Temporal-Spatial Global Locality Projections for Multimode Process Monitoring

Multimode is an important feature of modern processes, since various manufacturing strategies are needed to satisfy different demands of markets. Direct application of traditional multivariate statistical process monitoring methods cannot obtain satisfactory results, as the data set collected from multimode processes always follows multimodal distribution. To construct a single model which can monitor multimode processes directly, this paper proposes an original algorithm named temporal-spatial global locality projections. First, given that both temporal and spatial neighbors can express the similarity, the determination of the neighborhood is conducted in both the temporal and spatial scale. Second, an optimization objective function which preserves not only the local structure but also the global structure is defined. Third, the monitoring statistic is established via the local outlier factor. To certify the effectiveness, a numerical example, the multimode Tennessee Eastman process, and the CE117 process which is proposed by TecQuipment for process control are studied.