核脉冲波形甄别（PSD）在核物理实验粒子甄别（如中子和γ光子甄别）等中有重要应用。本文围绕PSD开展了中子和γ光子甄别，α和β粒子甄别以及核辐射探测器模拟电路混合故障智能诊断的研究，提出了新的PSD方法并检验了其性能。本研究主要工作如下：
⑴ 使用液闪探测器EJ-301和波形数字化仪CAEN DT5751搭建了探测器系统，对241Am-Be中子源的中子和γ光子进行了探测并采集了核脉冲波形，应用数字化的电荷比较法（CCM）和时间比较法（TCM）进行了中子和γ光子的PSD实验。实验结果表明，CCM甄别参数的设置对PSD效果影响很大，数字化的CCM可灵活地根据需要设置相关参数。CCM的最佳品质因子（FoM）为1.0606；TCM的最佳FoM值为0.9440，CCM的PSD能力强于TCM。
⑵ 使用1个专用的塑闪（探测器A，含ZnS掺杂）和1个非专用的塑闪探测器（探测器B，无ZnS掺杂）分别搭建了探测器系统，对241Am的α粒子和90Sr的β粒子进行探测并采集了核脉冲波形，应用CCM和TCM进行了α和β粒子的PSD实验。在合适的CCM甄别参数下，探测器A的总误分比例约为0.25%；探测器B的约为1.5%。
⑶ 参考常见的核辐射探测器模拟电路制作了26种元器件参数偏移30%的故障状态和正常状态的PCB电路，实验研究了基于支持向量机（SVM）的混合故障的智能诊断与定位。虽然27种不同状态对应的输出波形差异微小，但故障的漏报率为仅为0.2%，误报率为5%；故障定位正确率最低为76%。
⑷ 采集3组9种差异很小的脉冲信号，进行了基于SVM、决策树、主成分分析（PCA）和SVM、PCA和决策树的PSD研究，测试准确度分别为98.3%~100%、76.7%~96.3%、90.7%~100%、90.7%~100%。PCA有效地节省了计算时间，时间成本降低到不使用PCA时的1/14以下。PCA和决策树的测试准确度比SVM的稍低，但测试速度是SVM的36倍以上，用Intel i7-7500U CPU可以在1秒内分类处理7.5万个波形样本。
⑸ 使用PCA进行中子和γ光子甄别，FoM值为1.2957，比CCM的1.0606增加了22.17%。使用机器学习综合运用的新PSD方法，对γ光子甄别准确度为99.9%；对α和β粒子甄别的准确度为99.8%（探测器A）和98.7%（探测器B）。

Nuclear pulse shape discrimination (PSD) has important applications in particle discrimination (such as neutron and γ-photon discrimination) in nuclear physical experiments. In this paper, it was focused on neutron and γ-photon discrimination, α- and β-particle discrimination, hybrid fault diagnosis of analog circuits of nuclear radiation detector based on PSD. And then, new PSD methods were proposed and their performances were tested. The main work of this study is as follows:

⑴ The detector system was assembled with the liquid scintillator detector EJ-301 and the waveform digitizer CAEN DT5751. The neutrons and γ-photons from ²⁴¹Am-Be source were detected and the nuclear pulse waveforms were collected. The PSD experiment of neutron and γ-photon was carried out with the digital charge comparison method (CCM) and time comparison method (TCM).The experimental results showed that the discrimination parameters of CCM have a great influence on the effect of PSD, and the relevant parameters can be flexibly set according to the needs in the digital CCM. The best Figure of Merit (FoM) of CCM was 1.0606; the best FoM of TCM was 0.9440, and CCM performed better than TCM.

⑵ The detector system was constructed using a special plastic scintillator (Detector A, containing ZnS doping) and a non-special plastic scintillator (Detector B, without ZnS doping)， respectively. The pulse waveforms of the events of the α-particles from ²⁴¹Am and the β-particles from ⁹⁰Sr were collected by the digital detection system and discriminated with CCM and TCM. The total misidentification ratio of Detector A is about 0.25%, and the Detector B is about 1.5% with the appropriate CCM discrimination parameters.

⑶ Based on the common analog circuit of nuclear radiation detector, 26 kinds of PCB circuits with 30% offset of component parameters and one normal condition PCB were fabricated. And then, the intelligent diagnosis and location of hybrid fault based on support vector machine (SVM) was experimentally studied. Although the differences of output waveforms corresponding to 27 different states were small, the rate of alarm failure was only 0.2% and the rate of false alarm was 5%. The lowest accuracy of fault location was 76%.

⑷ Three groups (9 kinds) of pulse waveforms with little difference were collected. And then, PSD studies were conducted based on SVM, decision tree, principal component analysis (PCA) and SVM, PCA and decision tree, with test accuracies of 98.3%~100%, 76.7%~96.3%, 90.7%~100%, 90.7%~100%, respectively. The computing time was effectively saved with PCA, and the time costs were reduced to less than 1/14 of that without PCA. The test accuracy of PCA and decision tree was slightly lower than that of SVM, but the test speed was 36 times higher than that of SVM. 75 000 waveform samples can be discriminated by an Intel i7-7500u CPU in 1 second.

⑸ In the neutron and γ-photon discrimination, FoM of PCA was 1.2957, which was 22.17% higher than that of CCM 1.0606. With the new PSD method, the accuracy of γ-photon discrimination was 99.9%. The accuracies of α- and β-particle discrimination were 99.8% (Detector A) and 98.7% (Detector B).