Noise-robust acoustic signature recognition using nonlinear Hebbian learning.

Area of Science:

• Signal Processing
• Machine Learning
• Bio-inspired Computing

Background:

• Acoustic signal recognition in noisy environments is challenging due to overlapping signals and unknown noise characteristics.
• Conventional methods like Mel-frequency cepstral computation (MFCC) struggle with high noise levels and complex acoustic scenes.
• Existing Hebbian learning approaches, such as linear Hebbian learning (LHL), primarily utilize second-order statistics, limiting their ability to capture complex feature dependencies.

Purpose of the Study:

• To propose and evaluate a novel biologically inspired approach, nonlinear Hebbian learning (NHL), for enhanced acoustic signal recognition.
• To demonstrate NHL's capability in extracting statistically independent features from spectro-temporal representations (STRs) for improved signal detection and identification.
• To validate the system's robustness and real-time performance in practical, noisy acoustic environments.

Main Methods:

• Utilized auditory gammatone filterbanks for spectral analysis and analyzed filtered feature vectors over multiple temporal frames to create spectro-temporal representations (STRs).
• Employed nonlinear Hebbian learning (NHL) to extract representative, statistically independent features (signatures) from STRs, reducing dimensionality and attenuating noise.
• Compared NHL's performance against linear Hebbian learning (LHL) and Mel-frequency cepstral computation (MFCC) using metrics like error rate and improvement percentages.

Main Results:

• The proposed NHL system significantly decreased error rates compared to MFCC (up to 60.3%) and LHL (up to 20%) in detecting vehicle sounds contaminated by various noises at SNR=0 dB.
• Achieved substantial performance improvements over LHL in vehicle type identification, demonstrating a 40% enhancement for gasoline heavy wheeled cars in AWGN at SNR=5 dB.
• Real-time field testing over months showed a low missing rate (1-2%) and false alarm rate (around 1%) for detecting moving vehicles amidst diverse environmental noises (SNR=0-20 dB).

Conclusions:

• Nonlinear Hebbian learning (NHL) provides an efficient and robust method for extracting representative independent features from high-dimensional acoustic data.
• The NHL-based system exhibits superior performance and robustness against severe noise compared to conventional methods, making it suitable for real-world applications.
• This approach offers a promising solution for acoustic signal recognition tasks requiring high accuracy and reliability in challenging, noisy conditions.