用卷积循环网络提前预测噪声，实现更快的主动降噪响应。
Predictive Fixed-Filter Active Noise Control (PFANC) Using Convolutional Recurrent Neural Networks for Dynamic Noises
- 用多帧噪声数据训练CRNN模型，提前预测下一帧控制滤波器。
- 在真实动态噪声下，降噪性能比传统方法提升1.2~2.3 dB。
- 适合需要快速响应的工业降噪场景，如机械振动、车载环境。
现有生成式固定滤波主动降噪（GFANC）方法基于当前噪声帧生成控制滤波器，属于反应式设计，仅优化当前帧而无法预判未来变化，因而存在固有追踪延迟，难以应对快速变化的噪声。为此，本文提出一种前瞻式主动降噪方法——预测型固定滤波主动降噪（PFANC），利用卷积循环神经网络（CRNN）处理多个连续噪声帧，提前预测下一帧的控制滤波器。通过捕捉噪声帧间的时序相关性，实现对动态噪声变化的预判与快速跟踪。基于高阶马尔可夫链的理论分析表明，使用多帧信息能显著提升滤波器预测精度。数值仿真采用线性与对数啁啾信号及真实动态噪声，验证了PFANC方法的有效性及其相比GFANC及其变体的优越性。此外，该方法在不同声学路径间具有良好的迁移能力。
原文摘要 · Abstract (English)
The existing Generative Fixed-Filter Active Noise Control (GFANC) method generates a suitable control filter based on the current noise frame. This reactive design aims to estimate a control filter that is optimal for the present frame rather than the upcoming one. Consequently, it suffers from an inherent tracking lag and lacks the predictive capability to handle rapidly varying noises. To address this limitation, we propose the Predictive Fixed-Filter Active Noise Control (PFANC) method with a proactive control paradigm in this paper. In the PFANC method, multiple consecutive noise frames are processed by a Convolutional Recurrent Neural Network (CRNN) to predict the next-frame control filter. By utilizing temporal correlations across noise frames to anticipate the control filter in advance, the PFANC method can effectively track dynamic noise changes. Furthermore, the theoretical analysis based on a high-order Markov chain shows that incorporating multiple noise frames enhances the prediction of the control filter. Numerical simulations with linear and logarithmic chirp signals, as well as real-world dynamic noises, validate the effectiveness of the PFANC method and its superiority over GFANC and its variations. The PFANC method also exhibits good transferability across different acoustic paths.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。