For the physical layer security communication system based on artificial noise, traditional artificial noise is usually generated by using closed-form expressions derived from derivation or numerical optimization methods which both require accurate channel state information matrix to guarantee the secrecy of the communication system. However, the channel estimation error in the real scenarios causes the artificial noise precoding error to reduce the security capacity of the communication system. For this reason, this paper proposes an artificial noise precoding generation method based on deep learning. By taking the channel estimation information with estimation error as input and fitting it with the precoding matrix obtained by traditional numerical solution generated by perfect channel estimation, a well-trained deep neural network that can adapt to the channel estimation error is obtained. Simulation shows that the security performance and robustness of this method when there are errors in channel estimation are better than traditional artificial noise generation systems. Compared with other deep learning methods for physical layer security, the method proposed in this paper has faster convergence speed.