A framework for inspection of dies attachment on PCB utilizing machine learning techniques

Abstract

Decision Support Systems are considered as a robust technology able to provide an advantage to several manufacturing companies. As part of the Z-Fact0r EU project Early Stage-Decision Support System, a framework for the inspection of a printed circuit boards (PCB) and the inference of faults, regarding the excess or insufficient glue, is proposed. For the inspection of the PCB, a pixel-based vector of the regions of interest is utilized and several very popular in research community machine learning algorithms are tested on their performance on fault recognition. In order to determine the most efficient and effective classifier, a schema of Monte Carlo simulations for each classification algorithm and set of hyper-parameters was performed. Simulation results show a superiority of the support vector machine (SVM) classifier with polynomial and radial basis function kernels, compared to the rest. The best overall classifier was the SVM polynomial (accuracy: 81.39%, f-measure: 78.72%).