基于机器学习的核电仪控系统关键芯片温度预测

汪凡雨; 王东伟; 邓强; 赵阳; 严浩; 陈起

doi:10.3969/j.issn.0258-2724.20240332

基于机器学习的核电仪控系统关键芯片温度预测

doi: 10.3969/j.issn.0258-2724.20240332

中国核动力研究设计院核反应堆系统设计技术重点实验室，四川成都 610213

基金项目: 国家自然科学基金项目（52105220）；四川省自然科学基金项目（2022NSFSC1950）

详细信息

作者简介:
汪凡雨（1996—），女，工程师，研究方向为仪控设备可靠性，E-mail：wfanyu1201@163.com

通讯作者:
王东伟（1988—），男，高级工程师，博士，研究方向为仪控设备可靠性，机械系统动力学，E-mail：dongwei1013@sina.cn

中图分类号: TM623
计量
- 文章访问数: 52
- HTML全文浏览量: 43
- PDF下载量: 10
- 被引次数: 0
出版历程
- 收稿日期: 2024-07-09
- 修回日期: 2024-09-13
- 网络出版日期: 2025-11-07

Temperature Prediction of Key Chips in Nuclear Power Instrumentation and Control System Based on Machine Learning

Science and Technology on Reactor System Design Technology Laboratory, Nuclear Power Institute of China, Chengdu 610213, China

摘要

摘要:
为研究核安全级仪控系统中控制保护柜的内部热学特性及其关键芯片（CPU与FPGA）的稳态温度（SST）变化规律，在不同环境温度下对该控制保护柜开展了试验研究，并采用有限元方法模拟试验过程，通过对比试验结果验证了数值模型的准确性；在此基础上，利用有限元模型计算得到100组随机工况下CPU和FPGA的SST值，并采用M-SVR、XGBoost、ANN及BRR 4种算法对不同工况下CPU和FPGA的SST进行学习预测. 研究结果表明：1）在环境温度为20 ℃时，CPU和FPGA的SST分别为37.5 ℃和33.5 ℃；当环境温度为55 ℃时，相应的SST上升至为72 ℃和68 ℃；有限元分析结果能够很好地模拟试验现象，计算所得芯片的SST同试验结果吻合较好. 2） 4种算法模型均能够对芯片SST进行预测，其中ANN算法在测试集上的预测性能最佳，其MSE值小于0.15%，R²大于0.99，泛化能力最强；相比之下，其他3种模型对高SST样本表现出较好的预测性能，但对低SST样本的预测误差较大，尤其是XGBoost模型，其预测误差高达3.65 ℃. 本研究为核安全级控制系统的芯片SST预测提供了一种新方法.
- 核安全级仪控系统 /
- 热学特性 /
- 芯片温度 /
- 有限元分析 /
- 机器学习
Abstract:
To investigate the thermal characteristics inside the control and protection cabinet of nuclear safety-class instrumentation and control (I&C) systems and the variation patterns of the steady-state temperature (SST) of key chips (CPU and field programmable gate array (FPGA)), experimental studies were conducted on the cabinet under different ambient temperatures. The finite element method was employed to simulate the experimental process, and the accuracy of the numerical model was validated by comparing the experimental results. Furthermore, the SST values of CPU and FPGA under 100 sets of random working conditions were calculated by the finite element model, and the SST values of CPU and FPGA under different working conditions were learned and predicted using four algorithms of multi-output support vector regression (M-SVR), extreme gradient boosting (XGBoost), artificial neural network (ANN), and Bayesian ridge regression (BRR). Results show that when the ambient temperature is 20 ℃, the SST of CPU and FPGA is 37.5 ℃ and 33.5 ℃, respectively. When the ambient temperature is 55 ℃, the SST of the CPU and FPGA rises to 72 ℃ and 68 ℃, respectively. Finite element analysis can well simulate the test phenomenon, and the calculated chip SSTs are in good agreement with the experimental results. All the four algorithm models can be used to predict chip SST, among which the ANN algorithm exhibits the best prediction performance on the test set. It has a mean squared error (MSE) less than 0.15% and an R² value greater than 0.99 and exhibits the strongest generalization ability. In contrast, although the other three models show good prediction performance for samples with high SSTs, the prediction error for samples with low SSTs is large, especially for the XGBoost model, whose prediction error is as high as 3.65 ℃. The research provides a new method for SST prediction of chips in nuclear safety-class control systems.
- nuclear safety-class I&C system /
- thermal characteristics /
- chip temperature /
- finite element analysis /
- machine learning

