Abstract: In view of the problem that railway turnout switch machines lack a large number of abnormal samples to implement abnormal detection of their operation state, this paper proposed an improved model of abnormal detection based on improved support vector domain description. Taking ZYJ7 hydraulic turnout switch machine as the research object, the power data of turnout were collected by using the existing computer monitoring system. Clustering method was used to clean the power data, then unlock, transform and decompose the power data in time series. Statistical eigenvalues were extracted. Principal component analysis (PCA) method was used to reduce the dimension of the eigenvalues. The processed data were input to the abnormal detection classifier for model training and model testing. The experimental results show that the improved Support Vector Domain Description (SVDD) classifier has a strong recognition ability for abnormal detection of turnout operation state.