在单晶硅加工中,硅片多线切割质量检测耗时和检测成本高造成硅片质量检测难.因此,提出一种基于生成对抗网络(WGAN-GP)数据处理与自注意力残差网络(SeResNet)的硅片质量预测方法.分析多线切割的机制,确定影响硅片质量的工艺参数,建立数据样本,使用WGAN-GP对样本数据进行数据增强.在此基础上,建立基于SeResNet的硅片总体厚度偏差预测模型.以硅片的多线切割加工过程监控数据为模型验证数据,对构建的硅片总体厚度偏差预测模型进行验证.实验结果表明:该模型具有良好泛化性和高准确率,有效解决了小样本数据下的预测难题,实现了平均相对误差小于10%的硅片总体厚度偏差预测,所以基于数据驱动的硅片质量预测来代替硅片加工中的质量检测具有重要的现实意义.
In monocrystalline silicon processing,the time-consuming and high cost of quality detection make it difficult to detect the quality of silicon wafers.Therefore,a silicon wafer quality prediction method based on WGAN-GP data processing and SeResNet was proposed.The mechanism of multi-wire saw of silicon wafer was analyzed,the process factors affecting the quality of silicon wafer were identified,the data samples were established,and WGAN-GP was used to enhance the sample data.On this basis,a prediction model for total thickness variation of silicon wafers was established based on SeResNet convolutional neural network.Taking the monitoring data of multi-wire cutting process of silicon wafer as the model verification data,the prediction model of total thickness variation of silicon wafer was verified.The experimental results show that the model has good generalization and high accuracy,the prediction problem under the small sample data is effectively solved,and the overall thickness deviation prediction with the average relative error less than 10%is re-alized.So it is of great significance to replace the quality detection in silicon wafer processing with data-driven based on silicon wafer quality prediction.