HTML全文

图 1 控制保护柜设备及1U风扇示意

Figure 1. Diagram of control and protection cabinet and 1U fan

下载: 全尺寸图片幻灯片

图 2 20 ℃环境温度下设备内部温度云图

Figure 2. Cloud image of the internal temperature at ambient temperature of 20 ℃

下载: 全尺寸图片幻灯片

图 3 20 ℃环境温度下芯片温度变化曲线

Figure 3. Chip temperature change curves at ambient temperature of 20 ℃

下载: 全尺寸图片幻灯片

图 4 55 ℃环境温度下芯片温度变化曲线

Figure 4. Chip temperature change curves at ambient temperature of 55 ℃

下载: 全尺寸图片幻灯片

图 5 控制保护柜设备有限元模型

Figure 5. Finite element model of control and protection cabinet

下载: 全尺寸图片幻灯片

图 6 20 ℃环境温度下机柜内部温度分布与热流特性

Figure 6. Temperature distribution and heat flow characteristics inside the cabinet at ambient temperature of 20 ℃

下载: 全尺寸图片幻灯片

图 7 55 ℃环境温度下机柜内部温度分布与热流特性

Figure 7. Temperature distribution and heat flow characteristics inside the cabinet at ambient temperature of 55 ℃

下载: 全尺寸图片幻灯片

图 8 预测模型开发流程

Figure 8. Flowchart of predictive model development

下载: 全尺寸图片幻灯片

图 9 四个模型的10折交叉验证结果对比

Figure 9. Comparison of 10-fold cross validation results of the four models

下载: 全尺寸图片幻灯片

图 10 验证与测试结果对比

Figure 10. Comparison of validation and test results

下载: 全尺寸图片幻灯片

图 11 不同机器学习模型的预测结果

Figure 11. Prediction results of different machine learning models

下载: 全尺寸图片幻灯片

图 12 4种机器学习模型的预测性能对比

Figure 12. Prediction performance comparison of the four machine learning models

下载: 全尺寸图片幻灯片

表 1 部件材料参数值

Table 1. Material parameters of components

部件	材料	热导率/ （W·mK⁻¹）	比热容/ （J/Kg·K）	密度/ （kg·m⁻³）
机柜	Q235	50.2	460	7860
内部支撑件	不锈钢	16.3	500	7900
机箱	Al 6061	180	963	2700
PCB	FR4	0.3	880	1200
CPU	金属封装	15	395	2000
FPGA	金属封装	15	395	2000

下载: 导出CSV

表 2 特征的详细参数值

Table 2. Detailed parameter values of features

特征项	参数值
环境温度/℃	20, 30, 40, 55
CPU功耗/W	5, 6, 8, 10
FPGA功耗/W	3, 4, 5, 6
风扇风量/（m³·h⁻¹）	0.015, 0.018, 0.025, 0.030

下载: 导出CSV

表 3 4个模型的最优超参数

Table 3. Optimal hyperparameters of the four models

模型	参数值
M-SVR	kernel=“RBF”，C=700，gamma=0.03
XGBoost	learning_rate=0.1，n_estimators=300， max_depth=3，subsample=0.8， reg_alpha=0，reg_lambda=10
ANN	Dense （8, activation='sigmoid'）， Dense （4, activation='sigmoid'）， optimizer = “rmsprop”， batch_size = 5，epochs = 300
BRR	alpha_1=100，alpha_init=1， lambda_1=10，lambda_init=0.001

