Orthogonal Multi-Block Dynamic PLS for Quality-Related Process Monitoring

Project to latent structure (PLS) is a well-known data-driven method, which has been widely used in process monitoring. Several dynamic PLS algorithms have been proposed and applied in the dynamic process. However, the oblique decomposition induced by the PLS-based methods leads to redundant information in quality-related space, which results in missing alarms and false alarms. To cope with this problem, an orthogonal multi-block dynamic PLS (OMDPLS) algorithm is proposed. In the proposed method, a modified PLS (MDPLS) model is first built to provide an explicit inner dynamic model, which describes the relationship of latent variables between quality variables and time-delay inputs. Furthermore, based on the proposed MDPLS model, the orthogonal projection operator is constructed for each time-delay multi-block input matrix to reduce the quality-unrelated information in principal space. Afterward, the vector auto-regressive model is constructed to separate the dynamic interaction of the current sample to reflect the true state. Then, a completely dynamic process monitoring strategy is developed for quality-related faults. Finally, two experiments including a numerical simulation and the Tennessee Eastman process (TEP) are performed to demonstrate the effectiveness of the proposed method. Note to Practitioners—Controlling and measuring the quality state is challenging for advance industrial processes due to the dynamic correlation in the states. To ensure safe operation in the dynamic processes, avoiding false alarms of quality indicators has become a crucial task in modern industrial plants. This paper develops a novel dynamic process monitoring method, which reduces the false alarms caused by quality-unrelated faults and has good fault detection capability. The key advantage of OMDPLS is that the orthogonality in the sample space and the orthogonal decomposition in the variable space can be guaranteed. As a result, the latent variables extracted from the input are unrelated to the residuals, and the quality-related space and the quality-unrelated space are mutually exclusive. Driven by OMDPLS, the impact of quality-unrelated faults on the monitoring of key indicators will be reduced with fewer false alarms. In this way, unnecessary maintenance can be avoided. Therefore, the proposed method can be applied to processes with high requirements for low false alarms, such as large-scale process industries (metallurgy, steel rolling, chemical, and more). Meanwhile, the proposed method is purely data-driven and does not require complex process knowledge, which can supply more information for the monitoring needs of engineers. Two case studies show that the process monitoring strategy delivers more comprehensive performance than other methods.