PiGGO: Physics-Guided Learnable Graph Kalman Filters for Virtual Sensing of Nonlinear Dynamic Structures under Uncertainty

1. Problem Statement

The paper addresses the challenge of virtual sensing (estimating structural responses at unmeasured locations) in complex, nonlinear dynamic structures under conditions of model-form uncertainty and sparse sensing.

• Limitations of Current Approaches:
  • Purely Physics-Based Models: Struggle to accurately represent complex nonlinearities arising from material effects, joint mechanisms, or discrete feedback forces. They often fail to capture inherent operational variability.
  • Purely Data-Driven Models (e.g., standard GNNs): While flexible, they suffer from error accumulation in online settings, especially when generalizing to structures with different parameters or topologies. They also struggle with sparse measurements.
  • Standard Kalman Filters (EKF): Require accurate, differentiable state-transition models, which are often unavailable for complex nonlinear systems.
• The Core Challenge: Developing a framework that can perform reliable, online state estimation for nonlinear systems with unknown or partially known governing equations, while maintaining robustness against measurement noise and model uncertainty.

2. Methodology: The PiGGO Framework

The authors propose PiGGO (Physics-Guided Graph Neural ODE), a unified framework that integrates a Graph Neural Ordinary Differential Equation (GNODE) with an Extended Kalman Filter (EKF).

A. Graph Representation & State-Space

• The structural system is modeled as a graph $G = \{X, E, A\}$, where nodes represent structural components and edges represent connections (e.g., springs/dampers).
• State Variables: Node features include displacement, velocity, and acceleration; edge features include connection extensions, rates, and geometric properties (angles).
• Physics-Guided Inductive Biases: The architecture explicitly encodes structural mechanics:
  • Equivariance: Features are designed to be invariant to coordinate system transformations (using sine/cosine of angles).
  • Force Balance: The state evolution is decomposed into a known linear physics component ($C_{phy}$) and a learned nonlinear component ($C_{bb}$).

B. The Physics-Guided GNODE (Offline Training)

• State Evolution: The continuous-time state transition is defined by a learnable differential equation: $\dot{Z} = F_{ev}(Z, G)$.
• Hybrid Dynamics: The evolution function $F_{ev}$ combines:
  1. Physics-Based Term: A standard linear force balance derived from known mechanics (stiffness, damping).
  2. Black-Box Term: A Graph Neural Network (GNN) that learns the residual nonlinear forces (e.g., cubic stiffness, gap nonlinearity).
• Training Objective: Instead of minimizing observation error alone, the model is trained using a Physics-Informed Loss based on the force-balance residual ($r_i = m\ddot{u} + \xi - f$). This ensures the learned dynamics adhere to Newton's laws, reducing the need for massive datasets and manual loss-weight tuning.

C. Graph Extended Kalman Filter (Online Inference)

• Integration: The pre-trained PiGGO model serves as the state-transition model within an EKF loop.
• Uncertainty Quantification: The EKF handles model-form uncertainty and measurement noise by:
  • Propagating the state mean and covariance.
  • Calculating the Kalman Gain to correct predictions based on sparse, noisy measurements.
  • Providing uncertainty-aware estimates (confidence intervals) for all state variables.
• Strategy: The framework uses a decoupled training strategy:
  1. Offline: Train the PiGGO to learn a generalizable representation of the system dynamics.
  2. Online: Embed the trained model into the EKF to correct for real-time deviations and noise.

3. Key Contributions

1. Unified Framework: Introduces PiGGO, the first framework to combine physics-guided GNODEs with recursive Bayesian filtering for online virtual sensing of nonlinear structures.
2. Physics-Guided Inductive Bias: Demonstrates how embedding structural mechanics (force balances, geometric equivariance) into the GNN architecture improves generalization and reduces data dependency.
3. Robustness to Uncertainty: Shows that the hybrid approach outperforms both open-loop GNODEs and standard EKFs when dealing with:
   • Model-form uncertainty: Variations in structural parameters (e.g., stiffness, damping) not seen during training.
   • Sparse Sensing: Accurate estimation even when 75–87.5% of nodes are unmeasured.
4. Uncertainty-Aware Prediction: The EKF component provides posterior covariance estimates that effectively capture model error and noise, offering reliable confidence intervals for safety-critical applications.

4. Results

The framework was evaluated on two numerical case studies:

1. Random Truss Array: A system with cubic nonlinear stiffness.
2. Bridge Truss Array: A system with non-smooth angular gap nonlinearity.

Key Findings:

• Offline Performance: The PiGGO model accurately learned the dynamics during training, matching ground truth for displacement, velocity, and acceleration.
• Online Generalization:
  • Open-Loop GNODE: When applied to larger, unseen structures with varying parameters, the open-loop model suffered from significant phase drift and amplitude errors due to accumulated integration errors.
  • PiGGO-EKF: The filtering scheme significantly reduced errors. For the random array, the Normalized Mean Squared Error (NMSE) for displacement dropped from 24.5 (open-loop) to 0.3 (EKF).
  • Bridge Truss: Similar improvements were observed, with the EKF correcting the bias and drift inherent in the open-loop predictions.
• Uncertainty Estimation: The posterior covariance estimates from the EKF successfully enveloped the true system responses, whereas the open-loop model provided no uncertainty quantification.

5. Significance and Impact

• Digital Twins for SHM: The work provides a scalable, robust solution for Structural Health Monitoring (SHM), enabling the creation of digital twins that can operate reliably despite incomplete knowledge of system parameters and limited sensor data.
• Bridging the Gap: It effectively bridges the gap between the interpretability of physics-based models and the flexibility of data-driven learning.
• Practical Applicability: The decoupled training/inference strategy makes the approach computationally feasible for real-time applications, avoiding the high cost of end-to-end training of neural-EKFs.
• Future Directions: The authors suggest extending the framework to handle local non-smooth behaviors (e.g., localized damage) and exploring adaptive graph topologies where the adjacency matrix itself is a latent variable.

In summary, PiGGO represents a significant advancement in Physics-Enhanced Machine Learning (PEML), offering a reliable method for monitoring complex, nonlinear infrastructure where traditional methods fail due to uncertainty and sparsity.