A Multi-Fidelity Convolutional Autoencoder-Transfer Learning Framework for Guided-Wave-Based Damage Diagnosis Using Large Simulated and Limited Experimental Datasets
Santosh Kapuria, Abhishek
Read on arXiv →Key claim
Achieves R^2 scores over 0.93 for damage localization.
This study presents a new approach to structural health monitoring that effectively combines simulation data with limited experimental data to accurately detect and size damage in structures. The key result is that the proposed framework achieves R^2 scores over 0.93 for damage localization, making it a practical solution for real-world applications.
In plain English
Imagine you're responsible for maintaining large engineering structures like bridges or buildings. You want to catch any damage early to prevent catastrophic failures, but inspecting these structures can be time-consuming and expensive. Traditionally, engineers rely on extensive labeled data from experiments to train models that can identify damage, but gathering this data is often impractical and costly. This is where things can go wrong: without enough data, models can be inaccurate, leading to missed damage or false alarms. This issue is known as data scarcity in machine learning for structural health monitoring.
To address this, the authors propose a new method that cleverly combines lightweight physics-based simulations with deep learning techniques. Instead of needing vast amounts of experimental data, their approach uses a smaller set of real measurements alongside a large synthetic dataset generated from simulations. This allows the model to learn effectively even with limited real-world data. The framework employs a convolutional autoencoder to extract features from the data, which are then used in a feed-forward neural network for damage detection.
What’s exciting is that this new method significantly outperforms traditional convolutional neural networks in terms of accuracy for damage localization. The results show that the model can achieve R^2 scores exceeding 0.93 for localization and 0.99 for sizing, indicating it can make very accurate predictions even on new, unseen data. This means that engineers can deploy this framework in real-world scenarios with confidence, knowing it can accurately identify and size damage without needing extensive data collection efforts.
The paper introduces a multifidelity transfer learning framework that combines physics-based simulations with deep learning for damage detection, extending existing methods into a new application area.
The claims are supported by strong experimental results, including high predictive performance metrics and comparisons to existing methods.