Condition Monitoring: Based on "black-box"- and First Principal Models

Abstract

Summary and conclusionsThis master thesis investigates Principal Component Analysis (PCA) methods used in the field of Early Fault and Disturbance Detection (EFDD). Statoil and ABB have in collaboration developed an application for EFDD that among others consist of PCA methods. The applications are used to run condition monitoring on oil- and gas processes, and are a software prototype currently being tested in Statoil. To make the method more robust and convincing to the operators it is desirable to improve the application.This thesis will focus on the Principal Component Analysis (PCA) method and some extension to it based on Model Based PCA (MBPCA). The PCA in its simplest form have some severe restrictions due to linear and stationary data. The motivation will therefore be to see how PCA and extension based on MBPCA and Nonlinear PCA (NLPCA) methods operates when used on non-linear data. For the PCA to describe a process adequately a certain amount of data is required. In the industry the process is often sort of instrumentation. The next motivation would be to investigate if lack of instrumentation could be replaced by some estimates and then approve the ability for the PCA analysis. Another issue concerning instrumentation is the use of virtual tags. Virtual tags are mathematical functions based on already available measurements. The idea is based on process insight. If we know the process well and the cause of nonlinearities some additional nonlinear functions could be incorporated to increase the performance of the PCA method. To verify the factors mention above some of the methods would use data from a heat exchanger and Centrifugal pump process, and some using only one of the processes.The conclusion from the work is as follows: Based on the simulations preformed it s evident that using MBPCA do improve the PCA method for fault detection, even if the model is not entirely correct to the real process. When it comes to using virtual tags the simulations on centrifugal pump increased the performance of the PCA method. NLPCA here based on Autoassociative Neural Networks did not perform as well as MBPCA but the method is harder to tune and therefore it would be wrong to brush aside this method. The improvement of using estimates for missing measurement gave small improvement.