Abstract

The paper presents two readily implementable and computationally efficient approaches for sensor fault detection and identification (SFD/I) for group of sensors in complex systems. Specifically, hierarchical clustering (HC) and self-organizing map neural networks (SOMNNs) are demonstrated for use on industrial turbine systems. HC fingerprints are found for normal operation, and SFD/I is achieved by monitoring cluster changes occurring in the resulting dendrograms. Similarly, for the SOMNNs, a fingerprint of the classification map is found during normal operation, and SFD/I is performed according to the classification changes from the outputs. Unlike most existing methods that only monitor the condition of a single sensor, here, the proposed methods are shown to detect sensor faults from a large group of sensors. Moreover, by comparison with other methods that require an additional algorithm to identify which sensor is faulted, here, the proposed methods are shown to achieve SFD/I in a single stage. Whilst the SOMNN provides a numerical classification of the sensor-group condition, dendrograms from the HC method provide a more useful graphical interpretation for SFD/I—the presented techniques are now fully operational and monitoring a fleet of industrial turbines in real-time. Moreover, it is also shown that, after identifying a faulted sensor, it is possible in some circumstances to reconstruct the measurements expected from that sensor (using the remaining non-faulted sensor information) and thereby facilitate improved unit availability. The efficacy of the proposed approaches is demonstrated through the copious use of experimental measurements.