Condition Process Monitoring for Thermal Process Based on Local Normal Space Alignment

Dimension reduction analysis of high-dimensional process data is one of the main methods for process monitoring. Classical manifold learning methods struggle with problems including non-isotropic distribution of process data and preservation of local features in high-dimensional data. In this paper, we propose a different semi-supervised multi-manifold learning dimensionality reduction algorithm for typical thermal devices. The normal vectors of different manifolds are used to characterize the geometric features of high-dimensional data by constructing neighborhood of different data points in finite time period, and the dimensionality reduction results are verified by simulation datasets. Combined with the experimental data of the open literature, the simulation analysis of the condenser in a thermal power plant unit under different working conditions has been carried out to study the dimension reduction of the high-dimensional operation data based on different defect types and defect severity caused by aging. The consequences show that the proposed algorithm can effectively capture the distribution and geometric structure of high-dimensional data. The low-dimensional embedding results reflect that the load and malfunction speed up the deterioration of equipment operating status, consisting with the real phenomenon. Simulation experiments reveal that the proposed algorithm well preserves the geometric topology of the raw data, enhances the rationality of the low-dimensional embedding distribution, and effectively improves the condition monitoring accuracy of thermal devices. The approach in this paper can be used as a reference for monitoring and decision-making in thermal processes.