How to realize the accurate identification of low count and multi-class complex radionuclides is one of the problems that need to be solved urgently in the field of radiation monitoring. In this paper, convolution neural network is introduced to build a recognition model for low count and multi-multi-class complex radionuclides.The Monte Carlo simulation was adopted to create the radionuclide energy spectrum database. This database contained 63 different radioactive nuclear sources. This database encompassesed three different types of radioactive sources: single-source, dual-source, and triple-source, which were composed of 241Am, 133Ba, 57Co, 60Co, 137Cs, 152Eu, and 40K. The simulation training set and simulation verification set samples were used to complete the training and hyperparameter optimization of convolutional neural networks. The test set samples were used to verify the model performance. The result showed that, the macro precision, macro recall and macro F1 values obtained from the prediction results of the simulation test set are all greater than 0.99, and the macro precision, macro recall and macro F1 values obtained for the sample of the measurement test set were all greater than 0.90. The experimental results demonstrate the feasibility of convolutional neural networks in the identification of low count and multi-class complex radionuclides. Meanwhile, the visualization analysis of the transformation process of energy spectrum data in convolutional neural network was performed for the first time, which can reveal the processing mechanism of energy spectrum data of convolutional neural network.