A multiple heterogeneous kernel RVM approach for analog circuit fault prognostic

Abstract

This paper presents a multiple heterogeneous kernel relevance vector machine (MHKRVM) approach for analog circuit fault prognostic. Compared to other kernel learning methods, multiple kernel learning method produces the optimal kernel function because many effective kernels’ combination always generates better generalization performance. By using the multiple heterogeneous kernel learning method, the proposed MHKRVM method’s kernel function holds its diversification. Meanwhile, the sparse weights of consisted basic kernels in the MHKRVM are helpful in prediction accuracy, and they are yielded through particle swarm optimization algorithm. Six fault prognostic cases are conducted to demonstrate the whole prognostic procedure, and prove that the presented MHKRVM can predict the trend of falling circuit elements’ health degree trajectories closely and estimate the remaining useful performance of failing circuit elements accurately.