Information criterion-based source number estimation methods with comparison

To effectively evaluate the source number of mechanical systems from the measured mixed signals, and solve the calculating difficulty of BIC for large data points, three information criterion-based source number estimation methods, Akaike information criterion (AIC), minimum description length (MDL), and Bayesian information criterion (BIC), are comparatively studied and an improved BIC, named IBIC, is proposed following an exponential function modification, which transforms the multi-parameter exponential calculating to multiplications. Without decreasing the accuracy, IBIC obviously improves the calculating efficiency and engineering application performances. The numerical case study results show that AIC and MDL obtain the similar performances on source number estimation, and they are both very sensitive to the nonlinear modulation effects. In respect to signal energy ratios, the proposed method has a robustness tolerance on nonlinear modulation effects for 5.15%, which is higher than that of AIC (0.07%) and MDL (0.08%). The results of source number estimation for acoustical signals of a test bed with shell structures show that all the three methods are effective for the given acoustical signals. This work benefits model order selection, complexity analysis of a system, and applications of source separation to mechanical systems for the condition monitoring and fault diagnosis purposes.