Classification of non-stationary random signals using multiple hypotheses testing

In this paper we introduce a new time-frequency based method for classifying non-stationary random signals. The method involves dividing the signal into overlapping or nonoverlapping segments considered to be subpopulations of the entire population. From each sub-population we calculate a test statistic which can be used to construct a single hypothesis test. To control the global type-I error it is necessary to consider the hypotheses from all subpopulations simultaneously. We use the generalised sequentially rejective Bonferroni multiple hypothesis test which provides an efficient method to simultaneously test multiple hypotheses while maintaining the global type-1 error. Finally, we show the results of classifying time-dependent AR(1) processes which have identical expected instantaneous power and power spectral densities but different time-frequency representations.