A new method based on a general expression for linear and nonlinear auto-regressive model (GNAR model) is presented for state identification and fault diagnosis of mechanical systems. GNAR models are establised based on the system acquisition signal. Then by applying principal component analysis strategy, the eigenvectors are extracted from training samples which provides some reference patterns. Finally, the undetermined systems are detected and classified by Euclide distances between the undetermined system eigenvectors and the reference pattern centers. The results from lathe and air compressor test data analysis show that the proposed method is convenient and effective and can be applied to the engineering.