Fuzzy neural incident detection algorithms with rolling training procedure

Abstract

This paper attempts to establish a fuzzy neural automatic incident detection (FNAID) algorithm, using back-propagation training procedures. A rolling training procedure continuously updating the traffic flow parameters is proposed to enhance the adaptability of FNAID to different traffic flow conditions. A real incident case is deliberately generated to calibrate the traffic simulator-Paramics. To validate the FNAID with and without rolling training procedure, the calibrated Paramics is used to simulate sufficient incident samples. The off-line tests and statistic tests conclude that under various traffic flow conditions, the FNAID with rolling training procedure has better detection performance than the one without rolling.