A Nonparametric Error Model for Pulsed Waveform Feature Extractors

Simulated pulse features (PFs) are commonly employed to develop signal exploitation algorithms for classification, clustering, and other applications. Typically, an arbitrary constant error model is applied to the simulated PFs to represent errors in the feature extraction process. But we show that PF extractors can impart SNR-dependent errors on PFs, making the traditional naive approach unrealistic. To produce more realistic errors, raw signal data can be simulated and passed through a receiver model to produce PFs that are corrupted by realistic errors. However, the signal generation process is computationally expensive and could be avoided by leveraging an appropriate PF error model. To this end, we propose a nonparametric model for pulsed waveform feature extractors based on Gaussian process regression (GPR). Our proposed model is constructed a priori from a parametric sweep across pulse parameters and can be applied to future PF-based simulations. We additionally show that SNR-dependent errors can be modeled as a heteroscedastic random process. In this work, we employ the GPR framework to build a heteroscedastic data-driven error model that can be applied to PF simulations, empirically demonstrate the model's effectiveness, and provide suggestions for further improvements of the models in future efforts.