A Machine Learning Based Fuel Consumption Saving Method with Time and Environment Dependency Aware Management
A Machine Learning Based Fuel Consumption Saving Method with Time and Environment Dependency Aware Management
复制标题
DOI：
10.1145/3531028.3531035
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
P. Yu;Hiroyuki Yamauchi
中科院分区：
文献类型：
--
作者：
P. Yu;Hiroyuki Yamauchi
由于影响长途车的燃油消耗的关键因素在整个卡车运行路线上每小时都有变化，该路线长距离并延伸到各个环境区，因此在以下两件事将要限制节省油耗的量已经能够做到了。首先是根据训练和驾驶辅助系统提高驾驶员的技能，从而可以通过提高驾驶员的技能，从而可以找到自己的问题，从而可以找到自己的问题，从而可以找到自己的问题。第二个是更好的机械维护，以保持更好的燃油效率。因此，本文提出了一种基于机器学习的时间和环境条件依赖性意识到的燃料消耗跟踪方法，该方法可以及时向驾驶员和卡车维护设施提供反馈。通过更清楚的是，根据其时间和环境依赖性分析结果，将提供反馈。由于嵌入式传感器（CANBUS＆GPS）获得的各种数据很复杂，并且其变量数量超过100（比传统案例少于50的情况大得多），因此我们最初使用了LightGBM，但我们面临偏见与差异权衡问题。为了解决这个问题，我们新提出了堆叠计划，该计划通过将随机森林，SVM和LightGBM结合起来来优化权衡。多亏了此方案，方差（RMSE）误差从6.04减少到5.03，同时将最佳准确度（R2）保持在97.6％。我们提出了反馈系统，以根据对驾驶员造成的关键影响因素的分析来减少燃料消耗。据我们所知，这是第一次证明，如果可以将某些关键因素仔细地减少50％，则提议的反馈方案可以节省总体燃料消耗。
Since the key factors affecting the fuel consumption of the long-distance vehicle are varied hourly on the whole truck running routes, which are long distance and stretched to various environmental district, the amount of saving fuel consumption must be limited until the following two things would have become able to do. The first is a better fuel-efficient driving control in terms of the low/high gear, accel, and the idling by increasing the driver's skills based on the training and driving assist systems that makes it possible to find own problem. And the second is a better mechanical maintenance to keep a better fuel-efficiency. Thus, this paper proposes a machine learning based time and environment condition dependency aware fuel consumption tracking method, which enables to give the timely feedback to the drivers and truck maintenance facility. The feedback will be given by making clearer what are the key factors affecting fuel consumption every period based on their time and environment dependencies analysis results. Since the various data obtained by the embedded sensors (CANbus &GPS) are complicated and its number of variables is over 100 (much larger than the conventional cases of less than 50), we initially used the LightGBM but we faced the bias vs variance tradeoff issue. To address this issue, we newly proposed the stacking scheme, which optimizes the trade-off by combining the random forest, the SVM, and the LightGBM. Thanks to this scheme, the variance (RMSE) error was reduced from 6.04 to 5.03, while keeping the best accuracy (R2) of 97.6%. We proposed the feedback system to reduce the fuel consumption based on the analysis of the key influencing factors that caused by the drivers. To our best knowledge, it is the first time to demonstrate that the saving of overall fuel consumption by the proposed feedback scheme can be reduced by 8% if some key factors could be carefully reduced by 50%.