Tian Ying, Yao Heng, Li Zeqiu
Shanghai Key Lab of Modern Optical System, School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China.
Shanghai Key Lab of Modern Optical System, School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China.
ISA Trans. 2020 Jan;96:24-36. doi: 10.1016/j.isatra.2019.06.002. Epub 2019 Jul 9.
Multiblock methods have been proposed to capture the complex characteristics of plant-wide monitoring due to the enlargement of process industries. These methods based on automatic sub-block division and copula-correlation, which simultaneously describe the correlation degree and correlation patterns, are designed for sub-block partition. However, the selection of variables for each sub-block through copula-correlation analysis requires a pre-defined cutoff parameter which is difficult to be determined without sufficient prior knowledge, and a "bad" parameter leads to a degraded performance. Therefore, a weighted copula-correlation-based multiblock principal component analysis (WCMBPCA) is proposed. First, the variables in each sub-block are obtained through the copula-correlation analysis-based weighted strategy rather than the cutoff parameter, which highly avoids information loss and prevents "noisy" information. Second, a PCA model is established in each sub-block. Third, a Bayesian inference strategy is used to merge the monitoring results of all sub-blocks. Finally, an online-horizon Bayesian fault diagnosis system is established to identify the fault type of the system based on the statistics of each sub-block. The average detection rate and the average diagnosis rate for numerical example are 77.85% and 98.95%, and that for TE example are 80.63% and 89.50%. Comparing with other candidate methods, the proposed method achieves excellent detection and diagnostic performance.
由于流程工业的扩大，已经提出了多块方法来捕捉全厂监测的复杂特征。这些基于自动子块划分和Copula相关性的方法，同时描述了相关程度和相关模式，是为子块划分而设计的。然而，通过Copula相关性分析为每个子块选择变量需要一个预先定义的截止参数，在没有足够先验知识的情况下很难确定，并且一个“不好”的参数会导致性能下降。因此，提出了一种基于加权Copula相关性的多块主成分分析（WCMBPCA）方法。首先，通过基于Copula相关性分析的加权策略而不是截止参数来获得每个子块中的变量，这极大地避免了信息损失并防止了“噪声”信息。其次，在每个子块中建立主成分分析（PCA）模型。第三，使用贝叶斯推理策略来合并所有子块的监测结果。最后，建立一个在线贝叶斯故障诊断系统，基于每个子块的统计数据来识别系统的故障类型。数值示例的平均检测率和平均诊断率分别为77.85%和98.95%，TE示例的平均检测率和平均诊断率分别为80.63%和89.50%。与其他候选方法相比，所提出的方法具有出色的检测和诊断性能。
