Feature transfer projection with low-rank and dual graph embedding for multimode process monitoring

A complicated industrial process in general switches operation modes because of varying product specifications or raw materials, resulting in the distribution of multimodal process data may be inconsistent. It poses challenges to traditional process monitoring methods. In this paper, a feature transfer projection with low-rank and dual graph embedding (FTPLG) approach is proposed for multimode process monitoring, with the aim of capturing the informative representation among data. A multi-source weighted mean discrepancy alignment strategy is proposed for fine-scale distribution discrepancy reduction among multiple modes. A coefficient matrix imposed low-rank constraint is learned so that the proposed FTPLG captures shared information across different modes. Moreover, a dual graph regularization term is developed to preserve the relationships hidden in multiple modes, in which a reward graph and penalty graph focus on exploiting uniqueness within modes and connections between modes, respectively. With these distinctive characterizations, the FTPLG approach provides deeper insights into the multimodal processes. A simulated process and a real industrial process are carried out to demonstrate the effectiveness of the proposed approach. The average fault detection is improved by 9 %-28 % in the benchmark process.