Aiming at the problem of inaccurate recognition of current communication signals in low signal-to-noise ratio (SNR), a recognition algorithm combining wavelet feature and depth neural network is proposed. This method generates 10 kinds of common communication signals with Gauss white noise{MASK, MPSK, MFSK, OFDM, 16QAM, AM, FM}. A new kind of wavelet characteristic parameters are extracted from the signals by using the wavelet decomposition and reconstruction algorithm. The improved BP neural network with plenty hidden layers is studied and tested as classifier. The parameters of the neural network are trained by the elastic back propagation algorithm. The optimal layers of the neural network are determined by the identification results. The simulation results show that the minimum recognition rate of single modulated signals is more than 95% and the average recognition rate is more than 98%, when the signal-to-noise ratio is as low as 0 dB, which greatly improves the recognition rate of standard recognition under low signal-to-noise ratio, thus proving the effectiveness and practicability of this method.