Abstract: Both seismograph and deformation stations recorded low frequency pulselike signals before the Wenchuan earthquake. In order to further analyze and judge the authenticity of such pulse signals, one earthquake pulselike records detection method based on deep learning is proposed. Take the Wenchuan earthquake as an example, firstly, the continuous waveform data of 9 days before the Wenchuan earthquake in 49 stations in Sichuan province are collected and are taken as training data set. The two dimensional time ‐ spectrum of the waveform data are calculated by continuous wavelet transform method. Then, based on the deep autoencoder neural network and time‐spectrum of the waveform data, the network model of pulselike records recognition is obtained and used for automatic identification of earthquake pulselike signals. Experimental results show that our deep autoencoder network model has good stability and the recognition percentage for new data is over 93%. Finally, the spatial distribution of pulselike signals in Sich‐uan Province one month before the Wenchuan earthquake is preliminarily calculated, and a possible explanation is given from the perspective of slow slip motion of the faults.