Failure Prediction Using Sequential Pattern Mining in the Wire Bonding Process

Abstract

In the semiconductor packaging system composed of complex assembly processes, it is necessary to detect faults early to minimize the time and cost required to produce semiconductor chip scraps and to improve production yield. Early detection of faults is possible by performing a failure prediction analysis during the semiconductor packaging process. In this paper, we consider a statistical process control model for predicting the final failure of a printed circuit board lot based on observed event sequences in the wire bonding process step. To estimate the parameters efficiently in the predictive model, we propose a two-stage process. Here, irrelevant subsequences of events are deleted by a sequential mining approach at the first step, and a predictive model is constructed with the remaining subsequences of events using the logistic regression and bagged least absolute shrinkage and selection operator (LASSO) estimate, which we call the B-LASSO method. In particular, to resolve problems caused by unbalanced data, the B-LASSO uses case-control sampling rather than simple random sampling with replacement. The performance of the B-LASSO is compared with other competing methods by analyzing a work-site dataset to confirm that the B-LASSO is superior.