Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised learning based end-to-end delayless generative fixed-filter active noise control

Zhengding Luo, Dongyuan Shi, Xiaoyi Shen, Woon-Seng Gan

TL;DR

The paper tackles delayless fixed-filter ANC by removing the need for labelled data through an end-to-end unsupervised training of the 1D CNN within the GFANC framework, using the accumulated squared error loss $\mathcal{L}=\sum_{n=1}^{F_s} e^2(n)$ with $e(n)=d(n)-\mathbf{x}'^{T}(n)\mathbf{W}\mathbf{x}'(n)$ and $\mathbf{W}=\mathbf{g}\cdot\mathbf{S}_c$. By integrating the co-processor and real-time controller into a differentiable ANC system, the 1D CNN learns to produce a weight vector $\mathbf{g}$ that forms the control filter via a sub-filter bank $\mathbf{S}_c$ (i.e., $\mathbf{W}=\mathbf{g}\cdot\mathbf{S}_c$) without noise labels. Empirical results on synthetic and real-world noises show that unsupervised-GFANC achieves superior NMSE and noise reduction compared to supervised GFANC and FxLMS, and that training on synthetic paths transfers effectively to measured paths, reducing labelling costs and biases. The work demonstrates practical delayless ANC with good transferability and suggests applicability to other deep-learning-based ANC methods.

Abstract

Delayless noise control is achieved by our earlier generative fixed-filter active noise control (GFANC) framework through efficient coordination between the co-processor and real-time controller. However, the one-dimensional convolutional neural network (1D CNN) in the co-processor requires initial training using labelled noise datasets. Labelling noise data can be resource-intensive and may introduce some biases. In this paper, we propose an unsupervised-GFANC approach to simplify the 1D CNN training process and enhance its practicality. During training, the co-processor and real-time controller are integrated into an end-to-end differentiable ANC system. This enables us to use the accumulated squared error signal as the loss for training the 1D CNN. With this unsupervised learning paradigm, the unsupervised-GFANC method not only omits the labelling process but also exhibits better noise reduction performance compared to the supervised GFANC method in real noise experiments.

Unsupervised learning based end-to-end delayless generative fixed-filter active noise control

TL;DR

The paper tackles delayless fixed-filter ANC by removing the need for labelled data through an end-to-end unsupervised training of the 1D CNN within the GFANC framework, using the accumulated squared error loss with and . By integrating the co-processor and real-time controller into a differentiable ANC system, the 1D CNN learns to produce a weight vector that forms the control filter via a sub-filter bank (i.e., ) without noise labels. Empirical results on synthetic and real-world noises show that unsupervised-GFANC achieves superior NMSE and noise reduction compared to supervised GFANC and FxLMS, and that training on synthetic paths transfers effectively to measured paths, reducing labelling costs and biases. The work demonstrates practical delayless ANC with good transferability and suggests applicability to other deep-learning-based ANC methods.

Abstract

Delayless noise control is achieved by our earlier generative fixed-filter active noise control (GFANC) framework through efficient coordination between the co-processor and real-time controller. However, the one-dimensional convolutional neural network (1D CNN) in the co-processor requires initial training using labelled noise datasets. Labelling noise data can be resource-intensive and may introduce some biases. In this paper, we propose an unsupervised-GFANC approach to simplify the 1D CNN training process and enhance its practicality. During training, the co-processor and real-time controller are integrated into an end-to-end differentiable ANC system. This enables us to use the accumulated squared error signal as the loss for training the 1D CNN. With this unsupervised learning paradigm, the unsupervised-GFANC method not only omits the labelling process but also exhibits better noise reduction performance compared to the supervised GFANC method in real noise experiments.
Paper Structure (9 sections, 6 equations, 7 figures, 2 tables)

This paper contains 9 sections, 6 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The Unsupervised-GFANC Approach
  3. The 1D CNN
  4. Unsupervised Training
  5. Comparing Unsupervised and Supervised Training
  6. Numerical Simulations
  7. Real-world noise cancellation on measured paths
  8. Analysis of Power Spectral Density
  9. Conclusion

Figures (7)

  • Figure 1: Framework of the unsupervised-GFANC method. After unsupervised training, the 1D CNN collaborates with the real-time controller for delayless noise control.
  • Figure 2: Block diagram of training the 1D CNN in the unsupervised-GFANC approach, which is an end-to-end differentiable ANC system. Given each noise frame, the accumulated squared error signal is used as the training loss.
  • Figure 3: Block diagram of the labelling mechanism in the previous supervised GFANC method.
  • Figure 4: The magnitude responses and phase responses of the measured primary and secondary paths.
  • Figure 5: (a)-(c): Noise reduction results of different ANC algorithms, (d): Averaged noise reduction level in each second, on the aircraft noise.
  • ...and 2 more figures