Abstract
The multicrystalline silicon melting process of solar cell can be divided into 3 parts, denote as S1, S2, S3. The unmelted silicon height inside the crucible of each process has to be h1, h2, h3. During the melting process, operator would determine the time to dip a quartz rod to measure the unmelted silicon height depend on his or her own experiences and record time and height. During each process, operator has to dip more than 3 times, that’s totally 9 times more to get the target height, quartz rod can only use no more than 20 times so the cost of quartz rods are high. The motivation of our research is to decrease frequency of dipping to reduce cost of quartz rods. In others words, our target is to get h1, h2, h3 by fewer dipping. We firstly find the linear regression relationship between height and time from known data. The coefficient of determination of regression model is high. But in fact, if the regression model is used to estimate the target height, the estimated height will be far-off target from real height h1, h2, h3. So the original dipping data of time and height have to be moderately transformed. In our research, we made a clever transformation and proposed a prediction model to find out the hidden part of the transformed data. The result shows dipping frequency reduce to 4 to 5 times and predicted height error compared to traditional linear regression have improved 60% on the final process S3. Our research provides on-line operators a logical dipping method and can effectively decrease dipping frequency,reduce cost of quartz rod and lower operator’s workloads.