Multivariate Dynamic Process Fault Diagnosis Based on Multi-Block Convolutional Variational Information Bottleneck
-
摘要: 针对多变量动态过程的故障诊断,采用局部提取、全局整合的特征提取策略,提出了一种多块卷积变分信息瓶颈(Multi-Block Convolutional Variational Information Bottleneck,MBCVIB)模型。首先,根据过程机理,对所有变量分块,将同一操作单元的变量划分为同一子块,再利用一维卷积神经网络(One-Dimensional Convolutional Neural Network,1-D CNN)提取过程中每个子块的局部特征,从而考虑样本间的时序相关性;然后,整合局部特征得到全局特征表示,在全局特征的基础上,根据变分信息瓶颈(Variational Information Bottleneck,VIB)原理进一步提取与故障最相关的信息;最后,采用连续搅拌釜反应器(Continuous Stirred Tank Reactor,CSTR)和田纳西-伊士曼过程(Tennessee Eastman Process,TEP)对模型的有效性进行了验证。结果显示本文模型在CSTR上实现了0.983的平均故障诊断准确率,在TEP上实现了0.955的平均故障诊断准确率。Abstract: By using the feature extraction strategy of local extraction and global integration, this paper proposes a multi-block convolutional variational information bottleneck (MBCVIB) for the fault diagnosis of multivariate dynamic processes. Firstly, according to the process mechanism, all variables are divided into sub-blocks and the variables in the same operation unit will be put into the same block. Secondly, one-dimension convolutional neural network (1-D CNN) is used to extract the local dynamic features of each operating unit in the process, which considers the temporal correlation between samples. Besides, a global feature representation is constructed by concatenating the local dynamic features of all operating units. On the basis of global features, the most relevant fault information is further extracted according to the variational information bottleneck principle. Finally, the proposed model is validated via Continuous Stirred Tank Reactor (CSTR) and Tennessee Eastman Process (TEP), which achieves an average fault diagnosis accuracy of 0.983 on CSTR and 0.955 on TEP.
-
表 1 CSTR故障描述
Table 1. Faults description in CSTR
Fault number Fault description 1 Bias in the sensor of the output temperature T 2 Bias in the sensor of the inlet temperature T0 3 Bias in inlet reactant concentration CAA 4 Drift in the sensor of CAA 5 Slow drift in reaction kinetics 6 Drift in E/R 7 Drift in Uac 8 Drift in CAA, CAS, FS, and T0 9 Step change in T0 表 2 CSTR故障诊断模型结构
Table 2. Detailed fault diagnosis model structures in CSTR
Model Structure Stacked AE Input(1×(9·10))-FC(30)-FC(20)-FC(10)-Softmax LSTM Input(9×10)-LSTM(50)-LSTM(50)-FC(30)-FC(10)-Softmax 2-D CNN Input(9×10)-Conv2D(64)-Conv2D(64)-Maxpool(2×2)-Conv2D(128)-FC(128)-FC(10)-Softmax MBCVIB Input(9×10)- (Conv(60)·2)·4-Concatenation- FC(100)×2-FC(100)-FC(10)-Softmax 表 3 CSTR故障诊断模型参数
Table 3. Fault diagnosis model parameters in CSTR
Model Window width Activation function Convolutional kernel Learning rate Batch size Epoch Stacked AE 10 SELU − 1×10−3 24 20 LSTM 10 SELU − 1×10−3 24 20 2-D CNN 10 SELU 3×3 1×10−3 24 20 MBCVIB 10 SELU (Vb×3) 1×10−3 24 20 表 4 CSTR分块结果
Table 4. Block division result in CSTR
Block Process variables 1 T0, CAS, FS 2 FA, CAA 3 FC, TC 4 CA, T 表 5 不同模型在CSTR上的故障诊断准确率
Table 5. Fault diagnosis accuracy of different models on CSTR
Fault Accuracy SVM Stacked AE LSTM 2-D CNN MBCVIB Normal 0.299 0.688 0.480 1.000 0.992 1 0.573 0.938 0.830 0.919 1.000 2 0.992 0.994 0.992 0.992 0.994 3 0.998 0.993 0.977 0.996 0.997 4 0.525 0.710 0.931 0.850 0.876 5 0.973 0.973 0.984 0.983 0.995 6 0.912 0.958 0.956 0.968 0.998 7 0.811 0.938 0.932 0.965 0.985 8 0.994 0.994 0.989 0.994 0.998 9 0.945 0.961 0.988 0.960 1.000 Average 0.802 0.915 0.906 0.963 0.983 表 6 TEP实验时延数据维度
