Terahertz wireless communication has been regarded as an emerging technology to satisfy the ever-increasing demand of ultra-high-speed wireless communications.However,affected by the imperfec-tions of cheap and energy-efficient Terahertz devices,Terahertz signals suffer from serve hybrid distortions,including in-phase/quadrature imbalance,phase noise and nonlinearity,which degrade the demodulation per-formance significantly.To improve the robustness against these hybrid distortions,an improved autoen-coder is proposed,which includes coding the transmit-ted symbols at the transmitter and decoding the corre-sponding signals at the receiver.Moreover,due to the lack of information of Terahertz channel during the training of the autoencoder,a fitting network is pro-posed to approximate the characteristics of Terahertz channel,which provides an approximation of the gra-dients of loss.Simulation results show that our pro-posed autoencoder with fitting network can recover the transmitted symbols under serious hybrid distortions,and improves the demodulation performance signifi-cantly.