Machine learning inverse surrogates for damage identification in plates based on Lamb waves

Area of Science:

• Structural Health Monitoring
• Machine Learning Applications
• Non-Destructive Testing

Background:

• Accurate identification of structural damage is crucial for safety and maintenance.
• Existing methods for damage identification often rely on physics-based models or complex feature extraction.
• Data-driven approaches offer a promising alternative for efficient and accurate damage assessment.

Purpose of the Study:

• To develop a black-box, purely data-driven strategy for identifying localized damage in plate-like structures.
• To investigate the physical interpretability and accuracy of damage parameter estimation using Convolutional Neural Networks (CNNs).
• To evaluate the performance of CNN-based inverse surrogate models under different material property conditions.

Main Methods:

• A supervised learning regression task utilizing Convolution Neural Networks (CNNs).
• Employing Lamb waves, three actuators, and 16 sensors for data acquisition in an elastic plate model.
• Training two inverse surrogate models: one with homogeneous and another with non-homogeneous material properties.
• Evaluating damage recovery using overlap metrics to assess accuracy and limitations.

Main Results:

• CNNs effectively estimated damage positional parameters with high accuracy.
• Damage size and intensity estimation proved more challenging than localization.
• The inverse surrogate model trained with non-homogeneous material properties demonstrated robustness against system variability.
• Overlap metrics provided insights into the accuracy and limitations of the damage recovery.

Conclusions:

• The proposed data-driven CNN strategy is effective for structural damage identification, particularly for localization.
• Positional parameter estimation accuracy aligns with literature classifying damage localization as a classification problem.
• Robustness against system variability is achievable with surrogate models trained on non-homogeneous material properties.