• High Power Laser and Particle Beams
  • Vol. 37, Issue 5, 059001 (2025)
Chao Li1, Rui Shi1,2,*, Shuxin Zeng2, Xinhua Xu2..., Yuhong Wei2 and Xianguo Tuo1|Show fewer author(s)
Author Affiliations
  • 1College of Physics and Electronic Engineering, Sichuan University of Science and Engineering , Yibin 644000, China
  • 2School of Computer Science and Engineering, Sichuan University of Science and Engineering , Yibin 644000, China
  • show less
    DOI: 10.11884/HPLPB202537.240398 Cite this Article
    Chao Li, Rui Shi, Shuxin Zeng, Xinhua Xu, Yuhong Wei, Xianguo Tuo. Lightweight neural network model for nuclide recognition based on nuclear pulse peak sequence and its FPGA acceleration method[J]. High Power Laser and Particle Beams, 2025, 37(5): 059001 Copy Citation Text show less
    Detector modeling and simulation energy spectrum
    Fig. 1. Detector modeling and simulation energy spectrum
    DT5800D analog signal flow
    Fig. 2. DT5800D analog signal flow
    Convolution contrast and residual module contrast
    Fig. 3. Convolution contrast and residual module contrast
    Diagram of neural network structure
    Fig. 4. Diagram of neural network structure
    Model training process
    Fig. 5. Model training process
    Test set confusion matrix
    Fig. 6. Test set confusion matrix
    Pipeline diagram
    Fig. 7. Pipeline diagram
    Partition type
    Fig. 8. Partition type
    HLS partially optimized pseudocode
    Fig. 9. HLS partially optimized pseudocode
    Block diagram of nuclide recognition system
    Fig. 10. Block diagram of nuclide recognition system
    modelrecognition accuracy/%model parametermodel size/kbit
    my model98.3212825.9
    LSTM85.327536111
    GhostNet96.917100966711
    MobileNet-V194.123966889430
    ResNet-1893.8385691215151
    VGGNet-1693.034341200134156
    Table 1. Model comparison
    resource32 bit float (utilization rate)16 bit fixed-point number (utilization rate)
    BRAM_18K240(85%)79(28%)
    DSP162(73%)114(51%)
    FF46 657(43%)12 968(12%)
    LUT51 846(97%)21 483(40%)
    latency(cycles)38 10134 792
    Table 2. Quantifying resource consumption and delay
    resourcebefore fusion(utilization rate)after fusion(utilization rate)
    BRAM_18K252(90%)240(85%)
    DSP169(76%)162(73%)
    FF51006(47%)46657(43%)
    LUT54753(102%)51846(97%)
    latency(cycles)4701338101
    Table 3. Resource consumption and latency before and after fusion
    resourcebefore optimization (utilization rate)after optimization (utilization rate)
    BRAM_18K0(0%)0(0%)
    DSP4(1%)32(14%)
    FF328(~0%)1 808(1%)
    LUT427(~0%)802(1%)
    latency(cycles)3 208811
    Table 4. Resource consumption and delay before and after partial optimization
    resource16 bit fixed-point number (utilization rate)after-optimization (utilization rate)VIVADO (utilization rate)
    BRAM_18K79(28%)109(38%)95(68%)
    DSP114(51%)136(61%)154(70%)
    FF12968(12%)21099(19%)15372(14%)
    LUT21483(40%)39204(73%)12452(23%)
    latency(cycles)3479227334\
    Table 5. Resource consumption and delay before and after optimization
    Chao Li, Rui Shi, Shuxin Zeng, Xinhua Xu, Yuhong Wei, Xianguo Tuo. Lightweight neural network model for nuclide recognition based on nuclear pulse peak sequence and its FPGA acceleration method[J]. High Power Laser and Particle Beams, 2025, 37(5): 059001
    Download Citation