NUCLEAR TECHNIQUES, Volume. 47, Issue 11, 110502(2024)
Application of lightweight neural network models for nuclear pulse parameter prediction
Figures & Tables(13)
Fig. 2. Waveforms of negative exponential (a) and shaped (b) pulse sequence (color online)
Fig. 5. Illustration of negative exponential pulses in Dataset I (color online)
Fig. 6. Illustration of absolute error of amplitude parameters (a) and time parameters (b)
Fig. 7. Diagram of test set with additional white noise (color online)
Fig. 8. Diagram of test set with additional Gaussian noise (color online)
Fig. 9. Diagram of test set with additional flicker noise (color online)
Table 1. Comparison of pulse amplitude prediction results with different time parameters
View tableView in ArticleTable 1. Comparison of pulse amplitude prediction results with different time parameters
幅度真值
Ground truth of amplitude
时间真值
Ground truth of time
成形的幅度参数
Parameter A by shaping
幅度绝对误差
Absolute error of A
幅度的相对误差Relative error of A / % 成形的时间参数Parameter Tby shaping 时间参数的绝对误差Absolute error of T 时间参数的相对误差Relative error of T 20 40 6.87 13.13 65.65 40 0 0 20 45 9.41 10.59 52.95 45 0 0 20 50 11.65 8.35 41.75 50 0 0 20 55 13.61 6.39 31.95 55 0 0 20 60 15.32 4.68 23.40 60 0 0 20 65 16.81 3.19 15.95 65 0 0 20 70 18.11 1.89 9.45 70 0 0 20 75 19.22 0.78 3.90 75 0 0 20 80 20.00 0 0.00 80 0 0
Table 2. The details of data sets
View tableView in ArticleTable 2. The details of data sets
数据集Data set 数据集明细Data set details 脉冲类型Types of pulses 标签Labels 大小Size 数据集I
Data set I
― 模拟脉冲序列
Simulated pulse sequences
幅度，时间
A, T
121 770×256+121 770×2 数据集II
Data set II
训练集Train set 成形脉冲
Shaping pulses
97 416×256+97 416×2 验证集Validation set 12 177×256+12 177×2 测试集Test set 12 177×256+12 177×2
Table 3. The comparison of performance between different models
View tableView in ArticleTable 3. The comparison of performance between different models
模型
Models
批大小
Batch size
学习率
Learning rate
训练损失
Train loss
验证损失
Validation loss
参数量
Numbers of parameters
相对误差Relative error / % 幅度Amplitude / mV 时间Time / ns LeNet5 1 1×10-3 138.69 40.29 0.13M 1.12 4.34 LSTM 1 1×10-3 160.82 69.30 1.97M 2.04 6.69 CNN-LSTM 1 1×10-3 242.33 28.72 1.10M 0.98 3.93 GRU 1 1×10-3 210.48 30.66 2 471 1.51 5.02 CNN-GRU 1 1×10-3 367.14 148.74 1.11 M 1.56 8.10 UNet 1 1×10-3 96.50 18.84 2.11 M 0.57 3.51
Table 4. The parameter prediction performance of two models under different noise
View tableView in ArticleTable 4. The parameter prediction performance of two models under different noise
模型
Models
数据集Data set 信噪比SNR level / dB 白噪声White noise 高斯噪声Gaussian noise 闪烁噪声 1/f noise 幅度相对误差 of amplitude 时间相对误差 of time 幅度相对误差 of amplitude 时间相对误差 of time 幅度相对误差 of amplitude 时间相对误差 of time UNet 测试集 Test set ― 0.57 3.51 0.57 3.51 0.57 3.51 噪声测试集
Noise test set
100 0.99 3.01 0.57 3.52 0.57 3.52 80 1.01 3.02 0.97 3.73 0.97 3.69 60 4.20 5.00 10.09 9.35 8.04 7.92 50 10.70 12.00 29.13 25.17 23.65 20.56 40 24.76 21.33 47.01 40.50 44.75 39.34 30 33.51 25.84 48.95 38.61 58.17 50.74 20 37.54 29.83 37.53 29.82 57.10 47.15 10 64.67 73.10 108.45 111.73 53.58 41.18 CNN-LSTM 测试集 Test set ― 0.98 3.93 0.98 3.93 0.98 3.93 噪声测试集
Noise test set
100 2.21 14.21 2.21 14.21 2.20 14.21 80 2.27 14.20 2.27 14.19 2.28 14.21 60 4.35 14.25 4.34 14.24 5.79 14.63 50 19.94 17.32 19.94 17.32 26.15 19.74 40 59.75 34.40 59.75 34.40 63.12 37.98 30 82.30 56.09 82.30 56.09 83.93 59.77 20 92.57 61.96 92.57 61.95 93.09 62.28 10 96.97 59.02 96.97 59.02 95.93 62.81
Tools
Get Citation
Copy Citation Text
Lin TANG, Shuang ZHOU, Xianli LIAO, Bo LI. Application of lightweight neural network models for nuclear pulse parameter prediction[J]. NUCLEAR TECHNIQUES, 2024, 47(11): 110502
Paper Information
Category: NUCLEAR PHYSICS, INTERDISCIPLINARY RESEARCH
Received: Jun. 13, 2024
Accepted: --
Published Online: Jan. 2, 2025
The Author Email: LI Bo (LIBo)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20