The detection of weak signal submerged in chaos is discussed in this paper. In classical statistic detection theory the chaotic noise is regarded as a random signal, which will greatly reduce the performance of signal detection. To improve the performance based on chaotic dynamic mechanism, by using neural network to establish forecast model of chaotic time series and restructure its phase space, and based on the neural network's powerful ability of studying and nonlinear processing and local predictability of chaos, the method of neural network prediction and detection weak signal in chaotic time series is proposed, the results of theoretical analysis and simulation indicate that the method is effective.