MAntRA: A framework for model agnostic reliability analysis

Yogesh Chandrakant Mathpati, Kalpesh Sanjay More, Tapas Tripura, Rajdip Nayek, Souvik Chakraborty

TL;DR

MAntRA is a new data-driven framework that assesses the reliability of engineering systems even when their underlying physics is unknown. By combining Variational Bayesian inference with stochastic differential equations (SDEs), it can "discover" the governing laws of a system from just one second of noisy sensor data and accurately predict its probability of failure over much longer timeframes.

What Problem Does This Paper Address?

In engineering, reliability analysis is the art of calculating the "probability of failure." Traditionally, this is done at the design stage using blueprints and known mathematical models. However, we face a massive gap when dealing with existing structures, especially aging infrastructure or heritage buildings.

Over decades, structures degrade, and their original design models no longer reflect reality. If the governing physics has changed or was never fully known, traditional methods like Monte Carlo simulations fail because they lack an accurate model to simulate. While "black-box" machine learning models can help, they often fail to generalize beyond their training data. The authors propose MAntRA (Model-Agnostic Reliability Analysis) to bridge this gap by learning the physics directly from measurements and using that learned model for safety predictions.

Methods

MAntRA operates in a two-stage pipeline:

  1. Stochastic Equation Discovery: The framework uses the Kramers-Moyal expansion to decouple a system's "drift" (deterministic trends) from its "diffusion" (random noise/stochasticity) based on measured output data.
  2. Variational Bayesian (VB) Inference: To identify the specific equations, the authors employ a sparse Bayesian learning approach. They use a "spike and slab" prior—a technique that forces the model to choose only the most relevant mathematical terms from a library of candidates (like polynomials or sines), effectively "discovering" the simplest equation that explains the data.

What’s new? Previous versions of this approach relied on Gibbs sampling (MCMC), which is computationally expensive. This paper introduces a Variational Bayes approach, which is significantly faster while still accounting for both epistemic uncertainty (lack of data) and aleatoric uncertainty (inherent randomness in the environment).

Datasets

The authors validated MAntRA using three increasingly complex synthetic datasets generated with the Euler-Maruyama scheme, all corrupted with 5% Gaussian noise to simulate real-world sensor inaccuracies:

  • SDOF Duffing Oscillator: A single-degree-of-freedom system with 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 equipped with a vibration control device.

Key Insights and Ideas

  • Extrapolation Power: Perhaps the most impressive "aha" moment is that the model was trained on only 1 second of noisy data but could accurately predict the probability of failure up to 30 seconds into the future. This demonstrates that the framework is truly learning the underlying physics, not just memorizing patterns.
  • Interpretable Physics: Unlike a deep neural network, MAntRA produces an actual Stochastic Differential Equation. Engineers can look at the discovered terms and understand why a structure is behaving a certain way.
  • Efficiency via VB: By switching to Variational Inference, the authors made the "discovery" phase computationally feasible for real-time or near-real-time monitoring applications.

Key Results

  • Parameter Accuracy: In the Duffing oscillator example, the predicted non-linear parameter ($\alpha/m$) was 100,369.79, remarkably close to the actual value of 100,000.00.
  • Reliability Matching: Across all three examples, the "Probability of Failure" ($P_f$) curves generated by MAntRA matched the ground truth (the "True SDE") almost perfectly.
  • Sparsity: The model successfully ignored hundreds of "junk" candidate functions in the library, correctly identifying only the 3-5 terms that actually governed the physical systems.

Limitations and Future Work

  • Library Dependency: MAntRA relies on a manually designed library of candidate functions. If the true physics of a system involves a mathematical term not present in the library, the model may struggle or resort to a "surrogate" approximation.
  • Dimensionality: While the paper demonstrates success with a 5-DOF system, scaling this to extremely high-dimensional systems (like a full bridge model with thousands of nodes) remains a challenge for future research.
  • Real-World Validation: The current study uses high-quality synthetic data. The next logical step is applying MAntRA to experimental data from actual vibrating structures in the field.

Summary updated

Abstract

We propose a novel model agnostic data-driven reliability analysis framework for time-dependent reliability analysis. The proposed approach -- referred to as MAntRA -- combines interpretable machine learning, Bayesian statistics, and identifying stochastic dynamic equation to evaluate reliability of stochastically-excited dynamical systems for which the governing physics is \textit{apriori} unknown. A two-stage approach is adopted: in the first stage, an efficient variational Bayesian equation discovery algorithm is developed to determine the governing physics of an underlying stochastic differential equation (SDE) from measured output data. The developed algorithm is efficient and accounts for epistemic uncertainty due to limited and noisy data, and aleatoric uncertainty because of environmental effect and external excitation. In the second stage, the discovered SDE is solved using a stochastic integration scheme and the probability failure is computed. The efficacy of the proposed approach is illustrated on three numerical examples. The results obtained indicate the possible application of the proposed approach for reliability analysis of in-situ and heritage structures from on-site measurements.
BibTeX
@inproceedings{mathpati2023mantra,
  author = {Yogesh Chandrakant Mathpati and Kalpesh Sanjay More and Tapas Tripura and Rajdip Nayek and Souvik Chakraborty},
  title = {{MAntRA: A framework for model agnostic reliability analysis}},
  booktitle = {Reliab. Eng. Syst. Saf},
  year = {2023},
  url = {https://openreview.net/forum?id=ikrTddCgy5V},
}
Figure