Minimum classification error using time-frequency analysis

For certain classes of signals, such as time varying signals, classical classification algorithms are not suitable. Hence, time-frequency based techniques are employed for classification of these types of signals. In this paper we propose data-driven time frequency representations kernel optimization, that leads to the minimum classification error (MCE) for nonstationary signal classification. Our central issue is to determine the optimal kernel parameters and best distance measure to achieve the MCE performance measure. The minimum classification error achievable using optimized kernels is investigated for two types of nonstationary signals; namely simulated chirp signals and real-life newborn EEG signals. For the EEG signals a classification error as low as 4.6% was achieved.