Hierarchical Deep Learning System for Radio Signal Classification

This repository presents a hierarchical deep learning system for Automatic Modulation Classification (AMC), capable of identifying one of 11 modulation types from raw I/Q radio signals.
The system leverages three specialized neural network models — Binary CNN, Non-Phase CNN-LSTM, and Phase-based CNN — arranged in a pipeline to maximize accuracy and efficiency.

---

Model Summary

This is a multi-model hierarchical classifier designed for robust signal modulation recognition.

The system operates in three stages:

1. Binary Model (CNN):
   Determines whether the input signal uses phase-based or non-phase-based modulation.
2. Phase Model (CNN):
   Classifies phase-based modulations using constellation plot images.
3. Non-Phase Model (CNN-LSTM):
   Handles temporal features of non-phase modulations directly from I/Q data.

This divide-and-conquer architecture allows each model to specialize in its domain, leading to improved overall classification performance.

---

Usage

This AMC system can be applied in:

• Cognitive Radio Systems
• Spectrum Monitoring
• Interference and Signal Detection
• Wireless Research and Education

Primary failure mode:
A misclassification by the Binary model (rare, as it achieves >99% accuracy).
Performance degrades under very low SNR (Signal-to-Noise Ratio) conditions.

---

Full Prediction Pipeline

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image

# Define class labels
NON_PHASE_CLASSES = ['PAM4', 'GFSK', 'CPFSK', 'WBFM', 'AM-DSB', 'AM-SSB']
PHASE_CLASSES = ['QPSK', 'QAM16', 'QAM64', 'BPSK', '8PSK']

# Load trained models
binary_model = tf.keras.models.load_model('best_binary_model.keras')
non_phase_model = tf.keras.models.load_model('best_model_non_phase.keras')
phase_model = tf.keras.models.load_model('best_model.keras')

def create_constellation_image(signal_frame):
    I, Q = signal_frame[:, 0], signal_frame[:, 1]
    fig, ax = plt.subplots(figsize=(4, 3), dpi=80)
    ax.plot(I, Q, '.')
    ax.axis('off')
    fig.tight_layout(pad=0)
    fig.canvas.draw()
    img_data = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
    img = img_data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
    plt.close(fig)
    img_gray = Image.fromarray(img).convert('L')
    img_resized = img_gray.resize((192, 256))  # W, H
    img_array = np.array(img_resized) / 255.0
    return np.expand_dims(img_array, axis=[0, -1])

def predict_modulation(signal):
    binary_input = np.expand_dims(signal, axis=[0, -1])
    is_phase_prob = binary_model.predict(binary_input)[0][0]

    if is_phase_prob > 0.5:
        # Phase-based path
        constellation_img = create_constellation_image(signal)
        prediction = phase_model.predict(constellation_img)
        return PHASE_CLASSES[np.argmax(prediction)]
    else:
        # Non-phase path
        non_phase_input = np.expand_dims(signal, axis=[0, 1])
        prediction = non_phase_model.predict(non_phase_input)
        return NON_PHASE_CLASSES[np.argmax(prediction)]

Example

dummy_signal = np.random.randn(128, 2) predicted_class = predict_modulation(dummy_signal) print(f"Predicted Modulation Class: {predicted_class}")

System Overview

• Input: Raw In-phase and Quadrature (I/Q) signal, shape (128, 2)
• Output: Predicted modulation class (string), e.g., 'QPSK'
• Pipeline:
  Binary CNN → Phase CNN or Non-Phase CNN-LSTM → Final Class
• Dependencies: Python, TensorFlow, NumPy, Matplotlib, Pillow

---

Implementation Details

Component	Description
Software	Python 3.8+, TensorFlow (Keras API)
Hardware (Training)	GPU (NVIDIA V100 / A100) or TPU
Hardware (Inference)	Works efficiently on CPU or GPU
Training Duration	Few hours to a day (depends on hardware)
Inference Time	< 1 second per signal (CPU), faster on GPU

---

Model Characteristics

Model	Type	Description
Binary Model	CNN	2 Conv2D + 2 Dense layers; fast and accurate (>99%)
Non-Phase Model	CNN-LSTM	3 Conv2D + 2 LSTM layers; processes temporal patterns
Phase Model	CNN	3 Conv2D + 1 Dense layer; classifies constellation images

Initialization

All models were trained from scratch using Keras’ default Glorot uniform initialization.
No transfer learning or pre-trained weights were used.

---

Dataset

• Dataset: RML22 (based on RadioML 2016.10a)
• Source: Synthetic wireless signals generated using GNU Radio
• Classes (11 total): ['BPSK', 'QPSK', '8PSK', 'PAM4', 'QAM16', 'QAM64','GFSK', 'CPFSK', 'WBFM', 'AM-DSB', 'AM-SSB']
• SNR Range: -20 dB to +18 dB
• Total Samples: 462,000 (2,000 signals per modulation-SNR pair)
• Train/Validation Split: 80/20

---

Evaluation Results

Model	Accuracy	Notes
Binary Model	>99%	Nearly perfect separation of phase/non-phase
Non-Phase Model	~80%	PAM4: 75%, GFSK: 85%, CPFSK: 78%, WBFM: 64%, AM-DSB: 82%, AM-SSB: 85%
Phase Model	High	Vision-based constellation classification
Overall System	Robust	Hierarchical integration yields strong accuracy

---

Fairness & Ethics

• Fairness:
  Balanced class distribution, using sparse_categorical_crossentropy ensures no class bias.

• Usage Limitations:

• Not reliable at extremely low SNRs.

• Trained only on 11 modulation types (won’t recognize unseen ones).

• Ethical Considerations:
  This model is intended for academic and research purposes such as spectrum management.
  Misuse in unauthorized surveillance or military applications is strongly discouraged.

---

Data Source

Dataset: RML22: Realistic Dataset Generation for Wireless Modulation Classification
Reference Paper:

V. Sathyanarayanan, P. Gerstoft, and A. E. Gamal,
"RML22: Realistic Dataset Generation for Wireless Modulation Classification,"
IEEE Transactions on Wireless Communications, vol. 22, no. 11, pp. 7663–7675, Nov. 2023.
DOI: 10.1109/TWC.2023.3254490

---

Author

Suyash Kumar Bhagat
B.Tech, Electronics and Communication Engineering
Delhi Technological University (DTU)
📧 suyashkbhagat_ec22a18_53@dtu.ac.in

---

Kaggle Model

Explore the full implementation and documentation on Kaggle:
-> Kaggle Notebook (Signal Classification Models)

---

License

This project is released under the MIT License.

---