In the paper the usability of the Multiway PCA (MPCA) method for early detection of leakages in the pipeline system of a steam boiler in a thermal-electrical power plant is presented. A long segment of measurements of selected process variables was divided into a series of "batches" (representing daily recordings of normal behavior of the plant) and used to create the MPCA model of a "healthy" system in a reduced space of three principal components (PC). The periodically updated MPCA model was used to establish the confidence ellipsoid for the "healthy" system in the PC coordinates. [d=replaced]The staff's decision of the probable leak detection is supported by comparison of the current location of the operating point (on the "fault trajectory") with the boundaries of the confidence ellipsoid.The location of the process operating point created the "fault trajectory," which (if located outside the confidence ellipsoid) supported the decision of probable leak detection. It must be emphasized that due to daily and seasonal changes of heat/electricity demands, the process variables have substantially greater variability than in the examples of batch processes studied in literature. Despite those real challenges for the MPCA method, numerical examples confirmed that the presented approach was able to foresee the leaks earlier than the operator, typically 3-5 days before the boiler shutdown. The presented methodology may be useful in implementation of an on-line system, developed to improve safety and maintenance of boilers in a thermal-electrical power plant.
Keywords: fault detection; multiway PCA method; pipeline leaks; steam boiler.