Abstract

The paper presents two readily implementable methods for sensor fault detection and identification (SFD/I) for complex systems. Specifically, principal component analysis (PCA) and self-organizing map neural network (SOMNN) based algorithms are demonstrated for use on industrial gas turbine (IGT) systems. Two operational regimes are considered viz. steady-state operation and operation during transient conditions. For steady-state operation, PCA based squared prediction error (SPE) is used for SFD, and through the use of contribution plots, SFI. For SFD/I under operational conditions with transients, a proposed ‘y-index’ is introduced based on PCA with transposed input matrix that provides information on anomalies in the sensor domain (rather than in the time domain as with the traditional PCA approach). Moreover, using a SOMNN approach, during steady-state operation the estimation error (EE) is used for SFD and EE contribution plots for SFI. Additionally, during transient operation, SOMNN classification maps (CMs) are used through comparisons with ‘fingerprints’ taken during normal operation. Validation of the approaches is demonstrated through experimental trial data taken during the commissioning of IGTs. Although the attributes of the techniques are focused on a particular industrial sector in this case, ultimately their use is expected to be much more widely applicable to other fields and systems.