Shareability–Exclusivity Collaborative Representation and Its Application for Distributed Process Monitoring

Industrial processes in general consist of multiple operating units; complicated interactions are encapsulated within process variables, bringing a great challenge for monitoring these large-scale industrial processes. The existing distributed monitoring schemes fail to consider explicitly shared information across all blocks when modeling unique characterizations within each block, deteriorating monitoring performance. This article proposes a novel dual shared-specific distributed modeling approach termed shareability-exclusivity collaborative representation (SECR) to monitor large-scale industrial processes, which simultaneously captures the shared information across multiple blocks and distills the exclusive characteristics within each block. Optimal transport-spectral clustering is proposed to automatically divide variables into blocks, in which variable associations and distribution information are considered in terms of a probabilistic manner. The lowest rank representation among all blocks is learned to reveal the shared information across all blocks. Meanwhile, SECR can capture the exclusive representation within each block by introducing an exclusivity reconstruction term. Additionally, a structure-preserving matrix with neighboring and nonneighboring information is proposed to explore the intrinsic local and nonlocal structure of data. On the basis of these distinct modeling representations, the proposed approach is more powerful to capture faithful representations of given data and further strengthen monitoring capabilities. To achieve decision fusion of monitoring performance, Bayesian inference is introduced to integrate monitoring results of multiple blocks. The proposed approach is verified in a benchmark simulated processes and applied to a real fluid catalytic cracking (FCC) process. Compared with other monitoring methods, the average fault detection rate (FDR) of the proposed approach is improved by 3.56%–32.41% in the benchmark processes, demonstrating the effectiveness of the proposed distributed monitoring approach.