Detection of Electric Network Frequency in Audio Using Multi-HCNet.

Area of Science:

• Forensic Science
• Signal Processing
• Machine Learning

Background:

• Electrical network frequency (ENF) detection is crucial for forensic audio and video analysis.
• High-pass filtering in communication systems degrades ENF detection by removing fundamental frequency information.
• Existing ENF detection methods suffer performance loss in filtered environments.

Purpose of the Study:

• To develop an innovative deep learning model, Multi-HCNet, for robust ENF signal detection in high-pass filtered environments.
• To address the challenge of lost fundamental frequency information in modern communication scenarios.
• To enhance the accuracy and reliability of ENF detection for forensic applications.

Main Methods:

• Multi-HCNet model incorporating an array of high-order harmonic filters (AFB) to capture harmonic components.
• Grouped multi-channel adaptive attention mechanism (GMCAA) for distinguishing between multiple frequency signals (e.g., 50 Hz vs. 60 Hz).
• Sine activation function (SAF) to better capture the periodic nature of ENF signals.

Main Results:

• Multi-HCNet achieved a peak detection accuracy of 98.84% and maintained an average accuracy over 80% under high-pass filtering.
• The model effectively detects ENF signals even with the absence of fundamental frequency information.
• Successful differentiation between 50 Hz and 60 Hz ENF signals was demonstrated.

Conclusions:

• Multi-HCNet offers a novel solution for ENF signal detection in challenging environments with lost fundamental frequency data.
• The model's ability to distinguish between 50 Hz and 60 Hz signals supports practical forensic applications.
• This research significantly improves ENF detection accuracy and robustness for forensic analysis.