Support Vector Machine (SVM) has been extensively studied and have shown remarkable success in fault detection application. However the performance of traditional support vector machine drops significantly when it is applied to the problem of learning from imbalanced datasets where the normal instances heavily outnumbers the fault instances. To address this problem, a novel fault detection approach is proposed which is based on a variant of the Synthetic minority over-sample technique (SMOTE) combined with different error cost-sensitive SVM. As the SVM decision boundary is determined only by a small quantity of support vectors, Consequently, based on SMOTE, this paper presents a new minority over-sample method, in which only the minority examples near the borderline are over-sampled. In order to solve the noise effect, the different error cost-sensitive SVM based on K-Nearest Neighbors (KNN) is adopted to remedy the problem of noise positive instances. The proposed algorithm is applied in bearings fault detection application and is compared against these algorithms along with traditional SVM, different class cost-sensitive SVM (SVM-C), SVM+SMOTE. The experimental results show our approach can achieve better detection performance than other methods.