基于频域数据注意力机制的核电厂水泵故障模式识别模型研究

刘子铭; 罗能; 艾琼

doi:10.13832/j.jnpe.2021.06.0203

基于频域数据注意力机制的核电厂水泵故障模式识别模型研究

doi: 10.13832/j.jnpe.2021.06.0203

1.
成都信息工程大学，成都，610225
2.
中国核动力研究设计院，成都，610213

详细信息

作者简介:
刘子铭（1997—），男，学士，现从事信息化管理工作，E-mail: 446046726@qq.com

中图分类号: TL353⁺.12
计量
- 文章访问数: 193
- HTML全文浏览量: 121
- PDF下载量: 37
- 被引次数: 0
出版历程
- 收稿日期: 2021-09-22
- 修回日期: 2021-10-23
- 刊出日期: 2021-12-09

Research on Fault Pattern Recognition Model of Nuclear Power Plant Water Pump Based on Frequency-Domain Data Attention Mechanism

1.
Chengdu University of Information Technology, Chengdu, 610225, China
2.
Nuclear Power Institute of China, Chengdu, 610213, China

摘要

摘要: 针对核电厂水泵共性的异常振动、转子部件摩擦与磨损等故障模式，利用水泵最容易获取的泵壳加速度信号的频域数据为输入，提出了一种结合卷积神经网络和注意力网络的频域数据注意力机制方法，并建立了核电厂水泵故障模式识别模型。研究结果表明：相对于传统方法，利用频域数据作为输入、基于频域数据注意力网络算法建立的水泵故障模式识别模型输入的数据长度更短，能够有效提升模型训练的效率，该故障模式识别模型在测试集上的故障模式识别准确率达到100%，优于其他基于深度学习算法建立的故障诊断模型，证明了本文提出方法的优势。
- 核电厂水泵 /
- 频域数据 /
- 注意力机制 /
- 故障模式识别
Abstract: In view of the common fault modes of nuclear power plant pump, such as abnormal vibration, friction and abrasion of rotor parts, etc., this paper uses frequency domain data of the acceleration signal on pump shell which is easiest to be obtained as input, proposes a new method for frequency-domain data attention mechanism which combines convolutional neural network and attention network, and establishes the recognition model of fault mode of nuclear power plant water pump. The results show that: Compared with the traditional methods, the water pump fault pattern recognition model based on frequency domain data as input and based on frequency domain data attention network algorithm has a shorter input data length and can effectively improve the efficiency of model training. The fault pattern recognition accuracy of the fault pattern recognition model on the test set is 100%, which is better than other fault diagnosis models based on deep learning algorithm, which proves the advantages of the method proposed in this paper.
- Nuclear power plant water pump /
- Frequency domain data /
- Attention mechanism /
- Fault pattern recognition

HTML全文

图 1 频域数据注意力机制结构图

$ {s_1},{s_2}, \cdots {s_{{n_{\rm{s}}}}} $—截断后的每一段多通道一维数组；h—多头自注意力网络的头数

Figure 1. Structure of Frequency Domain Attention Mechanism　　　　　

下载: 全尺寸图片幻灯片

图 2 多头自注意力网络结构图

Figure 2. Structure of Mutihead Self-Attention Network

下载: 全尺寸图片幻灯片

图 3 核电厂水泵结构及传感器布置位置

Figure 3. Pump Structure and Sensor Location of Nuclear Power Plant

下载: 全尺寸图片幻灯片

图 4 模型训练过程图

Figure 4. Model Training Process Diagram

下载: 全尺寸图片幻灯片

图 5 本文提出模型预测结果混淆矩阵

Figure 5. Prediction Result Confusion Matrix Based on Proposed Model

下载: 全尺寸图片幻灯片

表 1 水泵故障模拟试验的故障类型、故障程度及数据条数　　　　　

Table 1. Fault Type, Degree and Data Number of Pump Fault Simulation Test

编号	故障码	故障类型	故障程度	数据条数
0	00	正常		3680
1	10	导轴承磨损	轻度	3720
2	11	导轴承磨损	中度	3870
3	12	导轴承磨损	重度	3820
4	20	止推轴承磨损	轻度	3760
5	21	止推轴承磨损	中度	4020
6	22	止推轴承磨损	重度	3700
7	30	口环刮磨	中度	3770
8	31	口环刮磨	重度	3660
9	41	转子偏心	中度	3740
10	42	转子偏心	重度	3720
注：空白表示无此项

下载: 导出CSV

表 2 频域数据注意力机制模型相关超参数

Table 2. Related Hyper-Parameter of Frequency Domain Data Attention Mechanism Model

所属部分	超参数名	超参数值
输入层	数据段数目n_s	50
输入层	截段后每个数据段长度l_s	500
卷积模块	卷积1核长度/通道数	5/32
卷积模块	卷积2核长度/通道数	3/1
卷积模块	池化长度l_p	2
多头自注意力网络	h/$ {d_{{K_i}}} $/$ {d_{{V_i}}} $	2/50/30

下载: 导出CSV

表 3 各个模型在时域数据集和频域数据集上的训练信息

Table 3. Training Information of Each Model on Time-Domain Dataset and Frequency-Domain Dataset

数据集	网络模型	单步时长/ms	模型待训练参数数目	验证集平均准确率/ 标准差^①	测试集准确率
频域数据集	本文提出的	16	107943	99.68%±0.38%	100.00%
	文献[6]提出的	29	127843	99.30%±1.02%	99.79%
	文献[7]提出的	30	173043	99.47%±1.19%	99.58%
时域数据集	本文提出的	26	172943	50.77%±1.69%	51.91%
	文献[6]提出的	36	227843	79.88%±1.70%	78.18%
	文献[7]提出的	36	323043	75.15%±2.05%	71.82%
注：①采用训练第20个迭代周期以后模型在验证集上的平均准确率和标准差

下载: 导出CSV

参考文献(8)

[1]	XIA M, LI T, XU L, et al. Fault diagnosis for rotating machinery using multiple sensors and convolutional neural networks[J]. IEEE/ASME Transactions on Mechatronics, 2018, 23(1): 101-110. doi: 10.1109/TMECH.2017.2728371
[2]	ZHAO M H, ZHONG S S, FU X Y, et al. Deep residual shrinkage networks for fault diagnosis[J]. IEEE Transactions on Industrial Informatics, 2020, 16(7): 4681-4690. doi: 10.1109/TII.2019.2943898
[3]	SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]//3rd International Conference on Learning Representations. San Diego, CA, USA: ICLR, 2015: 1-14.
[4]	ZHANG W, PENG G L, LI C H, et al. A new deep learning model for fault diagnosis with good anti-noise and domain adaptation ability on raw vibration signals[J]. Sensors, 2017, 17(2): 425. doi: 10.3390/s17020425
[5]	VAN DER MAATEN L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(86): 2579-2605.
[6]	LI X, ZHANG W, DING Q. Understanding and improving deep learning-based rolling bearing fault diagnosis with attention mechanism[J]. Signal Processing, 2019(161): 136-154. doi: 10.1016/j.sigpro.2019.03.019
[7]	YANG Z B, ZHANG J P, ZHAO Z B, et al. Interpreting network knowledge with attention mechanism for bearing fault diagnosis[J]. Applied Soft Computing Journal, 2020(97): 106829. doi: 10.1016/j.asoc.2020.106829
[8]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, CA, USA: NIPS, 2017: 5998-6008.