A Novel Method of Operation Condition Division based on Time Domain Constraint Kernel Representation

Modern industrial processes usually work continuously under varying operation conditions due to factors such as the fluctuation of raw material quality, changes of product specifications and reaction periods. The data characteristics and structures vary significantly under different operation conditions, so it is vital to recognize the multi-operation-condition characteristic of processes so as to develop corresponding decision schemes for engineers. The clustering-based method can help to understand the operating processes by grouping the history data. However, most of them is hardly to accurately characterize the relationships between samples based on the distance measure, especially when the data have complex structures and close distributions. Meanwhile, they tend to face the challenges of unknown number of operation conditions. Therefore, this paper proposes a novel method based on kernel sparse representation and complex network to accurately and automatically divide operation conditions. First, combining the data characteristics of time-series and self-expressiveness, time domain constraint kernel sparse representation (TDCKSR) is proposed to establish the similarity measure. Then the topological structure of complex network can be acquired by mapping all samples to the network based on the similarity matrix. Finally, community detection algorithm is introduced to reveal the aggregation behavior of samples under multi-operation-condition processes. Cases studies including a numerical case, the Tennessee Eastman process, and a real industrial roasting process have demonstrated the proposed method has good performance and application aspects.