Considering the nonlinea r, time-varying and ripple coupling properties in the hydraulic servo system, a two-stage Radial Basis Function （RBF） neural network model is proposed to realize the failure detection and fault localization. The first-stage RBF neural network is adopted as a failure observer to realize the failure detection. The trained RBF observer, working concurrently with the actual system, accepts the input voltage signal to the servo valve and the measurements of the ram displacements, rebuilds the system states, and estimates accurately the output of the system. By comparing the estimated outputs with the actual measurements, the residual signal is generated and then analyzed to report the occurrence of faults. The second-stage RBF neural network can locate the fault occurring through the residual and net parameters of the first-stage RBF observer. Considering the slow convergence speed of the K-means clustering algorithm, an improved K-means clustering algorithm and a self-adaptive adjustment algorithm of learning rate arc presented, which obtain the optimum learning rate by adjusting self-adaptive factor to guarantee the stability of the process and to quicken the convergence. The experimental results demonstrate that the two-stage RBF neural network model is effective in detecting and localizing the failure of the hydraulic position servo system.
