As on-chip computing power advances, it has allowed active noise control systems to cover larger regions as a zone of quiet with the use of a multi-input and output (MIMO) system. With the impracticality of having unlimited number of physical error microphones, virtual sensing techniques have been recently explored, but they also suffer from performance degradation in time-varying acoustic environments. Therefore, in this work, the application of a machine learning technique in virtual sensing to ensure robustness in time-varying environments is proposed. A data-driven machine learning model was designed and trained with the input of the cross-spectral matrix of acoustic signals from microphones at the reference points to predict the frequency response functions between reference points and target locations for virtual sensing. Training data was acquired using the Head and Torso simulator (HATS), placed between the noise source and the reference microphone arrays, with the microphones at the ear locations of HATS serving as target virtual sensing points. Measurements were conducted with different angles and locations of the HATS. The results show that the model can successfully predict the frequency response with an error that was considerably lower than the standard deviation of the frequency response from the measurements.
Kim et al. (Wed,) studied this question.