This document outlines a methodology for developing an artificial intelligence algorithm to detect head-in-pillow defects in solder ball inspection. It discusses acquiring 3D images of printed circuit boards using x-ray projection to better identify the location of these defects. The goals are to solve data imbalance issues from the rare defective samples and compare performance to other CNN models and commercial inspection software. Key challenges addressed include overfitting from limited data and properly preprocessing x-ray images to reduce noise before 3D reconstruction.

1. Advisor : Prof. Chiou-Shann Fuh 傅楸善教授
TsanKit:
Artificial Intelligence for Solder Ball
Head-In-Pillow Defect Inspection

2. 01 / Introduction
02 / Related Works
03 / Background
04 / Methodology
05 / Experimental Results
06 / Conclusion and Future Works
07 / References
OUTLINE

3. 1
Introduction

4. Introduction
⌬ On the production line, inspection [30] is essential to controlling the
quality of products.
⚉ It can help to fix the sources of detected defects immediately and
reduce defect rates.
⌬ Automated Inspection
⚉ Automated X-ray Inspection (AXI)
⚉ Automated Optical Inspection (AOI)
⚉ Solder Paste Inspection (SPI)
4

5. Introduction
⌬ Printed Circuit Board (PCB)
⚉ PCB plays an important role in many electronic products.
⚉ Solder balls provide the contact between the BGA and the PCB.
⚉ For multi-layer PCB, many overlapped electronic components make
the defect inspection more difficult and challenging.
5BGA: Ball Grid Array [6]
BGAPCB
Solder
Balls

6. Introduction
⌬ Due to the complexity of most images of electronic components,
traditional machine vision methods cannot solve the problem
completely.
⌬ Deep Learning has been widely used in many computer vision tasks.
⌬ Convolutional Neural Network (CNN) [29]
⚉ CNN has an outstanding performance in image recognition and
classification.
⚉ Different levels of features can be integrated by the deep network
structure.
⚉ The complicated high-level features can be combined with an
end-to-end network to predict the result.
6

7. Introduction
⌬ Problem 1: Overfitting
⚉ The trained models will easily over-fit due to insufficient data.
⚉ Even if the results of the training data are good, the results of testing
data are usually not as good.
Attempt
⚉ We try to combine machine learning algorithms to increase the
testing accuracy.
7

8. Introduction
⌬ Problem 2: Data Imbalance
⚉ It is hard to obtain defective data because defective products should
not appear on the production line.
⚉ The numbers between normal data and defective data vary greatly.
⚉ It may make the model blindly learn the characteristics of normal
data while ignoring defective data.
8

9. Introduction
⌬ Goal
⚉ Develop a solder ball Head-In-Pillow defect inspection algorithm by
AI techniques.
⚉ Aim to solve the data imbalance problem caused by rare defective
data.
⚉ Compare the performance and execution time with:
⚆ Several classic CNN models
⚆ Deep learning inspection software SuaKIT
9AI: Artificial Intelligence

10. Introduction: Head-in-Pillow Defect
⌬ Head-In-Pillow (HIP) is a latent solder ball defect occurring in the
soldering process [10, 12, 18, 31].
10[31] D. Xie, et al. “Head in Pillow (HIP) and Yield Study on SIP and PoP Assembly,” ECTC, 2009.

11. Introduction: Head-in-Pillow Defect
⌬ Difficult Points [31]
⚉ HIP defects often escape inspection and tests on the factory floor as
there may still be mechanical and electrical contact.
⚉ HIP defect will cause the unstable conductivity of the particular BGA
balls and lead to intermittent failures.
⚉ It is difficult to achieve zero miss detection rate.
⌬ The results of the inspection usually need to be further checked by
experts or FAE.
11FAE: Field Application Engineers

12. Introduction: Head-in-Pillow Defect
⌬ It is hard to find the location of HIP defects from 2D X-Rays images
due to the variable shape of the defect.
⌬ The 3D solder ball model can represent the location of HIP defects
more clearly and provide more information.
12
X-ray images of PCB [5, 28]

13. Introduction: Acquisition of PCB Images
⌬ The 3D PCB images are reconstructed with 2D X-ray images.
13
Theta 𝜃𝜃 : the angle between the tube to the detector and vertical line (typically 31 degrees)
Phi 𝜑𝜑 : the angle of rotation ( = 360º / #projected images ) (e.g. 2.81 degrees = 360 degrees / 128)
X-ray Projection System
SOD : Source to Object Distance SID : Source to Image Distance

14. 14
PCB1
Normal Defect

15. 15
PCB2
Normal Defect

16. 16
PCB3
Normal Defect

17. Introduction: Acquisition of PCB Images
⌬ Under the influence of X-rays, there is a large amount of white
noise in the projected images.
⌬ We need to optimize the images before reconstruction; otherwise,
we will get inferior quality images.
⌬ White noise is removed by averaging multiple (9) images over time.
17