Abstract: Aiming at the problem that it is difficult to achieve fault decision-making due to an unknown failure mechanism of planetary roller screw mechanism (PRSM)and a shortage of faulty types of PRSM in practice,this paper proposes a one-class model called deep support vector data description (deep SVDD) to determine whether PRSM is normal or not. Firstly,vibration signals of PRSM are collected on a PRSM test bench under three working states, including normal state, failure of lubrication and failure of teeth on one side of the roller. Then, the data are normalized and enhanced by window cropping to expand the number of samples. After that,wavelet packet transform is used to initially extract features of the data by decomposing signals. Finally, deep SVDD is used to complete fault diagnosis of PRSM.Meanwhile, it is compared with one-class support vector machine (OCSVM) and support vector data description (SVDD). The results show that deep SVDD has a better classification ability and higher training efficiency, so it is suitable for fault diagnosis of PRSM.