Semisupervised Kernel Independent Component Analysis and Its Application for Fault Diagnosis

Abstract

Fault reconstruction, a common approach for fault diagnosis, involves extracting fault subspaces from measurement data and using them to identify the type of online fault samples. These subspaces, however, are often challenged by insufficient fault representation and weak discriminability among different fault types. To address these issues, a semisupervised kernel independent component analysis (SsKICA) algorithm is proposed and applied in fault diagnosis. First, an innovative feature extraction method is developed that considers both the non-Gaussianity and discriminability of the features in parallel. Its objective function is reformulated into a dual-maximization criterion function that incorporates the supervised information of sample categories and independence into the same mathematical frame. Second, Newton’s methods and fixed-point iteration are employed to derive iterative solutions for this dual-maximization function, and the convergence of these iterative solutions is analyzed. Third, a fine-grained fault subspace extraction method is investigated by extending fault reconstruction to the Hilbert space. Finally, a complete fault diagnosis strategy based on SsKICA is designed that can provide feedback for unseen faults and easily interpretable diagnostics. The simulation results, on the Tennessee Eastman process (TEP) and the PROcess Networks Optimization (PRONTO), demonstrate that the proposed method provides effective feedback mechanisms for unseen faults and enhanced diagnostic accuracy for known faults.