We quantify the impact of input sanitation strategies—nan_to_num, zero-padding, and linear interpolation—on
classification error and latency under controlled NaN corruption
of IQ streams. We integrate sanitation hooks in temporal and
spectral feature builders and systematically evaluate robustness
across corruption ratios. Our analysis reveals that linear interpolation typically dominates at low-to-moderate corruption levels,
while nan_to_num offers the fastest processing but introduces
the most spectral distortion. We provide quantitative guidance for
selecting appropriate sanitation strategies based on corruption
characteristics and performance requirements.
Index Terms—RF signal processing, robustness, input sanitation, ensemble methods, spectral analysis
Radio frequency (RF) signal classification systems frequently encounter corrupted input data due to hardware failures, interference, or transmission errors. Missing samples,
represented as NaN (Not a Number) values in digital signal
processing pipelines, can propagate through feature extraction
and classification stages, leading to degraded performance or
complete system failures.
This paper systematically evaluates the robustness of RF
ensemble classifiers to input corruption, specifically focusing
on the impact of different NaN sanitation strategies. We inject
controlled corruption patterns into IQ data streams and measure the resulting effects on classification accuracy, processing
latency, and spectral feature quality.