Skip to Main Content
In many bearing fault anomaly detection application, only positive (normal) samples are available for training purposes, other abnormal samples are difficult to be available. In order to solve these practical application problems, a novel model of one-class bearing fault detection based on SVDD and genetic algorithm is presented in this paper. The time domain statistics features are processed as inputs to SVDD for one-class (normal) recognition. Then SVDD is used to describe the normal data distribution characteristics with high data description ability. The SVDD is trained only with a subset of normal samples. This paper also analyzes the behavior of the classifier based on parameter selection and proposes a novel way based on genetic algorithm to determine the optimal threshold parameters. The hybrid one-class classification model of SVDD and genetic algorithm is determined to address the problem of difficultly collecting abnormal samples in bearing fault detection. Comparison of the performance of detection of SVDD with different kernel parameters is experimented. This hybrid approach is compared against other MLP detection techniques. The results show the relative effectiveness of the proposed classifiers in detection of the bearing condition with some concluding remarks.