Data driven fault detection using multi-block PLS based path modeling approach

Early detection of process faults (while the plant is still operating in a controllable region) can save billions of dollars and enhance safety by minimizing the loss of productivity and preventing the occurrence of process mishaps. This has encouraged researchers to develop methods for improved monitoring of industrial units. This work is based on a pathway modeling approach using multi-block PLS as the mathematical machinery. The proposed approach is a multivariate data analysis procedure which divides the process data into different blocks, determines relationships among the blocks, which are then used for fault detection. In the developed methodology, an extension of multi-block PLS, a special modification of the PLS technique is realized by incorporating the H-Principle in the algorithm. This renders it different from PLS as the analysis is done in steps, maximizing the product of size of improvement of fit and associated precision at each step. The T ² statistic is primarily used as an indicator of normalcy or fault in the system. This new technique is illustrated via application to two industrial-scale, high-fidelity simulated systems namely the Tennessee Eastman Process (TEP) and a Depropanizer Process (DPP).