In the semiconductor/TFT-LCD manufacturing industry, the timely quality estimation and control of products are the key solutions for the throughput improvement and cost reduction. Therefore, a virtual metrology model is present for high-mixed manufacturing processes with variety products under small quantity. By using the stepwise regression, the key variables are selected from the process variables monitored by the fault detection & classification(FDC) system. Then, combined with the product-effect factors, the virtual metrology model is built for multi-products by using the analysis of variance(ANOVA) algorithm. To reduce the disturbance effect, a state estimation method based on ANOVA is developed to estimate the relative states of each product. The method is formulated in the form of a recursive state estimation by using the Kalman filter. Numerical simulations are implemented by using practical production data from a wet etching process of TFT-LCD industry. The results show that proper variable selection and dynamic MANCOVA can improve the precision of prediction models effectively.