Modified Receiver

Architecture in SDR for

Real-Time Modulation
Classification
GUIDED BY A K S H AY A J
P R O F. N I S S A S U R L I N G T RV 2 1 E C 0 0 5
D E P T. O F E C E , G E C B A R TO N H I L L G E C B A R TO N H I L L
Contents
• Introduction
• Points of Discussion
• Traditional AMC & DL Approaches
• SDR Overview & Related Work
• Proposed DL Models & Receiver
• Dataset & Training Setup
• Results & Performance Analysis
• Conclusion
• References

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 2

Introduction
• Automatic Modulation Classification (AMC) is key for systems like Cognitive
Radio and secure communications.

• AMC enables blind recognition of signal modulations.

• Traditional methods face hardware and channel challenges.

• Software-Defied Radio (SDR) and Deep Learning (DL) offer better real-time AMC
performance with improved classification accuracy.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 3

Points of Discussion

• Analyse the development of real-time AMC system using SDR and DL.

• Discussion on implementation of CLDNN and RSTM models for improved accuracy.

• Validate the system through real-world tests and simulations.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 4

Existing DL Architecture &
Modulation Classification
• CNN Model
o Architecture: Two convolutional layers with 2M+ parameters.
o Performance: 73% accuracy at 16 dB SNR.
o Limitations: Large number of parameters, room for optimization.

• CLDNN Model
o Architecture: Hybrid CNN + LSTM, 3 CNN layers with 50 LSTM units.
o Performance: 91.5% accuracy at 14 dB SNR with 2.5M+ parameters.
o Limitations: High complexity, needs parameter reduction for practical use.

• ResNet Model
o Architecture: Residual stacks with six 1D convolutional layers, avoiding gradient vanishing.
o Performance: 96% accuracy at 14 dB SNR, with only 155K parameters.
o Advantages: Efficient with fewer parameters, high accuracy.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 5

SDR Overview & Related Work
• Software-Defined Radio
o Definition: A communication system where functions like modulation, demodulation, and
signal processing are handled by software.
o Application: Widely used in cognitive radio, 5G testing, spectrum sensing, and real-time
AMC.

• Related Works
o Previous Studies: Earlier works focused on integrating DL models (CNN, ResNet) with SDR
for AMC.
o Current Research Focus: Developing a more robust SDR-based AMC system with DL models
that improve real-time classification accuracy and handle complex scenarios.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 6

Proposed DL Classifier Models
• Convolutional Long Short-Term Memory Deep Neural Network (CLDNN)
o Effective for real-time modulation classification, especially in complex and noisy
environments.

• Residual Long Short-Term Memory Neural Network (RSTM)

o Suitable for real-time classification with stable performance in varying conditions.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 7

Convolutional Long Short-Term Memory
Deep Neural Network (CLDNN)

Figure 1

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 8

Convolutional Long Short-Term Memory
Deep Neural Network (CLDNN)
• Architecture: Combines CNN for feature extraction and LSTM for temporal sequence learning.
• Layers:
◦ 3 CNN layers for feature extraction (256, 256, 80 filters).
◦ LSTM layer with 50 units to capture long-term dependencies.
◦ Fully Connected Layer for prediction.
• Total Parameters: 309,450 (was 2M+ for Existing DL models)
• Advantages:
◦ Learns both spatial and temporal features of signals.
◦ High classification accuracy (93% at high SNR levels).
◦ Reduced processing time for real-time applications.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 9

Residual Long Short-Term Memory
Neural Network (RSTM)

Figure 2
MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 10
Residual Long Short-Term Memory
Neural Network (RSTM)
• Architecture: Combines Residual Networks (ResNet) with LSTM for improved feature
learning.
• Layers:
o Residual Stack with 5 one-dimensional CNN layers (64 filters each).
o LSTM layers with 128 hidden units for sequence learning.
o Dropout Layers for preventing overfitting.
• Total Parameters: 428,234, slightly higher but with better memory retention.
• Advantages:
o Residual connections avoid gradient vanishing and improve training.
o High accuracy across a wide range of SNR levels.
o Effective for long-term signal dependencies.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 11

