Estimation of system assembly and test manufacturing yields through product complexity normalization

Abstract

Cisco Systems, Inc. (Cisco) has recently adopted Six Sigma as the main platform to drive quality improvements in its manufacturing operations. A key component of the improvement strategy is the ability to define appropriate manufacturing yield goals. Cisco's manufacturing operations can be divided, at a very high level, in two major steps: Printed Circuit Board Assembly (PCBA) and System Assembly and Test. The company has already deployed a global yield goal definition methodology for the PCBA operation, but the creation of a similar methodology for the System Assembly and Test operation proved difficult: Cisco lacked a universal methodology to determine the expected variation on manufacturing performance resulting from differences on product design and manufacturing processes attributes. This thesis addresses this gap by demonstrating a methodology to relate relevant design and process attributes to the System Assembly and Test manufacturing yield performance of all products. The methodology uses statistical analysis, in particular Artificial Neural Networks, to generate a yield prediction model that achieves excellent prediction accuracy (4.8% RMS error). Although this study was performed using Cisco Systems' product and manufacturing data, the general process outlined in this exercise should be applicable to solve similar problems in other companies and industries. The core components of the methodology outlined can be easily reproduced: 1) identify the key complexity attributes, 2) design and execute a data collection plan and 3) generate statistical models to test the validity and impact of the selected factors.