posterior sample variances.The Gibbs sampler draws from a small number of parameter groups, so its efficiency does not degrade with larger uncertain parameter spaces.Source: §7 · p. 21
  • Stiffness identification is scale-invariant, independent of mass and stiffness matrix units and measured mode-shape scaling.Source: §7 · pp. 21–22
  • The proposed framework is suggested for future sequential tracking of structural parameters, incorporating previous updates into current ones.Source: §7 · p. 22

Nomenclature

  • The nomenclature defines stiffness-scaling parameters and distinguishes undamaged-stage parameters, modal quantities, and measured and model dimensions.Source: Nomenclature · p. 22

Appendix A: Derivation of conditional posterior PDFs for Algorithm 1

The appendix derives the conditional posterior distributions used in Algorithm 1’s Gibbs sampler. It details a Laplace approximation for integrating nuisance hyperparameters and the MAP calculations underlying that approximation.

Appendix A: Derivation of conditional posterior PDFs for Algorithm 1

  • The appendix derives conditional posterior PDFs for the mode shapes, natural frequencies, and stiffness parameters used in Gibbs sampling.Source: Appendix A · p. 25
  • The likelihood formulation treats available data as a linear prediction plus Gaussian error with covariance determined by uncertain hyperparameters.Source: Appendix A · p. 25
  • For a parameter vector, nuisance hyperparameters are integrated using a Laplace approximation at their MAP value, yielding a Gaussian conditional posterior.The approximation assumes the hyperparameter posterior has a unique maximum.Source: Appendix A · pp. 25–26
  • The approximation’s hyperparameter MAP value is obtained by maximizing a log-posterior objective and setting its derivative to zero for updates.Source: Appendix A · pp. 26–27

Appendix B: Derivation of conditional posterior PDFs for Algorithm 2

Appendix B derives Algorithm 2’s posterior distributions and hyperparameter updates, including treatment of equation-error precision β. It also integrates over β to form marginal posteriors and a likelihood for MAP hyperparameter estimation.

Appendix B: Derivation of conditional posterior PDFs for Algorithm 2

  • The derivation specifies a Gamma posterior for equation-error precision β, with shape and rate determined by the model dimensions and prior parameter b0.The data dimensionality K appears in the shape parameter.Source: Appendix B · pp. 27–28
  • The derivation rewrites the quadratic expression as the sum of a modal-data residual term and a model-equation residual norm.This identity is used in the expression for the likelihood involving β and b0.Source: Appendix B · pp. 27–28
  • Integrating over β yields a marginal posterior for x1, expressed as a mixture of conditional Gaussian distributions weighted by β’s posterior.The conditional mean and covariance are given in terms of the model matrices and posterior parameters.Source: Appendix B · p. 28
  • The likelihood for the remaining parameter groups is obtained by integrating over β, then used to form the MAP objective for ν and b0 under broad uniform priors.The derivation rescales prediction-error covariance parameters using ν = βκ and Φ = βL.Source: Appendix B · pp. 27–28
  • Setting derivatives of the MAP objective to zero gives update formulas for b0 and each component ν_m.The derivation uses the determinant identity in B12 and notes that terms involving derivatives of μ drop out.Source: Appendix B · p. 29
Loading 1701.03550v1…