【作者】 王宁；

【导师】 宫洵；

【作者基本信息】 吉林大学 ， 控制科学与工程， 2024， 硕士

【摘要】 柴油车尾气排放是大气污染的主要来源,为此国家出台了一系列法规以加强对商用车的监管,而油品监管是商用车监管中的重要一环。尽管我国有严格的车用标准柴油质量要求,但高硫柴油因其成本低廉的优势仍然存在于柴油市场。使用高硫柴油会破坏车辆后处理系统,造成车辆排放升高并加重企业运维部门的负担,因此高硫柴油识别问题受到企业和学术界的广泛关注。然而,油品识别技术的开发存在着高硫柴油异常数据获取难、瞬态异常数据识别难、缺乏系统级部署验证等问题。随着大数据、云计算和人工智能技术的飞速发展,商用车企业的网联化、智能化水平不断提升,为解决高硫柴油识别问题带来了新的机遇。针对上述挑战,本文致力于建立网联环境下的高精度和可部署的高硫柴油识别方法和系统,为企业提供技术支持。主要研究内容如下:(1)针对高硫柴油异常数据获取难的问题,依托一汽解放车云一体化平台,进行长达一年的整车实验,构建了面向高硫柴油识别问题的多特征、多工况、多环境的1000小时以上的自然行驶数据集,考虑网联数据集中冗余、断续和丢包等现象,基于选择性催化氧化器(Selective Catalytic Reduction,SCR)硫中毒机理进行特征选择并设计了连续性分割结合滑动窗口的数据预处理方法,为后续高硫柴油识别模型的开发提供数据支撑。(2)针对样本平衡条件下异常数据识别问题,提出了基于Transformer的正常-高硫柴油分类方法,选取SCR的上下游NOx浓度、尿素喷射量、排气温度和流量等特征,通过嵌入层、堆叠的多头注意力(Multi-Head Attention,MHA)层和油品检测层挖掘不同油品序列间的差异,实现区别正常和高硫柴油的功能,实验表明该方法能够实现94.23%的高识别精度,并通过灵敏度分析证明所选参数和特征的合理性,为企业应用奠定了基础。(3)针对样本不平衡条件下异常数据识别问题,提出了数据知识融合的高硫柴油异常检测方法,采用仅基于正常样本数据训练的SCR下游NOx浓度回归模型计算重构残差以量化序列异常程度,通过基于知识的双阈值残差规则完成油品识别,实验表明该方法能够捕捉到SCR内部的反应规律,在仅由正常样本训练模型时也能取得同二分类方法相近的优良的识别精度。(4)面对系统级部署和验证,开发了包含“终端-网络-平台-应用”架构的高硫柴油识别系统,引入“局部-全局”的序列油品推理机制以解决不定长序列的油品识别问题,实现油品信息的可回溯、可监控和可统计的功能。考虑流数据场景下面临的数据漂移问题,提出了系统模型和参数的在线自适应更新方法。目前该系统通过企业验证测试并已在一汽解放车辆远程数据服务平台上部署运行。更多还原

【Abstract】 Diesel vehicle exhaust emissions are a major source of atmospheric pollution.To address this,the government has implemented a series of regulations to strengthen supervision of commercial vehicles,with fuel regulation being a crucial aspect.Despite stringent diesel quality requirements for vehicles in our country,high-sulfur diesel remains prevalent in the market due to its low cost advantage.The use of high-sulfur diesel can damage vehicle after-treatment systems,leading to increased emissions and exacerbating the burden on enterprise operations and maintenance departments.Consequently,the identification of high-sulfur diesel has garnered widespread attention from both businesses and the academic community.However,the development of fuel identification technology faces challenges such as difficulty in obtaining abnormal data related to high-sulfur diesel,identifying transient abnormal data,and lacking systematic deployment verification.With the rapid advancement of big data,cloud computing,and artificial intelligence technologies,the level of connectivity and intelligence in commercial vehicle enterprises continues to improve,presenting new opportunities for addressing the identification of high-sulfur diesel.Addressing the aforementioned challenges,this paper aims to establish a high-precision and deployable method and system for identifying high-sulfur diesel in a connected environment,providing technical support to enterprises.The main research contents are as follows:(1)To address the challenge of obtaining abnormal data for high-sulfur diesel,a year-long full-vehicle experiment was conducted using the FAW Jiefang cloudintegrated platform.This experiment resulted in the creation of a natural driving dataset exceeding 1000 hours,covering multiple features,conditions,and environments relevant to the high-sulfur diesel identification problem.Considering issues such as redundancy,intermittency,and packet loss in connected vehicle data,a feature selection method based on the sulfur poisoning mechanism of Selective Catalytic Reduction(SCR)was implemented.Additionally,a data preprocessing method combining continuous segmentation with a sliding window was designed.This provides the necessary data support for the development of subsequent high-sulfur diesel identification models.(2)To address abnormal data identification under sample balance conditions,a Transformer-based normal-high-sulfur diesel classification method was proposed.This method utilizes embedding layers,stacked multi-head attention(MHA)layers,and fuel detection layers to explore differences between different fuel sequences,achieving differentiation between normal and high-sulfur diesel.Experimental results demonstrate a high identification accuracy of 94.23%,and sensitivity analysis validates the rationality of selected parameters and features,laying the foundation for enterprise applications.(3)To address abnormal data identification under sample imbalance conditions,a high-sulfur diesel anomaly detection method incorporating data knowledge fusion was proposed.This method utilizes a SCR downstream NOx concentration regression model trained solely on normal sample data to calculate reconstruction residuals,quantifying sequence anomaly severity.Oil identification is then completed through a knowledge-based dual-threshold residual rule.Experimental results show that this method accurately captures SCR reaction patterns and achieves excellent identification accuracy comparable to binary classification methods when trained only on normal sample data.(4)To address system-level deployment and verification,a high-sulfur diesel identification system was developed,featuring a "terminal-network-platformapplication" architecture.A "local-global" sequence fuel reasoning mechanism was introduced to address the issue of identifying fuel in sequences of varying lengths,enabling traceability,monitoring,and statistical analysis of fuel information.To account for data drift in streaming data scenarios,an online adaptive update method for system models and parameters was proposed.The system has undergone enterprise verification testing and is currently deployed and operational on the FAW Jiefang Vehicle Remote Data Service Platform.更多还原