EI、Scopus 收录
中文核心期刊
赵健, 刘彦辰, 朱冰, 李扬, 李雅欣, 孔德成, 姜泓屹. 基于SHAP-RF框架的越野车辆路面识别算法研究. 力学学报, 2022, 54(10): 2922-2935. DOI: 10.6052/0459-1879-22-229
引用本文: 赵健, 刘彦辰, 朱冰, 李扬, 李雅欣, 孔德成, 姜泓屹. 基于SHAP-RF框架的越野车辆路面识别算法研究. 力学学报, 2022, 54(10): 2922-2935. DOI: 10.6052/0459-1879-22-229
Zhao Jian, Liu Yanchen, Zhu Bing, Li Yang, Li Yaxin, Kong Decheng, Jiang Hongyi. Research on road recognition algorithm of off-road vehicle based on SHAP-RF framework. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(10): 2922-2935. DOI: 10.6052/0459-1879-22-229
Citation: Zhao Jian, Liu Yanchen, Zhu Bing, Li Yang, Li Yaxin, Kong Decheng, Jiang Hongyi. Research on road recognition algorithm of off-road vehicle based on SHAP-RF framework. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(10): 2922-2935. DOI: 10.6052/0459-1879-22-229

基于SHAP-RF框架的越野车辆路面识别算法研究

RESEARCH ON ROAD RECOGNITION ALGORITHM OF OFF-ROAD VEHICLE BASED ON SHAP-RF FRAMEWORK

  • 摘要: 根据越野车辆在不同路面上行驶时的动力学响应特征, 可以实现路面类型的在线识别, 为面向路面特征调整底盘控制子系统参数从而获取更好的行驶性能奠定基础. 但越野环境地面特征复杂, 车辆响应机理分析困难, 给基于车辆动力学响应进行路面准确识别带来挑战. 提出了一种SHAP-RF路面识别算法设计框架, 通过SHAP (Shapley additive explanations)模型解释方法实现高维随机森林(random forest, RF)路面识别模型的降维化: 首先采集了试验车在压实土路、沙地、良好沥青路与冰雪路4种路面上的行驶数据并计算了3个次级行驶特征; 进一步计算了行驶数据的共计105个时域特征和频域特征, 并以此为输入特征建立了高维随机森林路面识别模型; 利用SHAP解释法分析高维模型输入特征对识别结果的影响从而提炼出各个特征与路面类型的关联性, 完成特征筛选; 最后, 利用筛选后的特征设计降维随机森林路面分类器. 基于实车数据的算法验证试验表明, 设计的降维路面识别模型对4种路面的识别精确率在94%以上, 召回率在93%以上, 相比高维的随机森林路面识别模型, 各种路面上的精确率和召回率最大降幅不超过3.2%, 证明本文提出的SHAP-RF路面识别算法设计框架能够在选用较少特征的情况下依然保证车辆行驶路面类别的准确识别.

     

    Abstract: The road identification of off-road vehicles can be carried out according to the dynamic response on different road surfaces, which lays a foundation for adjusting the parameters of the chassis control subsystem to obtain better driving performance. However, it is difficult to analyze the response mechanism of vehicles on different road surfaces due to the complexity of the off-road environment, which brings challenges to accurate road recognition based on vehicle dynamic response. In this paper, an SHAP-RF road recognition algorithm design framework is proposed, which realizes dimensionality reduction of the high-dimensional RF (random forest) road recognition model through the SHAP (Shapley additive explanations) model interpretation method. Firstly, we collected the driving data of the test vehicle on soil road, sand road, good asphalt road, and snow-icing road, and then three secondary driving features were calculated. Furthermore, a total of 105 features of driving data were calculated, including time domain features and frequency domain features. A high-dimensional RF road recognition model was established with all the features as input. The SHAP interpretation method was used to analyze the influence of input features on the recognition results in the high-dimensional model, and the correlation between each feature and road type was extracted to complete feature screening. Finally, a dimensional-reduction RF road recognition model is designed using the selected features. The validation test of the algorithm based on real vehicle data shows that the identification accuracy rate of the dimensional-reduction road recognition model is above 94% for all four kinds of road, and the recall rate is above 93%. Compared with the high-dimensional RF road recognition model, the accuracy rate and recall rate on all kinds of roads drop by no more than 3.2%. This proves that the proposed SHAP-RF road recognition algorithm design framework can reduce the number of input features while ensuring the recognition accuracy of road categories.

     

/

返回文章
返回