Chen-Fu Chien, Yung-Chen Tai, and Yun-Siang Lin
As key component of electronic products, the importance of printed circuit boards (PCBs) has increased with the development of technology. With the rise of new technologies, the demand and importance for PCBs is also rising. With development of electronic products towards size shrinking, combined with the demand of transmission rate and computing power, PCB manufacturing becomes sophisticated, thus requiring more effective quality control methods. The quality of PCBs will directly affect the performance and reliability of back-end products, more precise manufacturing and more effective quality control are required. Solder resist is a surface coating of PCB and is one of the important processes. Above it, some parts need to be excluded from being covered, called solder resist opening (SRO). The size of SRO is a critical quality characteristic, as PCBs become increasingly shrinking, the size of the SRO is also becoming smaller. If the manufactured SRO size is out of specification, it may cause abnormalities during subsequent soldering, resulting in product scrap and yield loss. However, quality control of SRO size is complicated. SRO is affected by multiple processes and can only be measured after the last process, so the inspection must go through a pilot production, leading to material cost for inspection. Besides, the variance of incoming material, solder resist ink, is comparatively high, thus different SRO size compensation values are required for each incoming material batch. To control the quality and confirm the compensation value setting, manufacturers need to carry out stages of pilot runs and inspections, which consumes material costs and time.
In response to actual measurement problems, virtual metrology (VM) has been proposed to achieve in-line prediction of target quality characteristic, yet little research exists for PCB manufacturing. Little VM research has been conducted to propose VM framework for PCB manufacturing. In practical application, the uncertainty of VM model prediction and the cost of model update are faced, so the VM framework needs to be integrated with the process of batch compensation value determination based on domain knowledge. If VM is used instead of actual inspection, the pilot run material cost can be reduced and the capacity loss due to inspection can be decreased. This study develops a decision-based VM framework for SRO size, which combined feature selection, prediction model, and the decision process of compensation value determination. In addition, a confidence indicator is proposed to evaluate the reliability of predicted values, which also used as a reference for the model update mechanism to achieve a trade-off between model stability and update cost. An empirical study was conducted in a PCB manufacturing company. The result shows that the framework can help engineers to make decision on compensation values, further replace physical inspections to improve the efficiency of quality control and help PCB manufacturing industry to move towards intelligent production.