下载: 导出CSV

表 4 4种模型的10折交叉验证结果

Table 4. 10-fold cross validation results of the four models

验证次数/次	MSE/%
验证次数/次	M-SVR	XGBoost	ANN	BRR
1	0.0326	0.0994	0.0178	0.0358
2	0.0030	0.0167	0.0342	0.0051
3	0.0082	0.0372	0.0075	0.0061
4	0.0023	0.0117	0.0375	0.0032
5	0.0094	0.0104	0.0305	0.0065
6	0.0125	0.0391	0.0401	0.0034
7	0.0046	0.0154	0.0178	0.0016
8	0.0029	0.0174	0.0269	0.0018
9	0.0051	0.0214	0.0430	0.0044
10	0.0155	0.0250	0.0346	0.0142

下载: 导出CSV

表 5 测试集样本数据

Table 5. Sample data of the test set

环境温度/℃	CPU功耗/W	FPGA 功耗/W	风扇风量/ （m³·h⁻¹）	T_CPU/℃	T_FPGA/℃
20	8	5	0.015	36.29	33.65
20	8	5	0.025	34.54	32.41
20	8	6	0.030	33.95	31.83
20	10	6	0.015	38.43	34.95
30	6	5	0.030	42.25	41.28
30	8	4	0.030	43.44	40.60
30	8	5	0.015	45.66	42.93
30	10	3	0.030	44.71	40.46
40	6	5	0.030	50.35	48.91
40	8	5	0.025	54.51	51.97
40	8	6	0.030	54.05	51.92
40	10	3	0.015	57.15	52.55
40	10	6	0.030	56.84	53.62
55	6	5	0.030	67.21	65.85
55	8	4	0.025	69.20	66.30
55	8	4	0.030	68.46	65.62
55	8	5	0.030	68.6	66.11
55	10	4	0.025	70.66	66.68
55	10	5	0.018	71.90	68.40
55	10	6	0.015	73.24	69.80

下载: 导出CSV

表 6 4种模型的预测性能对比

Table 6. Comparison of prediction performance of the four models

模型	MSE/%	R²
M-SVR	0.1793	0.9847
XGBoost	0.2346	0.9800
ANN	0.1352	0.9885
BRR	0.1773	0.9849

下载: 导出CSV

表 7 4种模型的详细性能指标

Table 7. Detailed performance metrics of the four models

模型	MSE/%		R²
模型	CPU	FPGA	CPU	FPGA
M-SVR	0.2063	0.1524	0.9822	0.9873
XGBoost	0.2816	0.1875	0.9756	0.9844
ANN	0.1651	0.1053	0.9857	0.9912
BRR	0.2035	0.1511	0.9824	0.9874

下载: 导出CSV

参考文献(17)

