Industrial Process Monitoring Based on Collaborative Structure Projection

Monitoring an industrial process with satisfactory accuracy is crucial to ensure product quality, improve production efficiency, and assure human’s operations safely. However, most process monitoring methods based on local and global structures exhibit two limitations: 1) a Euclidean distance does not represent the actual distance of two samples in the manifold structure accurately and 2) the K-nearest neighbors (KNN) method is adopted, but the number (k) of the nearest neighbors is chosen empirically or experimentally. To overcome the above limitations, we propose the collaborative structure projection (CSP) to monitor an industrial process, and it is innovative in three aspects: 1) an adaptive neighbor strategy is introduced to select more appropriate neighbors in capturing and representing the local structure of the data automatically; 2) a geodesic distance is used to replace the Euclidean distance in representing an actual distance of two samples in a manifold structure for an enhanced accuracy; and 3) the orthogonality constraint is applied to recognize features and avoid potential singularities for an improved discriminability. The proposed CSP is verified by using the Tennessee Eastman process (TEP) benchmark dataset and the data collected from an actual magnesium smelting process (MSP), and the verification shows the effectiveness of the proposed method.