In order to identify the tilt direction of the self-mixing signals under weak feedback regime interfered by noise, a deep learning method is proposed. The one-dimensional U-Net (1D U-Net) neural network can identify the direction of the self-mixing fringes accurately and quickly. In the process of measurement, the measurement signal can be normalized and then the neural network can be used to discriminate the direction. Simulation and experimental results show that the proposed method is suitable for self-mixing interference signals with noise in the whole weak feedback regime, and can maintain a high discrimination accuracy for signals interfered by 5 dB large noise. Combined with fringe counting method, accurate and rapid displacement reconstruction can be realized.