Scalable fault models for diagnosis of synchronous generators

Abstract

In this paper, we experiment with a small working model SWM, where we can inject faults and learn the intelligence about the system, then scale up this fault models to monitor the condition of an actual/complex system, without injecting faults in the actual system. We refer to this approach as scalable fault models. We check the effectiveness of our approach using 3 kVA and 5 kVA synchronous generators to emulate the behaviour of SWM and actual system, respectively. We linearise the features from the SWM and actual system in a higher-dimensional space using locality constrained linear coding LLC to make them linearly separable. Subsequently, the system-independent features are selected using principal component analysis PCA to make the fault models robust across the systems. Support vector machine SVM is used as a back-end classifier. Experiments and results show that proposed LLC-PCA system outperforms the baseline system.