Research on Application of RST and Artificial Neural Network in Multi-class Package Product Classification

• Main Authors: HSIN-CHIAO LIU, 柳馨喬
• Other Authors: YUNG-HSIANG HUNG
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/52759303862131687215

碩士 === 國立勤益科技大學 === 工業工程與管理系 === 97 === The semiconductor industry is an important industry in Taiwan. The IC design, fabrication and package are all focused on light-and-thin style. Since customer orders may be various, and there are numerous product types and applications, the purchase-order to production-order process requires more labors to respond to the demands. However, the current operational process is human communication. Therefore, it is very important to provide package information to designer efficiently. Only by doing so could the subsequent design method and fabrication process of IC chip be properly arranged, and minimize processing problems. This study treated five types of common IC package products (TFBGA, LGA, PBGA, CFBGA, and QFP) as the subjects. IC package dataset contained 63 end products, and 2496 pieces of data records, including 14 product attribute variables and 1 decision-making variable. Taking TFBGA cluster for example, it included 16 end products, 632 data records in total; QFP cluster includes 8 end products, and 272 data records. Since the current IC package product application is extensive, and product types are enormous, this study adopted two-stage classification method to probe into the IC package product database for multi-class imbalanced data structure, and propose a classification method suitable for multi-class IC package pattern. Rough set theory was also applied in clustering IC package products. The experiment showed the product cluster corresponding to result of first-stage classification. BPNN was employed to study the end product classification accuracy. The experiment results showed that, when using one-stage Artificial Neural Network(ANN) on five kinds of IC package products, and two-stage (rough set and ANN) study on dataset, multi-class IC package classification accuracy on 63 end products based on two-stage method was higher than that by BPNN. For experiments using BPNN only, the overall training accuracy was 77.24%, test accuracy was 75.80%; but two-stage experiments on end products obtained better results than one-stage method. Taking QFP package cluster for example, its training accuracy was 85.97%, test accuracy was 79.05%. This study could efficiently integrate the important geometry, property, design criteria and application criteria of IC package, in order to improve the operation efficiency of IC designers, save labor cost, and increase profits.