Research on Fault Diagnosis of Improved Kernel Fisher Based on Mahalanobis Distance in the Field of Chemical Industry
-
摘要: 针对化工故障诊断数据存在非线性分布、 数据类别复杂、数据量大且故障特征不易区分等问题, 本文提出一种基于马氏距离的改进核Fisher故障诊断方法(Mahalanobis distance-based kernel Fisher discrimination, MKFD). 首先, 针对数据非线性分布的特点, 本文将核Fisher判别分析算法改进, 改进后的算法可以有效解决原始样本在投影后出现的因类间距离差异过大、类内距离不够紧凑造成的样本混叠现象. 除此之外, 利用Euclidean距离对类间距做加权处理时, 用组平均距离取代质心距离, 提升了运算效率, 降低了时间复杂度; 其次, 根据高斯径向基核函数(Radial basis function, RBF)在MKFD中所呈现出的诊断精度的规律, 本文采用一种新的核参数选择方法: 区间三分法, 用以取代在实际应用中依靠经验的交叉验证法; 最后, 本文采用马氏距离对故障进行分类, 基于田纳西伊—斯特曼过程(Tennessee-Eastman, TE)数据将本方法与其他改进核Fisher算法进行仿真验证对比. 结果表明新提出MKFD算法不仅可以提高故障诊断的运算效率, 也能有效提高诊断的精度.Abstract: Aiming at the problems of the non-linear distribution, complex category, large amount of fault diagnosis data in chemical industry and the difficulty of distinguishing fault features, a improved kernel Fisher fault diagnosis method based on Mahalanobis distance is proposed in this paper. Firstly, due to the data with non-linear property, a new improved kernel Fisher discriminant analysis method is proposed, which can effectively solve the sample aliasing phenomenon caused by large difference between classes and insufficient compact distance between classes after projection of original samples. In addition, using the Euclidean distance in class spacing, the group average distance is used to replace the center of mass distance, which improves the efficiency of operation and reduces the time complexity. Secondly, according to the rule of diagnostic accuracy presented by the (RBF) in Fisher discriminant analysis (MKFD), a new method, interval “three-point method”, of selecting nuclear parameters is proposed in this paper, which is used to replace the cross-validation method relying on experience in practical application. Finally, faults are classified based on Mahalanobis distance using Tennessee-Sterman process. The proposed method is compared with other improved kernel Fisher algorithm. The results show that (MKFD) can not only improve the calculation efficiency of fault diagnosis, but also improve the accuracy of diagnosis.
-
Key words:
- Kernel Fisher /
- fault diagnosis /
- interval three-point method /
- Tennessee-Sterman process /
- optimization
-
表 1 故障类型描述
Table 1 Description of the selected fault sample sets
Fault Number Fault description Fault type 3 物料 D 的温度的异变 阶跃 4 反应器冷却水入口温度的异变 阶跃 5 泠凝器冷却水入口温度的异变 阶跃 7 物料 C 压力下降 阶跃 表 2 选取不同核参数σ下故障诊断的准确率 (KFD)
Table 2 The fault diagnosis accuracy based on different kernel parameter σ(KFD)
The value of the parameter σ Test accuracy (%) The value of the parameter σ Test accuracy (%) 0.1 25 30 81.25 0.2 30.31 40 80.94 0.8 50 70 53.13 2 66.88 90 51.56 4 75.63 100 45 8 78.44 160 43.75 10 79.38 180 33.44 表 3 利用区间三分法求解最优核参数σ对应的故障诊断的准确率 (KFD)
Table 3 The accuracy of fault diagnosis of optimal kernel parameter by using the interval three-part method (KFD)
迭代次数 对应区间 三分点 1 三分点 2 三分点 3 三分点 4 ${X_1}$ $D({X_1})$ ${X_2}$ $D({X_2})$ ${X_3}$ $D({X_3})$ ${X_3}$ $D({X_4})$ 1 [1, 100] 1 50 % 34 79 % 67 51 % 100 45 % 2 [1, 67] 1 50 % 23 80 % 45 73.8 % 67 51 % 3 [1, 45] 1 50 % 15.7 79.4 % 30.3 81.25 % 45 73.8 % 4 [15.7, 45] 15.7 79.4 % 25.5 80 % 35.2 78.8 % 45 73.8 % 5 [15.7, 35.2] 15.7 79.4 % 22.2 80.3 % 28.7 80.4 % 35.2 78.8 % 6 [22.2, 35.2] 22.2 80.3 % 26.5 80 % 30.9 81.25 % 35.2 78.8 % 表 4 KFD算法和MKFD算法中不同核参数的故障诊断结果
Table 4 The fault diagnosis with different kernel parameters in KFD algorithm and MKFD algorithm
