Updating the secondary path using deep learning models to enhance the performance of active road noise control

Abstract

This research investigates the importance of accurately updating the secondary path estimation to enhance the performance of active road noise control (ARNC), particularly in response to variable conditions within a vehicle. The study analyzes the impact of four spatial conditions on the secondary path and introduces prediction models using Delaunay triangulation-based interpolation and deep learning methods, using multi-layer perceptron (MLP) networks. MLP networks were constructed based on the representation domain chosen for the secondary path. The deep learning methods demonstrated higher prediction accuracy and smaller data storage requirements compared to the interpolation method. Among these deep learning methods, the approach that stood out as the best performer was representing the secondary path through principal component analysis (PCA) and learning the weight of each basis vector. Through this technique, it was experimentally confirmed that the performance of ARNC under variable conditions in the vehicle improved through the accurate update of real-time secondary path estimation. These advantages are particularly prominent when dealing with high frequencies, which are likely to be the foundational technology for the frequency expansion of ARNC.