Nuclide identification is one of the key techniques researched in the field of nuclear detection.Traditional nuclide identification instruments based on energy spectrum analysis algorithms have poor real-time performance and high power consumption,which limit the identification efficiency in practical applications.This work proposes an FPGA-based convolutional neural network hardware acceleration method for nuclide identification to accelerate the qualitative analysis of radionuclides.Firstly,a lightweight one-dimensional convolutional neural network model for nuclide classification is constructed,and then deployed in a Xilinx ZYNQ7020 heterogeneous chip using hardware acceleration strategies such as parallel pipelines and adder trees according to the computing features of the convolutional,pooling and fully connected layers of the model.The experimental results show that the average recognition accuracy of the model in FPGA reaches 98.41％,and the single recognition takes only 1.57 ms.Compared with the desktop CPU,this hardware acceleration method achieves 64 times acceleration effect,and the power consumption is only 2.115 W.Measured at a distance of 30 cm,the identification accuracy of 137Cs single source is 98％,and the identification accuracy of 137Cs and 60Co mixed source reaches 98.17％.This hardware acceleration method satisfies the requirements of low latency and low power consumption,and is suitable for the scene of fast nuclide detection,and is of reference value for the development of portable nuclide recognition instruments.
关键词
能谱数据/核素识别/FPGA/卷积神经网络/硬件加速
Key words
energy spectrum data/nuclide identification/field programmable gate array/convolutional neural network/hardware acceleration
维普
万方数据