The value of the
parameter σ in KFDTrain
accuracy (%)Test
accuracy (%)The value of the
parameter σ in MKFDTrain
accuracy (%)Test
accuracy (%)0.1 100 25 0.1 100 25 1 100 50 1 100 50 10 99.8 79.4 4 100 76.9 30 99.8 81.3 8 100 99.69 60 70.5 44.7 12 99.9 92.5 90 27.7 25.3 16 99.9 80.6 表 5 选取不同核参数σ下故障诊断的准确率(按照区间三分法做纵向表)
Table 5 The fault diagnosis accuracy based on different kernel parameters σ (Make the longitudinal table according to the interval three-part method)
Ionosphere Breast cancer The value of the parameter σ Test accuracy (%) The value of the parameter σ Test accuracy (%) 1 78.9 1 31.7 34 91.6 149 95.1 49 92 223 94.9 56 92.4 248 95.4 59 92.8 297 95.4 63 92.8 334 95.4 67 92.4 346 95.4 68 92 445 94.6 78 90.8 667 94 100 86.1 1000 93.2 表 6 区间三分法迭代求解最优核参数σ (MKFD)
Table 6 The iterative solution of the optimal kernel parameters σ using interval partition method
迭代次数 对应区间 三分点 1 三分点 2 三分点 3 三分点 4 ${X_1}$ $D({X_1})$ ${X_2}$ $D({X_2})$ ${X_3}$ $D({X_3})$ ${X_3}$ $D({X_4})$ 1 [1, 100] 1 50.9 % 34 60.6 % 67 57.5 % 100 58.1 % 2 [1, 67] 1 50.9 % 23 76.6 % 45 58.1 % 67 57.5 % 3 [1, 45] 1 50 % 15.7 96.3 % 30.3 63.8 % 45 58.1 % 4 [1, 30.3] 1 50 % 10.8 99.69 % 20.5 84.69 % 30.3 63.8 % 5 [1, 20.5] 1 50 % 7.5 99.38 % 14 97.81 % 20.5 84.69 % 6 [1, 14] 1 50 % 5.3 81.56 % 9.7 99.69 % 14 97.81 % 表 7 交叉验证法选取不同核参数σ下故障诊断的准确率(FDGLPP)
Table 7 The fault diagnosis accuracy based on different kernel parameters σ by cross validation method
The value of the
parameter σTest
accuracy (%)The value of the
parameter σTest
accuracy (%)The value of the
parameter σTest
accuracy (%)0.1 25 0.5 68.13 3 55.31 1 52.19 5 75.31 6 79.38 50 28.44 25 25.0 9 99.69 100 41.25 50 28.44 12 25.0 500 39.06 75 34.69 15 55.94 1000 38.75 95 40.0 18 25.0 表 8 四种模型的故障诊断结果与运行时间
Table 8 Fault diagnosis results and running time of the four models
Model Optimal value of parameter σ Test accuracy (%) Test time (s) KFD 30 81.25 3.90072 CKFD 8 97.81 4.14769 FDGLPP 10 99.69 9.30612 MKFD 9 99.69 3.86806 -
[1] 张妮, 车立志, 吴小进. 基于数据驱动的故障诊断技术研究现状及展望. 计算机科学, 2017, 44(S1): 47−52Zhang Ni, Che Li-Zhi, Wu Xiao-Jin. Present situation and prospect of data-driven based fault diagnosis technique. Computer Science, 2017, 44(S1): 47−52 [2] 郭一帆, 唐家银. 基于机器学习算法的寿命预测与故障诊断技术的发展综述. 计算机测量与控制, 2019, 27(3): 13−19Guo Yi-Fan, Tang Jia-Yin. A review of the development of life prediction and fault diagnosis technology based on machine learning algorithm. Computer Measurement & Control, 2019, 27(3): 13−19 [3] Zhang Y, Zhang Y, Zhang J, et al. Collaborative representation cascade for single-image super-resolution. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2017: 1−16 [4] Kock N, Hadaya P. Minimum sample size estimation in PLS‐SEM: The inverse square root and gamma‐exponential methods. Information Systems Journal, 2018, 28(1): 227−261 doi: 10.1111/isj.12131 [5] Maisinger K, Hobson M P, Lasenby A N. A maximum entropy method for reconstructing interferometer maps of fluctuations in the cosmic microwave background radiation. Monthly Notices of the Royal Astronomical Society, 2018, 290(2): 313−326 [6] Hwang J S, Noh J T, Lee S H, et al. Experimental verification of modal identification of a high-rise building using independent component analysis. International Journal of Concrete Structures and Materials, 2019, 13(1): 4 doi: 10.1186/s40069-018-0319-7 [7] 胡志新. 基于深度学习的化工故障诊断方法研究. 