Reproduce “Anomaly Detection for Solder Joints Using $β$-VAE” on OpenResearch

About this paper

Furkan Ulger, Seniha Esen Yuksel, Atila Yilmaz

Researchers at Hacettepe University developed a β-Variational Autoencoder (β-VAE) for unsupervised anomaly detection in solder joints, achieving an F1-score of 0.809 ± 0.04 by training solely on normal samples and leveraging a gradient-based anomaly scoring technique (GradCon). The method eliminates the need for specialized hardware, feature engineering, and labeled defect data, and inspects both IC and non-IC components using a single architecture.

Key insights

• Framing solder joint inspection as an unsupervised anomaly detection task using β-VAE effectively addresses the data scarcity problem by only requiring normal samples for training.
• The gradient-based anomaly scoring technique (GradCon) consistently improves detection accuracy across various autoencoder models, outperforming methods based solely on reconstruction error or ELBO.
• Optimizing the β hyperparameter in β-VAE is crucial for inducing beneficial disentangled representations, with an optimal value of β=3 yielding the highest detection performance.

Summary by alphaXiv.