Table 6. Time-delay data dimension of TEP
Status index Training Testing Normal 481×(52×20) 981×(52×20) Fault 1~5, 7~20 461×(52×20) 781×(52×20) Fault 6 121×(52×20) 121×(52×20) 表 7 TEP故障诊断模型结构
Table 7. Detailed fault diagnosis model structures in TEP
Model Structure Stacked AE Input((1×(52·20))-FC(500)-FC(250)-FC(18)-Softmax LSTM Input(52×20)-LSTM(128)-LSTM(128)-FC(128)-FC(64)-FC(18)-Softmax 2-D CNN Input(52×20)-Conv2D(64)-Conv2D(256)-Maxpool(2×2)-Conv2D(128)-Maxpool(2×2)-FC(384)-FC(18)-Softmax MBCVIB Input(52×20)- (Conv(60)·6)×4-Concatenation-FC(100)×2 -FC(100)-FC(18)-Softmax 表 8 TEP故障诊断模型参数
Table 8. Detailed fault diagnosis model parameters in TEP
Model Window width Activation function Convolutional kernel Learning rate Batch size Epoch Stacked AE 20 SELU − ${\rm{1}} \times {\rm{1}}{{\rm{0}}^{ - 3}}$ 24 20 LSTM 20 SELU − ${\rm{1}} \times {\rm{1}}{{\rm{0}}^{ - 3}}$ 24 20 2-D CNN 20 SELU 3×3 ${\rm{1}} \times {\rm{1}}{{\rm{0}}^{ - 3}}$ 24 20 MBCVIB 20 SELU (${V_b}$×3) ${\rm{1}} \times {\rm{1}}{{\rm{0}}^{ - 3}}$ 24 20 表 9 TEP中每个操作单元的变量
Table 9. Variables of each operation unit in TEP
Operation unit Process variables Reactor XMEAS(1, 2, 3, 4, 5, 6, 7, 8, 9, 21, 23, 24, 25, 26, 27, 28); XMV(1, 2, 3, 10) Condenser XMV(11) Recycle compressor XMEAS(20); XMV(5) Separator XMEAS(10, 11, 12, 13, 14, 22, 29, 30, 31, 32, 33, 34, 35, 36); XMV(6, 7) Stripper XMEAS(15, 16, 17, 18, 19, 37, 38, 39, 40, 41; XMV(4, 8, 9) 表 10 TEP分块结果
Table 10. Block division result of TEP
Block Process variables 1 XMEAS(1, 2, 3, 4, 5, 6, 7, 8, 9, 21, 23, 24, 25, 26, 27, 28); XMV(1, 2, 3, 10) 2 XMV(5, 11); XMEAS(20) 3 XMEAS(10, 11, 12, 13, 14, 22, 29, 30, 31, 32, 33, 34, 35, 36); XMV(6, 7) 4 XMEAS(15, 16, 17, 18, 19, 37, 38, 39, 40, 41); XMV(4, 8, 9) 表 11 不同模型在TEP中的故障诊断准确率
Table 11. Fault diagnosis accuracy of different models on TEP
Fault Accuracy SVM Stacked AE LSTM 2-D CNN MBCVIB Normal 0.428 0.330 0.243 0.414 0.942 1 1.000 0.997 1.000 1.000 0.997 2 0.996 0.997 0.995 0.995 0.990 4 1.000 0.996 0.858 0.319 1.000 5 1.000 0.992 0.855 0.223 0.996 6 1.000 1.000 0.975 1.000 1.000 7 1.000 1.000 0.967 1.000 1.000 8 0.406 0.530 0.535 0.738 0.813 10 0.269 0.272 0.073 0.426 0.887 11 0.1 0.460 0.538 0.876 0.982 12 0.690 0.640 0.697 0.817 0.942 13 0.877 0.916 0.939 0.917 0.939 14 1 1 0.999 0.999 1.000 16 0.347 0.384 0.292 0.364 0.955 17 0.876 0.899 0.867 0.841 0.941 18 0.932 0.933 0.935 0.931 0.949 19 0.620 0.444 0.670 0.887 0.947 20 0.70 0.835 0.618 0.694 0.910 Average 0.736 0.757 0.725 0.747 0.955 -
[1] 薛敏, 杨健, 谭帅, 等. 基于多数据结构的集成质量监控方法[J]. 华东理工大学学报(自然科学版), 2019, 45(6): 938-945. [2] 吕铮, 杨健, 侍洪波, 等. 基于T-TSNPR的动态过程质量监控[J]. 华东理工大学学报(自然科学版), 2019, 45(6): 946-953. [3] YANG J, LYU Z, SHI H, et al. Performance monitoring method based on balanced partial least square and statistics pattern analysis[J]. ISA Transactions, 2018, 81: 121-131. doi: 10.1016/j.isatra.2018.07.038 [4] TAO Y, SHI H B, SONG B, et al. Parallel quality-related dynamic principal component regression method for chemical process monitoring[J]. Journal of Process Control, 2019, 73: 33-45. doi: 10.1016/j.jprocont.2018.08.009 [5] AMIN M T, IMTIAZ S, KHAN F. Process system fault detection and diagnosis using a hybrid technique[J]. Chemical