Towards a physics-informed network paradigm with data generation and background noise removal for different...

Area of Science:

• Geophysics
• Signal Processing
• Artificial Intelligence

Background:

• Distributed Acoustic Sensing (DAS) is crucial for event recognition and noise reduction.
• Current AI models for DAS require extensive real-world data, which is often scarce.
• Limited data availability and high noise levels hinder practical DAS applications.

Purpose of the Study:

• To propose a novel physics-informed neural network paradigm for DAS data analysis.
• To eliminate the need for real-world event data during the training of AI models.
• To develop a method for synthesizing realistic DAS event data and effectively removing noise.

Main Methods:

• Physically modeling target events and system constraints to derive training functions.
• Utilizing a generative network to synthesize DAS event data.
• Training a noise-removal network with the generated synthetic data.
• Validating the approach on event identification and belt conveyor fault monitoring tasks.

Main Results:

• The proposed paradigm achieves performance comparable or superior to traditional data-driven methods.
• Achieved 91.8% fault diagnosis accuracy on a real belt conveyor using transferred networks from a simulation site.
• Demonstrated strong generalization capabilities across different sites within the same application.
• Effectively eliminated background noise in DAS measurements.

Conclusions:

• The physics-informed DAS neural network paradigm addresses the critical challenges of limited data and noise.
• This approach offers a viable solution for practical DAS applications requiring robust event recognition and denoising.
• The method shows significant potential for real-world deployment without extensive site-specific data collection.