In this paper, a soft measurement model has been proposed by combining recursive fixed-memory principal component analysis (RFMPCA) with least squares support vector machines (LS-SVM). To solve outliers, missing data points of the outliers and deviation from normal values are detected. The RFMPCA was applied to the model, which not only solved drawbacks of conventional PCA and data saturation, but also simplified the LS-SVM structure and improved the training speed. The proposed model has been successfully applied to the decomposition process of Jiuganghongda Cement Plant in China. Industrial application results have shown that the soft measurement model has high accuracy and guidance to calciner temperature setting. (C) 2011 Elsevier Ltd. All rights reserved.