Internal defect detection model based on laser ultrasonic signal decomposition and deep learning

It is essential to accurately detect the internal defects of metal equipment. The accuracy of traditional detection methods is affected by noise. This paper proposes an innovative genetic algorithm-weighted permutation entropy (GA-WPE) optimized Variational mode Decomposition (VMD), continuous wavelet transform (CWT) and deep learning fusion model, which aims to achieve efficient automatic defect detection. Firstly, the laser ultrasonic experimental system was built, and the hole defects with different diameters and depths were prepared. The GA-WPE-VMD was used to decompose and reconstruct the ultrasound signal, and CWT converted the reconstructed signal into a time–frequency scalogram. Three deep learning models were used for defect detection. ViT model achieved 99.88%, 97.01%, and 91.7% accuracy on the training, validation, and test set, respectively, significantly better than DenseNet121 and ResNet50. The AUC of ViT on the test set was close to 100%. The results verify the feasibility and effectiveness of the proposed method. © 2024