In channel estimation tasks for intelligent reflecting surface （IRS）-assisted communication systems when line-of-sight communication between user equipment and base station （BS） is blocked， this paper proposes a neural network based on noise suppression in the latent feature space， which can realize accurate channel estimation. The neural network combines the variational auto-encoder （VAE） and UNet to reduce the noise in the input signal while performing channel estimation. Firstly， the VAE model takes noise-free BS received signals as input， with the objective of minimizing the error between the estimated noise-free BS received signals and their true value， so that the encoder of the VAE model maps a feature vector as a potential representation of the pure received signal. Secondly， the VAE model part is fixed. The entire network is trained using noisy BS received signals as input to the UNet model， in which the noise-free latent feature vectors learned by the VAE assist the encoder of the UNet model in learning noise-free feature representations. Subsequently， the pure feature representations are fed into the decoder of the UNet model to achieve the channel estimation task. Finally， during the estimation phase， only the UNet model part is utilized， which effectively reduces computational complexity. The results of simulation experiments demonstrate that the proposed channel estimation method can effectively suppress noisy information in the feature space， and can estimate the channel information more accurately with lower time complexity.