Abstract: Support vector machine（SVM）is a new learning machine based on the statistical learning theory.A regression algorithm based on least squares vector machine（LS-SVM）within the Bayesian evidence framework was proposed by considering the noise parameter β.Within the evidence framework，the regularization and kernel parameters could be adjusted automatically，to achieve a better tradeoff between the minimum error and model’s complexity.The proposed method was applied to predict the concentration of 4-carboxybenzaldhyde（4-CBA）in an industrial purified terephthalic acid（PTA）oxidation process.The estimated outputs of LS-SVM match the real values and followed the change of the 4-CBA content very well.The results show that the proposed method provides a new and effective method for on-line 4-CBA prediction.