杭州电子科技大学, 中国, 2018Hu Zhi-Xin. Research on Chemical Fault Diagnosis Methods Based on Deep Learning [Master thesis], Hangzhou Dianzi University, China 2018 [8] Mika S, Ratsch G, Weston J, et al. Fisher discriminant analysis with kernel. In: Proceedings of the Neural Networks for Signal Processing IX, the 1999 IEEE Signal Processing Society Workshop. Madison, WI, USA: IEEE, 1999: 41−48 [9] 张曦, 赵旭, 刘振亚, 等. 基于核Fisher子空间特征提取的汽轮发电机组过程监控与故障诊断. 中国电机工程学报, 2007, (20): 1−6 doi: 10.3321/j.issn:0258-8013.2007.20.001Zhang Xi, Zhao Xu, Liu Zhen-Ya, et al. Process monitoring and fault diagnosis of turbine generator unit based on feature extraction in kernel Fisher subspace. Proceedings of the CSEE, 2007, (20): 1−6 doi: 10.3321/j.issn:0258-8013.2007.20.001 [10] 马立玲, 徐发富, 王军政. 一种基于改进核Fisher的故障诊断方法. 化工学报, 2017, 68(3): 1041−1048Ma Li-Ling, Xu Fa-Fu, Wang Jun-Zheng. A fault diagnosis method based on improved kernel Fisher. CIESC Journal, 2017, 68(3): 1041−1048 [11] Chikr-Elmezouar Z, Almanjahie I M, Laksaci A, et al. FDA: strong consistency of the kNN local linear estimation of the functional conditional density and mode. Journal of Nonparametric Statistics, 2019, 31(1): 175−195 doi: 10.1080/10485252.2018.1538450 [12] 李普煌, 李敏, 范新南, 等. 迭代分析相对密度的高光谱异常检测. 中国图象图形学报, 2018, 23(2): 219−228 doi: 10.11834/jig.170243Li Pu-Huang, Li Min, Fan Xin-Nan, et al. Hyperspectral anomaly detection algorithm based on iterative analysis with relative density. Journal of Image and Graphics, 2018, 23(2): 219−228 doi: 10.11834/jig.170243 [13] 赵忠盖, 刘飞. 基于马氏距离统计监控指标的应用研究. 自动化学报, 2008, 34(4): 493−495Zhao Zhong-Gai, Liu Fei. Application research of statistical monitoring index based on Mahalanobis distance. Acta Automatica Sinica, 2008, 34(4): 493−495 [14] 卫芬. 旋转机械多传感器信息融合智能故障诊断关键技术研究. 哈尔滨工业大学, 中国, 2018Wei Fen. Research on Key Technologies of Intelligent Fault Diagnosis Based on Multi-sensor Information Fusion for Rotating machinery [Ph. D. dissertation]. Harbin Institute of Technology, China, 2018 [15] 杜伟, 房立清, 齐子元. 一种邻域自适应半监督局部Fisher判别分析算法. 计算机应用研究, 2019, 36(1): 105−108Du Wei, Fang Li-Qing, Qi Zi-Yuan. Neighborhood adaptive semi-supervised local Fisher discriminant analysis algorithm. Application Research of Computers, 2019, 36(1): 105−108 [16] 杨武夷, 梁伟, 辛乐, 等. 子空间半监督Fisher判别分析. 自动化学报, 2009, 35(12): 1513−1519 doi: 10.1016/S1874-1029(08)60120-2Yang Wu-Yi, Liang Wei, Xin Le, et al. Subspace semi-supervised Fisher discriminant analysis. Acta Automatica Sinica, 2009, 35(12): 1513−1519 doi: 10.1016/S1874-1029(08)60120-2 [17] 郑建炜, 王万良, 姚晓敏, 等. 张量局部Fisher判别分析的人脸识别. 自动化学报, 2012, 38(9): 1485−1495 doi: 10.3724/SP.J.1004.2012.01485Zheng Jian-Wei, Wang Wan-Liang, Yao Xiao-Min, et al. Face recognition using tensor local Fisher discriminant analysis. Acta Automatica Sinica, 2012, 38(9): 1485−1495 doi: 10.3724/SP.J.1004.2012.01485 [18] 温廷新, 于凤俄. 基于KPCA-Fisher判别分析的煤炭自燃预测研究. 矿业安全与环保, 2018, 45(2): 49−53 doi: 10.3969/j.issn.1008-4495.2018.02.011Wen Ting-Xin, Yu Feng-E. Research on prediction of coal spontaneous combustion based on KPCA-Fisher discriminant analysis. Mining Safety & Environmental Protection, 2018, 45(2): 49−53 doi: 10.3969/j.issn.1008-4495.2018.02.011 [19] Wen T, Jia Y, Huang D, et al. Feature extraction of electronic nose signals using QPSO-based multiple KFDA signal processing. Sensors, 2018, 18(2): 388 doi: 10.3390/s18020388 [20] 刘廷瑞, 常林. 弯扭耦合风力机叶片的准稳态响应及LLTR控制. 