Okay, here is the abstract and summary for the provided transcript on Physics-Informed Guided Wave Ultrasonic Machine Learning.
Abstract:
This lecture by Professor Joel B. Harley discusses the challenges and strategies for applying machine learning (ML) to ultrasonic guided wave testing for nondestructive evaluation (NDE) and structural health monitoring (SHM). While ML excels in data-rich domains, its use in industrial applications like infrastructure monitoring is hindered by (1) data scarcity, (2) high data diversity across materials and conditions, and (3) the need for engineering assurances often lacking in purely data-driven models. The talk presents three paradigms to bridge the gap between imperfect physics-based models and complex experimental data: Model Fitting (Sparse Wave Number Analysis), Dictionary Learning (Transfer Learning), and Physics-Informed Learning. These approaches leverage techniques like sparse regression, transfer learning from simulations or surrogate data, and incorporating physical constraints (like the wave equation) directly into the ML framework. Demonstrations show how these methods can interpret complex wave propagation, interpolate sparse data, create synthetic baselines, isolate signals from damage, and locate defects, even in the presence of dispersion, anisotropy, and inhomogeneities, ultimately aiming to create more robust and interpretable models for engineering applications.
Physics-Informed Machine Learning for Guided Wave Ultrasonics
0:12 Introduction: Welcome and introduction to the IEEE UFFC Virtual Education Series.
1:57 Speaker Introduction: Introduction of Professor Joel B. Harley from the University of Florida.
4:23 Talk Start & Motivation: Prof. Harley introduces the topic, acknowledging his students and lab.
5:40 Core Problem: Imperfect Models: Discussion on the persistent mismatch between physics-based models (like the wave equation) and real-world experimental ultrasonic data, hindering direct model use for data analysis.
6:50 Guided Wave SHM Basics: Explanation of guided wave structural health monitoring using sensors, the advantages (sensitivity, low attenuation), and the inherent complexities (reflections, multimodal and dispersive propagation).
8:00 Signal Variability Challenge: Environmental effects (like temperature) cause significant changes in received signals, making comparisons difficult.
8:56 Simulation Limitations: Even validated simulations often show poor correlation (e.g., 18%) with experimental data when compared numerically, insufficient for tasks like signal subtraction.
10:51 Proposed Strategy: Combining Models: Introduction of the idea to improve models by combining three types: physics-based, experimental realizations (baselines), and learned (ML) models.
12:41 Paradigm 1: Model Fitting (Sparse Wave Number Analysis - SWA): Combines physics models with experimental data. Uses sparse regression to estimate dispersion curves (wave number vs. frequency) from sensor measurements.
13:00 Guided Wave Physics: Details on how waves propagate in plates (Lam waves), leading to multiple modes and dispersion (frequency-dependent velocity).
16:59 SWA Formulation: Representing wave propagation as a linear equation (X = ΦV) where V (containing dispersion curve info) is sparse.
19:11 SWA Process & Example (Interpolation): Using SWA to reconstruct a full wavefield from very sparse measurements (50 points -> 40,000 points) with high accuracy by leveraging the physics model.
21:50 SWA for Localization: Demonstrating SWA's ability to locate damage more effectively than traditional methods (like Delay and Sum) by using the full wave complexity.
23:11 SWA Adaptation (Anisotropy): Showing SWA can be adapted for complex propagation like anisotropy found in composites.
26:06 SWA Strengths/Limitations: Good at interpolation/extrapolation, adaptable model; but struggles with unmodeled geometry, inhomogeneity, and reflections.
27:55 Transfer Learning Process: Learn a "dictionary" (basis functions, Φ matrix) from surrogate data (simulations or other experiments), then use it with sparse regression on new test data.
28:53 DL Example (Synthetic Baseline): Using a simple simulation to learn geometry, then applying it to experimental data to predict a damage-free baseline, enabling damage isolation via subtraction.
32:02 DL Strengths/Limitations: Can handle complex geometry if simulated, less prior knowledge needed; but not physically interpretable, hard to extrapolate, doesn't inherently use physics.
32:47 Paradigm 3: Physics-Informed Learning: Combines all three model types (Physics, Learned, Experimental).
33:11 Physics-Informed Concept: Embeds physical laws (like the wave equation) as constraints within the ML optimization problem (similar idea to Physics-Informed Neural Networks but using linear models).
36:13 Physics-Informed Process Example: Solves an optimization problem to find wave modes (dictionary D) and coefficients (V) that fit the data (X ≈ DV) while satisfying the wave equation. Can handle inhomogeneities by identifying regions with different wave properties.
41:31 Physics-Informed Strengths/Limitations: Potential to handle complex data/inhomogeneity with little prior knowledge, incorporates physics for better generalization; but optimization is more complex, and handling boundary conditions is challenging.
42:29 Conclusion: Models are imperfect, but combining physics-based models, learned models, and experimental data through tailored ML approaches can create better, more reliable tools for engineering data analysis in NDE/SHM.
43:32 Q&A Session: Addressing audience questions on dispersion curve estimation (Lasso regression), training complexity (no neural networks needed, fast for SWA/DL, manageable for Physics-Informed), comparison to SINDy, computational cost (low), regularization, spatial resolution, and scalability to complex, inhomogeneous structures.