LCD 面板 C / FOG 工艺制造虚拟计量方法研究

针对液晶显示器(LCD)面板的“Chip/FPC on Glass”(C/FOG)工艺生产制造过程中存在的计量延迟大、生产异常无法提前预测的问题，本文提出一种基于神经网络的C/FOG工艺生产制造虚拟计量方法。该方法利用生产机台上的传感器采集生产过程中的过程状态数据，构建基于多尺度一维卷积及通道注意力模型(MS1DC-CA)的虚拟计量模型。通过多个尺度的卷积核提取不同尺度范围内的状态数据特征。在对含有缺失值的原始数据预处理中，提出了基于粒子群算法改进的K近邻填补方法(PSO-KNN Imputation)进行缺失值填充，保留特征的同时，减少因填充值引入的干扰。最后在实际生产采集的数据上进行实验对比分析，实际不良率主要集中在0.1%～0.5%,该虚拟计量模型的拟合均方误差为0.397 7■,低于其他现有拟合模型，在平均绝对误差、对称平均绝对百分比误差和拟合优度3种评价指标下也均优于其他现有的拟合模型，具有良好的预测性能。

Research on virtual metrology method for LCD panel C / FOG process manufacturing

Aiming at the problems of large measurement delay and unpredictable production abnormalities that cannot be predicted in themanufacturing process of “ Chip / FPC on Glass” ( C/ FOG) for liquid crystal display ( LCD) panels, this paper proposes a virtualmetrology method for C/ FOG manufacturing. This method uses sensors installed on the production machine to collect process state dataduring the production process and constructs a virtual metrology model based on multi-scale one-dimensional convolution and channelattention network (MS1DC-CA). Where the state data features in different scale ranges are extracted through multiple scale convolutionkernels. In the preprocessing of original data containing missing values, an improved K-nearest neighbor interpolation method based onthe particle swarm optimization algorithm (PSO-KNN Interpolation) is proposed to fill in the missing values. This method can reduce theinterference introduced by filling values while retaining the features. Finally, it′s found that the actual defect rate is concentratedbetween 0. 1% and 0. 5% according to the experimental comparative analysis carried out for the data collected in actual production. Thefitting mean square error of virtual metrology model is 0. 397 7 , which is lower than other existing fitting models. It also outperformsother existing fitting models under the evaluation indexes including mean absolute error, symmetric mean absolute percentage error andgoodness of fit, which prouides the good predictive performance.