Statistical data pre-processing for fuzzy modeling of semiconductor manufacturing process

A systematic algorithm is proposed to design a fuzzy inference system through statistical data pre-processing. This approach is appropriate in modeling the qualitative aspects of a semiconductor manufacturing process, when extensive training data are often limited or difficult to collect due to the high cost of conducting experiments. With the limited number of data sets from a designed experiment, our system employs a proper statistical analysis to extract simple fuzzy inference rules of input-output relationships and initialize the corresponding membership functions. The output process variable can be continuous or categorical, and the fuzzy system can be further tuned to accommodate newly acquired experimental data.<>