Engineering Science, 2018, 189: 191-211. doi: 10.1016/j.ces.2018.05.045 [6] ZHANG Y. Fault detection and diagnosis of nonlinear processes using improved kernel independent component analysis (KICA) and support vector machine (SVM)[J]. Industrial & Engineering Chemistry Research, 2008, 47(18): 6961-6971. [7] RATO T, RESI M, SCHMITT E, et al. A systematic comparison of PCA-based statistical process monitoring methods for high-dimensional, time-dependent processes[J]. AIChE Journal, 2016, 62(5): 1478-1493. doi: 10.1002/aic.15062 [8] ZHU Z B, SONG Z H. Fault diagnosis based on imbalance modified kernel Fisher discriminant analysis[J]. Chemical Engineering Research & Design, 2010, 88(8): 936-951. [9] YIN G, ZHANG Y T, LI Z N, et al. Online fault diagnosis method based on Incremental support vector data description and extreme learning machine with incremental output structure[J]. Neurocomputing, 2014, 128: 224-231. doi: 10.1016/j.neucom.2013.01.061 [10] 张祥, 崔哲, 董玉玺, 等. 基于VAE-DBN的故障分类方法在化工过程中的应用[J]. 过程工程学报, 2018, 18(3): 590-594. [11] 王翔, 任佳. 基于多注意力机制的深度神经网络故障诊断算法[J]. 浙江理工大学学报(自然科学版), 2020, 43(2): 224-231. [12] 王翔, 柯飂挺, 任佳. 样本重构多尺度孪生卷积网络的化工过程故障检测[J]. 仪器仪表学报, 2019, 40(11): 184-191. [13] ZHU J Z, SHI H B, SONG B, et al. Deep neural network based recursive feature learning for nonlinear dynamic process monitoring[J]. The Canadian Journal of Chemical and Engineering, 2020, 98(4): 919-933. doi: 10.1002/cjce.23669 [14] ZHANG Z H, JIANG T, ZHAN C J, et al. Gaussian feature learning based on variational autoencoder for improving nonlinear process monitoring[J]. Journal of Process Control, 2019, 75: 136-155. doi: 10.1016/j.jprocont.2019.01.008 [15] 于志强, 余正涛, 黄于欣, 等. 基于变分信息瓶颈的半监督神经机器翻译[J/OL]. 自动化学报, 2020, 46: 1-12 [2020-10-20]. https://doi.org/10.16383/j.aas.c190477. [16] ALEMI A, FISCHER I, DILLON J V, et al. Deep variational information bottleneck[EB/OL]. (2016-12-01) [2020-10-21]. https://www.arxiv-vanity.com/papers/1612.00410. [17] WU H, ZHAO J. Deep convolutional neural network model based chemical process fault diagnosis[J]. Computers & Chemical Engineering, 2018, 115: 185-197. [18] CHEN S M, YU J B, WANG S J. One-dimensional convolutional auto-encoder-based feature learning for fault diagnosis of multivariate processes[J]. Journal of Process Control, 2020, 87: 54-67. doi: 10.1016/j.jprocont.2020.01.004 [19] YIN J, YAN X F. Mutual information−dynamic stacked sparse autoencoders for fault detection[J]. Industrial & Engineering Chemistry Research, 2019, 58: 21614-21624. [20] TISHBY N, ZASLAVSKY N. Deep learning and the information bottleneck principle[C]// IEEE Information Theory Workshop (ITW). Jerusalem: IEEE, 2015: 1-5. [21] ALCALA C F, QIN S J. Reconstruction-based contribution for process monitoring[J]. Automatica, 2009, 45(7): 1593-1600. doi: 10.1016/j.automatica.2009.02.027 [22] 杨健, 宋冰, 谭帅, 等. 时序约束NPE算法在化工过程故障检测中的应用[J]. 化工学报, 2016, 67(12): 5131-5139. [23] ZHANG Z, ZHAO J S. A deep belief network based fault diagnosis model for complex chemical processes[J]. Computers & Chemical Engineering, 2017, 107: 395-407. [24] CHIANG L H, RUSSELL E L, BRAATZ R D. Fault Detection and Diagnosis in Industrial Systems[M]. London: Springer, 2001. [25] HINTON G E. Visualizing high-dimensional data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(2): 2579-2605. -