Scalable fault models for diagnosis in a synchronous generator using feature mapping and transformation techniques

Abstract

Condition based maintenance (CBM) needs data acquired during healthy and faulty conditions to develop intelligent system for fault diagnosis. However, fault injection is not allowed/possible in a highly expensive components of complex/critical systems to collect fault condition data. Therefore, proto-type/small working models are used to conduct experiments for abnormal/fault conditions, to obtain and scale the intelligence of the system for effective health monitoring of complex system. This methodology is referred as scalable fault models. For proof of concept, in this work, we considered two different capacity synchronous generators with rating of 3 kVA and 5 kVA to emulate the behavior of prototype/small working model and complex system respectively, for scalable fault models. We explored feature mapping and transformation techniques to achieve effective scalability.From the preliminary experiments, it is observed that the baseline system performance deteriorated due to the changes in the system (capacity) and its characteristics with load changes.We therefore, expressed the input features in terms of load and system independent manner, to make the features less dependent on load and system variations. We explored localityconstrained linear coding (LLC) to express the features load/system independently. It is observed that experimenting LLC with the backend support vector machine (SVM) classifier gave the best fault classification performance for linear kernel, suggesting that the faults are linearly separable in the new feature space.Since the LLC mapped feature space is linearly separable, we then explored linear feature transformation technique, nuisance attribute projection (NAP) on the LLC mapped feature space to further minimize the load/system specific variations. We observed that LLC-NAP improved the overall accuracy and sensitivity of the classifier significantly. We also noted that the performance of NAP was limited in the original feature space since the feature space (NAP without LLC) is nonlinear with load/system variations.