A fault detection method for complex electromechanical system using improved integrated approach

In view of certain limitations to applications of conventional multivariate statistical methods under increasingly complicated industrial process system, this paper proposes a fault detection method for complex electromechanical system using ILLE-SVDD data dimensionality reduction. The method combines traditional ILLE and SVDD. First of all, sample points in the original high-dimensional space are linearly represented with their adjacent points' Euclidean distance, using locally linear embedding (LLE). Build a local reconstruction relationship matrix and keep its reconstruction relationship unchanged as far as possible in low-dimensional space, to reduce dimensionality. Secondly, by containing most of the samples in space within the space of sphere, separate a few exceptional samples outside the sphere, to detect exceptions and classify. Finally, in the process of TE, carry out a simulation experiment. Results show that ILLE has a better manifold unfolding effect. SVDD can detect exceptions effectively and is easier to implement in application.