A subspace learning-based feature fusion and open-set fault diagnosis approach for machinery components

Open-set fault diagnosis is an important but often neglected issue in machinery components, as in practical industrial applications, the failure data are in most cases unavailable or incomplete at the training stage, leading to the failure of most closed-set methods based on fault classifiers. Thus, based on the subspace learning methods, this paper proposes an open-set fault diagnosis approach with self-adaptive ability. First, for feature fusion, without using traditional dimensionality reduction methods, a data visualization method based on t-distributed stochastic neighbor embedding is employed for its ability in mining and enhancing the fault feature separability, which is the key in fault recognition. Then, for open-set fault diagnosis, to detect unknown fault classes and recognize known health states in only one model, the kernel null Foley-Sammon transform is applied to build a null space. To reduce the misjudgment rate and increase the detection accuracy, a self-adaptive threshold is automatically set according to the testing data. Moreover, the final recognition results are described as distances, which helps the operators to make maintenance decision. Case studies based on vibration datasets of a plunger pump, a centrifugal pump and a gearbox demonstrate the effectiveness of the proposed approach.