Multineighbor Ensemble Preserving Embedding for Process Monitoring

The effectiveness of process monitoring techniques is crucial to ensure industrial safety. Data-driven process monitoring methods can be advanced by considering both global and local data structures comprehensively, ensuring the capture of the entire spatial data structure. We propose a novel approach called multineighbor ensemble preserving embedding (MNEPE) with three key contributions. First, MNEPE considers both global and local data structures, building upon the neighborhood preserving embedding (NPE) by integrating various reconstruction modes. MNEPE incorporates orthogonal constraints to prevent singularity issues in solving generalized eigenvalue decomposition. Second, an adaptive approach is introduced for commonly used nearest neighbor modes to select nearest neighbors for capturing a local manifold structure. This enhancement improves the reliability of adjacency graphs and the accuracy of extracted local structures. Third, MNEPE computes or optimizes the target weights of various nearest neighbor modes to reduce the modeling cost in adjusting hyperparameters. This study implements MNEPE and verifies its effectiveness with a benchmark dataset and an actual dataset of the fused magnesia smelting process (FMSP). The experimental results confirm that MNEPE can be effectively applied to monitoring FMSP.