【作者】 吴浩中；

【导师】 王开文；

【作者基本信息】 西南交通大学 ， 车辆工程， 2002， 博士

【摘要】 我国是一个幅员辽阔、山区铁路众多的发展中国家，摆式列车技术的应用具有巨大的潜力。倾摆控制系统是摆式列车的核心部分，其可靠性对列车的安全运行至关重要，故障诊断技术为提高系统的可靠性开辟了一条新的途径。论文结合铁道部科技开发项目：“摆式列车倾摆控制系统的研制”，对摆式列车倾摆控制系统的故障诊断进行了探索性研究。 摆式列车倾摆控制系统是一个复杂的、具有诸多不确定性因素的动态系统，采用常规的故障诊断方法很难达到有效的诊断。神经网络作为一种新的方法和手段，已被广泛地应用于诸多领域的故障诊断中，但其在机车车辆故障诊断中的应用少有报道。论文在国内首次系统地将神经网络引入摆式列车倾摆控制系统的故障诊断中，从工程应用的角度深入研究了基于神经网络的故障诊断理论，探讨了其在摆式列车倾摆控制系统设计中的应用问题。 论文的主要创新成果有： (1)在摆式列车实际运行中，检测子系统的加速度传感器存在严重零偏，使用硬件冗余法进行故障诊断的设计初衷面临着巨大的困难。论文采用RBF神经网络构成传感器输出预测器，通过检测预测值与实际值之差来在线检测故障。对每个加速度传感器构造输出预测器，然后通过一复合故障决策策略来实现多传感器故障的同时诊断，该方法不仅消除了零偏对故障诊断的影响，而且提高了故障检测率及可靠性。 (2)根据检测子系统采用双冗余陀螺仪的实际情况，利用自适应噪声消除器的基本原理构造了双余度传感器的故障诊断模型，给出了双余度传感器的故障诊断原理和递推算法，实现了双冗余陀螺仪的故障诊断。 (3)作动子系统是一个复杂的非线性系统，其故障存在模糊性和相关性。论文将模糊逻辑和神经网络结合在一起，给出了模糊BP神经网络的结构、隶属度函数确定方法，并提出了基于阈值原则的故障判别方法，用模糊神经网络对作动子系统的常见故障进行了诊断。 (4)针对小波基的选择缺乏完善理论指导的实际情况，通过对常用的正交、半正交、双正交小波基提取信号特性的比较，确定出了适合摆式列车倾摆控制系统故障诊断的小波基——半正交小波基，并提出了一种新的确定故障检测阈值的方法，可以快速地检测出故障发生时刻。 (5)将小波分析技术与神经网络技术相结合，形成广义上的小波神经网 西南交通大学博士研究生学位论文 第！I页络，提出了一种基于小波包预处理的神经网络故障诊断方法。利用小波包分析，将摆式车体试验台上采集到的振动加速度信号分解在相互独立的频带之 p内，各频带内的能量值形成一个向量，将其作为神经网络的输入特征向量，＼。然后用神经网络对摆式列车倾摆控制系统的常见故障进行了识别和诊断。 实验和研究结果表明，论文提出的基于神经网络的摆式列车倾摆控制系统故障诊断方法在理论上是可行的，在工程上也是可实现的。采用神经网络理论诊断摆式列车倾摆控制系统故障，为我国的机车车辆尤其是高速列车的故障检测和诊断指出了新的研究方向，论文的研究结果对摆式列车倾摆控制系统的设计具有理论指导意义。更多还原

【Abstract】 The tilting train is a way of high-speed railways suiting for the situation ofour railways and country, which has great economic potential. The tilting colltrolsystem is the key technique of tilting train, whose safety is very important fortilting train to transport normally. Fault diagnosis technology has provided a newmeans to enhance the reliability of system. Fault diagnosis of tilting controlsystem in tilting train is discussed in this thesis, which is the main substance ofdeveloping the tilting control system supported by the R&D Program of Ministryof Railways of China.Considering the fact that the tilting control system of tilting train is workingin quite an uncertain environment, a conventional fault diagnosis method is hardto perfOrm effectively. Neural network has been widely used in fault diagnosis inmany fields as a novel method or measure. However, there are very fewpublished studies dealing in neural network used in fault diagnosis of thelocomotive and rolling stock. Neural network has been systematically introducedinto fault diagnosis of tilting contro1 system in tilting train for the first time inour country by this thesis. Fault diagnosis theory based on neural network isstudied in detail. Some practical issues of the tilting control system design areaddressed by applying neural network.The main contributions are as follows.(1) In view of practice that the data obtained from accelerometer has seriouszero deviation, it is very difficult to diagnosis fault by hardware redundant. Afault diagnosis scheme for multi-accelerometers employing predictor based onradial basis function (RBF) neural network is proposed. The strategy eliminatesthe disturbance with zero deviation and improves the reliability of faultdiagnosis.(2) In the light of reality adopting double redundant gyroscope, the faultdiagnosis model for double redundant sensors taking the basic principle ofadaptive noise cleaner is established. The diagnosis method, recurrence algorithmand implement are given.(3) On the basis of the fuzzy property and correlation characteristics existingin the fault of tilting driving subsystem, several typical faults in tilting drivingsystem are analyzed and identified with the fiizzy neural network (FNN). The structure of FNN and the membership function are given. Meanwhile a fault discriminate method for threshold vector is addressed.(4) Since there is no perfect theoretical guide, the characteristics of orthogonal wavelet, semi-orthogonal wavelet and biorthogonal wavelet in signal decomposition and reconstruction are compared, and the conclusion of the study demonstrates that semi-orthogonal wavelet is a good wavelet for fault diagnosis in tilting control system of tilting train. A new approach is presented for deciding the fault threshold.(5) A generalized wavelet neural network is put forward by integrating wavelet analysis with neural network. A fault diagnosis method using wavelet packet as its pretreatment is proposed. The method using wavelet packet analysis is proposed to extract fault information from vibration signal obtained from testing jig of tilting train. The vector comprised of the energy of signal in all spectrum bands is input to a feed forward neural network. The trained neural network can explore the typical faults of tilting control system.Trial and research show that the fault diagnosis method based on neural network is practicable in terms of theory. A direction to set up the fault diagnosis of the locomotive and rolling stock, especially, high-speed train by means of neural network is pointed out. The research results in this thesis lay the theoretical foundation of fault diagnosis in tilting control system of the first tilting train of our country.更多还原