NUCLEAR TECHNIQUES, Volume. 46, Issue 9, 090505(2023)
Application of multi-head attention mechanism with embedded positional encoding in amplitude estimation of stacked pulses
Figures & Tables(10)
Fig. 1. Principle of stacking pulse separation and pulse amplitude estimation
Fig. 4. Principle of positional encoding (color online) (a) Binary encoding, (b) Sine function encoding
Fig. 6. Iterative graph of the loss and accuracy for the training and validation sets during model training
Fig. 7. (a) Measured pulse sequence diagram, (b) Triangulation results for the pulse sequence, (c) Comparison chart of the estimated random pulse amplitude (color online)
Table 1. Specific characteristics of the blocks
View tableView in ArticleTable 1. Specific characteristics of the blocks
模块名Name 内容Content 矩阵计算Matrix calculation 矩阵维度Matrix dimension 模块A Block A Input matrix S S ×W i=Q 2×4 模块B Block B Weight matrix Wi 4×3 模块C Block C Attention matrix Q 2×3 模块D Block D Attention head Z i×W =Z 2×24 模块E Block E Weight matrix W 24×4 模块F Block F Output matrix Z 2×4
Table 2. Comparison of estimated and real values of overlapping pulses in the deep learning model
View tableView in ArticleTable 2. Comparison of estimated and real values of overlapping pulses in the deep learning model
脉冲
Pulse
是否堆积
Pile-up or not
真实值
Real value
三角成形值
Triangulation
成形相对误差Relative error of Triangulation / % 模型估计值Estimated value of model 模型相对误差Relative error of model / % P1 No 322.307 7 308.87 4.17 320.23 0.64 P2 Yes 615.384 6 568.21 7.67 610.48 0.80 P3 1 167.692 618.74 47.01 1 164.98 0.23 P4 Yes 910.769 3 849.85 6.69 908.56 0.24 P5 710.000 1 432.60 39.07 708.52 0.21 P6 Yes 644.615 4 557.88 13.46 643.19 0.22 P7 1 076.923 627.63 41.72 1 070.48 0.60 P8 No 386.153 9 363.08 5.98 380.86 1.37 P9 No 622.307 7 571.75 8.12 621.64 0.11 P10 Yes 611.538 5 556.17 9.05 607.77 0.62 P11 992.307 7 666.04 32.88 989.54 0.28 P12 Yes 780.769 3 732.13 6.23 777.57 0.41 P13 956.923 1 829.89 13.28 949.87 0.74 P14 No 730.769 3 675.82 7.52 722.75 1.10 P15 No 639.230 8 569.35 10.93 634.71 0.71
Table 3. Comparison table of parameter estimation effects of the deep learning model
View tableView in ArticleTable 3. Comparison table of parameter estimation effects of the deep learning model
序号Number 样品类型Sample type 堆积程度Stacking degree 堆积类型Stacking type 平均相对误差Average relative error / % 1 粉末铁矿样品
Powder iron ore sample
10% 双脉冲Double pulse 0.21 2 30% 双脉冲Double pulse 0.32 3 60% 双脉冲Double pulse 3.42 4 90% 双脉冲Double pulse 0.85 5 粉末岩石样品
Powder rock samples
10% 双脉冲Double pulse 0.19 6 30% 双脉冲Double pulse 0.48 7 60% 双脉冲Double pulse 0.65 8 90% 双脉冲Double pulse 0.98
Tools
Get Citation
Copy Citation Text
Lin TANG, Shuang ZHOU, Yong LI, Xianli LIAO, Yuepeng LI. Application of multi-head attention mechanism with embedded positional encoding in amplitude estimation of stacked pulses[J]. NUCLEAR TECHNIQUES, 2023, 46(9): 090505
Paper Information
Category: Research Articles
Received: Mar. 23, 2023
Accepted: --
Published Online: Oct. 8, 2023
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20