CNN-based acoustic spectrogram analysis for quality assessment of laser-welded screws

Laser welding offers high precision and efficiency but remains challenging to inspect for internal defects using conventional Non-Destructive Testing (NDT) methods. This study presents a contact-free, post-process inspection framework combining laser-induced acoustic excitation with Deep Learning (DL)-based spectrogram analysis for automated weld quality assessment. Acoustic responses from welded screw elements were recorded using a membranefree optical microphone and converted into time-frequency spectrograms. A Convolutional Neural Network (CNN) was trained to classify weld integrity based on these spectrograms, outperforming classical Machine Learning (ML) baselines that relied on manually engineered features. The optimized CNN achieved 93.5% accuracy at the segment level and 92.9% at the specimen level, demonstrating reliable discrimination between defect-free and defective welds. Hyperparameter optimization revealed that spectrogram resolution and network depth were the primary performance drivers. The proposed method establishes a foundation for scalable, AI-driven, and non-contact Quality Assurance (QA) in laser-based manufacturing. Future work will extend the system toward multi-class defect characterization and real-time deployment.