| Summary: | 碩士 === 國立成功大學 === 製造工程研究所碩博士班 === 94 === Incorrectness, fragment and asynchrony of collected data may lead to inaccurate virtual metrology results. To improve virtual metrology accuracy, data preprocess is extremely essential. Data preprocess deals with the processes of data examination and sifting. Considering advices from process engineers, important variables of the production process are selected and anomalous data are deleted before conducting sifting rules so as to improve the accuracy of virtual metrology.
This work investigates five sifting rules of independent variables in data preprocess. They are stepwise selection algorithm (SS), modified stepwise selection algorithm (Modified SS), sensitivity analysis of neural network (SA), SS rule on neural network, and expert-recommended rule. Analyses and experiments of two illustrative examples are executed to evaluate the performance of these sifting rules. This work further attempts to find out the key variables that affect the VMS accuracy most by classifying and sifting all the variables. The experimental results show that, due to missing a lot of important data, the sensitivity analysis of neural network is proved to be the worst in conjecture accuracy of all the sifting rules. In addition, the results also show that the stepwise selection algorithm, modified stepwise selection algorithm, and SS rule on neural network perform well in conjecture accuracy when the process-data sample size is small. However, if the sample size becomes large, the conjecture accuracy may perform poorly. On the contrary, when the rule of considering all of expert-recommended variables is applied, in the case of small sample size, its conjecture accuracy may be slightly worse than those of the SS rules, but still acceptable; while, in the case of large sample size, the conjecture accuracy of the expert-recommended rule is still acceptable but not those of the SS rules. In conclusion, the rule of considering all of expert-recommended variables is recommended.
|
|---|
