Neural network-based methods for spread spectrum signal classification in radio monitoring systems

Abstract

Automatic classification of spread spectrum signals — frequency-hopping, direct-sequence, and chirp — is a key task in modern radio monitoring systems, particularly relevant for distributed sensor networks with constrained computational resources. A critical review of existing approaches shows that none of the three generations of classification methods — classical deterministic, feature-based machine learning, and deep learning on time-frequency representations — simultaneously meets three essential requirements: high accuracy at negative signal-to-noise ratios, computational complexity below 10⁵ multiply–accumulate operations per realization, and compatibility with integer arithmetic for embedded deployment. This paper proposes a method that addresses this gap through
a compact, informative feature vector combining frequency-domain, time-frequency, and statistical characteristics. The informativeness of Hjorth parameters is theoretically justified via their analytical link to spectral moments of the power spectral density, enabling O(N) time-domain computation equivalent to frequency-domain analysis. A formalized ablation analysis with three quantitative selection criteria (individual significance, pairwise correlation below 0.7, and absence of negative contribution) yields a reduced vector of five components. Complexity analysis confirms approximately 3·10⁴ operations per realization and 2 kB model memory in integer configuration — four orders of magnitude less than convolutional network–based approaches. Experimental evaluation of a multilayer perceptron classifier demonstrates stable accuracy above 91% across a wide SNR range, 93.2% at 20 dB and 87% at −10 dB, with negligible degradation under integer quantization, confirming practical applicability to embedded and distributed radio monitoring systems.