To solve the problem of low accuracy and poor timeliness of in-vehicle CAN bus intrusion detection, an intrusion detection method based on gated recurrent unit(GRU) is proposed after analyzing the characteristics of intrusive data frame in CAN bus. This method built an intrusion detection model which consists of five-layer neural network, and constructed DoS（Denial of Service）attack, replay attack, fuzzy attack and virtual node attack data based on real CAN data, which was used for extracting 11-feature vector sequences. The experiment verified the influence of model parameters on detection results, then studied the accuracy and efficiency of binary and multiclass classification detection. The results show that the accuracy of this method reaches 99. 981 6% and 99. 894 2% in binary and multiclass classification, the corresponding recall rates are 0. 999 9 and 0. 999 1 respectively, which can achieve the equal accuracy of LSTM(long short-term memory model), and has shorter training and detection time. In all, this method improves the efficiency and robustness of intrusion detection, which has great significance for improving the safety of vehicles.