Anomaly Detection for Solder Joints Using <i>?</i>-VAE

In the assembly process of printed circuit boards (PCB), most errors are caused by solder joints in Surface Mount Devices (SMD). literature, traditional feature extraction based methods require designing hand-crafted features and rely on tiered RGB illumination to detect joint errors, whereas supervised Convolutional Neural Network (CNN) approaches a lot labelled abnormal samples (defective joints) achieve high accuracy. To solve optical inspection problem unrestricted environments with no special lighting without existence error-free reference boards, we propose new beta-Variational Autoencoders (beta-VAE) architecture for anomaly detection that can work both IC non-IC components. We show proposed model learns disentangled representation data, leading more independent improved latent space representations. compare activation gradient-based representations used characterize anomalies; observe effect different beta parameters accuracy untwining beta-VAE. Finally, anomalies be detected via trained directly normal designated hardware or engineering.