Fault detection using RBFN- and AR-based general parameter methods

This paper compares the performance of nonlinear radial basis function network-based (RBFN) and linear autoregressive (AR) model-based general parameter (GP) methods in a fault detection application. We use the efficient GP approach for initializing the weights of the RBFN model in the beginning of the off-line system identification phase, as well as for fine-tuning the modeling accuracy of RBFN and AR models online. Our fault detection scheme is based on monitoring the expectation value of the scalar general parameter. This provides improved robustness and detection sensitivity over such methods where the online prediction error is used directly in the decision-making process. In order to illustrate the performance of the efficient nonlinear and linear schemes, they are applied to fault detection of automobile transmission gears. As the acoustic sound level time-series, providing the necessary basis information for fault detection, is slightly nonlinear, the GP-RBFN outperformed the linear methods: the GP-AR method and conventional AR inverse filtering. Both of the GP-based methods provide competitive alternatives for real-world fault detection and diagnosis