Fault Diagnosis Of Complex Industrial Process Based On Multivariate Statistical Method

With the development of industry technology, modern industrial process tends to be large-scaled constantly and equipment to be more and more complex, which leads to the mutual coupling between every part of system. Once a part of such system malfunction, the whole system will be likely broken, which leads to stopping production and huge property damage. Therefore, the safety plays an important role in industrial production process. The fault diagnosis technology is obviously important method to improve process safety, so it is necessary and practical significance for research of fault diagnosis method. Among the fault diagnosis technology, the method based on multivariate statistical method has a good application prospects for its’ merits of the process data-driven,real-time monitoring,instead of relying on mathematical model.In this paper, several methods of fault diagnosis are studied systematically by using multivariate statistical theory. These methods are used to perform diagnosis for several faults chosen from Tennessee Eastman process(TEP), which are used to be the platform of simulate and verification for the following studies:1、Firstly, fault diagnosis is studied based on principal component analysis(PCA). The results show that fault diagnosis using PCA only have good performance of diagnosis for one faulted source.2、Secondly, considering the industrial process characteristic of dynamic and nonlinearity, the dynamic kernel principal component analysis(DKPCA)method is introduced. Because that DKPCA is time-consuming, a feature vector selection(FS) algorithm is introduced based on DKPCA, and a new diagnosis method called FS-DKPCA is proposed, which can reduce computation time significantly. The results show that fault diagnosis using FS-DKPCA have improved performance of diagnosis for several faulted sources.3、Lastly, considering real industrial process measurement data is non-Gaussian distribution, the independent component analysis(ICA) method based on blind source separation is introduced. Due to the shortcoming of slow convergence when ICA is applied to the data-lagged matrix, a new diagnosis technology called DPCA-ICA is proposed. The simulation results show that the speed of independent component extraction by ICA become faster after data preprocessing by DPCA. The performances of DPCA-ICA is promising most of the diagnosed faults.All the results in this paper show that Compared to PCA, the new fault diagnosis methods proposed have less false and missing alarm rates. The new methods also have good performance of fault identification, which can ensure the reliable operation of the system.