Although autonomous vehicles(AV) have made great progress in line planning and driving controlling, it is still difficult to make certain and reasonable decision-making in ethical dilemmas, which causes people's doubts and worry about AV's safety. Therefore, it is necessary to design a self-driving ethical decision-making model and mechanism to make reasonable decision-makings in ethical dilemmas for AV. For them, this paper provided an ethical decision-making model based on stacking fusion mechanism, in which machine learning and deep learning were deeply integrated. Firstly, naive Bayes model based on attribute correlation(ACNB), weighted average first-order Bayes model(WADOE) and adaptive fuzzy model(AFD) were constructed into basis learners of stacking fusion mechanism. According to the accuracy in previous verification, their weights were set in decision-making, and the decision-making results in all dilemmas should be calculated by weighted average method. Then the decision-making data by the basis learner was regarded as training set for the meta-learner in the stacking fusion mechanism to construct stacking fusion decision-making model. Lastly, all dilemmas in verification set were used to verify deep learning model and stacking fusion mechanism respectively. The effect of stacking fusion mechanism was evaluated and compared based on the average loss rate and accuracy in the verification and the correction rate in the testing. The results show that the average loss rate is more 0.64, the maximum average accuracy is 0.7, and the highest correction rate is 0.61 for the deep learning model. However, the lowest average loss rate is 0.35, the maximum average accuracy is 0.9, and the highest correction rate is 0.75 for stacking fusion mechanism, which indicates the stacking fusion mechanism is greatly improved compared with the deep learning model.