Deep Learning OFDM Classifier

The classifier operates directly on raw in-phase and quadrature (IQ) samples, bypassing traditional expert-defined feature extraction. A convolutional neural network (CNN) learns hierarchical representations from the time-domain waveform, automatically discovering discriminative patterns such as cyclic prefix correlation peaks, subcarrier spacing signatures, and pilot structures. This eliminates the information bottleneck of handcrafted features like cumulants or spectral correlation functions.

Many architectures preprocess IQ streams into spectrograms—2D time-frequency representations—before feeding them to the network. This transforms the classification problem into an image recognition task, where CNNs detect visual patterns:

• Vertical ridges indicating subcarrier energy
• Horizontal gaps revealing symbol boundaries
• Pilot subcarrier grids unique to specific OFDM numerologies Residual networks (ResNets) and vision transformers are commonly applied to these representations.

Beyond modulation type, the classifier jointly estimates physical-layer parameters through a multi-head architecture. A shared backbone network feeds separate output branches that simultaneously predict:

• FFT size (e.g., 128, 256, 512, 1024, 2048)
• Cyclic prefix length (normal vs. extended)
• Subcarrier spacing (15 kHz, 30 kHz, 60 kHz)
• Center frequency offset This shared representation learning improves generalization across tasks.

Training regimens incorporate augmented datasets simulating real-world channel conditions to ensure operational reliability:

• Additive white Gaussian noise (AWGN) across a wide SNR range
• Multipath fading profiles (EPA, EVA, ETU models)
• Carrier frequency offset (CFO) and sampling clock offset (SCO)
• Narrowband interference and adjacent channel leakage Dropout and batch normalization layers further regularize against overfitting to specific channel conditions.

Production classifiers incorporate open set recognition to handle unknown OFDM variants not seen during training. Techniques include:

• Softmax thresholding with calibrated confidence scores
• Extreme value theory (EVT) modeling of known-class activation vectors
• Distance-based rejection in the penultimate embedding space This prevents misclassification of novel waveforms as known types, critical for spectrum monitoring and electronic warfare applications.

Trained models undergo model compression for real-time inference on FPGAs and embedded SDR platforms:

• Post-training quantization to INT8 precision
• Structured pruning removing redundant convolutional filters
• Knowledge distillation from larger teacher networks to compact student models Inference latencies below 1 millisecond are achievable on modern RFSoC devices, enabling in-situ spectrum awareness.

The unique spectral correlation pattern generated by the cyclic prefix and pilot subcarriers in OFDM signals. Deep learning classifiers can learn to detect these signatures without explicit cyclic frequency calculation.

• Spectral Correlation Density (SCD): 2D function revealing hidden periodicities
• Blind CP Detection: Autocorrelation lag profiles identify guard interval length
• Robust to Low SNR: Cyclostationary features persist below noise floor

The identification of scalable physical-layer parameters in 5G NR, including subcarrier spacing (15, 30, 60, 120 kHz) and cyclic prefix length. Convolutional neural networks trained on spectrograms can distinguish between numerologies.

• μ Value Detection: Maps to SCS = 15 × 2^μ kHz
• Frame Structure Inference: Slot duration derived from numerology
• Frequency Range Discrimination: FR1 vs FR2 parameter sets

A time-frequency analysis technique that identifies energy ridges corresponding to individual subcarriers. Deep learning models process spectrograms as 2D images, applying convolutional layers to detect subcarrier spacing and symbol duration.

• STFT-Based Representations: Short-time Fourier transform as model input
• Subcarrier Spacing Estimation: Ridge separation in frequency dimension
• Symbol Boundary Detection: Temporal energy transitions

The identification of specific implementation details—pilot patterns, preamble structures, and frame timing—to determine the wireless standard (LTE, 5G NR, WiFi) or vendor-specific configuration. Deep classifiers learn protocol-specific signatures.

• PSS/SSS Detection: Zadoff-Chu and m-sequence identification
• DMRS Pattern Recognition: Pilot density and positioning
• Vendor Fingerprinting: Hardware-specific implementation artifacts

Preprocessing and model-based techniques for mitigating fading, noise, and frequency offset effects before classification. Robust deep learning classifiers incorporate channel estimation as an auxiliary task or learn invariant representations.

• Carrier Frequency Offset (CFO) Correction: Blind estimation from CP correlation
• Multipath Resilience: Learned equalization within network layers
• Data Augmentation: Synthetic channel impairments for training robustness