Underdetermined Blind Source Separation of Audio Signals for Group Reared Pigs Based on Sparse Component Analysis

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Separating mixed audio signals in complex environments like pig farms is challenging.
• Existing underdetermined blind source separation (UBSS) methods struggle with accuracy in such scenarios.

Purpose of the Study:

• To develop an improved UBSS method for pig environments using sparsification theory.
• To enhance the accuracy of mixing matrix estimation and audio signal reconstruction.

Main Methods:

• Proposed an underdetermined blind source separation (UBSS) method based on sparsification theory.
• Employed an improved AP clustering method for mixing matrix estimation.
• Utilized L1-paradigm separation for reconstructing individual pig audio signals.

Main Results:

• Achieved high performance with similarity coefficients above 0.8.
• Obtained an average recovered signal-to-noise ratio exceeding 8 dB.
• Demonstrated normalized mean square error below 0.02.
• The proposed algorithm outperformed classical UBSS methods in matrix estimation and signal separation.

Conclusions:

• The developed UBSS method effectively separates pig audio signals, improving reconstruction quality.
• Audio duration and mixing matrix significantly impact UBSS algorithm performance.
• Practical applications require careful consideration of recording duration and device spatial location.