Probabilistic Graphical Model using Graph Neural Networks for Bayesian Inversion of Discrete Structural Component States

Abstract: The health condition of components in civil infrastructures can be described by various discrete states according to their performance degradation. Inferring these states from measurable responses is typically an ill-posed inverse problem. Although Bayesian methods are well-suited to tackle such problems, computing the posterior probability density function (PDF) presents challenges. The likelihood function cannot be analytically formulated due to the unclear relationship between discrete states and structural responses, and the high-dimensional state parameters resulting from numerous components severely complicates the computation of the marginal likelihood function. To address these challenges, this study proposes a novel Bayesian inversion paradigm for discrete variables based on Probabilistic Graphical Models (PGMs). The Markov networks are employed as modeling tools, with model parameters learned from data and structural topology prior. It has been proved that inferring this PGM produces the same probabilistic estimation as the posterior PDF derived from Bayesian inference, which effectively solves the above challenges. The inference is accomplished by Graph Neural Networks (GNNs), and a graph property-based GNN training strategy is developed to enable accurate inference across varying graph scales, thereby significantly reducing the computational overhead in high-dimensional problems. Both synthetic and experimental data are used to validate the proposed framework

• The paper presents a novel method for Bayesian inversion by formulating the problem as inference in pairwise Markov Random Fields, solved via a GNN-based message-passing engine.
• It efficiently estimates node and edge potentials using Gaussian Mixture Models and normalized mutual information to capture inter-component dependencies.
• Experimental results demonstrate superior accuracy and speed (sub-5ms per structure) compared to traditional inference methods in structural health monitoring.

Probabilistic Graphical Model using Graph Neural Networks for Bayesian Inversion of Discrete Structural Component States

Overview

The paper, "Probabilistic Graphical Model using Graph Neural Networks for Bayesian Inversion of Discrete Structural Component States" (2604.23514), addresses the Bayesian inversion problem for inferring discrete health states (e.g., intact or damaged) of structural components from sparse, noisy monitoring data in civil infrastructure systems. It introduces a paradigm that models the inference task as probabilistic graphical model (PGM) inference—specifically, as marginalization in pairwise Markov Random Fields (MRFs), and solves the associated inference tractability and likelihood modeling bottlenecks by deploying a graph neural network (GNN) inference engine that demonstrates strong size generalization and computational efficiency.

Bayesian Inversion in Structural Health Monitoring

Conventional Bayesian inversion frameworks in structural health monitoring contend with two major issues when applied to discrete-state estimation:

1. Lack of explicit likelihoods: For discrete component states, there does not exist an explicit physical or empirical model that links the states directly to observed monitoring data, making analytical formulation of $p(\mathcal{D}|\theta)$ infeasible.
2. High-dimensional marginalization: Modern engineered structures exhibit large numbers (often hundreds) of interdependent components, rendering direct computation of marginal likelihoods over all discrete configurations intractable due to the combinatorial explosion in the state space.

Covariance priors and traditional approaches underserve these tasks: covariance is suboptimal for nominal/discrete variables; most inversion methods inadequately encode complex dependency patterns.

PGM Representation and Learning

The authors circumvent these obstacles by formulating the joint distribution over discrete states $\mathbf{\theta}$ and observations $\mathcal{D}$ as a pairwise MRF (Ising-type graphical model):

$$p(\mathbf{\theta}, \mathcal{D}) \propto \exp\left[-\sum_{(i, j)\in E} W_{ij} \theta_i \theta_j - \sum_{i\in V} b_i \theta_i\right]$$

Here, $W_{ij}$ encode inter-component dependencies, and $b_i$ (a function of monitoring data $d_i$) capture the local evidence. Structure learning is grounded in the physical system topology—mapping sensorized locations to vertices and mechanical adjacencies to edges—preserving the semantics of load transfer and failure propagation.

Key elements in parameter learning:

• Node potentials ($b_i$): Estimated via Gaussian Mixture Models (GMMs) for intact states, and either GMMs (if damage data exist) or maximum entropy (uniform) for damage states; uses data domain estimation with BIC to tune mixture complexity.
• Edge potentials ($W_{ij}$): Inferred from normalized mutual information (MI) between adjacent sensor streams, supporting both dense sensor networks and scenarios with sparse failure propagation statistics.

