Detect PCB Stack-up Error with Machine Learning Methods



Detect PCB Stack-up Error with Machine Learning Methods
With high frequency and high-speed transmission in information industry, PCB customers demand stricter signal transmission integrity of PCB.
Analysis Lab

DOWNLOAD

Authored By:


Ta Chang Chen, Gold Circuit Electronics Ltd., Taoyuan, Taiwan
Fei Fei Kao PhD, Ming Chuan University, Taoyuan, Taiwan
Huang Yu Chen PhD, Taoyuan, Taiwan

Summary


In manufacturing multilayer printed circuit boards (PCBs), the PCB layer stack-up usually includes multiple plies of various types of prepreg along with the inner layer core material. Most of the layer stack-up process is manual operation relying on an operator to place the correct type and quantity of prepreg at designated dielectric openings. Missing or adding additional one or more plies of prepreg can result in a overall board thickness which will possibly still fall within its thickness specification limits but, its electrical property will be greatly affected (e.g. impedance, loss etc.). In this case, this abnormal board could possibly escape to customer unless detected by more expensive electrical property testing. Impedance measurement may catch the missing / extra prepreg for the dielectric layers related to impedance. Other layers without impedance control could still possibly escape. Considering impedance TDR measurement is time consuming and costly measurement process, a new lower cost approach was developed to detect extra / missing prepreg and validate the correctness of the stack-up.

Our first approach is looking for statistical outliers for each given lot board thickness measurements after lamination. This approach is able to detect extra / missing prepreg, but its false alarm is too high due to missing one ply prepreg can be less than 0.002" thick which is generally less than one sigma of total board thickness. Thus, second approach using machine learning technique was developed with board thickness data (known bad and known good board thickness) and other features (e.g. prepreg type, lamination parameters etc.). From the model we developed and validated that our escape rate is still 0% and its false alarm rate was reduced by 85%.

Conclusions


Although it’s not a great benefit in this case, we just used the data in our database originally and coded with a free open source software Python to reduce waste in manufacturing environment. In the past few years, we’ve heard a lot of topics about Industry 4.0 and smart manufacturing, and then we started to collect data online. Until now, I believe most of those data are stored in database and have not been used well, this case showed us the benefits of correctly using these data. There are lot of training programs and training materials on the internet and in real life, just get your foot in the door and follow the guidelines as follows [2], machine learning is not unreachable.

Initially Published in the SMTA Proceedings

Comments

No comments have been submitted to date.

Submit A Comment


Comments are reviewed prior to posting. You must include your full name to have your comments posted. We will not post your email address.

Your Name


Your Company
Your E-mail


Your Country
Your Comments