Lou Diming, Zhao Yinghua, Fang Liang
College of Automotive Studies, Tongji University, Shanghai 201804, China.
ACS Omega. 2021 Aug 17;6(34):21971-21984. doi: 10.1021/acsomega.1c02403. eCollection 2021 Aug 31.
Fuel consumption is the most important parameter that characterizes the fuel economy of the engines. Instead of manual fuel consumption calibration based on the experience of engineers, the establishment of a fuel consumption model greatly reduces the time and cost of multiparameter calibration and optimization of modern engines and realizes the further exploration of the engine fuel economy potential. Based on the bench test, one-dimensional engine simulation, and design of experiment, a partially shared neural network with its sampling and training method to establish the engine fuel consumption model is proposed in this paper in view of the lack of discrete working conditions in the traditional neural network model. The results show that the proposed partially shared neural network applying Gauss distribution sampling and the frozen training method, after an analysis of the number of hidden neurons and epochs, showed optimal prediction accuracy and excellent robustness in full coverage over the whole load region on the test data set obtained through the bench test. Eighty-seven percent of the prediction errors are less than 3%, all prediction errors are less than 10%, and the value is improved to 0.954 on the test data set.
燃油消耗是表征发动机燃油经济性的最重要参数。建立燃油消耗模型，而非基于工程师经验进行手动燃油消耗校准，极大地减少了现代发动机多参数校准和优化的时间和成本，并实现了对发动机燃油经济性潜力的进一步探索。鉴于传统神经网络模型缺乏离散工况，本文基于台架试验、一维发动机仿真和实验设计，提出了一种具有采样和训练方法的部分共享神经网络来建立发动机燃油消耗模型。结果表明，所提出的采用高斯分布采样和冻结训练方法的部分共享神经网络，在对通过台架试验获得的测试数据集的全负荷区域进行全覆盖分析时，在分析隐藏神经元数量和训练轮次后，显示出最佳的预测精度和出色的鲁棒性。在测试数据集上，87%的预测误差小于3%，所有预测误差均小于10%， 值提高到0.954。
