Abstract: The diagnosability problems of the large-scale complex electromechanical systems are studied, and the main influencing factors and evaluation criteria of the fault diagnosability are proposed, then the applications of Hilbert/Huang time-frequency analysis to fault diagnosis are investigated in order to improve the quality of fault information. Subsequently, to correctly recognize some unknown faults, some unsupervising learning algorithms such as self-organizing feature maps, generative topographic mapping and curvilinear component analysis are used to classify different working modes of the large-scale complex electromechanical systems. Finally, based on the latest research finds in machine learning, to extract nonlinear features from fault signals and to deal with the problems of non-separated mechanical faults, some kernel-based methods are presented, and an exploratory study for machine running trend analysis using kernel methods has been executed.

Key words: Condition monitoring, Diagnosability, Fault diagnosis, Kernel methods, Unsupervised learning