A Defect Prediction Case Study for Printed Circuit Board Assemblies Containing Ball Grid Array Package Types
Authors: Phillip M. LaCasse, LTC, USA, Ph.D., Wilkistar Otieno, Ph.D., Gregory J. Vance, Francisco P. Maturana, Ph.D., Mikica Cvijetinovic Company: Air Force Institute of Technology, University of Wisconsin – Milwaukee, and Rockwell Automation Date Published: 9/22/2019
Abstract: This technical paper describes an ongoing project employing machine learning models to correlate defects identified at the in-circuit testing (ICT) station with parametric data on solder paste deposits measured at the upstream solder paste inspection (SPI) machine. The scope of the project is restricted to printed circuit boards (PCB) containing ball grid array (BGA) package types. BGA package types are of interest because, after the PCB has passed through the reflow oven, there is a dramatic increase in the cost associated with identification and rework of defective components. The data used to train the machine learning model is organized as arrays associated with the solder paste deposits in a single PCB location. An automated feature extraction tool converts the data arrays into a set of possible predictor variables that is then used to train the supervised machine learning models. The feature extraction step is exploratory, with a total of 3,970 features extracted from the SPI parametric data. Subsequently, a feature reduction and prioritization tool removes features that exhibit poor predictive ability. The final, reduced set of features is then tested iteratively through all possible combinations to identify the optimal subset of features to train the final model.
Decision tree models were executed and scored using four metrics: accuracy, precision, recall, and f1 score. Scores for accuracy exceeded 96.16%, precision exceeded 82.35%, recall exceeded 50%, and f1 score exceeded 62.22%. The results of the pilot study are encouraging but additional work is necessary to determine if results are due to generalizable relationships or specific circumstances in the piloted datasets.