Multi-sample-distances-fusion- and generalized-Pareto-distribution-based open-set fault diagnosis of rolling bearing

It is not so easy to obtain the complete fault data for mechanical equipment toward certain variable working conditions. Traditional deep learning algorithms based on limited training data are difficult to accurately identify unknown faults, which seriously restricts the development of intelligent fault diagnosis. Open-set recognition can identify the faults that do not exist in the training data and label them as “unknown,” so it is becoming a powerful tool for solving the above problems. A novel l 2,1 regularized sparse filtering and multi-sample distances fusion-based open-set fault diagnosis method (RSDOS) is developed in this article. Firstly, the l 2,1 regularized term is induced in a traditional sparse filtering model to learn a more representative feature representation. Then, multi-sample distances fusion is applied to calculate the distances of intra-class and inter-class samples. Thirdly, the generalized Pareto distributions (GPD) concerning the tail information of those two distance distributions are modeled by the statistical extreme value theory (EVT). Finally, hypothesis testing models and the inference threshold are established to identify unknown faults. The diagnosis results of two rolling bearing datasets verify the progressiveness and robustness of RSDOS.