Abstract: An improved de-noising auto-encoder is proposed to improve the recognition success rate of transmission characteristic in noisy power line channel. The proposed scheme uses one-dimensional time series instead of two-dimensional images as input, which improves the processing data network structure of traditional self-encoder. The introduction of Z-score standardization and corresponding anti-standardization to the input and output signal processing can improve the de-noising ability and speed up the convergence speed. One 4 layers neural network with two hidden layers is selected to verify the effectiveness of the proposed method from two aspects of software model and hardware implementation.