Research On Information Mining In The Feature Space For Fault Detection

Because of the increasing demand for the maximize economic benefits, safety and stability are key factors in the modern industrial production process. So, process monitoring and early fault detection has important theory significance and engineering application value. The traditional basic data-driven fault detection methods directly establish the monitoring model by projecting the data onto a lower-dimensional feature space, mainly focusing on global structure of samples and variables in the space. However, the ignorance of local structure may lead to miss important information. Therefore, it is essential to research how to extract the useful information of data in the feature space to improve fault detection rates.In this paper, we focus on the research about the samples and variables structure in the space based on the multivariate statistical process methods. In respect of preserving the samples structure, to solve the problem that the traditional kernel principal component analysis only concentrates on global structure of the samples, a new method named modified kernel principal component analysis is proposed based on the local structure. The idea of locality preserving is incorporated into the optimization goal of kernel principal component analysis. The new projection space enjoys the similar global structure and the local structure, so, more feature information can be extracted. Next, the feature information is classified through Fisher discriminant analysis. A monitoring statistic is established using the distance of each sample. In respect of preserving the variables structure, on the other hand, to solve the problem of missing the different contribution of each kernel independent component for the system fault, a weighted kernel independent component analysis for fault detection is proposed. First, kernel independent component analysis is performed to extract the kernel independent components. Kernel density estimation is used to evaluate the contribution of each kernel independent component. Then, set different weighting values to highlight the kernel independent components with more useful information. Finally, a monitoring statistic is established using the local outlier factor. Simulation results on Tennessee Eastman process and numerical example show that the proposed method can exploit structures in the feature space and improves the fault detection accuracy as a result.