多输入混合深度学习网络的健康因子构建方法

doi:10.3969/j.issn.1001-506X.2020.10.30

摘要/Abstract

摘要：

工业大数据具有多类型、多维度的特点,单一类型的深度学习网络结构无法充分提取数据中包含的性能退化特征。针对上述问题,提出一种可同时融合处理一维时间序列数据和二维图像数据的多输入混合深度学习网络健康因子构建模型。根据输入数据类型特点搭建的混合深度学习网络包含时间特征提取层、空间特征提取层、融合层和全连接层。时间特征提取层主要由叠加的多个长短时记忆(long short-term memory, LSTM)网络构成,用于提取一维时间序列数据中蕴含的时间特征。空间特征提取层主要由深度卷积神经网络(deep convolutional neural network, DCNN)构成,用于提取二维图像数据中的空间特征。融合层将时间特征与空间特征融合。最后,利用全连接层输出健康因子值。滚动轴承全寿命周期试验结果表明:本文提出的多输入混合深度学习网络的健康因子构建方法能够深度挖掘不同数据类型包含的性能退化信息,有效降低了性能退化曲线的离散性,有助于减小剩余寿命预测结果的不确定性,同时在一定程度上提高了单调性和趋势性,提高了剩余寿命预测精度。

关键词: 工业大数据, 深度学习, 健康因子, 评估准则, 离散性

Abstract:

Due to the multi-type and multi-dimension characteristics of the industrial big data, the single deep learning network can not extract the performance degradation characteristics effectively from the data. To solve the problem, the health indicator construction model based on the multi-input hybrid deep learning network is proposed, which can fuse and process the one-dimensional time series data and two-dimensional pictorial data simultaneously. According to the characteristic of the input data, the constructed hybrid deep learning network mainly is consisted of the time features extraction layer, spatial features extraction layer, fusion layer and fully connected layer. The time features extraction layer is constructed by stacking the multiple long short-term memory (LSTM) network to extract temporal features from the one-dimensional time series data. The spatial features extraction layer is constructed by deep convolutional neural network (DCNN) which is used to extract the spatial features from the two-dimensional pictorial data. The fusion layer is designed to fuse the time features and the spatial features, and the fully connected layer output the health indicator which could reflect the performance degradation process. The rolling element bearing life-cycle experiment indicates that the health indicator construction method based on the multi-input hybrid deep learning network could extracts and fuses different data type and reduces the discreteness of the degradation curve, which reduces the uncertainty of the remaining useful life prediction results. What is more, the proposed method could promote the monotonicity and trendability. The remaining useful life prediction result shows that the proposed health indicator could promote the remaining useful life prediction accuracy effectively.

Key words: industrial big data, deep learning, health indicator, evaluation criterion, discreteness

中图分类号:

TN911.72

孙世岩, 张钢, 田福庆, 梁伟阁. 多输入混合深度学习网络的健康因子构建方法[J]. 系统工程与电子技术, 2020, 42(10): 2390-2398.

Shiyan SUN, Gang ZHANG, Fuqing TIAN, Weige LIANG. Health indicator construction method of multi-input hybrid deep learning network[J]. Systems Engineering and Electronics, 2020, 42(10): 2390-2398.

