Power spectral density estimation via universal truncated order statistics filtering.

Area of Science:

• Signal Processing
• Underwater Acoustics
• Data Analysis

Background:

• Loud transient signals in underwater acoustic data corrupt background noise power spectral density (PSD) estimates.
• Existing order statistics filters (OSFs) require careful, static rank selection to mitigate transient impacts.
• Dynamic environments necessitate adaptive OSF rank adjustments for real-time applications to maintain low bias and variance.

Purpose of the Study:

• To develop a novel method for mitigating the impact of loud transients on PSD estimates in underwater acoustics.
• To address the limitations of static rank selection in existing OSFs for real-time, dynamic environments.
• To propose an adaptive OSF approach that automatically adjusts to changing transient rates.

Main Methods:

• Proposed a convex sum of order statistics filter (OSF) ranks with dynamically adjusted blending weights.
• The blending weights are sequentially optimized to favor OSF ranks with the lowest variance over a recent time window.
• Evaluated the performance using simulations and real underwater acoustic data.

Main Results:

• The proposed blended OSF provably approaches the performance of the optimal fixed-rank OSF.
• Demonstrated effective filtering of loud transients from spectrograms without explicit threshold rank selection.
• Confirmed the method's ability to maintain low bias and variance in PSD estimates.

Conclusions:

• The blended order statistics filter offers an effective and adaptive solution for transient signal removal in underwater acoustics.
• This approach overcomes the challenges of real-time rank selection in dynamic acoustic environments.
• The method provides robust and reliable power spectral density estimation.