Abstract: Nuclear detectors play important roles in radioactive monitoring. In order to keep nuclear detectors stable, an on-line intelligent fault diagnosis method based on BP neural network was proposed for scintillation detectors. By transforming detector’s output signals to frequency domain from time domain, characteristic vectors were obtained from wavelet packet transform, then these vectors were treated as input of BP neural network fault diagnosis model, and the parameters of fault diagnosis model were optimized by error gradient descent method. Finally, the optimal fault diagnosis model was employed to identify and classify fault types intelligently, which was also compared with a statistical model and another models based on support vector machine. Experimental results show that the outcome of proposed method is more accurate than the two methods above. Therefore, an application of this method can effectively improve the accuracy of nuclear detector fault diagnosis.