Graph-based sensor fusion for classification of transient acoustic signals

Area of Science:

• Acoustic Sensing and Signal Processing
• Machine Learning for Defense Applications
• Munitions Detection and Classification

Background:

• Simultaneous acoustic sensing enables multiple measurements of the same event.
• Correlations between sensor data offer potential for improved signal classification.
• Identifying munition launches (rockets, mortars) is critical for defense.

Purpose of the Study:

• To develop a novel acoustic signal classification framework for munition detection.
• To exploit correlations in multisensor acoustic data for enhanced identification.
• To improve robustness to signal distortions and reduce sensitivity to training data limitations.

Main Methods:

• Proposed a probabilistic graphical model to learn class conditional correlations.
• Utilized cepstral features and symbolic dynamic filtering-based features.
• Employed multisensor acoustic data for classification experiments.

Main Results:

• The proposed algorithm significantly outperforms conventional classifiers.
• Demonstrated superior performance compared to joint sparsity models for multisensor classification.
• Showcased improved robustness to signal distortions and reduced training sample requirements.

Conclusions:

• The developed probabilistic graphical model framework effectively classifies acoustic signals from multisensor data.
• The method offers a robust and data-efficient approach for identifying munition launches.
• This technique advances acoustic sensing capabilities for defense and security applications.