Statistical process monitoring and neural networks for early leak detection in a pipeline system of a steam boiler

World Congress on Industrial Automation

July 20-22, 2015 San Francisco, USA

Miroslaw Swiercz and Halina Mroczkowska

Posters-Accepted Abstracts: Adv Robot Autom

Abstract :

The paper presents the comparison between two approaches to detection of leakages in the pipeline system of a steam boiler, which works in a thermal-electrical power plant in Bialystok. The first methodology employs the Principal Component Decomposition (PCA) of a segment of real data, which contain historical measurements of 12 selected process variables. The second approach utilizes specific architectures of Artificial Neural Networks (ANN): Feed forward Multilayer Perceptron (MLP) and Learning Vector Quantization (LVQ) structures. Both methods belong to a class of data-driven approaches, as due to technical and economic limitations building a model of the plant is not feasible. The goal of our studies was to detect the symptoms of arising leak as early as possible, to warn the process operator and allow him to perform any planned actions, instead of an emergency shutdown of a boiler. The segment of historical data from normal plant operation was used to create the PCA model of a �healthy� system in a reduced space of three principal components. The length of the segment and time delay (in reference to the current time moment) of data employed for model development, were determined experimentally. The PCA model, periodically updated, was used to establish the confidence ellipsoid, i.e. the feasible area occupied by the process variables of the �healthy� system in PC coordinates. Similarly, selected and verified historical data segments, which represented both normal and faulty conditions, were used for the ANN training. Then the trained networks were used to classify current data segment as �normal� or �faulty�. In the series of numerical experiments we confirmed the ability of the above methods to detect leakages 4-5 days before the shutdown.