A New Singular Value Decomposition Method for AR Model Order Selection via Vibration Signal Analysis

Type selection and operational suitability test of model are two basic aspects of time series modeling. Ascertaining the order of time series model is the key problem of suitability test. Some traditional order selection criteria haven´t yet been adapted for ascertaining an optimal model order. In this paper, a new Singular Value Decomposition (SVD) method for determining the order of an autoregressive (AR) model was presented and compared with traditional order selection methods, i.e. Final Prediction Error(FPE), Akaike Information Criterion(AIC), Bayesian Information Criterion(BIC) and SVD Slicing, according to AR bispectrum analysis of vibration signals derived from the faults of the hydraulic valves. With simulation experiments, the order determined by traditional order selection methods was too low and the fault information could not be discriminated clearly in comparison with Frobenius normalized norm method of SVD. Considering the above situation, it has been drawn as a conclusion that the proposed new method outperforms the traditional order selection methods.