A HYBRID LEARNING-BASED ACTIVE NOISE CONTROL ALGORITHM FOR ENCAPSULATED STRUCTURES WITH LINEAR TIME-VARYING CHARACTERISTICS

This paper proposes a hybrid approach that integrates deep learning with ANC algorithms to enhance noise attenuation in these environments. A simulated test environment is developed using the finite element method (FEM) to model an encapsulated linear time varying (LTV) system. This model enables the extraction of frequency responses from the reference, primary, and secondary paths, revealing non-minimum phase (NMP) characteristics that further complicate noise control due to phase delays and inverted dynamics. To address these challenges, a hybrid ANC framework is proposed, integrating a two-dimensional convolutional neural network (2D CNN) with selective fixed-filter ANC (SFANC) and a normalized least mean square (NLMS) algorithm. The 2D CNN utilizes spectrogram inputs to dynamically select the most suitable pre-trained control filters for varying primary noise conditions, while the NLMS algorithm continuously updates the filter coefficients in real time to adapt to the LTV system. The results demonstrate that the proposed hybrid approach outperforms conventional SFANC methods, achieving more effective noise reduction, improved robustness, and optimized response time.