A data mining approach to the causal analysis of product faults in multi-stage PCB manufacturing

It is difficult for manufacturers of printed circuit boards (PCBs) to remain competitive because of the ever-increasing complexity of circuit board designs and processes that increase the product cost while decreasing the product yield. It is particularly difficult to ensure high yields because the products are made through sequential nano-scale processes, and the quality of each process may be affected by the results of the upstream processes. This type of cumulative effect makes it difficult to determine the machines that introduce product faults. In this paper, we develop a data mining approach to which large amounts of trace data are inputted to infer fault-introducing machines in the form of a L ⇒ R rule, where R contains the fault type and L contains a machine sequence that is the primary cause of the fault type. We tested our approach with industrial lot trace data collected from a PCB manufacturing line with 33 machines and six fault types. From the work-site experiment, we found 26 composite rules that showed a significant cumulative effect. The average fault detection accuracy of the rules was 87.2%. In addition, we found 13 rules that affected more than one fault type.