End-to-end virtual sensing with head tracking using a temporal convolutional network for active road noise control

Head-tracking active road noise control requires accurate virtual sensing of sound pressure at the occupants’ ears. Conventional approaches based on the remote microphone method (RMM) remain limited because the use of one observation filter per head position does not account for dynamic driving conditions and provides insufficient information for robust model training. This study proposed an end-to-end virtual sensing framework with head tracking that uses a temporal convolutional network (TCN) conditioned on the head-position information via feature-wise linear modulation (FiLM). The proposed model predicts time-domain virtual microphone signals from eight-channel physical microphone measurements and real-time head positions. This formulation enables the segmentation of long recordings into numerous overlapping input–output pairs, thereby providing richer supervision and allowing the model to learn temporal variations in road noise within the tested speed range. Experimental results indicated that the proposed method significantly outperformed the existing learning-based approach that directly updates the observation filters, reducing the virtual sensing error from − 14.79 to − 19.05 dB in terms of NMSE and from − 7.35 to − 12.34 dBA in terms of A-weighted spectral error. Moreover, the training-data duration per head-pose configuration could be reduced without a considerable loss in accuracy when sufficient temporal variability was preserved. Furthermore, the encoder–decoder structure effectively reduced the computational cost while maintaining comparable frequency-domain performance. Active noise control (ANC) simulations demonstrated that the improved virtual sensing accuracy resulted in enhanced control performance, and the proposed method achieved ANC results close to those obtained using actual measured signals without virtual sensing.