A Classifier for Nuclear Pulse Detection based on CNN

In the field of nuclear detection, there is a basic task to identify the category and orientation of radioactive source. The category and orientation depend on the energy spectrum and flood histogram which is the statistic of qualified pulses. In this paper, we propose a novel method to distinguish qualified pulses and disqualified pulses with a classification based on convolution neural network (CNN). Although many other methods also focus on this task, they have many defects, for instance they can only get energy spectrum, or they cannot get counting rate accurately. By analyzing the problem, we can find that convolution neural network is a good method to solve this problem. The advantages of our methods are: 1) Our model can reach high accuracy even though the training dataset is not very large. 2) Because convolution layer can ensure the consistency of displacement and scale, it is a good idea to extract the feature of pulses with 1D convolution kernel. 3) This model is small, the precision can reach more than 95% and the recall can reach 98% in few iterations. The result is good enough to recall final results. After the qualified pulses are reserved, we can get the energy spectrum and flood histogram at the same time. We use two different radioactive sources to test the method. The results show that both the energy spectrum and flood histogram are revised no matter in low radiation intensity nor in high radiation intensity. The inference is realized on computer to generate correct energy spectrum and flood histogram which would help to track the orientation of category of radioactive source.