Reduced KICA reconstruction modeling approach with dual attributes for fault diagnosis

Fault reconstruction identifies fault types by seeking a particular fault subspace that can effectively eliminate the alarm signal. A reconstruction model based on kernel independent component analysis (KICA) is proposed to address the non-Gaussian and nonlinear characteristics in fault diagnosis. However, nonlinear faults are characterized by nonlinear subspaces, which typically exhibit high dimensionality, contributing to increased spatial requirements and computational complexity. Moreover, commonalities among different subspaces may cause the same alarm signal to be eliminated by multiple subspaces, introducing uncertainty into the diagnostic process. To address these issues, a reduced KICA reconstruction modeling approach with dual attributes (RD-KICA) for fault diagnosis is proposed in this paper. An undersampling method is investigated to construct a less numerous but informative training set, such that the dimensionality of extracted subspaces is significantly reduced. The ideas of fault reconstruction and pattern classification are incorporated within the same framework, allowing their advantages to be complementary. Furthermore, the fault magnitude is supplemented as another attribute and used to train Bayesian classifiers for further diagnosis. Finally, several experiments on a numerical simulation, Tennessee Eastman process (TEP), and a rocket servo system are performed to validate the efficiency and benefits of the proposed method.