Development of Node-Decoupled Extended Kalman Filter (NDEKF) Training Method to Design Neural Network Diagnostic/Prognostic Reasoners

摘要

In this paper, we have proposed diagnostic techniques using a multilayered neural network where the weights in the network are updated using node-decoupled extended Kalman filter (NDEKF) training method. Sensor signals in both time domain and frequency domain are analyzed to show the effectiveness of the NDEKF algorithm in each domain. Comparisons of the NDEKF algorithm with other popular neural network training algorithms such as extended Kalman filter (EKF) and backpropagation (BP) will be discussed in the paper through a system identification problem. First, the simulation results reveal the comparison of outputs from actual system and trained neural network. Secondly, the ability of diagnosing a system with one normal condition and three known fault conditions is demonstrated. Thirdly, the robustness of the machine condition monitoring when the inputs to the system vary is shown. The proposed technique works even when there is noise in sensor signals as well.