Radionuclide identification method of two-dimensional convolutional neural network based on nuclear pulse peak sequence

Geng LUO; Rui SHI; Shuxin ZENG; Yanan SHANG; Guang YANG; Yixian DING; Yadong WU; Xianguo TUO

doi:10.11889/j.0253-3219.2025.hjs.48.240301

Journals >NUCLEAR TECHNIQUES >Volume 48 >Issue 3 >Page 030502 > Article

NUCLEAR TECHNIQUES
Vol. 48, Issue 3, 030502 (2025)

Radionuclide identification method of two-dimensional convolutional neural network based on nuclear pulse peak sequence

Geng LUO¹, Rui SHI^1,2,*, Shuxin ZENG¹, Yanan SHANG¹..., Guang YANG¹, Yixian DING³, Yadong WU^1,2 and Xianguo TUO^1,2,**|Show fewer author(s)

Author Affiliations

¹School of Computer Science and Engineering, Sichuan University of Science & EngineeringZigong 643000, China

²Perception and Control Key Laboratory of Sichuan Province, Sichuan University of Science & Engineering,, Yibin 644005, China

³Hongya Ecological Environment Bureau, Meishan 620010, China

show less

DOI: 10.11889/j.0253-3219.2025.hjs.48.240301 Cite this Article

Geng LUO, Rui SHI, Shuxin ZENG, Yanan SHANG, Guang YANG, Yixian DING, Yadong WU, Xianguo TUO. Radionuclide identification method of two-dimensional convolutional neural network based on nuclear pulse peak sequence[J]. NUCLEAR TECHNIQUES, 2025, 48(3): 030502 Copy Citation Text

show less

Fig. 1. Diagram of experimental conditions and data acquisition process

Download full size | View in the Article

Sequence scatter plot of 137Cs (a), 60Co (b), 155Eu-22Na (c), 137Cs-60Co (d), 60Co-155Eu-22Na (e), 137Cs-155Eu-22Na (f), 137Cs-60Co-155Eu-22Na (g), background (h)

Fig. 2. Sequence scatter plot of ¹³⁷Cs (a), ⁶⁰Co (b), ¹⁵⁵Eu-²²Na (c), ¹³⁷Cs-⁶⁰Co (d), ⁶⁰Co-¹⁵⁵Eu-²²Na (e), ¹³⁷Cs-¹⁵⁵Eu-²²Na (f), ¹³⁷Cs-⁶⁰Co-¹⁵⁵Eu-²²Na (g), background (h)

Download full size | View in the Article

Fig. 3. Schematic diagram of the mapping of nuclear pulse sequence from one-dimensional to two-dimensional (color online)

Download full size | View in the Article

Fig. 4. Structure diagram of 2D-CNN network model (color online)

Download full size | View in the Article

Fig. 5. Train 2D-CNN model loss line graphs with different parameter structures, filling method: 'Same' means zero filling, 'Valid' means no filling, 'BN' means batch normalization, 'NoneBN' means no batch normalization (color online)

Download full size | View in the Article

Fig. 6. Stratified 10-fold cross-validation accuracy curve: 4 categories (color online)

Download full size | View in the Article

Fig. 7. Confusion matrix (4 nuclide combinations) (a) Sequence length: 400, (b) Sequence length: 900

Download full size | View in the Article

Fig. 8. Stratified 10-fold cross-validation accuracy curve: 5 radionuclide combinations (color online)

Download full size | View in the Article

Fig. 9. Confusion matrix (5 categories) (a) Sequence length: 400, (b) Sequence length: 900

Download full size | View in the Article

Fig. 10. Stratified 10-fold cross-validation accuracy curve: 8 categories (color online)

Download full size | View in the Article

Fig. 11. Confusion matrix (8 categories) (a) Sequence length: 400, (b) Sequence length: 900

