Data-driven fault diagnosis of satellite power system using fuzzy Bayes risk and SVM

Data-driven fault diagnosis is more suitable than model-based methods for diagnosing the complicated spacecraft systems, e.g., the satellite power system, because of its simplicity and convenience. Nevertheless, some redundant and irrelevant features in monitoring data are usually not conducive to identify fault state but significantly reduce the correct rate of diagnosis and increase the computational complexity and memory storage space. The existing filter feature selection approaches usually need to provide the number of selected features in advance, which brings an extra burden to decision makers. To make up for this deficiency, this paper proposes a feature selection method based on fuzzy Bayes risk (FBR) to generate an optimal feature subset automatically without having to preset the feature number. A heuristic forward greedy feature selection algorithm based on the proposed fuzzy Bayes risk theory is designed. Subsequently, a data-driven fault diagnosis strategy is designed by employing FBR and Support Vector Machine (SVM). Finally, numerical experiments on UCI data and the fault diagnoses of satellite power system are carried out to illustrate the superiority and applicability of the proposed method. The results of comparison experiments show that both classification accuracy of UCI data and diagnosis effect of satellite power system are better than the other state-of-the-art methods. (C) 2018 Elsevier Masson SAS. All rights reserved.