Improved Non-Negative Matrix Factorization-Based Noise Reduction of Leakage Acoustic Signals

Yongsheng Yu, Yongwen Hu, Yingming Wang, Zhuoran Cai2024 · DOI

TL;DR

This paper introduces an improved Non-Negative Matrix Factorization (INMF) algorithm for denoising acoustic signals from gas leaks. By refining the objective function with Kullback-Leibler divergence and incorporating adaptive spectral estimation, the INMF method significantly outperforms existing techniques in challenging noisy environments.

What Problem Does This Paper Address?

Detecting gas leaks using acoustic signals is crucial for safety and environmental protection, especially for potent greenhouse gases like SF6. However, environmental noise frequently contaminates these acoustic signals, making accurate identification difficult and leading to potential false positives or negatives. Traditional noise reduction methods often struggle with non-stationary noise and can even amplify noise anomalies. Existing Non-Negative Matrix Factorization (NMF) based approaches, while promising, typically use Euclidean distance as their objective function, which can be sensitive to noise spikes, and rely on simpler spectral estimation methods like Wiener filtering, which are less effective for dynamic, non-stationary signals. This paper aims to address these limitations by developing a more robust and adaptive noise reduction technique for leakage acoustic signals.

Methods

The core of this work is the development of an Improved Non-Negative Matrix Factorization (INMF) algorithm. Key methodological advancements include:

  • Refined Objective Function: The traditional Euclidean distance is replaced with Kullback-Leibler (KL) divergence. This is further enhanced by incorporating:
  • A noise term (E) to mitigate the impact of anomalies.
  • A sparse constraint term on the noise term (using L1 regularization) to encourage sparsity in the noise component.
  • A sparsity constraint term on the coefficient matrix (H) to improve signal reconstruction and control audio distortion.
  • Adaptive Spectral Estimation: Instead of traditional methods like Wiener filtering, the paper employs an adaptive Minimum Mean-Square Error-Log Spectral Amplitude (MMSE-LSA) algorithm. This method dynamically adjusts its parameters to estimate the amplitude spectrum of non-stationary signals, adapting to varying signal-to-noise ratios (SNRs).
  • Two-Phase Approach: The INMF algorithm operates in two main phases:
  • Training Phase: A "pure" air leakage signal and a "noise" signal are processed using Short-Time Fourier Transform (STFT) to obtain their amplitude spectra. These are then used to train the INMF algorithm to decompose the noisy signal into its constituent pure signal and noise components, generating "pure" and "noise" dictionaries.
  • Noise Reduction Phase: The noisy audio signal's amplitude spectrum is obtained via STFT. The pre-trained joint dictionary is then used with the INMF algorithm to estimate the pure signal and noise spectra. Finally, the adaptive MMSE-LSA algorithm is applied to refine the amplitude spectrum, and the inverse STFT (ISTFT) reconstructs the denoised signal.

Datasets

The experiments were conducted using:

  • Simulated Gas Leakage Audio: Recorded in a semi-anechoic chamber.
  • Typical Disturbance Noises: Sourced from a field substation environment, including wind, rain, birdsong, and cicada chirping.
  • Synthesized Noisy Data: The pure leakage audio was mixed with noise at various signal-to-noise ratios (-10 dB, -5 dB, 0 dB, 5 dB, and 10 dB) to create realistic test scenarios.

The paper does not introduce a new benchmark dataset but rather uses a combination of real and synthesized data for comprehensive evaluation.

Key Insights and Ideas

  • KL Divergence for Robustness: Shifting from Euclidean distance to KL divergence in the objective function is crucial for handling the inherent anomalies and non-Gaussian nature of noise in acoustic signals, preventing noise amplification.
  • Sparsity is Key: Incorporating sparsity constraints on both the noise term and the coefficient matrix significantly improves the algorithm's ability to disentangle the desired signal from noise and control the fidelity of the reconstructed signal.
  • Adaptability is Essential: The adaptive MMSE-LSA method is a critical component, allowing the algorithm to effectively estimate spectral amplitudes even when the noise characteristics change dynamically, which is common in real-world environments.
  • NMF's Strength in Non-Stationary Noise: The paper highlights that NMF-based methods, particularly the proposed INMF, excel at denoising non-stationary noises (like birdsong) compared to methods like Variational Mode Decomposition (VMD), by effectively leveraging training samples and learned dictionary representations.

Key Results

  • Superior Noise Reduction: The proposed INMF algorithm consistently demonstrated superior noise reduction performance across various background noises and signal-to-noise ratios compared to traditional methods (Wiener filter) and other NMF variants (RNMF, SRNMF, HINMF).
  • Quantitative Improvements:
    • SDR (Source-to-Distortion Ratio): The INMF algorithm achieved higher SDR values, indicating less residual noise and better signal preservation. For example, at an input SNR of 0 dB with birdsong noise, INMF achieved an SDR of 6.82 dB, significantly better than VMD (-0.05 dB) and RNMF (0.25 dB). At 10 dB SNR with cicada noise, INMF achieved 12.85 dB, outperforming VMD (10.25 dB) and RNMF (10.23 dB).
  • PESQ (Perceptual Evaluation of Speech Quality): Higher PESQ scores were observed with INMF, signifying enhanced signal quality after noise reduction. For instance, at -10 dB SNR with cicada noise, INMF achieved a PESQ of 1.50, compared to VMD's 1.09 and RNMF's 1.13.
  • Qualitative Improvements: Time-frequency diagrams (Figure 4) visually confirm that INMF effectively suppresses non-stationary noise, leaving the target leakage signal more clearly defined with less residual noise compared to other methods.

Limitations and Future Work

  • Computational Cost: The authors acknowledge that the complexity of the INMF algorithm leads to a higher computational cost, which might limit its suitability for real-time applications with constrained processing power.
  • Real-World Validation: The current study primarily used synthetic noise data for evaluation. The authors suggest that further validation is needed with diverse real-world noise environments to fully assess the algorithm's effectiveness.
  • Generalizability: While effective for leakage acoustic signals, the paper could benefit from broader testing on other types of acoustic signals to understand its generalizability.

Future work should focus on optimizing the algorithm for real-time performance, conducting extensive testing with diverse real-world acoustic data, and further analyzing the computational complexity.

Summary updated

BibTeX
@misc{yu2024improved,
  author = {Yongsheng Yu and Yongwen Hu and Yingming Wang and Zhuoran Cai},
  title = {{Improved Non-Negative Matrix Factorization-Based Noise Reduction of Leakage Acoustic Signals}},
  year = {2024},
  doi = {10.3390/s24165146},
}
Figure