振动与冲击, 2018, 37(13): 123−129Liu Ting-Rui, Chang Lin. Quasi-steady response and LLTR control of a wind turbine blade with bending-torsion coupled. Journal of Vibration and Shock, 2018, 37(13): 123−129 [21] Shi H, Liu J, Wu Y, et al. Fault diagnosis of nonlinear and large-scale processes using novel modified kernel Fisher discriminant analysis approach. International Journal of Systems Science, 2016, 47(5): 1−15 [22] 郭金玉, 韩建斌, 李元, 等. 基于局部Fisher判别分析的复杂化工过程故障诊断. 计算机应用研究, 2018, 35(4): 1122−1125 doi: 10.3969/j.issn.1001-3695.2018.04.035Guo Jin-Yu, Han Jian-Bin, Li Yuan, et al. Fault diagnosis of complex chemical process based on local Fisher discriminant analysis. Application Research of Computers, 2018, 35(4): 1122−1125 doi: 10.3969/j.issn.1001-3695.2018.04.035 [23] 林信川, 游贵荣. 基于iBeacon的室内定位算法优化研究. 陕西理工大学学报(自然科学版), 2017, 33(3): 67−73Lin Xin-Chuan, You Gui-Rong. Study on optimization of indoor location algorithm based on iBeacon. Journal of Shanxi University of Technology (Natural Science Edition), 2017, 33(3): 67−73 [24] Liu C, Sun Y. The research and application of learning program in adaptive learning system. Applied Mechanics & Materials, 2014, 347-350: 3109−3113 [25] 孟亚辉. 基于最优特征集和马氏距离KNN分类的机械故障分类方法研究. 机械设计与制造, 2017, (7): 104−108Meng Ya-Hui. A method of mechanical fault classification based on optimal feature subset and K-nearest neighbor using mahalanobis distance. Machinery Design & Manufacture, 2017, (7): 104−108 [26] 林升梁, 刘志. 基于RBF核函数的支持向量机参数选择. 浙江工业大学学报, 2007, 35(2): 163−167 doi: 10.3969/j.issn.1006-4303.2007.02.010Lin Sheng-Liang, Liu Zhi. Parameter selection in SVM with RBF kernel function. Journal of Zhejiang University of Technology, 2007, 35(2): 163−167 doi: 10.3969/j.issn.1006-4303.2007.02.010 [27] 魏国, 刘剑, 孙金玮, 等. 基于LS-SVM的非线性多功能传感器信号重构方法研究. 自动化学报, 2008, 34(8): 869−875Wei Guo, Liu Jian, Sun Jin-Wei, et al. Study on nonlinear multifunctional sensor signal reconstruction method based on LS-SVM. Acta Automatica Sinica, 2008, 34(8): 869−875 [28] 黄心汉, 杜克林, 王敏, 等. 基于阻抗控制的动态装配过程仿真研究. 自动化学报, 2000, 26(2): 169−175Huang Xin-Han, Du Ke-Lin, Wang Min, et al. A simulation investigation on dynamic assembly process based on impedance control. Acta Automatica Sinica, 2000, 26(2): 169−175 [29] 张小云, 刘允才. 高斯核支撑向量机的性能分析. 计算机工程, 2003, 29(8): 22−25 doi: 10.3969/j.issn.1000-3428.2003.08.009Zhang Xiao-Yun, Liu Yun-Cai. Performance analysis of support vector machines with Gauss kernel. Computer Engineering, 2003, 29(8): 22−25 doi: 10.3969/j.issn.1000-3428.2003.08.009 [30] Lawn S D, Kerkhoff A D, Burton R, et al. Diagnostic accuracy, incremental yield and prognostic value of determine TB-LAM for routine diagnostic testing for tuberculosis in HIV-infected patients requiring acute hospital admission in South Africa: A prospective cohort. Bmc Medicine, 2017, 15(1): 67 doi: 10.1186/s12916-017-0822-8 [31] 彭泽栋, 冯毅萍, 杨胜蓝, 等. 基于多智能体的TE过程扩展仿真系统设计与实现. 高校化学工程学报, 2019, 33(5): 1195−1205 doi: 10.3969/j.issn.1003-9015.2019.05.022Peng Ze-Dong, Feng Yi-Ping, Yang Sheng-Lan, et al. Design and implementation of extended TE process simulation system based on multi-agent. Journal of Chemical Engineering of Chinese Universities, 2019, 33(5): 1195−1205 doi: 10.3969/j.issn.1003-9015.2019.05.022 [32] Tang Q, Chai Y, Xu J, Fang X. Tang Q, Chai Y, Qu J, et al. Industrial process monitoring based on Fisher discriminant global-local preserving projection. Journal of Process Control, 2019, 81: 76−86 doi: 10.1016/j.jprocont.2019.05.010 -