Download full size | View in the Article

Fig. 12. Energy spectrum (1) ((a) nuc7, nuc2, (b) nuc6, nuc3) and scatter plot (2) ((c) nuc7, nuc2, (d) nuc6, nuc3) when sequence length is 900, detection distance is 40 cm (color online)

Download full size | View in the Article

放射源 Radioactive sources	活度 Radioactivity / Bq	标签 Label
¹³⁷Cs	2.940×10⁵	nuc0
⁶⁰Co	9.900×10⁴	nuc1
¹⁵⁵Eu-²²Na	1.120×10⁴	nuc2
¹³⁷Cs-⁶⁰Co	3.930×10⁵	nuc3
⁶⁰Co-¹⁵⁵Eu-²²Na	3.052×10⁵	nuc4
¹³⁷Cs-¹⁵⁵Eu-²²Na	1.102×10⁵	nuc5
¹³⁷Cs-⁶⁰Co-¹⁵⁵Eu-²²Na	4.042×10⁵	nuc6
环境本底 Environmental background	无 None	nuc7

Table 1. Information about the radioactive sources used in the experiment

样本标签 Label	核脉冲序列长度划分 Nuclear pulse sequence length division
样本标签 Label	200:600:100	900	1 225	1 600
nuc0	8×1 000	8×1 000	8×800	8×600
nuc1	8×1 000	8×1 000	8×800	8×600
nuc2	8×1 000	8×1 000	8×800	8×600
nuc3	8×1 000	8×1 000	8×800	8×600
nuc4	8×1 000	8×1 000	8×800	8×600
nuc5	8×1 000	8×1 000	8×800	8×600
nuc6	8×1 000	8×1 000	8×800	8×600
nuc7	8 000	7 600	5 600	4 200
合计Total	64 000	63 600	50 400	37 800

Table 2. The number of samples divided into each length sequence set (8 refers to 8 detection points)

序列长度 Sequence length	$x_{m a x}$
200	700
300	1 000
400	1 200
500	1 700
600	2 800
900	4 200
1 225	5 500
1 600	7 200

Table 3. The relationship between sequence set length and the corresponding real-time arrival time layer values

类别

Category

预测值Predicted

\hat{y} = c

\hat{y} \neq c

真实值

True

y = c

T P_{c}

F N_{c}

y \neq c

F P_{c}

T N_{c}

Table 4. Confusion matrix

模型参数Model parameter	10次测试平均精度Average accuracy of 10 tests / %	方差Variance
2×2 + Valid	80.58	3.96
2×2 + Same	83.75	1.38
2×2 + Valid + BN	82.00	2.07

Table 5. Mean recognition accuracy for different model parameters and variance between accuracies

输入维度 Input size		网络层Layer
输入维度 Input size		CL1	PL1	CL2	PL2	CL3	PL3	Output
[10,20,2]	Dim	[10,20,32]	[10,5,32]	[9,4,64]	[64]	―	―	4/5/8
	K-num	32	―	64	―	―	―	―
	Field	[2, 2]	[2, 2]	[2, 2]	GAP	―	―	―
	Stride	1	1	1	―	―	―	―
[20,15,2]	Dim	[20,15,32]	[10,7,32]	[9,6,64]	[4,3,64]	[3,2,128]	128	4/5/8
	K-num	32	―	64	―	128	―	―
	Field	[2, 2]	[2, 2]	[2, 2]	[2, 2]	[2, 2]	GAP	―
	Stride	1	1	1	1	1	―	―
[20,20,2]	Dim	[20,20,32]	[10,10,32]	[9,9,64]	[4,4,64]	[3,3,128]	128	4/5/8
[20,25,2]	Dim	[20,25,32]	[10,12,32]	[9,11,64]	[4-5,64]	[3-4,128]	128	4/5/8
[20,30,2]	Dim	[20,30,32]	[10,15,32]	[9,14,64]	[4, 7,64]	[3,6,128]	128	4/5/8
[30,30,2]	Dim	[30,30,32]	[15,15,32]	[14,14,64]	[7,7,64]	[6,6,128]	128	4/5/8
[35,35,2]	Dim	[35,35,32]	[17,17,32]	[16,16,64]	[8,8,64]	[7,7,128]	128	4/5/8
[40,40,2]	Dim	[40,40,32]	[20,20,32]	[19,19,64]	[9,9,64]	[8,8,128]	128	4/5/8

