Support vector data description (SVDD) has been applied to fault detection of batch processes without any restriction of process data distribution. However, the SVDD based traditional fault detection method of batch processes just simply divides the phase of batch processes using cluster analysis and model identification, and leads to rough phase division results which may compromise the accuracy of fault detection in multiphase batch processes. To address this issue, a SVDD based fault detection method for multiphase batch processes is presented. Firstly, the different phases are divided according to the change of hypersphere radius and support vectors of SVDD. Then, the SVDDbased fault detection models of different phases are established. Finally, fault detection is achieved by judging whether the difference between hypersphere radius of corresponding phase and the distance from the sample point to center of corresponding hypersphere exceeds the control limit. The experimental results of the fedbatch penicillin fermentation process show that the presented method can achieve phase division of multiphase batch processes with a better accuracy, and further realize multiphase fault detection of batch processes with a higher fault detection rate.