[1]	王丛林, 柴晓明, 杨博, 等. 先进核能技术发展及展望[J]. 核动力工程, 2023, 44(5): 1-5. WANG Conglin, CAI Xiaoming, YANG Bo, et al. Development and prospect of advanced nuclear energy technology[J]. Nuclear Power Engineering, 2023, 44(5): 1-5.
[2]	张军, 张犁, 成瑜. IGBT模块寿命评估研究综述[J]. 电工技术学报, 2021, 36(12): 2560-2575. ZHANG Jun, ZHANG Li, CHENG Yu. Review of the lifetime evaluation for the IGBT module[J]. Transactions of China Electrotechnical Society, 2021, 36(12): 2560-2575.
[3]	DING B, ZHANG Z H, GONG L, et al. A novel thermal management scheme for 3D-IC chips with multi-cores and high power density[J]. Applied Thermal Engineering, 2020, 168: 114832. doi: 10.1016/j.applthermaleng.2019.114832
[4]	薛志虎, 艾邦成, 曲伟. 基于LHP的数据中心服务器芯片级散热技术[J/OL]. 华中科技大学学报(自然科学版), https://doi.org/10.13245/j.hust.240339. XUE Zhihu, AI Bangcheng, QU Wei. Server chip-level cooling technology in data center based on LHP[J/OL]. Journal of Huazhong University of Science and Technology, https://doi.org/10.13245/j.hust.240339.
[5]	李宇婷, 徐超, 陆规. 极高热流密度的降级散热及耦合优化[J]. 工程热物理学报, 2022, 43(6): 1580-1587. LI Yuting, XU Chao, LU Gui. Degraded heat dissipation and coupling optimization with extremely high heat flux density[J]. Journal of Engineering Thermophysics, 2022, 43(6): 1580-1587.
[6]	ALSHEHRY A H, ALSHAHRY S M, ALHAZMI A K, et al. A study on the effect of temperature variations on FPGA-based multi-channel time-to-digital converters[J]. Sensors, 2023, 23(18): 7672. doi: 10.3390/s23187672
[7]	盘朝奉, 刘兵, 陈龙, 等. 锂离子电池温升特性分析及液冷结构设计[J]. 西南交通大学学报, 2020, 55(1): 68-75. PAN Chaofeng, LIU Bing, CHEN Long, et al. Temperature rise characteristic analysis and liquid cooling structure design of lithium battery[J]. Journal of Southwest Jiaotong University, 2020, 55(1): 68-75.
[8]	赵泽宇, 杜明星. 多数据驱动人工神经网络的IGBT结温在线估计方法[J]. 电子测量与仪器学报, 2022, 36(7): 223-229. ZHAO Zeyu, DU Mingxing. On line estimation of IGBT junction temperature based on multi data driven artificial neural network[J]. Journal of Electronic Measurement and Instrumentation, 2022, 36(7): 223-229.
[9]	BHATASANA M, MARCONNET A. Deep learning for real-time chip temperature and power predictions[C]//2023 22nd IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (ITherm). Orlando: IEEE, 2023: 1-7.
[10]	WANG C, VAFAI K. Heat transfer enhancement for 3D chip thermal simulation and prediction[J]. Applied Thermal Engineering, 2024, 236: 121499. doi: 10.1016/j.applthermaleng.2023.121499
[11]	SADIQBATCHA S I, ZHANG J W, AMROUCH H, et al. Real-time full-chip thermal tracking: a post-silicon, machine learning perspective[J]. IEEE Transactions on Computers, 2021, 71(6): 1411-1424.
[12]	RAJPUT D, WANG W J, CHEN C C. Evaluation of a decided sample size in machine learning applications[J]. BMC Bioinformatics, 2023, 24(1): 48. doi: 10.1186/s12859-023-05156-9
[13]	DOU B Z, ZHU Z L, MERKURJEV E, et al. Machine learning methods for small data challenges in molecular science[J]. Chemical Reviews, 2023, 123(13): 8736-8780. doi: 10.1021/acs.chemrev.3c00189
[14]	BAO Y K, XIONG T, HU Z Y. Multi-step-ahead time series prediction using multiple-output support vector regression[J]. Neurocomputing, 2014, 129: 482-493. doi: 10.1016/j.neucom.2013.09.010
[15]	JOHNSTONE C, SULUNGU E D. Application of neural network in prediction of temperature: a review[J]. Neural Computing and Applications, 2021, 33(18): 11487-11498. doi: 10.1007/s00521-020-05582-3
[16]	ASSELMAN A, KHALDI M, AAMMOU S. Enhancing the prediction of student performance based on the machine learning XGBoost algorithm[J]. Interactive Learning Environments, 2023, 31(6): 3360-3379. doi: 10.1080/10494820.2021.1928235
[17]	MASSAOUDI M, REFAAT S S, ABU-RUB H, et al. A hybrid Bayesian ridge regression-CWT-catboost model for PV power forecasting[C]//2020 IEEE Kansas Power and Energy Conference (KPEC). Manhattan: IEEE, 2020: 1-5.