Abstract: In order to recognize drunk driving and identify the level of drunk driving states accurately, and improve the efficiencies of drunk driving governances, all drivers' driving behavior data (including the data on human, vehicle and environment) of three driving states including normal, tipsy and intoxicated were gathered by human factors engineering experiments and driving simulation experiments, the experimental data were pretreated. The original parameters were analyzed by factor analysis and the characteristic parameters were extracted as the vectors to be input in a multilayer neural network. The multilayer neural network was trained, and an identification model was built for drunk driving behaviors based on factor analysis and multi-layer neural network. The research shows that the training time of the model is 0.905 seconds, the optimal validation mean square error(MSE) of the model is 0.034, and recognition accuracy is 92.41%, and the model could be used to identify drunk driving behaviors quickly and accurately .

Key words: drunk driving, driving behavior, characteristic parameter, factor analysis, multilayer neural network