Discriminative kernel learning in ambiguity domain

Research in stochastic signal analysis is targeted towards two main objectives: (i) to obtain an overall dimensionality reduction and (ii) to provide reasonable characteristic estimates for quantification applications. Owing to the improved performance characteristics, time-frequency (TF) transformation tools are commonly used for such analysis. In this article, we propose a one-step characterization approach that exploits the collective advantages of TF analysis and discriminative kernels in the intermediate ambiguity domain (AD) for non-stationary signal analysis. Here, a machine learning kernel is used to suitably model the AD-map, following which certain robust AD-based features are extracted from the signal- and cross-term (generated during TF transformation) components. The novelty of the work is also geared towards finding out the usefulness of cross-terms for non-stationary signal classification applications. The proposed technique is evaluated for a multiclass quantification problem using one of the challenging stochastic triangular waveform datasets. Obtained misclassification accuracies are close to the theoretical minimum, previously reported using a Bayes classifier. Results indicate that this scheme shows great potential and can be extended in design of robust tools for real-life signal non-stationary analysis.