Error Origin Analysis for Adaptive Time-Frequency Distribution Thresholding of Nonstationary Signals

This paper introduces a criterion for the classification of the Time-Frequency Distribution (TFD) coefficients considered as wrong estimates of the TFD automatic thresholding. The method for the automatic thresholding of TFDs is based on an initial one-dimensional segmentation of the TFD, which creates K subsets, and is followed by a local Renyi entropy evaluation of each subset to distinguish those containing useful information coefficients from those containing noise. The evaluation of the algorithm's performance has been assessed on test signals, introducing different error categories, based on the origin of the mistakenly classified coefficient. The coefficient position in the time-frequency plane determines its role in the thresholded TFD as "true", "false positive" or "false negative". The proposed error classification offers new insights into the algorithm's strengths and weaknesses when applied to signals embedded in stationary and nonstationary AWGN.