Autonomous Assessment of Delamination Using Scarce Raw Structural Vibration and Transfer Learning

Area of Science:

• Materials Science
• Mechanical Engineering
• Artificial Intelligence

Background:

• Deep learning models require extensive data, which is often unavailable for faulty states in composite materials.
• Developing robust diagnostic strategies for laminated composites is challenging due to data scarcity.

Purpose of the Study:

• To introduce a transfer learning framework for autonomous detection, isolation, and quantification of delamination in laminated composites.
• To utilize scarce low-frequency structural vibration data for damage assessment.
• To compare deep learning-extracted autonomous features with handcrafted statistical features.

Main Methods:

• Structural vibration data from simulations and experiments were encoded into time-frequency images using SynchroExtracting Transforms (SETs).
• Pretrained deep learning models (AlexNet, GoogleNet, SqueezeNet, ResNet-18, VGG-16) were used to extract autonomous features.
• A support vector machine (SVM) algorithm was employed to assess the extracted features for delamination analysis.

Main Results:

• The transfer learning framework successfully extracted low- and high-level autonomous features from limited vibration data.
• Autonomous features demonstrated effectiveness in detecting, isolating, and quantifying delamination in laminated composites.
• Performance using autonomous features was comparable or superior to handcrafted statistical features.

Conclusions:

• The proposed transfer learning framework enables autonomous damage assessment of laminated composites even with scarce raw structural vibration data.
• This approach provides a promising direction for advancing structural health monitoring in composite materials.
• The method overcomes data limitations inherent in deep learning applications for damage diagnostics.