Blast noise classification with common sound level meter metrics.

Area of Science:

• Acoustics and Signal Processing
• Machine Learning Applications
• Defense and Security Technologies

Background:

• Distinguishing military blast sounds from ambient noise is crucial for situational awareness and security.
• Existing methods may lack the precision needed for reliable acoustic event detection in complex environments.
• Acoustic signal analysis offers a non-intrusive method for identifying specific sound events.

Purpose of the Study:

• To analyze common signal features measurable by sound level meters for discriminating military blast sounds.
• To evaluate the information quality within feature subsets for accurate blast sound classification.
• To implement and compare machine learning classifiers for blast sound identification.

Main Methods:

• Analysis of over 120,000 human-classified signals from seven diverse datasets.
• Implementation of linear and Gaussian radial basis function (RBF) support vector machines (SVM).
• Application of orthogonal centroid dimension reduction and recursive feature elimination (SVM-RFE) for feature selection and ranking.

Main Results:

• Identified specific sound features with high discriminatory power for military blast sounds.
• Recursive feature elimination effectively reduced feature redundancy and ranked features by classification importance.
• Achieved accurate classification of blast sounds using both linear and RBF SVM models across multiple datasets.

Conclusions:

• A defined set of acoustic features can reliably differentiate military blast sounds from non-blast sounds.
• Support vector machine classifiers, particularly with optimized feature sets, are effective tools for acoustic blast detection.
• The findings contribute to developing more robust systems for acoustic monitoring and threat identification.