Table 6. The detailed structure parameters of the model

标签 Labels	长度 Length / cm	5	10	15	20	25	30	35	40
nuc0	400	40.18	68.49	105.55	153.73	200.98	252.66	297.35	340.07
nuc0	900	97.27	154.26	237.88	346.44	453.25	568.56	669.28	766.61
nuc1	400	39.06	69.21	104.11	147.01	191.48	244.58	292.79	344.53
nuc1	900	98.21	155.76	234.17	331.75	432.38	551.44	659.17	774.70
nuc2	400	449.83	610.92	714.32	777.13	806.52	843.04	858.08	869.49
nuc2	900	1 010.89	1 378.85	1 613.96	1 753.75	1 818.56	1 893.11	1 927.83	1 951.92
nuc3	400	30.33	39.59	57.87	78.93	103.97	135.57	168.41	207.66
nuc3	900	95.09	98.49	129.41	177.71	234.39	305.61	380.21	468.16
nuc4	400	47.71	68.88	112.77	144.31	202.04	233.65	290.96	342.26
nuc4	900	106.86	155.78	254.07	325.60	439.70	525.41	655.11	772.39
nuc5	400	47.14	59.02	92.53	125.15	176.06	216.98	259.01	311.06
nuc5	900	107.78	132.57	208.17	281.54	382.06	488.61	584.71	700.99
nuc6	400	29.91	38.58	55.29	79.65	105.57	133.81	169.23	201.72
nuc6	900	94.44	98.95	124.85	179.47	237.36	301.35	382.67	456.34
nuc7	400	925.07 2 087.88
nuc7	900	925.07 2 087.88

Table 7. The mean time (ms) to collect 400 and 900 pulse sequence points at various detection distances for different radiation sources

方法/数据集 Methods/DataSet	样本类别 Category	5 cm	10 cm	15 cm	20 cm	25 cm	30 cm	35 cm	40 cm	测试精度 Test accuacy
2D-CNN/序列 2D-CNN/Sequence	4	99.99%	100.00%	99.95%	100.00%	99.85%	100.00%	100.00%	99.99%	100.00%
	5	99.92%	99.88%	99.76%	99.56%	99.03%	96.76%	93.76%	90.92%	98.61%
	8	99.03%	79.22%	90.88%	78.85%	85.47%	81.95%	76.45%	76.45%	84.45%
2D-CNN/能谱 2D-CNN/Spectrum	4	99.92%	99.94%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	99.84%
	5	99.90%	99.98%	99.94%	99.60%	99.16%	97.90%	96.88%	96.10%	97.02%
	8	88.74%	81.00%	84.14%	81.34%	81.03%	79.25%	77.40%	73.77%	79.37%
BPNN+PCA/脉冲幅值 BPNN+PCA/ Pulse amplitude	4	81.35%	98.20%	99.05%	99.15%	98.95%	98.45%	96.40%	92.15%	95.81%
	5	70.68%	84.80%	84.40%	85.48%	85.00%	84.00%	84.00%	80.56%	81.99%
	8	69.48%	69.48%	72.26%	70.37%	73.37%	71.34%	66.00%	65.42%	68.83%
SVM+PCA/脉冲幅值 SVM+PCA/ Pulse amplitude	4	81.2%	99.40%	95.90%	91.85%	85.65%	78.55%	72.2%	66.45%	84.13%
	5	59.28%	59.32%	50.36%	42.84%	41.00%	41.04%	40.76%	40.48%	47.04%
	8	54.83%	42.57%	38.08%	31.77%	29.37%	28.51%	28.37%	28.10%	34.07%

Table 8. Performance comparison of different methods

Download Citation

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享