A Levitation Condition Awareness Architecture for Low-Speed Maglev Train Based on Data-Driven Random Matrix Analysis

Modern low-speed maglev trains typically use multi-node decentralized levitation control modules, which results in a complex levitation control system with coupling interaction. Conducting systematic levitation condition awareness of the levitation control system is still a promising challenge. In this paper, under the hypothesis of levitation residuals following normal distribution, a levitation condition awareness architecture for the levitation control system is proposed based on data-driven random matrix analysis. The proposed architecture consists of an engineering procedure followed by a cascaded mathematical procedure. In the decentralized engineering procedure, the data-driven modeling for individual levitation control modules is achieved by nonlinear autoregressive modeling with an exogenous input neural network, and the unknown parameters are identified by a modified combinatorial genetic algorithm. On this basis, high-dimensional analysis of streaming residual random matrices for the levitation control system is conducted aided by large-dimensional random matrix theory, and the control limits of the constructed indicators are well-designed using the theorical distributions. Based on the comparative analysis of the experimental datasets, the proposed awareness architecture is verified to show the effectiveness of the systematic condition evaluation of the levitation system, and incipient train-guideway interaction vibration abnormalities can be detected in a timely manner.