Since permutation entropy (PE) algorithm can better magnify tiny change of a time series data, is computed simply and shows good quality in real-time application, which has given us a good application prospect in sudden change detection of a signal, but whose parameters that are embedding dimension and delay time are still determined by experience or trial, this problem has been a bottle-neck of the PE algorithm for engineering application. For theory of the PE algorithm, method based on reconstructing optimal phase space of time series is put forward to determine model parameters. Considering to two views on phase space reconstruction, basic theories of independent and joint determination methods are introduced to determine delay time and embedding dimension. Then validation and comparison of methods are carried out by simulated signal and whole life data of rolling bearing, it is concluded that the independent determination of model parameters was better than joint determination for abnormality detection.