A hybrid neural network and attention mechanism (Att-CNN-GRU) is presented to solve the problems of fast attenuation,strong interference,ambiguous disturbance characteristics and ineffective signal detection in magnetic field proximity detection of underwater targets.A method for detecting time series disturbance signal of underwater target with power frequency magnetic field is presented.The method combines CNN,GRU neural network and Attention mechanism to fit the signal,and constructs a classification neural network to classify and identify the target signal.The method is compared with the prediction and detection performance of CNN-LSTM model without attention mechanism and single CNN and LSTM network model.The results show that the error of signal fitting is reduced by [BF]36.24%[BFQ],[BF]14.44%[BFQ],[BF]4.878%[BFQ] and the target detection accuracy is [BF]83.3%[BFQ] compared with the traditional methods.Therefore,the CNN-GRU model with Attention mechanism has better performance than CNN,LSTM and CNN-GRU models.As an auxiliary means,it can effectively solve the problems of weak disturbance signal,unclear disturbance law and more background noise in power frequency magnetic field detection,to realize the fitting and detection of power frequency magnetic disturbance signal to underwater target.