Feature-Based AMC

Engineers manually design and select discriminative statistical features from the received signal before classification. Common features include:

• Higher-order cumulants (HOC) that are theoretically immune to Gaussian noise
• Spectral correlation density functions from cyclostationary analysis
• Instantaneous amplitude, phase, and frequency statistics
• Signal moments up to the 8th order

This manual process requires deep domain expertise in signal processing but produces features with clear physical interpretations.

Higher-order cumulants are the most widely used features in feature-based AMC due to their theoretical insensitivity to additive white Gaussian noise. Key properties:

• 2nd-order cumulants capture variance (power)
• 4th-order cumulants distinguish between modulation families (PSK vs. QAM)
• 6th and 8th-order cumulants enable intra-class discrimination (e.g., 16-QAM vs. 64-QAM)

Cumulant values form a unique fingerprint for each modulation scheme, making them robust features for hierarchical decision trees.

Feature-based AMC systems commonly employ hierarchical decision trees that mirror the logical structure of modulation taxonomy:

• Root node: Separates analog from digital modulations
• Intermediate nodes: Distinguish modulation families (ASK, PSK, FSK, QAM)
• Leaf nodes: Identify specific modulation orders (BPSK, QPSK, 16-QAM)

Each decision boundary is a threshold on a specific cumulant or spectral feature, making the entire classification process fully auditable and explainable.

When decision boundaries become non-linear, Support Vector Machines (SVMs) with kernel functions replace simple thresholding:

• Radial Basis Function (RBF) kernels map features to higher-dimensional spaces
• SVMs find the maximum-margin hyperplane between modulation classes
• Particularly effective for distinguishing higher-order QAM schemes (64-QAM, 256-QAM)

SVMs maintain the interpretability advantage while handling more complex feature distributions than decision trees alone.

Feature-based AMC is highly sensitive to front-end preprocessing quality. Critical steps include:

• Carrier frequency offset (CFO) estimation and compensation to prevent constellation rotation
• Symbol timing recovery to align samples with symbol centers
• Blind equalization to mitigate multipath channel distortion
• Noise power estimation for SNR-aware feature normalization

Errors in any preprocessing stage propagate directly into feature computation, degrading classification accuracy. This creates a brittleness not present in end-to-end deep learning approaches.

Feature-based methods offer significant deployment advantages on resource-constrained hardware:

• Feature computation requires only statistical moments, not matrix multiplications
• Decision trees and SVMs have minimal inference latency compared to neural networks
• No GPU requirement—runs efficiently on embedded DSPs and FPGAs
• Memory footprint measured in kilobytes, not megabytes

This makes feature-based AMC ideal for real-time tactical SIGINT and low-SWAP cognitive radio applications.

Higher-order statistics (HOS) that form the mathematical backbone of feature-based AMC. Cumulants of order greater than two are theoretically immune to additive white Gaussian noise, making them robust discriminators. Key cumulants used include:

• C₄₀: Distinguishes between PSK and QAM families
• C₄₂: Separates QAM orders (16-QAM vs. 64-QAM)
• C₆₃: Resolves intra-class ambiguities in higher-order PSK These hand-crafted features are computed from the signal's empirical moments and fed directly into a downstream classifier.

A signal processing technique that exploits the periodic statistical properties inherent in modulated signals. Unlike stationary noise, modulated carriers exhibit spectral correlation at specific cycle frequencies related to the symbol rate and carrier offset. The Spectral Correlation Density (SCD) function serves as a rich, noise-robust feature map. Key advantages:

• Discriminates signals with overlapping power spectra
• Provides blind estimates of symbol rate and carrier frequency
• Resilient to low SNR conditions where constellation-based methods fail

The dominant shallow classifier paired with hand-crafted features in traditional AMC pipelines. SVMs find the optimal separating hyperplane that maximizes the margin between modulation classes in the cumulant feature space. Common configurations:

• RBF kernel: Handles non-linear decision boundaries between higher-order QAM schemes
• One-vs-One strategy: Decomposes multi-class AMC into binary sub-problems
• Soft-margin parameter C: Controls the trade-off between training accuracy and generalization to unseen SNR conditions SVMs offer strong theoretical guarantees with limited training data compared to deep learning approaches.

Hierarchical, rule-based classifiers that partition the feature space through a sequence of binary decisions on individual cumulant or statistical features. Ensemble variants dominate practical deployments:

• Random Forests: Aggregate multiple de-correlated trees trained on bootstrap samples to reduce variance
• Gradient Boosted Trees (XGBoost): Sequentially build trees that correct the residual errors of the ensemble, achieving state-of-the-art performance on tabular feature sets These methods provide inherent interpretability through feature importance scores, a critical advantage over black-box neural networks in electronic warfare applications.

A critical blind synchronization step required before feature extraction. Feature-based AMC assumes the signal has been properly sampled at the correct symbol rate. Key estimation techniques:

• Cyclic cumulant methods: Exploit the cyclostationary property at integer multiples of the symbol rate
• Haar wavelet transform: Detects transient edges at symbol transitions
• Delay-and-multiply nonlinearity: Generates spectral lines at the symbol rate Without accurate symbol rate estimation, cumulant features become corrupted and classification accuracy degrades catastrophically.

The fundamental theoretical limit below which a specific feature-based classifier can no longer reliably discriminate between signal and noise, regardless of observation length. This phenomenon arises because:

• Cumulant estimators have finite-sample variance that increases at low SNR
• Decision boundaries in feature space collapse as noise dominates the signal component
• Each modulation family exhibits a distinct SNR wall determined by its constellation geometry Understanding this limit is essential for defining the operational envelope of a deployed AMC system.