# Detecting Manufacturing Defects in PCBs via Data-Centric Machine Learning on Solder Paste Inspection Features

**Authors:** Jubilee Prasad-Rao, Roohollah Heidary, Jesse Williams  
**arXiv:** [2309.03113](https://arxiv.org/abs/2309.03113)  
**Date:** 2023-09-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.03113v1/

## Key points

- The paper explores automated detection of defects in PCB manufacturing using Solder Paste Inspection (SPI) and Automated Optical Inspection (AOI) machines.
- Researchers leverage SPI-extracted features from 6 million pins to train Machine Learning (ML) models for improved defect detection.
- Dataset used includes 6 million PCB pins representing 2 million components across 15,387 PCBs.
- Base extreme gradient boosting (XGBoost) ML model is employed with refined data pre-processing techniques.
- Training instances are created at pin, component, and PCB levels to capture potential inter-pin, inter-component, or spatial effects.
- ML models trained at various levels are combined to effectively identify defective components.
- Data-centric approach aims to improve accuracy in identifying manufacturing defects in PCBs and enhance operational efficiency while reducing manual intervention.

## Summary

The paper titled "Detecting Manufacturing Defects in PCBs via Data-Centric Machine Learning on Solder Paste Inspection Features" by authors Jubilee Prasad-Rao, Roohollah Heidary, and Jesse Williams explores the automated detection of defects in Printed Circuit Board (PCB) manufacturing. The study utilizes Solder Paste Inspection (SPI) and Automated Optical Inspection (AOI) machines to detect defects at different stages of production. By leveraging SPI-extracted features from 6 million pins, the researchers demonstrate a data-centric approach to training Machine Learning (ML) models for improved defect detection. The dataset used in this research comprises 6 million PCB pins representing 2 million components across 15,387 PCBs. A base extreme gradient boosting (XGBoost) ML model is employed and data pre-processing techniques are refined to enhance performance. Training instances are created at both the component and PCB levels by combining pin-level SPI features with component and PCB IDs. This allows the ML model to capture potential inter-pin, inter-component, or spatial effects that may not be readily apparent at the pin level. The study trains ML models at various levels - pin, component, and PCB - and combines their results to effectively identify defective components. By integrating information from multiple levels within the dataset, this data-centric approach aims to improve accuracy in identifying manufacturing defects in PCBs. The findings suggest that this method can significantly enhance operational efficiency and reduce manual intervention in PCB manufacturing processes.