Multivariate Process Monitoring And Fault Diagnosis Based On Distribution Characteristics Of The Data

Process monitoring and fault diagnosis for industrial processes are becoming more and more important in order to ensure production safety, improve product quality and increase the economic benefits. PCA and ICA, which only use the process data, can remove redundancy effectively. And feature information extracted using them can well characterize the system. Therefore, multivariate process monitoring and fault diagnosis technology based on PCA and ICA has been widely used in all kinds of industrial processes. However, both of them have certain constraints in application. For example, PCA method can not extract the high-order characteristics well, while ICA only extracts information from the non-gaussian data. If the monitoring method is chosen without considering these constraints, it will get wrong conclusions. However, in most industry processes, there is no prior knowledge of the data distribution. The problem brings a great challenge to process monitoring technology.In order to deal with the above-mentioned problem, a multivariable process monitoring method based on data distribution characteristics is proposed in this thesis. Specific details are as follows:(1) According to relationship between the Mahalanobis distances and F distribution, an improved method for multivariate gaussianity test—Q-Q plot method of F statistics is proposed. The method does not need PCA to process the data. It can be used when the sample covariance matrix is irreversible.(2) A multivariate statistical monitoring method based on data distribution characteristics is proposed. The appropriate modeling method will be chosen according to the distribution of the data. Off-line model and control limit are calculated. For the linear data, PCA model will be chosen if the data obey the Gaussian distribution; ICA model will be chosen while the data distribution is non-gaussian; if the data contains both Gaussian and non-gaussian information, ICA-PCA model will be chosen.(3) The above-mentioned algorithm has been extended to nonlinear process:First of all, map the original data into high dimensional feature space, and use KPCA to whiten the data. Then test the distribution characteristics of whitened data, and choose appropriate monitoring method according to the test result (KPCA, KICA or Kernel ICA-PCA).(4) The algorithm is successfully applied to monitor continuous annealing line. ICA-PCA model is chosen automatically. And statistics T2, S2 and SPE are used for process monitoring. If the monitoring statistics exceed the control limits, the fault variables will be tracked by analyzing the contribution of statistics T2, S2 and SPE.