Transferable Selective Virtual Sensing Active Noise Control Technique Based on Metric Learning

Boxiang Wang, Dongyuan Shi, Zhengding Luo, Xiaoyi Shen, Junwei Ji, Woon-Seng GanICASSP 2025 · DOI

TL;DR

This paper introduces a novel "Transferable Selective Virtual Sensing" (TSVS) technique for active noise control (ANC). By leveraging metric learning with a pre-trained convolutional neural network (CNN) feature extractor, TSVS allows ANC systems to adapt to new acoustic environments and noise types without requiring retraining, significantly improving efficiency and practicality.

What Problem Does This Paper Address?

Active noise control (ANC) systems aim to cancel unwanted noise by generating an anti-noise signal. However, traditional ANC systems are limited to the immediate vicinity of physical error microphones, creating a "zone of quietness" (ZoQ) that is often too small. Virtual sensing (VS) techniques extend this ZoQ to desired virtual locations, but conventional VS methods struggle when noise characteristics or acoustic paths change between training and deployment. While CNN-based VS approaches can adapt to these changes, they typically require extensive retraining for each new system, which is time-consuming and data-intensive, and can lead to overfitting. This paper addresses the challenge of creating a VS system that can generalize across different ANC setups and noise conditions without costly retraining.

Methods

The core of this work is the "Transferable Selective Virtual Sensing" (TSVS) technique. It builds upon existing CNN-based selective VS methods but introduces metric learning to enhance transferability.

Here's a breakdown:

  • Pre-trained CNN Feature Extractor: Instead of training a full CNN for each new VS system, TSVS utilizes the feature extraction layers of a pre-trained CNN. This module acts as a powerful encoder, transforming raw audio signals into meaningful "embedding codes."
  • Metric Learning for Similarity: The key innovation is using these embedding codes to measure the similarity between incoming noise signals and pre-defined noise classes (represented by auxiliary filters, AFs). This is done by calculating the cosine similarity between the embedding of the current noise and the embeddings of known AFs.
  • AF Selection: The AF whose corresponding noise embedding is most similar to the current noise's embedding is selected. This allows the system to dynamically choose the most appropriate AF for the current acoustic environment.
  • Transferability: Because the feature extraction module is pre-trained and not retrained, it can be directly applied to new VS systems. This means a CNN trained on one set of noise conditions can effectively classify and select AFs for entirely different noise types and acoustic paths in a new system.
  • Efficiency: The feature extraction module has significantly fewer parameters than a full CNN, making it computationally efficient and suitable for deployment on co-processors, reducing the burden on the real-time controller.

Datasets

The paper describes simulations using synthesized noise data.

  • VS System I: 15 broadband noises spanning the 20-2020 Hz spectrum were used to pre-train AFs. For training the CNN, 80,000 1-second noise waveforms were synthesized, with 1,000 for validation and 1,000 for testing.
  • VS System II: Five different broadband noises (20-160 Hz, 150-340 Hz, 330-770 Hz, 765-1275 Hz, and 1270-2020 Hz) were used to pre-train AFs. To test transferability, 1,000 synthesized noise samples not present in the initial training set were used for classification.
  • Real-world Noise: The paper also mentions using concatenated recordings of genset noise and compressor noise for evaluation, neither of which were in the training set.

Key Insights and Ideas

  • Metric Learning as a Transfer Bridge: The paper demonstrates that metric learning, when applied to CNN feature embeddings, can effectively bridge the gap between different VS systems and noise conditions, enabling true transferability.
  • Feature Extractor as a Universal Classifier: By isolating the feature extraction module of a CNN, the authors show it can act as a general-purpose noise classifier, adaptable to unseen noise types without retraining.
  • Computational Efficiency for Practicality: The significant reduction in parameters by using only the feature extraction module makes the approach practical for embedded systems and co-processors, a crucial step for real-world ANC applications.
  • Handling Unseen Noise: The method's ability to perform well even with noise classes not encountered during initial training is a major advantage over traditional selective VS techniques.

Key Results

  • High Accuracy with Fewer Parameters: The proposed TSVS method achieved higher noise classification accuracy in VS system II compared to several other 1D CNN models, while using significantly fewer parameters (12,955 vs. 2160 for the feature extractor alone).
  • Superior Noise Reduction: Numerical simulations showed that TSVS achieved noise reduction performance closely matching optimal control at the virtual microphone, outperforming both full-band AF and conventional selective VS techniques.
  • Effective on Varying and Real-world Noises: The technique demonstrated effectiveness in attenuating both sudden-varying broadband noises and real-world noises (genset and compressor noise) that were not part of the training set.

Limitations and Future Work

  • Limitations: The paper focuses on simulations. While real-world noise was used for evaluation, the primary validation is through simulated environments. The performance might be affected by more complex real-world acoustic phenomena not fully captured in the simulations.
  • Future Work: The authors suggest that the precomputed embedding codes for AFs can further reduce computational load. The efficiency and transferability of the proposed method make it a promising direction for developing more robust and adaptable ANC systems for diverse applications.

Summary updated

BibTeX
@inproceedings{wang2025transferable,
  author = {Boxiang Wang and Dongyuan Shi and Zhengding Luo and Xiaoyi Shen and Junwei Ji and Woon-Seng Gan},
  title = {{Transferable Selective Virtual Sensing Active Noise Control Technique Based on Metric Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10887958},
}
Figure