图/表 13

图1

图2

图3

图4

图5

图6

表1

图7

图8

图9

图10

表2

表3

参考文献 18

1	INSUN S , JUNMIN L , JUM Y L , et al. A framework for prognostics and health management applications toward smart manufacturing systems[J]. International Journal of Precision Engineering and Manufacturing-green Technology, 2018, 5 (4): 535- 554. doi: 10.1007/s40684-018-0055-0
2	LIN Y H , LI X D , HU Y , et al. Deep diagnostics and prognostics: an integrated hierarchical learning framework in PHM applications[J]. Applied Soft Computing, 2018, 72, 555- 564. doi: 10.1016/j.asoc.2018.01.036
3	LEI Y G , LI N P , GUO L , et al. Machinery health prognostics: a systematic review from data acquisition to RUL prediction[J]. Mechanical Systems and Signal Processing, 2018, 104, 799- 834. doi: 10.1016/j.ymssp.2017.11.016
4	ZHAO R , YAN R Q , CHEN Z H , et al. Deep learning and its applications to machine health monitoring[J]. Mechanical Systems and Signal Processing, 2019, 115, 213- 237. doi: 10.1016/j.ymssp.2018.05.050
5	丁红卫, 万良, 龙廷艳. 深度自编码网络在入侵检测中的应用研究[J]. 哈尔滨工业大学学报, 2019, 51 (5): 185- 194.
	DING H W , WAN L , LONG T Y . Rearch on the application of deep auto-encoder network in intrusion detection[J]. Journal of Harbin Institute of Technology, 2019, 51 (5): 185- 194.
6	GUO L , LI N P , JIA F , et al. A recurrent neural network based health indicator for remaining useful life prediction[J]. Neurocomputing, 2017, 240, 98- 109. doi: 10.1016/j.neucom.2017.02.045
7	TANG G, ZHOU Y G, WANG H Q, et al. Prediction of bearing performance degradation with bottleneck fearure based on LSTM network[C]//Proc.of the IEEE International Instrumentation and Measurement Technology Conference, 2018.
8	REN L , SUN Y Q , WANG H , et al. Prediction of bearing remaining useful life with deep concolutional neural network[J]. Special Section on Cyber-physical-social Computing and Networking, 2018, 6, 13041- 13049.
9	YOO Y J , BAEK J G . A novel image feature for the remaining useful lifetime prediction of bearings based on continuous wavelet transform and convolutional neural network[J]. Applied Sciences, 2018, 8, 1102- 1119. doi: 10.3390/app8071102
10	ZHAO R , YAN R Q , WANG J J , et al. Learning to monitor machine health with convolutional bi-directional Lstm networks[J]. Sensors, 2017, 17 (2): 273- 290.
11	HINCHI A Z , TKIOUAT M . Rolling element bearing remaining useful life estimation based on a convolutional long-short-term memory network[J]. Procedia Computer Science, 2018, 127, 123- 132. doi: 10.1016/j.procs.2018.01.106
12	GERS F A , SCHRAUDOLPH N N , SCHMIDHUBER J . Learning precise timing with LSTM recurrent networks[J]. Journal of Machince Learning Rearch, 2002, 3, 115- 143.
13	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[C]//Proc.of the Neural Information Processing Systems, 2012: 1097-1105.
14	NECTOUX P, GOURIVEAU R, MEDJAHER K, et al. PRONOSTIA: an experimental platform for bearings accelerated degradation tests[C]//Proc.of the IEEE International Conference on Prognostics and Health Management, 2012.
15	ALI A D , SOHEIL Z , AMIR A , et al. A multimodal and hybrid deep neural network model for remaining useful life estimation[J]. Computers in Industrial, 2019, 108, 186- 196. doi: 10.1016/j.compind.2019.02.004
16	KRIZHEVSKY A , SUTSKEVER I , HINTON G . Imagenet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems, 2012, 25 (2): 1097- 1105.
17	赵光权, 刘小勇, 姜泽东, 等. 基于深度学习的轴承健康因子无监督构建方法[J]. 仪器仪表学报, 2018, 39 (6): 82- 88.
	ZHAO G Q , LIU X Y , JIANG Z D , et al. Unsupervised health indicator of bearing based on deep learning[J]. Chinese Journal of Scientific Instrument, 2018, 39 (6): 82- 88.
18	HU Y, PALME T, FINK O. Deep health indicator extraction: a method based on auto-encoders and extreme learning machines[C]//Proc.of the Annual Conference of the Prognostics and Health Management Society, 2016.

网络层	参数
网络层	输入图像大小/(像素×像素)	通道数/个	卷积核大小/(个×个)	步长/单元	单元数/个
输入层	[100×100]	3	-	-	-
卷积层1	-	96	[11×11]	4	-
池化层1	-	96	[3×3]	2	-
卷积层2	-	256	[5×5]	1	-
池化层2	-	256	[3×3]	2	-
卷积层3	-	384	[3×3]	1	-
卷积层4	-	384	[3×3]	1	-
卷积层5	-	256	[3×3]	1	-
池化层5	-	256	[3×3]	2	-
CNN展平层	-	7 424(6 400+1 024)	-	-	-
LSTM层1	-	-	-	-	80
LSTM层2	-	-	-	-	60
LSTM层3	-	-	-	-	30
LSTM展平层	-	-	-	-	30
CNN+LSTM连接层	-	-	-	-	7 454(7 424+30)
全连接层1	-	4 096	-	-	-
全连接层2	-	1 000	-	-	-
输出层	-	1	-	-	-

轴承	RMS			PCA			ELM-AE			SDAE-SOM			HYB-HI
轴承	趋势性	单调性	离散性	趋势性	单调性	离散性	趋势性	单调性	离散性	趋势性	单调性	离散性	趋势性	单调性	离散性
Bearing1_3	0.77	0.14	0.15	0.78	0.15	0.56	0.77	0.14	0.44	0.67	0.12	0.62	0.96	0.88	0.40
Bearing1_4	0.32	0.16	0.33	0.32	0.15	0.53	0.31	0.15	0.70	0.41	0.17	0.76	0.95	0.78	0.58
Bearing1_5	0.16	0.12	0.21	0.23	0.11	0.66	0.28	0.12	0.63	0.26	0.14	0.55	0.93	0.77	0.48
Bearing1_6	0.10	0.10	0.12	0.16	0.11	0.72	0.18	0.11	0.86	0.31	0.13	0.79	0.86	0.53	0.60
Bearing1_7	0.23	0.11	0.26	0.33	0.13	0.61	0.34	0.10	0.77	0.35	0.14	0.58	0.95	0.87	0.57

