GFANC-RL: Reinforcement Learning-based Generative Fixed-filter Active Noise Control.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Generative Fixed-filter Active Noise Control (GFANC) offers a balance between noise reduction and stability.
• Training GFANC's Convolutional Neural Network (CNN) requires extensive, accurately labeled noise data, which is resource-intensive and prone to errors.
• Labeling inaccuracies can significantly impair the CNN's filter-generation precision in GFANC systems.

Purpose of the Study:

• To propose a novel Reinforcement Learning-based GFANC (GFANC-RL) approach.
• To eliminate the need for manual noise data labeling in GFANC systems.
• To enhance the accuracy and efficiency of the GFANC method through automated learning.

Main Methods:

• Developed a GFANC-RL framework utilizing Reinforcement Learning (RL) to automate CNN parameter updates.
• Employed the exploring capabilities of RL to bypass the data labeling process.
• Addressed the non-differentiability challenge associated with binary combination weights in GFANC using RL algorithms.

Main Results:

• Simulation results confirm the effectiveness of the GFANC-RL method.
• Demonstrated the transferability of GFANC-RL in processing real-world recorded noises.
• Validated performance across diverse acoustic paths, showcasing robustness.

Conclusions:

• GFANC-RL successfully removes the dependency on labeled data for GFANC training.
• The proposed method achieves accurate noise reduction and maintains system stability.
• GFANC-RL offers a more efficient and robust solution for active noise control applications.