Correct feature selection is critically important to any feature-based diagnostic techniques, but it is not always easy to achieve for systems with complex fault modes. This paper proposes an artificial intelligence methodology for mechanical fault detection using vibration data, which incorporates intelligent feature optimization. After preliminary feature extraction through spectrum analysis of measured vibration signals, this approach uses backpropagation neural network twice, first for feature reselection and then for fault detection. Applications of this method to over fifty lubrication pumps proved its effectiveness.