A Novel Global-Local Feature Preserving Projection Method Based on Adaptive Linear Local Tangent Space Alignment for Process Monitoring

Abstract

Abstract Due to the rapid increase of complexity in industrial process, feature extraction plays an important role in process monitoring. In conventional process monitoring methods, represented by multivariate statistical process monitoring (MSPM) method, global feature of process data is mainly considered, such as variance information. By contrast, the local feature of process data, obtained by preserving neighbourhood information, is preferred in newly emerging manifold learning method. Process monitoring methods with both global and local features of data were then proposed. Theoretically, better monitoring results can be expected with the consideration of both global and local features, but methods can be significantly different for the extraction of global or local feature, which requires better mathematical description of the global and local features ahead. However, there is no specific discussion on this topic in literature yet. To address this issue, A novel global-local feature preserving projection (NGLFPP) method based on adaptive linear local tangent space alignment is proposed. The neighbourhood of each sample is given adaptively by considering the local distribution of data. The local feature of data is obtained by modified tangent estimation. On the basis of the adaptive neighbourhood, geodesic distance is introduced to represent global feature of data. In order to extract global and local features of data simultaneously, a dual objective optimization function is constructed. Tennessee Eastman (TE) process is employed to validate the proposed method.