Blind Source Separation Method Based on Neural Network with Bias Term and Maximum Likelihood Estimation Criterion.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Blind source separation (BSS) is vital for engineering applications.
• Improving convergence speed and steady-state performance of BSS algorithms is essential.

Purpose of the Study:

• To propose a novel BSS method combining maximum likelihood estimation and neural networks (NNs).
• To enhance algorithm performance by modifying the loss function and optimizing the algorithm.
• To improve convergence speed and steady-state performance for practical engineering use.

Main Methods:

• Developed a BSS method integrating maximum likelihood estimation with a biased NN.
• Incorporated L2 regularization terms into the loss function for improved steady-state performance.
• Designed a novel optimization algorithm featuring a dual acceleration strategy to boost convergence speed.

Main Results:

• The dual acceleration strategy accelerates convergence by four times compared to competing algorithms.
• Achieved a 96% improvement in the steady-state performance index.
• Validated performance using both simulated and measured data, demonstrating superior convergence and steady-state results.

Conclusions:

• The proposed BSS method offers significant improvements in convergence and steady-state performance.
• The novel optimization strategy makes the algorithm more suitable for real-world engineering applications.
• Demonstrated effectiveness across diverse datasets, highlighting its robustness and applicability.