Abstract: The nuclear detector is a key component in the reliability of nuclear radiation monitoring. It faces possible accident that involves high risks due to frequent operation in a radiation environment, which seriously affects the measurement accuracy. To ensure the detector reliability, a method for diagnosing scintillator detector fault based on the Knearest neighbor (KNN) algorithm was proposed in this paper. The key points in a pattern recognition are the extraction of nuclear detector fault characteristic signal and how to establish fault diagnosis models. In order to accurately analyze the output signal and improve the signaltonoise ratio, a method based on timedomain statistical analysis to extract characteristics of the detector output signal was used. This method extracted partial statistical features from output pulse waveform, which includes amplitude and falling edge time, and the residual statistical ones like a peak channel address and total count of low channel address are from the energy spectrum. A fault diagnosis method of nuclear detector based on KNN algorithm was proposed. After the feature extractions by the timedomain statistical analysis method as mentioned above, a feature reference sample set of detector normal signal and fault signal was established in this method. The similarity function was optimized by adding the weighted value of feature attribute to improve the classification performance of KNN algorithm fault diagnosis model for different typical faults. And a system was designed for collecting and validating fault data, extracting the performance information of the detector output signal, and was put into a fault diagnosis model for diagnostic experiments. The diagnostic accuracy of this method for the fault category is 100%, and the diagnostic accuracy of the degree of failure is more than 92.5%. The method is fast, intuitive, and insensitive to abnormal values, and can analyze nuclear pulse signal and energy spectrum signal at the same time. The research results show that this method not only allows intelligent diagnostics of the detector malfunction category, but also distinguishes well the severity of various malfunctions. Finally, the application comparison of this method with SVM and BP neural network in nuclear detector fault diagnosis was verified by experiments. In practical application, these methods can be selected corresponding to various needs to fulfill the purpose of automatic and intelligent fault diagnosis, identification and positioning, and provide a theoretical basis and a technical reference to some extent for the application of digital nuclear instrument in the field of intelligent fault detection.