【作者】 刘应吉；

【导师】 张天侠；

【作者基本信息】 东北大学 ， 车辆工程， 2008， 博士

【摘要】 作为国民经济的支柱产业,汽车工业的发展受到了世界各国的高度重视,激烈的市场竞争促进汽车生产和研发水平不断提高。车辆系统结构、功能日趋复杂,车辆故障种类也日益多样化,这些都对汽车故障诊断和监控技术提出了更高要求。本文以汽车最关键部件—发动机为研究对象,在分析汽车发动机故障诊断研究现状及存在问题的基础上,重点研究其状态监测和故障诊断的理论和方法,包括信号采集、信号处理、神经网络、模糊推理系统、信息融合理论、车上网络通信技术以及虚拟仪器等技术。在此研究基础上提出并设计了一种综合上述理论和技术方法的发动机综合故障诊断测试平台,解决了传统发动机故障诊断方法中存在的一些问题。研究的主要内容包括以下几部分：(1)在故障诊断特征提取方法方面,针对于发动机缸体采集的振动信号,研究时域分析、频域分析及小波变换等三种故障信号特征提取方法。对于利用振动信号进行发动机机械故障诊断存在的问题和现代汽车自诊断系统只适用于对车辆电控单元故障诊断的局限性,确定以发动机多种运行状态参数作为故障诊断模型输入特征向量。提出并设计了基于CAN总线和SAE J1939协议的发动机在线故障诊断系统,完成系统软硬件的设计,实现诊断信息提取和传输。(2)在故障诊断技术方法方面,着重研究了三种典型神经网络(BP网络、RBF网络和PNN网络)及自适应神经模糊推理系统(ANFIS)的基本原理、模型结构和算法设计。针对于BP网络进行了多种算法的改进研究,并对相应的改进结果进行了对比分析,提出了合理选择这些算法的指导思想。针对不同的特征向量提取方式(频域分析、小波分析、发动机运行状态参数),分别建立发动机神经网络及ANFIS故障诊断模型。针对发动机运行状态参数故障诊断特征向量存在较高相关性的问题,应用主成分分析法实现降维和去相关,确定能够表征故障的主要特征状态参数。通过诊断结果比较分析,选取和确定每种特征提取方式下的较优推理诊断模型。(3)建立了适用于发动机故障诊断的信息融合结构模型。对发动机故障诊断的多源信息,采用主成分分析进行特征级融合,采用D-S证据理论进行决策级融合。针对D-S证据理论在信息融合过程中存在的对高冲突证据失效问题,提出一种改进的D-S融合方案,将由BP、RBF、ANIFS模型获得的发动机故障诊断结果进行融合,能够有效解决失效问题,提高诊断结果准确率、确定度和实时性。(4)根据论文理论研究成果,研发了一套完整发动机状态监测与故障诊断综合系统,利用该系统在发动机试验台架上分别完成了发动机无负荷测功实验、基于发动机振动信号的故障诊断实验和基于发动机运行状态参数的故障诊断实验,实现了对论文提出的故障特征提取和诊断理论方法的全面验证。更多还原

【Abstract】 As a pillar industry of national economy, automotive industry has been paid more and more attention all over the world. Fierce market competition promotes automobile production and scientific research to improve constantly. Increasing complications of vehicle structure and function bring diversification of fault type and then put forward a higher requirement for fault diagnosis and detection technique of vehicles. As the core component of automobile, engine was selected as the study object in this thesis, and the theory and methods of engine condition monitoring and fault diagnosis were mainly concerned. The study includes signal collection, signal processing, neural networks, fuzzy inference system, information fusion theory, network communication technique of vehicle and virtual instrument technology etc. An integrated test platform of engine fault diagnosis was presented and designed based on the study, and some problems of the traditional engine fault diagnosis were solved. The main contents of the study are as follows:(1) As for the method of diagnosis characteristic extraction, aiming at the vibration signals obtained from engine cylinder, three types of characteristic extraction including time domain analysis, frequency domain analysis and wavelet transformation were studied. Considering that the problems existing in fault diagnosis of engine mechanical structure by vibration signals, and the limitation of self-diagnosis system in engine which can only be applied to diagnosis of the electronic control unit (ECU) fault, multiple condition parameters of engine were selected as feature vector of fault diagnosis model. An on-line fault diagnosis system based on CAN bus and the SAE J1939 protocol was proposed, and then design of the system hardware and software was performed to realize extraction and transmission of diagnosis information.(2) As for the method of fault diagnosis, three classical neural networks (BP network, RBF network, PNN network) and adaptive neural-fuzzy inference system(ANFIS) were mainly studied about their basic principle, model structure and algorithm design. BP network was researched with several improved algorithms, then the improved results were comparably analyzed and guiding ideology was given for algorithm selection reasonably. According to different methods of feature extraction (frequency domain analysis, wavelet transformation and engine condition parameters data), fault diagnosis models of engine have been built based on the neural networks and ANFIS respectively. Aiming at the problem that the condition parameters of engine have correlations, principal component analysis was adopted to realize the dimension reduction and decorrelation. The major characteristics condition parameters which can be used to indicate faults were confirmed. Through the comparative analysis of diagnosis results, the most appropriate diagnosis model of each feature extraction method can finally be determined.(3) An information fusion structural model suitable for engine condition monitoring and fault diagnosis was built. For the multi-sources information of engine fault diagnosis, principal component analysis was adopted for feature level fusion, while D-S evidence theory for decision level fusion. Aiming at the evidence conflict phenomenon existing in the information fusion process, an improved information fusion method based on D-S theory was presented; results of engine fault diagnosis which were obtained separately by BP network, RBF network and ANFIS model have been fused by the improved method, this approach increased diagnosis accuracy, certainty degree and real time, and solved the evidence conflict effectively.(4) According to the research results of this thesis, an integrated system of engine condition monitoring and fault diagnosis was developed. a series of experiments including engine unloaded power detection, fault diagnosis based on the vibration signals of engine cylinder, fault diagnosis based on the condition parameters of engine, had been made on engine experimental bench. The experiment results verify the methods of fault feature extraction and fault diagnosis theory which proposed in this thesis.更多还原