Novel deep-learning based approach for detection of single channel signal over high frequency (HF) channels

• Main Author: Lei, Mingyuan
• Other Authors: Teh Kah Chan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering::Computer hardware, software and systems; Engineering::Electrical and electronic engineering::Antennas, wave guides, microwaves, radar, radio
• Online Access: https://hdl.handle.net/10356/158022
• Institution: Nanyang Technological University

This report is on the Final Year Project “Novel deep-learning based approach for detection of single channel signal over high frequency (HF) channels”. Deep Learning (DL) is a powerful machine learning technique which has been recently implemented in many fields such as image classification, Natural language processing, recommendation system and achieves great success. It also shows great power in communication-system based classification and generation tasks and a series of deep-learning based noise reduction and signal classification techniques have been well developed. This report investigates some novel deep-learning algorithms for single channel high frequency (HF) signal noise detection. Three types of classifiers are used in this report: (i)Residual Network (ResNet) based on Convolutional Neural network (CNN); (ii) Recurrent neural network (RNNs) and (iii) Long-Short Term Memory (LSTM). Their performance is first evaluated by data samples with different SNR separately and then an Autoencoder denoiser is trained and combined with Resnet 34 model which is the best backbone classifier shown by the empirical results. The empirical results also show that the exact choice of digital modulation methods will not affect the performance. To fairly compare the performance of different models, only 4-ary Frequency shift keying (4-FSK) and Quadrature Phase Shift Keying (QPSK) will be used in our signal dataset. Finally, the model is trained and evaluated on a mixed dataset and obtained accuracy of 95% when SNR is higher than -10dB and 75% when SNR is low to -20dB.