Session: 13-19-02: Scientific Machine Learning (SciML) for Characterization, Modeling, and Design of Structures and Materials II

Paper Number: 172875

Acoustic Signal-Based Defect Detection Through Lstm Networks for Non-Destructive Evaluation 

Accurate characterization of internal defects in solid materials and structures is essential for ensuring the integrity, functionality and reliability of engineering systems. Traditional non-destructive evaluation (NDE) techniques such as ultrasonic inspection, thermography, or X-ray imaging can effectively detect defects at the subsurface of the structure. However, these conventional methods often require expensive equipments and instrumentations, experienced operators, intense data interpretations, and long evaluation time. In contrast, data-driven machine learning and modelling enable a promising route for in situ automated service life monitoring and defect prediction.  With a well-trained machine learning model, it is possible to pursue on-site structural health monitoring and in situ decision-making on maintenance scheduling in an effective and efficient manner.

This study generates a set of time-series acoustic signals from high fidelity finite element analysis to train a recurrent neural network (RNN) model based on a Long Short-Term Memory (LSTM) architecture for achieving defect detection in a two-dimensional (2D) solid domain.  COMSOL Multiphysics is adopted to generate the pressure wave propagations in a 2D solid main (30 mm × 50 mm) with defects of varying size, shape and location.  A Gaussian pressure pulse is introduced at one end of the domain, and the resulting wave reflections are recorded at five locations. The multi-channel pressure-time signals produced by 1500 simulations and the corresponding defect information (size, shape and location) form the LSTM model's input. The LSTM model was trained over 200 epochs using a weighted mean squared error loss to emphasize accurate prediction of defect radius. It is optimized using the Adam algorithm with learning rate scheduling. Adopting the circular defect as example, this regression-only architecture focuses on precise inverse prediction of defect geometry from sparse sensor probe data. The trained model achieved average prediction errors of 6.3% for radius, 2.4% for x-location, and 4.3% for y-location on a completely independent validation set, demonstrating strong generalization and accuracy. These findings underscore the model’s strong ability to estimate geometric parameters, making it highly effective for inverse characterization of internal defects. Our results demonstrate that pressure-time signals alone can serve as a rich source of information for inferring the location and size of internal defects without requiring full-field imaging.

This work provides a foundation for future non-destructive testing systems leveraging time-series modeling. It also highlights the broader potential of LSTM-based inverse frameworks across other areas of physical sensing. Future directions include expanding into more complex shapes, integrating physics-informed loss functions, and validating experimental datasets.

 

Presenting Author: Huijuan Zhao Clemson University

Presenting Author Biography: Dr. Huijuan Zhao's research area is computational materials, computational mechanics and multi-scale/multi-physics modeling. Her current research focus is to understand new physics/mechanisms behind high-performance materials and structures at the nano-scale and their potential engineering applications through various computational modeling techniques.

Authors:

Tanuj Gupta Clemson University
Md Arif Sadik Chomok Clemson University
Hai Xiao Clemson University
Huijuan Zhao Clemson University