Robust Non-negative Matrix Projection Method for Process Monitoring with Incomplete Data

The conventional statistical process monitoring methods usually require that the measurement data on all variables must be present at all times. In practice, however, most large process data sets contain missing data. In this paper, a new method called robust non-negative matrix projection (RNMP) based on the concept of incomplete process data, is proposed to detect and diagnose the faults in industrial process. Non-negative matrix factorization (NMF) is a widely-used method for low-rank approximation where the non-negativity constraints are imposed on factor matrices in the decomposition. The proposed RNMP method is a new variant of NMF based on positively constrained projections, which can obtain an accurate prediction of the missing values in the data set. Then, we use RNMP to extract the latent variables from incomplete data that drive a process and to combine them with process monitoring techniques for fault detection and isolation. Afterwards, the proposed method is applied to the continuous-stirred tank reactor (CSTR) process to evaluate the monitoring performance. The experiment results clearly illustrate the feasibility of the proposed method.