# Anomaly Detection for Solder Joints Using $β$-VAE

**Authors:** Furkan Ulger, Seniha Esen Yuksel, Atila Yilmaz  
**arXiv:** [2104.11927](https://arxiv.org/abs/2104.11927)  
**Date:** 2021-04-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.11927v2/

## Key points

- Traditional methods for detecting solder joint errors rely on hand-crafted features and specialized lighting
- Supervised Convolutional Neural Network (CNN) approaches require a large number of labeled abnormal samples
- The authors propose a new beta-Variational Autoencoders (beta-VAE) architecture for anomaly detection in solder joints
- The proposed model can work on both Integrated Circuit (IC) and non-IC components without the need for specialized lighting or error-free reference boards
- The model learns disentangled representations of data, leading to more independent features and improved latent space representations
- Different anomaly scoring techniques are compared and the effect of different beta parameters on accuracy and feature representations in beta-VAE is investigated
- Anomalies on solder joints can be detected with high accuracy using a model trained on normal samples without designated hardware or feature engineering
- This study addresses a unique inspection problem in unrestricted domains and makes significant contributions by designing a tailored beta-VAE architecture, inspecting both IC and non-IC component solder joints with a single model, comparing different anomaly scoring techniques, and investigating the effect of different beta parameters on accuracy and feature representations.

## Summary

This study addresses the inspection problem of solder joints in unrestricted environments during the assembly process of printed circuit boards (PCB). Traditional methods for detecting solder joint errors rely on hand-crafted features and specialized lighting, while supervised Convolutional Neural Network (CNN) approaches require a large number of labeled abnormal samples. To overcome these limitations, the authors propose a new beta-Variational Autoencoders (beta-VAE) architecture for anomaly detection that can work on both Integrated Circuit (IC) and non-IC components without the need for specialized lighting or error-free reference boards. The proposed model learns disentangled representations of data, leading to more independent features and improved latent space representations. The authors compare different anomaly scoring techniques and investigate the effect of different beta parameters on accuracy and feature representations in beta-VAE. They demonstrate that anomalies on solder joints can be detected with high accuracy using a model trained on normal samples without designated hardware or feature engineering. This study makes significant contributions by addressing a unique inspection problem in unrestricted domains, designing a tailored beta-VAE architecture, inspecting both IC and non-IC component solder joints with a single model, comparing different anomaly scoring techniques and investigating the effect of different beta parameters on accuracy and feature representations in beta-VAE. The related work in solder joint inspection is discussed followed by a description of the proposed method, dataset, parameter settings, experiments and results.