Parallel Frequency Function-Deep Neural Network for Efficient Approximation of Complex Broadband Signals

Area of Science:

• Signal Processing
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) are increasingly used for complex signal approximation.
• DNNs exhibit spectral bias, hindering their efficiency in fitting broadband signals.
• Existing methods like PhaseDNN use frequency selection but can be inefficient for smooth spectrums.

Purpose of the Study:

• To address the inefficiency of current methods in fitting broadband signals with smooth spectrums.
• To propose a novel deep neural network architecture for improved broadband signal fitting efficiency.
• To leverage frequency domain analysis and spectral bias for enhanced performance.

Main Methods:

• Developed a novel parallel frequency function-deep neural network (PFF-DNN) architecture.
• Utilized frequency domain analysis of broadband signals.
• Incorporated the spectral bias nature of neural networks into the model design.

Main Results:

• The proposed PFF-DNN method demonstrated substantial improvements in fitting efficiency for broadband signals.
• Extensive numerical experiments validated the enhanced performance of PFF-DNN.
• The method proved particularly effective for signals with smooth spectrums.

Conclusions:

• The PFF-DNN method offers a significant efficiency improvement for broadband signal fitting.
• PFF-DNN is a promising alternative solution to existing spectral bias mitigation techniques.
• This work contributes to the advancement of deep learning applications in signal processing.