In the process of collecting laser self-mixing interference signal, it is interfered by environment and circuit noise, resulting in signal distortion. In order to remove the noise and preserve the original signal features to the maximum extent, a self-mixing interference filtering method based on deep learning is proposed, which is suitable for weak feedback conditions. An autoencoder is used as a neural network, and a noisy signal is used as input and an unnoisy signal as output to train the network. The simulation results show that this method can not only improve the signal-to-noise ratio of the noisy self-mixing interference signal, but also preserve the waveform characteristics of the interference fringe, namely, the inclination direction of the fringe. In the experiment, the deep learning method is used to filter, and then the fringe counting method is used to reconstruct the displacement. The results show that this method has a good filtering effect on the self-mixing interference signal under the weak feedback condition.