Receiver

Figure 3

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 12

Receiver
• Architecture:
• Signal Reception: SDR hardware receives and downconverts signals.
• Preprocessing: AGC and FLL correct amplitude and frequency offsets.
• AMC Block: Uses CLDNN/RSTM for real-time modulation classification.
• Demodulation:
• Based on AMC output, the receiver selects the appropriate demodulation method.
• Real-Time: Automatic and adaptive demodulation for fast signal decoding.
• Key Features:
• Fast, accurate classification with deep learning.
• Adaptive demodulation based on predicted modulation type.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 13

Dataset
•RadioML-2016.10b Dataset:
oPurpose: Used for optimizing model architectures.
oContent: 10 modulation schemes (BPSK, QPSK, 8PSK, PAM4, QAM16, QAM64, CPFSK, GMSK,
GFSK, WBFM).
oSNR Range: -20 dB to 18 dB.
oSize: 1.2 million samples, each with 128 IQ pairs (2×128 format).
•GRA Dataset (Generated):
oPurpose: For real-world validation of models.
oContent: 8 modulation schemes (8PSK, BPSK, GFSK, GMSK, PAM4, PAM8, QAM16, QPSK).
o SNR Levels: 24 dB and 10 dB for testing.
oSize: 102.4 million samples, with 50,000 frames per modulation

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 14

Training Setup
•Training Framework:
• Tools: Implemented using GNU Radio for SDR setup and TensorFlow for DL model training.
• Models: CLDNN and RSTM trained on RadioML dataset, validated on GRA dataset.

•Training Strategy:
• Data Augmentation: Applied noise, frequency shifts, and fading effects to improve model
robustness.
• Cross-Validation: 80% training, 20% validation split on RadioML dataset.
• Real-World Testing: Models validated on real SDR transmission, simulating varying distances
and SNR conditions.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 15

Results & Performance Analysis

Training Results

• CLDNN achieves 93% accuracy;

RSTM reaches 90% at high SNR levels.
• Both models outperform traditional
methods in accuracy and efficiency.

Figure 4

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 16

Results & Performance Analysis
Bit Error Rate Simulation
• Simulated BER under AWGN approaches theoretical values.

Figure 5
MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 17
Results & Performance Analysis
Real-World Testing
• CLDNN achieves 98% accuracy and
RSTM reaches 96% at high SNR levels
in real transmission.
• CLDNN shows minimum inference and
decision time over RSTM

Model Inference Decision

Time (in s) Time (in s)
CLDNN 2 x 10-4 2

RSTM 4 x 10-4 4
Figure 6 Table 1

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 18

Conclusion
• CLDNN achieves up to 98% accuracy in real-world scenarios.

• RSTM demonstrates robust performance with effective handling of complex

modulations.

• Significant reduction in inference and decision times, suitable for practical

applications.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 19

References
[1] Q. N. Le, T. Q. Huynh, H. Q. Ta, P. V. Tan, and L. L. Nguyen, "Modified receiver architecture
in] software-defined radio for real-time modulation classification," EURASIP J. Adv. Signal
Process., vol. 2024, no. 77, 2024,pp. 1-16.
[2] T. J. O'Shea, J. Corgan, and T. C. Clancy, "Convolutional radio modulation recognition
networks," in Proc. GNU Radio Conf., 2016, pp. 1-6.
[3] A. Emam, M. Shalaby, M. A. Aboelazm, H. E. A. Bakr, and H. A. Mansour, "A comparative
study between CNN, LSTM, and CLDNN models in the context of radio modulation
classification," in Proc. 12th Int. Conf. Elect. Eng. (ICEENG), 2020, pp. 190-195.
[4] C. Lin, W. Yan, L. Zhang, and W. Wang, "A real-time modulation recognition system based
on software-defined radio and multi-skip residual neural network," IEEE Access, vol. 8,2020,
pp. 221235-221245.
[5] T. T. An and B. M. Lee, "Robust automatic modulation classification in low signal to noise
ratio," IEEE Access, vol. 11,2023, pp. 7860-7872.

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 20

 THANK YOU

MODIFIED RECEIVER ARCHITECTURE IN SDR FOR REAL-TIME MODULATION CLASSIFICATION 21