Feature Extraction of Nonlinear Spectrum and Fault Diagnosis for Multivariable Dynamic System

The fault diagnosis of multivariable dynamic system is researched based on nonlinear spectrum data and support vector machine. In order to overcome the problem of calculated amount expansion of generalized frequency response function description, the nonlinear spectrum feature is obtained based on one dimensional nonlinear output frequency response function. A frequency domain variable step size normalized LMS adaptive identification algorithm is proposed. The step size is changed instantaneously by using estimation error, so the convergence rate and steady state error are both considered. After obtain nonlinear spectrum feature, least square support vector machine is used to construct multi-fault classifier for fault identification. In order to reduce training time, support vector machine is trained by conjugate gradient algorithm based on simplified formula. The fault diagnosis of a vibration system with two inputs and four outputs is studied. The experiment results indicate that the proposed fault diagnosis method has short training time and high recognition rate so that it can meet the demand of online diagnosis.