Probabilistic Updating of Structural Models for Damage Assessment Using Approximate Bayesian Computation

Zhouquan Feng, Yang Lin, Wenzan Wang, Xugang Hua, Zhengqing Chen2020 · DOI

TL;DR

This paper introduces a novel method for assessing structural damage by updating computational models using vibration data. It leverages Approximate Bayesian Computation with Subset Simulation (ABC-SubSim) to overcome limitations in traditional Bayesian methods, offering a more robust and efficient approach to identifying structural damage even with noisy data.

What Problem Does This Paper Address?

Ensuring the safety and serviceability of structures is paramount, and Structural Health Monitoring (SHM) plays a crucial role in detecting and characterizing damage. Vibration-based methods, which analyze a structure's dynamic behavior (like modal frequencies and shapes), are a popular global approach for SHM. A common technique is "model updating," where a computational model (often a finite element model) is calibrated against experimental data to better reflect the structure's actual condition. This process is essential for damage assessment, as damage typically alters a structure's stiffness.

However, traditional model updating methods face challenges:

  • Uncertainty: Real-world structures and measurements are subject to various uncertainties (modeling errors, sensor noise), which deterministic methods struggle to handle effectively.
  • Likelihood Function: Probabilistic methods, particularly Bayesian inference, are better at handling uncertainty by estimating the probability distributions of model parameters. However, constructing an explicit "likelihood function" (which describes the probability of observing the data given a model) can be mathematically complex or even impossible for sophisticated models.
  • Computational Cost: Even with probabilistic approaches, obtaining accurate posterior distributions of model parameters can be computationally intensive, especially when dealing with rare events (like precise matches between simulated and observed data).

This paper aims to address these issues by proposing a probabilistic model updating approach that avoids explicit likelihood functions and improves computational efficiency.

Methods

The core of this paper is the development and application of a Probabilistic Updating of Structural Models using Approximate Bayesian Computation with Subset Simulation (ABC-SubSim). Here's a breakdown of the key methodological contributions:

  • Approximate Bayesian Computation (ABC): This is a "likelihood-free" Bayesian approach. Instead of directly calculating the likelihood function, ABC simulates data from the model and compares it to the observed data using a "metric" and a "tolerance level" ($\epsilon$). If the simulated data is "close enough" to the observed data (within the tolerance), the model parameters used for simulation are considered plausible.
  • Subset Simulation (SubSim): To improve the efficiency of ABC, especially for rare events (i.e., when $\epsilon$ is small), the authors incorporate subset simulation. This technique breaks down the problem into a sequence of nested regions of increasing similarity to the observed data. Samples are generated in stages, with each stage building upon the previous one, effectively guiding the search towards regions of interest and reducing computational cost compared to direct Monte Carlo simulation.
  • Novel Stopping Criteria: A significant contribution is the introduction of new stopping criteria for the subset simulation. Instead of pre-defining a fixed tolerance ($\epsilon$), the algorithm adaptively determines when to stop based on the convergence of the model parameters and the metric values across simulation levels. This makes the process more robust and less reliant on subjective parameter choices.
  • Hybrid Optimization Scheme: To refine the optimal model parameters identified by ABC-SubSim, a hybrid approach is used. After the subset simulation converges, a local optimization search operator (specifically, the BFGS method) is applied to fine-tune the parameters, leveraging the efficiency of the global search of ABC-SubSim and the precision of local optimization.
  • Iterative Weighting Factor Determination: The metric used to compare model predictions and observations often involves weighting factors for modal frequencies and mode shapes. The paper proposes an iterative approach to determine these optimal weighting factors. This is crucial because the optimal weights depend on the model parameters themselves, creating a circular dependency that is resolved through an iterative process.

Datasets

The paper does not introduce new benchmark datasets. Instead, it uses simulated modal data for three different structural examples:

  1. An eight-story shear building.
  2. A 21-bar planar truss.
  3. A simply supported beam.

These datasets are generated by simulating modal frequencies and mode shapes and then adding controlled levels of white noise (e.g., 1% and 5% coefficient of variation) to mimic real-world measurement uncertainties.

Key Insights and Ideas

  • Likelihood-Free is Powerful: The success of ABC-SubSim demonstrates that complex structural models can be effectively updated and used for damage assessment even when their underlying likelihood functions are intractable. This opens doors for applying these methods to a wider range of real-world engineering problems.
  • Adaptive Stopping is Key: Moving away from fixed tolerance levels for ABC and instead using convergence criteria for model parameters and metrics makes the process more practical and less sensitive to arbitrary parameter choices. This is a crucial step towards making ABC methods more user-friendly.
  • Hybridization Boosts Performance: Combining the global search capabilities of subset simulation with the local refinement of BFGS optimization provides a powerful strategy for efficiently finding accurate optimal model parameters.
  • Weighting Matters (and can be learned): The iterative determination of weighting factors for modal frequencies and mode shapes highlights the importance of how different types of data are combined. Learning these weights adaptively allows the algorithm to better balance the information from different sources, leading to more reliable updates.
  • Sensor Placement Impacts Uncertainty: The truss example shows that even with a good updating method, the inherent uncertainty in identified parameters can vary significantly depending on the sensor layout. This underscores the importance of sensor placement in SHM.

Key Results

The proposed ABC-SubSim method, combined with the hybrid optimization and iterative weighting, demonstrated high accuracy in updating structural model parameters across various noise levels and structural types.

  • Accurate Parameter Identification: In the shear building example, identified optimal parameters were very close to their true values, even with up to 10% noise.
  • Damage Assessment Effectiveness: The method successfully identified damage extents in the truss and beam examples, with probability curves accurately reflecting the simulated damage locations and severities. For instance, damage extents of 21% and 42% were correctly identified for specific bars in the truss, and 10%, 20%, and 30% for elements in the beam.
  • Reduced Uncertainty: The use of the BFGS fine-tuning and stricter stopping criteria led to reduced coefficients of variation (CV) for the identified model parameters, indicating lower uncertainty in the updated model.
  • Fast Convergence: The iterative process for determining weighting factors converged quickly, and the subset simulation also showed efficient convergence to the desired levels.

Limitations and Future Work

  • Stopping Criteria Justification: While the proposed stopping criteria are more operable, the authors acknowledge that a deeper theoretical study is needed to fully understand the relationship between likelihood-informed Bayesian inference and likelihood-free ABC, which could lead to more theoretically grounded parameter selection.
  • Sensor Placement: The truss example highlighted that sensor configuration significantly impacts the uncertainty in identified parameters. While this is not the primary focus of the paper, it suggests that future work could explore integrating sensor placement optimization with the proposed model updating framework.
  • Generalizability: While three illustrative examples were used, further validation on more complex and diverse real-world structures would strengthen the findings.
  • Computational Cost: Although ABC-SubSim is more efficient than direct ABC, it can still be computationally demanding for very large-scale structures or extensive simulations. Exploring further algorithmic optimizations or parallelization could be beneficial.

Summary updated

BibTeX
@misc{feng2020probabilistic,
  author = {Zhouquan Feng and Yang Lin and Wenzan Wang and Xugang Hua and Zhengqing Chen},
  title = {{Probabilistic Updating of Structural Models for Damage Assessment Using Approximate Bayesian Computation}},
  year = {2020},
  doi = {10.3390/s20113197},
}
Figure