Aiming at the system instability caused by the DC offset of the sensors in the road noise active control system， a new modified multichannel adaptive filtering road noise active control algorithm is established. Using multi-coherence as the evaluation function， a comprehensive evaluation index is constructed， and 4 acceleration signals are selected as reference signals using a multi-island genetic algorithm from 24 acceleration signals measured under real vehicle road tests. With two car speakers as secondary sound sources and one headrest position in the front row as control targets， a multi-channel in-car road noise active control system with eight reference signals， two speakers， and two microphones are built in the Simulink environment. Simulations are performed using data collected at different vehicle speeds and different pavement （Belgian pavement， Rough asphalt pavement）. On the basis of simulation model， a corresponding hardware-in-the-loop test platform for the active noise control system is built based on dSPACE. The experimental results are consistent with the simulation results under various operating conditions， both of which can achieve stable and effective noise reduction.