测试数据集	当前时刻/s	实际剩余寿命/s	预测剩余寿命/s
测试数据集	当前时刻/s	实际剩余寿命/s	HYB-HI	RMS	PCA	ELM-AE	LSTM-HI	CNN-HI	SOM-HI
Bearing1_3	18 010	5 730	4 752	5 970	3 250	4 683	3 250	4 731	5 790
Bearing1_4	11 390	2 900	2 170	1 200	1 100	2 438	1 100	2 590	410
Bearing1_5	23 010	1 610	1 210	5 040	1 980	4 237	1 980	3 996	6 080
Bearing1_6	23 010	1 460	1 230	1 230	1 150	1 848	1 150	1 744	1 180
Bearing1_7	15 010	7 570	7 050	9 120	6 220	6 035	6 220	6 090	8 110
MAE	-	-	571.6	1 430	1 262	1 211.8	1 262	1 091.8	1 568
NRMSE	-	-	0.191 4	0.410 7	0.552 2	0.380 5	0.552 2	0.351 4	0.534 2

[1]	韩啸, 陈世文, 陈蒙, 杨锦程. 基于互易点学习的LPI信号开集识别[J]. 系统工程与电子技术, 2022, 44(9): 2752-2759.
[2]	张立民, 谭凯文, 闫文君, 张聿远. 基于多级跳线残差网络的雷达辐射源识别[J]. 系统工程与电子技术, 2022, 44(7): 2148-2156.
[3]	金国栋, 薛远亮, 谭力宁, 许剑锟. 基于孪生神经网络的目标跟踪算法进展研究[J]. 系统工程与电子技术, 2022, 44(6): 1805-1822.
[4]	赵晓枫, 徐叶斌, 吴飞, 牛家辉, 蔡伟, 张志利. 基于全局感知机制的地面红外目标检测方法[J]. 系统工程与电子技术, 2022, 44(5): 1461-1467.
[5]	邹虹, 白陈阳, 何鹏, 崔亚平, 王汝言, 吴大鹏. 基于分布式深度学习的边缘服务放置策略[J]. 系统工程与电子技术, 2022, 44(5): 1728-1737.
[6]	陈冬, 句彦伟. 基于语义分割实现的SAR图像舰船目标检测[J]. 系统工程与电子技术, 2022, 44(4): 1195-1201.
[7]	孙晶明, 虞盛康, 孙俊. 基于深度学习的HRRP识别姿态敏感性分析[J]. 系统工程与电子技术, 2022, 44(3): 802-807.
[8]	姚云翔, 陈莹. 注意力机制下双模态交互融合的目标跟踪网络[J]. 系统工程与电子技术, 2022, 44(2): 410-419.
[9]	高涌荇, 王旭东, 汪玲, 朱岱寅, 郭军, 孟凡旺. 基于RCNN的双极化气象雷达天气信号检测[J]. 系统工程与电子技术, 2022, 44(11): 3380-3387.
[10]	周毅恒, 杨军, 夏赛强, 吕明久. 闪烁现象下旋翼目标微动参数估计方法[J]. 系统工程与电子技术, 2022, 44(1): 54-63.
[11]	付蓉, 黄天耀, 刘一民. 基于深度学习的捷变相参雷达1-Bit块稀疏重构[J]. 系统工程与电子技术, 2022, 44(1): 70-75.
[12]	张聿远, 张立民, 闫文君. 基于互相关特征图和扩张稠密卷积网络的SFBC-OFDM识别方法[J]. 系统工程与电子技术, 2021, 43(9): 2657-2664.
[13]	王彩云, 李阳雨, 李晓飞, 王佳宁, 魏文怡. 基于稀疏性深度学习的航拍图像超分辨重构[J]. 系统工程与电子技术, 2021, 43(8): 2045-2050.
[14]	黄月平, 李小锋, 杨小冈, 齐乃新, 卢瑞涛, 张胜修. 基于相关滤波的视觉目标跟踪算法新进展[J]. 系统工程与电子技术, 2021, 43(8): 2051-2065.
[15]	王彬, 王国宇. 基于改进的深度学习网络的SAR图像瞬时海岸线自动提取算法[J]. 系统工程与电子技术, 2021, 43(8): 2108-2115.