Abstract
MAntRA (Model-Agnostic Reliability Analysis) is a novel data-driven framework designed to assess the reliability of engineering systems even when their underlying physics is unknown or has evolved over time. By combining Variational Bayesian inference with stochastic differential equations (SDEs), MAntRA can "discover" the governing laws of a system from limited, noisy sensor data. This allows for accurate predictions of failure probabilities over extended timeframes, bridging the gap between traditional model-based reliability analysis and purely black-box machine learning.
Introduction: The Challenge of Aging Infrastructure
In structural engineering, reliability analysis—calculating the probability of failure—is a critical task. Traditionally, this relies on known mathematical models established during the design phase. However, as structures like bridges and heritage buildings age, they degrade, and their original models often cease to reflect reality.
When the governing physics changes or is unknown, traditional simulation methods like Monte Carlo fail. While machine learning offers a potential solution, many "black-box" models struggle to generalize beyond their training data. MAntRA addresses this by learning the underlying physics directly from measurements, creating an interpretable model for safety predictions.
Methodology: Stochastic Equation Discovery
MAntRA operates through a two-stage pipeline to identify the dynamics of a system:
- Stochastic Equation Discovery: The framework utilizes the Kramers-Moyal expansion to decouple a system's "drift" (deterministic trends) from its "diffusion" (random noise) based on observed output data.
- Variational Bayesian (VB) Inference: To identify specific governing equations, a sparse Bayesian learning approach is employed. Using a "spike and slab" prior, the model is forced to select only the most relevant mathematical terms from a library of candidates (e.g., polynomials, trigonometric functions).
Unlike previous methods that relied on computationally expensive Gibbs sampling, MAntRA introduces a Variational Bayes approach. This method is significantly faster and effectively accounts for both epistemic uncertainty (data scarcity) and aleatoric uncertainty (environmental randomness).
Experimental Validation
The framework was validated using three synthetic datasets of increasing complexity, each corrupted with 5% Gaussian noise to simulate real-world sensor inaccuracies:
- SDOF Duffing Oscillator: A single-degree-of-freedom system featuring cubic non-linearity.
- 3-DOF Nonlinear Oscillator: A multi-component system with complex restoring forces.
- 5-DOF Linear System with Tuned Mass Damper (TMD): A model representing a tall building with vibration control.
Key Insights and Results
MAntRA demonstrated remarkable performance across several metrics:
- Extrapolation and Prediction: Trained on only one second of noisy data, the model accurately predicted failure probabilities up to 30 seconds into the future. This confirms that the framework learns the underlying physics rather than just memorizing patterns.
- Parameter Accuracy: In the Duffing oscillator test case, the predicted non-linear parameter ($\alpha/m$) was 100,369.79, extremely close to the ground truth of 100,000.00.
- Sparsity and Interpretability: The model successfully ignored hundreds of irrelevant candidate functions, correctly identifying only the 3-5 terms governing the physical system. This results in an interpretable Stochastic Differential Equation that engineers can analyze.
Conclusion and Future Directions
MAntRA represents a significant step forward in real-time structural health monitoring. By making the "discovery" phase computationally feasible through Variational Inference, it enables the assessment of complex systems with minimal data.
However, challenges remain. The framework's success depends on the inclusion of the "true" physics within its library of candidate functions. Future work will focus on scaling the approach to high-dimensional systems (e.g., full-scale bridge models) and validating the framework using experimental data from actual vibrating structures in the field.
Summary updated