Data-Based Fault Detection for Safe and Reliable Process Start-up and Recovery after Natural Disasters

Natural disasters such as tornados, hurricanes, and earthquakes have a devastating effect across impacted areas. In the chemical industry, these events often force the shutdown of plants in the affected area to minimize harm to personnel and lower the risk of any environmental accidents. Recently, many industries along the Gulf coast, including key providers of natural gas, petroleum, petrochemical, refining, and commodity chemicals, have experienced mandatory shutdowns and damage during hurricane Katrina and more recently by hurricanes Harvey, Irma, and Maria [1]. During these shutdown periods, chemical plants are subject to high winds, rain, and debris that lead to physical damage of plant equipment and machinery. While much of this damage can be obvious upon inspection, additional damage may occur that is not visually noticeable but may have dramatic effects on plant operation and safety [2]. These not so obvious process changes lead to increased risk to operators, and cause a decrease in plant operability and performance during start-up and subsequent post-disaster operation. In the wake of a natural disaster, it is important to accurately, reliably, and quickly detect changes in operability within critical loops of a process, determine the cause of these changes within each loop, and apply corrective actions to mitigate risk and quickly bring a product back to market.

To achieve this goal, we propose the use of novel data-based fault detection methods to provide early detection of changes (faults) in a process that are incurred during a natural disaster, increasing process safety and ensuring reliable operation. Our novel diagnostic method combines nonlinear fault detection methods with advanced statistical composite hypothesis testing to accurately and reliably detect equipment damage during plant start-up. Our novel framework substantially improves the detection ability over previously developed methods for process monitoring found in the literature by reducing the missed detection and false alarm rates, leading to higher accuracy and dependability [3]. Additionally, the developed multiscale kernel partial least square (MS-KPLS) based generalized likelihood ratio test (GLRT) method incorporates an ability to independently detect and isolate multiple simultaneous faults within a process, a critical factor when catastrophic events can cause damage to different critical loops simultaneously currently missing from traditional fault detection software. Our MS-KPLS based GLRT algorithm has also been developed to handle process noise and correlated data sets to increase reliability when determining faults within a chemical system [4, 5]. Going further, we will demonstrate abnormal event management capabilities with state-of-the-art techniques in state observer design to provide a first-of-its-kind isolation method aimed toward speeding up repair efforts and helping industrial processes recover after natural disaster events. While this work is not specific to any industrial process, relying only on the accessibility of data from prior to and after the occurrence of the disaster, we will demonstrate the improved fault detection algorithm using the Tennessee Eastman Process problem as a case study. This work will demonstrate how databased models can be used to ensure plant safety post-disaster with an accurate and reliable detection and isolation strategy.

1. Hogue, C., Reisch, M. S., Schulz, W. G. & Wilson, E. K. Chemical & Engineering News: Latest News - Natural Disaster Impacts Chemical Enterprise. (2005). Available at: http://pubs.acs.org/cen/news/83/i36/8336disaster.html. (Accessed: 11th September 2017)
2. Collette, M. & Dempsey, M. Government ill-equipped to monitor industrial plants damaged by Hurricane Harvey. Houston Chronicle (2017). Available at: http://www.houstonchronicle.com/news/houston-texas/houston/article/Gover.... (Accessed: 5th October 2017)
3. Bakshi, B. R. Multiscale PCA with application to multivariate statistical process monitoring. AIChE J. 44, 1596–1610 (1998).
4. Botre, C., Mansouri, M., Nounou, M., Nounou, H. & Karim, M.N. Kernel PLS-based GLRT method for fault detection of chemical processes. Loss Prev. Process Ind. 43, 212–224 (2016).
5. Botre, C., Mansouri, M., Karim, M.N., Nounou, H. & Nounou, M. Multiscale PLS-based GLRT for fault detection of chemical processes. Loss Prev. Process Ind. 46, 143–153 (2017).