This approach additionally provides a constructive proof of equivalence between the marginalization performed by the graphical model and the posterior distribution obtained from traditional Bayesian updates under appropriate potential encoding.

GNN-based Inference and Size Generalization

Traditional PGM inference schemes (variable elimination, BP, MCMC) suffer from exponential/superlinear complexity and poor scalability for cyclic, high-order graphs. The proposed framework operationalizes a message-passing GNN as a learned inference engine, drawing on the architectural similarity between BP and recurrent node-wise updates in MPNNs.

The GNN operates as follows:

• Inputs: Edge and node potentials $(W_{ij}, b_i)$ derived from data and prior.
• Message passing: At each iteration, each node aggregates messages from its neighbors via shared neural encoders (analogous to BP), with GRU-based state updating.
• Output: After $\mathbf{\theta}$0 iterations, hidden states are decoded to yield marginal state probabilities.

Size generalization—the ability to transfer GNN inference accuracy from small to large graph instances—is a central focus. Training on smaller graphs with matching average unique node degree yields robust performance for larger structures, provided the degree distribution is aligned. The theoretical underpinnings are informed by Rademacher complexity analysis and results from recent GNN generalization literature. Empirical studies systematically dissect the impact of training set size, graph order, and node degree on extrapolative inference accuracy, with Kullback-Leibler divergence as the principal metric.

Experimental Validation and Numerical Results

The framework's validity and competitiveness are established via synthetic and real-data experiments on truss-type structures:

• Synthetic studies: Demonstrate that GNN outperforms BP and MCMC in terms of both inference accuracy (mean K-L divergence) and computational time, especially as structure size increases. The analysis confirms that inference accuracy degrades principally with mismatched node degree distributions rather than graph order per se. When average unique node degree between training and application graphs is well-matched, GNN achieves strong accuracy even with minimal training data.
• Physical truss testbed: The method is tested on video-derived modal data from a multi-bar truss. Results show that the GNN achieves low false positive rates, higher F1 scores ($\mathbf{\theta}$1 mean), and the highest accuracy ($\mathbf{\theta}$2 mean) among approximate algorithms. GNNs complete inference in sub-5ms per structure, compared to 10–100ms for sampling-based methods.

Notably, explicit modeling of spatial dependencies among states is shown to be essential—control experiments in which dependencies are ignored yield markedly elevated false positives and undetected damages.

Implications and Future Directions

This paradigm offers a methodologically robust approach for Bayesian inversion of discrete component states in large-scale, high-dimensional engineered systems:

• Practical implications: Enables real-time, scalable, and uncertainty-quantified diagnosis for civil infrastructures using networked sensing, even in the presence of incomplete or noisy data. The size generalization capability facilitates rapid transfer to new assets without exhaustive retraining or expensive labeled data.
• Theoretical implications: Systematic use of physical priors in PGM structure, data-driven parameterization of potentials, and leveraging GNN generalization theory narrow the gap between statistical machine learning and domain-driven engineering modeling for SHM.

For extending industrial adoption and scientific reach:

• Development of multi-class component state representations (beyond binary) would allow richer civil diagnosis and prognosis.
• Integration with spatio-temporal PGMs may enable dynamic modeling under varying operational and environmental conditions.
• Hierarchical or non-local dependency structures could be incorporated for improved modeling of long-range couplings.

Conclusion

By integrating PGMs with GNN-based scalable inference and leveraging size generalization, this work provides a formalized and efficient solution to discrete-state Bayesian inversion for large and complex engineering systems. The proposed framework achieves superior inference accuracy and efficiency compared to classical approximate methods in both synthetic and real-world scenarios, while offering strong extensibility to high-dimensional, incomplete, or uncertain data regimes (2604.23514). This positions the methodology as a significant advance for structural health monitoring and broader engineering diagnostics involving high-dimensional discrete variables.