Unsupervised Learning Based End-to-End Delayless Generative Fixed-Filter Active Noise Control

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

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.

Introduction

Active Noise Control (ANC) operates on the principle of destructive interference, generating an "anti-noise" signal to cancel out unwanted primary noise. While traditional adaptive algorithms like Filtered-x Least Mean Square (FxLMS) are computationally efficient, they often struggle with slow convergence and stability issues in dynamic environments.

To address these challenges, the Generative Fixed-Filter ANC (GFANC) framework was developed. GFANC utilizes a neural network to predict optimal weights for a bank of pre-defined filters. However, the original GFANC relied on supervised learning, which requires extensive labeled datasets—ideal filter weights for every noise sample. This labeling process is not only resource-intensive but can also inherit biases from the algorithms used to generate the labels. This paper introduces unsupervised-GFANC, which eliminates the need for manual labeling by treating the entire ANC process as a differentiable pipeline.

Proposed Method: Unsupervised-GFANC

System Architecture

The framework consists of two main components: a lightweight 1D CNN co-processor and a real-time controller. The co-processor analyzes noise frames to output "soft weights" ($g_1, ..., g_M$), which are then used to linearly combine a set of sub-control filters into a single generated control filter ($W$). This architecture, shown in Figure 1, allows the system to update filter coefficients at a frame rate while the controller operates at the sampling rate, ensuring zero-latency performance.

The GFANC framework architecture featuring a 1D CNN co-processor and a real-time controller.

End-to-End Differentiable Training

The core innovation is the unsupervised training paradigm. By integrating the co-processor and the physical noise-cancellation process into a single differentiable pipeline, the system can be trained end-to-end. As illustrated in Figure 2, the loss function is defined as the accumulated squared error signal ($e(n)$), which represents the residual noise.

The proposed unsupervised training paradigm showing the end-to-end differentiable pipeline and loss function.

Because the relationship between the CNN's output and the final error signal is mathematically defined, standard backpropagation can be used to update the network weights. In this paradigm, the physical outcome—silence—serves as the direct optimization target, bypassing the need for intermediate algorithmic labels and their associated biases.

Experimental Results

The model was trained on a synthetic dataset of 80,000 noise instances and evaluated using real-world aircraft and drill noise recordings. The primary and secondary acoustic paths were measured from a real noise chamber to ensure realistic simulation.

Performance was measured using Normalized Mean Squared Error (NMSE). As shown in the table below, the unsupervised approach outperformed both the supervised GFANC and traditional FxLMS methods.

Method Aircraft Noise (dB) Drill Noise (dB)
Unsupervised-GFANC -12.65 -9.04
Supervised GFANC -11.68 -7.97
Traditional FxLMS -3.54 -4.32

Unsupervised-GFANC provided approximately 1 dB better noise reduction than the supervised version. Spectral analysis further confirmed its effectiveness in attenuating noise below 2,000 Hz, a frequency range where passive noise control methods typically underperform.

Conclusion

Unsupervised-GFANC simplifies the deployment of neural ANC systems by removing the labeling bottleneck. By optimizing directly for noise reduction in an end-to-end differentiable framework, the system achieves superior performance compared to supervised methods. Future work will focus on reducing the initial startup delay and exploring more diverse training datasets to further bridge the gap between synthetic training and real-world application.

Summary updated

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.
BibTeX
@inproceedings{luo2024unsupervised,
  author = {Zhengding Luo and Dongyuan Shi and Xiaoyi Shen and Woon‐Seng Gan},
  title = {{Unsupervised Learning Based End-to-End Delayless Generative Fixed-Filter Active Noise Control}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10448277},
}
Figure