A modified AdaBoost algorithm with updating sample weight by dual threshold technique was proposed to model a soft sensor for estimating 4-CBA concentration

which could not be measured on-line in PX oxidation process. In this method

weak learners of BP neural networks were trained by part of samples selected by their weights and roulette wheel mechanism. The absolute values of last round training relative errors in weak learners were adopted to update weights of all training samples. Then

a second round updating on sample weights were completed by the product of original sample value and its weighting factor

which was defined by ratio of error range over dual thresholds. In the second updating process

weights were decreased for samples with gross errors but were increased for those with medium error. Consequently

probability of selecting outliers was reduced in following iteration of the training process. Five different methods were applied to model soft sensor of 4-CBA concentration with industrial data. Simulation results showed that the modified AdaBoost algorithm can improve soft sensor performance of 4-CBA concentration with predicting error less than that of other models.