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

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

TL;DR

This paper introduces unsupervised-GFANC, a novel framework for Active Noise Control (ANC) that eliminates the need for expensive, manually labeled datasets. By integrating a 1D CNN co-processor with a real-time controller into an end-to-end differentiable system, the model learns to minimize noise directly from the error signal, outperforming previous supervised methods in real-world scenarios.

What Problem Does This Paper Address?

Active Noise Control (ANC) works by generating "anti-noise"—a sound wave with the same amplitude but opposite phase to the target noise—to cancel it out through destructive interference.

Traditional adaptive algorithms like FxLMS are computationally efficient but often suffer from slow convergence and potential instability. To solve this, researchers previously developed Generative Fixed-Filter ANC (GFANC), which uses a neural network to predict the best combination of pre-defined filters for a given noise. However, GFANC traditionally relied on supervised learning, requiring "labels" (the ideal filter weights) for every noise sample. Generating these labels is resource-intensive, slow, and often introduces biases from the labeling algorithm itself.

Methods

The authors propose an unsupervised learning paradigm that treats the entire ANC process as a single, differentiable pipeline.

  1. The 1D CNN Co-processor: A lightweight convolutional network takes a frame of noise and outputs "soft weights" (values between 0 and 1).
  2. Filter Generation: These weights are used to combine a bank of "sub-control filters" into a single, optimized control filter ($W$).
  3. End-to-End Differentiability: Unlike previous versions where the CNN was trained separately, here the CNN and the physical noise-cancellation process are linked. The "Loss Function" is simply the accumulated squared error signal (the residual noise).
  4. Backpropagation through Physics: Because the relationship between the CNN's weights and the final error signal is mathematically defined and differentiable, the system can use standard backpropagation to update the CNN. In essence, the residual noise is the teacher.

Datasets

  • Training: A synthetic dataset of 80,000 noise instances created by passing white noise through bandpass filters with random center frequencies and bandwidths.
  • Evaluation: Real-world recordings of aircraft noise and drill noise.
  • Acoustic Paths: The primary and secondary acoustic paths (how sound travels through the physical environment) were measured from a real noise chamber’s vent to ensure realistic simulation.

Key Insights and Ideas

  • The "Error is the Label": The most elegant shift here is moving away from trying to guess what the "correct" weights are. By making the system end-to-end, the physical outcome (silence) becomes the direct optimization target.
  • Eliminating Algorithmic Bias: Supervised GFANC usually gets its labels from an LMS algorithm. If the LMS algorithm converges slowly or poorly, the CNN learns those same flaws. Unsupervised learning bypasses this "middleman."
  • Zero-Latency Coordination: The system maintains a "delayless" architecture by having the co-processor update filter weights at a frame-based rate while the controller processes audio at the sampling rate.

Key Results

The unsupervised approach didn't just simplify training; it actually worked better. The researchers measured performance using Normalized Mean Squared Error (NMSE):

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
  • Superior Attenuation: Unsupervised-GFANC provided roughly 1 dB better noise reduction than the supervised version and was significantly more effective than traditional FxLMS.
  • Spectral Performance: Power Spectral Density (PSD) analysis showed that the method is particularly effective at attenuating noise components below 2,000 Hz, where passive noise control (like earplugs or foam) typically fails.

Limitations and Future Work

  • Initial Startup: Like many fixed-filter generative methods, there is no noise cancellation in the very first second of operation because the initial filter coefficients are set to zero.
  • Synthetic-to-Real Gap: While the model transferred well to real noise, the training still relies on synthetic bandpass noise. Future work could explore training on more diverse, non-stationary synthetic datasets.
  • Scaling: The authors suggest this unsupervised, end-to-end differentiable approach could be applied to other, more complex deep learning-based ANC architectures beyond 1D CNNs.

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},
}

Loading related content...

Figure