基于小样本的GDI涡轮增压发动机性能预测方法比较分析

为了降低建立发动机性能预测模型消耗的实验成本,利用具有强大的非线性问题分析能力的机器学习算法:广义回归神经网络(general regression neural network,GRNN)和支持向量回归(support vector regression,SVR),预测GDI(gasoline direct injection)涡轮增压发动机性能。首先采用田口正交实验法和拉丁超立方算法确定训练数据和测试数据的操作点,然后基于仅包含25个样本的相同训练数据训练模型,使用100组相同的测试数据测试GRNN和SVR的预测性能并进行了对比研究。对比分析表明,在实验数据有限的情况下,GRNN有收敛到局部最小值的风险,而SVR可以找到最优的全局解,并具有良好的预测精度和泛化能力,因此SVR非常适合应用于GDI涡轮增压发动机性能预测,并将显著降低实验成本。

Comparison and analysis of performance prediction methods for GDI turbocharged engine based on limited data

In order to reduce the experimental cost of establishing engine performance prediction model, machine learning algorithms with powerful non-linear problem analysis ability were employed to predict the performance of gasoline direct injection (GDI) turbocharged engine, such as general regression neural network (GRNN) and support vector regression (SVR). The Taguchi orthogonal experiment method and the Latin hypercube sampling (LHS) were introduced to determine the operation points of the training data and the test data. And then based on the training models containing the same training data of only 25 samples, the prediction performance of both GRNN and SVR was tested and compared using the same 100 sets of testing data. The comparative analysis shows that GRNN has the risk of converging to the local minimum when the experimental data is limited, whereas SVR can find the optimal global solution with good prediction accuracy and generalization ability. So SVR is very suitable for GDI turbocharged engine performance predictions and significantly reduces experimental costs.