【摘要】 为解决发动机在恶劣工况下因传感器故障导致的可靠性下降与维护成本增加问题，提出并验证了一种基于数据驱动的轻量化空气流量虚拟传感器方案．该方案首先评估了回归树、集成树、支持向量机和神经网络4类机器学习模型在空气流量预测任务上的性能，选取性能表现较好且在算法原理上具有典型差异的集成树与神经网络作为重点优化对象，采用贝叶斯优化算法对其参数进行调优，并针对存储资源与预测精度之间的平衡提出了一种面向资源约束的混合优化策略．在五折交叉验证基础上，神经网络模型文件大小仅为0.062 5 MB，在56种工况下均表现出良好的建模能力与部署优势，在测试集上平均相对误差为0.56%，预测时间小于1 ms；集成树虽相对在未压缩状态下性能最好，但其对结构压缩更为敏感，压缩后损失的性能较大．最后实现了模型核心逻辑通过统一的参数解析格式在不同硬件平台与不同机型之间的快速移植和验证，增强了模型部署的可解释性．更多还原

【Abstract】 To address the issues of reduced engine reliability and increased maintenance costs caused by sensor failures under harsh operating conditions,a data-driven lightweight virtual airflow sensor solution was proposed and validated. This approach first evaluated the performance of four machine learning models including regression trees,ensemble trees,support vector machines,and neural networks in airflow prediction tasks. Ensemble trees and neural networks,selected for their superior performance and fundamentally distinct algorithms,were prioritized for optimization. Bayesian optimization algorithms were employed to tune their parameters,and a resourceconstrained hybrid optimization strategy was proposed to balance storage requirements and prediction accuracy. Based on five-fold cross-validation,the neural network model file size is only 0.062 5 MB. It demonstrates strong modeling capabilities and deployment advantages across 56 operating conditions,achieving an average relative error of 0.56% on the test set with prediction times less than 1 ms. The ensemble tree model performs best in its uncompressed state,but it is more sensitive to structural compression and suffers significant performance loss after compression. Finally,the model’s core logic achieves rapid portability and verification across different hardware platforms and vehicle models through a unified parameter parsing format,enhancing the interpretability of model deployment.更多还原