Deep Learning versus Spectral Techniques for Frequency Estimation of Single Tones: Reduced Complexity for...

Area of Science:

• Signal Processing
• Artificial Intelligence
• Machine Learning

Background:

• Frequency estimation (FE) is crucial in signal processing.
• Classical methods like Discrete Fourier Transform (DFT) have limitations in accuracy and computational complexity.
• The application of artificial intelligence (AI) and deep learning (DL) to FE remains underexplored.

Purpose of the Study:

• To comprehensively analyze a deep learning (DL) approach for single-tone frequency estimation.
• To compare the performance of DL-based FE against classical techniques under various conditions.
• To evaluate the impact of signal-to-noise ratio (SNR), network architecture, and input data size on DL-based FE accuracy.

Main Methods:

• Development and analysis of a two-layer deep learning network for frequency estimation.
• Systematic evaluation of the DL model's performance across different signal-to-noise ratios (SNRs).
• Assessment of the model's accuracy with varying numbers of network nodes and input samples.

Main Results:

• The DL approach achieved higher accuracy than classical methods, even with limited input data.
• DL-based FE demonstrated robustness against SNR variations and network size.
• The DL model successfully estimated frequencies at minimal input node counts where classical methods failed.
• Computational complexity was reduced to O{N} for DL, compared to O{Nlog2 (N)} for DFT-based FE.

Conclusions:

• Deep learning provides a superior and more efficient alternative for frequency estimation.
• The reduced complexity, memory, and power requirements of DL make it suitable for resource-constrained systems like IoT devices.
• The findings pave the way for hardware-efficient implementations in software-defined radio (SDR) and other embedded systems.