Automatic Modulation Classification (AMC)

The defining characteristic of AMC is the ability to identify the modulation scheme without any prior knowledge of the transmitter's parameters. Unlike traditional demodulators that require pre-shared configuration, AMC systems extract features directly from the raw IQ samples to classify the signal. This is essential for spectrum monitoring, electronic warfare support, and cognitive radio applications where the receiver must autonomously adapt to unknown emitters.

AMC systems are broadly categorized into two architectural paradigms:

• Likelihood-Based (LB): Derives a test statistic from the probability density function of the received signal. Offers theoretical optimality under matched conditions but suffers from high computational complexity and sensitivity to model mismatches like phase offset.
• Feature-Based (FB): Extracts discriminating signal characteristics such as higher-order cumulants, cyclostationary signatures, or wavelet transforms, then feeds them into a classifier. This approach is more robust to real-world impairments.

Modern AMC systems leverage deep neural networks to replace hand-crafted feature extraction with end-to-end learned representations. Architectures include:

• Convolutional Neural Networks (CNNs): Treat IQ samples as 2D inputs to learn spatial hierarchies of features.
• Recurrent Neural Networks (RNNs/LSTMs): Model temporal dependencies in symbol sequences for time-varying modulations.
• Transformer Networks: Apply self-attention mechanisms to capture long-range dependencies in complex signal bursts, achieving state-of-the-art accuracy at low signal-to-noise ratios (SNRs).

Operational AMC systems must maintain high classification accuracy despite severe channel conditions. Key robustness requirements include:

• Multipath Fading: Resilience to frequency-selective fading that distorts the signal envelope.
• Carrier Frequency Offset (CFO) and Phase Noise: Tolerance to oscillator mismatches between transmitter and receiver.
• Low SNR Performance: Ability to classify signals below 0 dB SNR, often achieved through noise-robust feature extraction or denoising autoencoders.
• Co-Channel Interference: Discrimination of the target signal in the presence of overlapping transmissions.

Advanced AMC systems perform classification at multiple levels of granularity:

• Inter-Class Classification: Distinguishing between broad families such as PSK, QAM, and FSK.
• Intra-Class Classification: Identifying the specific modulation order within a family, e.g., differentiating QPSK from 8-PSK or 16-QAM from 64-QAM.
• Composite Modulation Recognition: Deconstructing hybrid schemes like OFDM with subcarrier modulation, identifying both the multiplexing technique and the underlying constellation.

Deployment in tactical software-defined radios (SDRs) imposes strict latency and resource constraints. Key performance metrics include:

• Classification Latency: The time from signal acquisition to modulation decision, typically required to be under 1 millisecond for time-sensitive applications.
• Model Footprint: Compressed models using quantization-aware training and weight pruning to fit within the limited FPGA or embedded processor resources of fielded systems.
• Incremental Learning: The ability to update the classification model on-device to recognize new modulation types without catastrophic forgetting of previously learned schemes.

A link adaptation mechanism that dynamically adjusts the modulation order and channel coding rate based on real-time channel conditions. While often confused with Automatic Modulation Classification, AMC is the transmitter-side decision to change schemes, whereas AMC (classification) is the receiver-side blind identification of that scheme. The two are symbiotic in a closed-loop cognitive radio system.

The direct manipulation of raw In-phase and Quadrature (IQ) data streams, which form the complex baseband representation of a modulated signal. AMC models typically operate on these raw IQ samples or derived features. Key preprocessing steps include:

• I/Q Imbalance Correction: Compensating for gain and phase mismatches
• DC Offset Removal: Eliminating carrier feedthrough
• Normalization: Scaling to unit power for consistent neural network input

A feature extraction technique that exploits the periodic statistical properties of modulated signals. Unlike stationary noise, modulated signals exhibit cyclostationarity at symbol rates and carrier frequencies. The Spectral Correlation Function (SCF) reveals these hidden periodicities, providing highly discriminative features for AMC that are robust to low Signal-to-Noise Ratio (SNR) conditions.

A classical non-ML approach that uses Higher-Order Statistics (HOS) to distinguish modulation formats. Fourth-order cumulants (e.g., C40, C42) are particularly effective at differentiating QAM, PSK, and ASK constellations because they capture the distribution shape of the signal's amplitude and phase. These features are often used as baseline inputs for lightweight ML classifiers or as a benchmark against deep learning methods.

A physical-layer security technique that identifies unique hardware impairments (e.g., power amplifier non-linearity, oscillator phase noise) embedded in transmitted waveforms. While AMC identifies the modulation type, RF fingerprinting identifies the specific device. Both can operate on the same raw IQ input, and a combined system can classify both the signal format and its emitter for threat assessment.