Joint Feature Classification for Wire Bond Joint Based on KPCA and Random Forest

We present a feasible algorithm to automatic identify and classify the quality of bonding joint in wire bonding via machine learning, named as KPCA (Kernel Principal Analysis) and Random Forest. The result can be acceptable in calculation time and accuracy, which is possible to use as feedback to control bonding parameters such as ultrasonic power and pressure, to strength the bonding reliability in production. First, the bonding joint images are mapped to a high dimension space, where KPCA is applied to decrease the image dimension for less calculation consumption and to eliminate high correlation features. The joint defect are then automatically identified and classified by Random Forest algorithm. Several strategies are adopted for improvement of accuracy. Our experiment result shows that the joint classification based on KPCA and Random Forest algorithm are better than conventional SVM and CNN algorithm on efficiency and accuracy.