Fault detection based on modified query by committee(MQBC) and cost-sensitive support vector machine(CS-
SVM) is proposed to solve three difficulties of fault detection, including class-imbalanced dataset, expensive labeled cost and outlier of sample set. The definition of information is given and the MQBC is proposed. The score of unlabeled sample is evaluated by using information, and the high score of unlabeled sample is selected to be labeled and added to the training set. Different misclassification types of samples are given to different misclassification cost in CS-SVM, so that the fault detection rate is increased. Finally, fault detection for copper flash smelting process is studied to illustrate the effectiveness of the proposed approach.