【作者】 雷亚国；

【导师】 何正嘉； 訾艳阳；

【作者基本信息】 西安交通大学 ， 机械工程， 2007， 博士

【摘要】 大型复杂机械设备的故障往往表现为复杂性、不确定性、多故障并发性等,运用单一的智能故障诊断技术,存在精度不高、泛化能力弱等问题,难以获得满意的诊断效果,故急需一种新的思路和方法来解决这些工程实际问题。利用人工神经网络、模糊逻辑、遗传算法等单一智能技术之间的差异性和互补性,扬长避短,优势互补,并结合不同的现代信号处理技术和特征提取方法,将它们以某种方式综合、集成或融合,提出混合智能诊断技术,能够有效地提高诊断系统的敏感性、鲁棒性、精确性,降低它的不确定性,准确定位故障发生的位置,估计其严重程度。因此,研究混合智能技术及其在故障诊断中的应用,具有重要的科学理论意义和工程应用价值。论文正是围绕这一艰难而又诱人的主题,以机械设备的早期、微弱和复合故障的诊断为目的,对混合智能故障诊断技术的基本原理和工程应用进行了深入的研究。论文介绍了模糊逻辑、神经网络、聚类算法和遗传算法等技术的基本概念和原理,针对每种技术各举一例,说明其使用方法和有效性。描述了两种适合于处理非平稳、非线性信号的现代信号处理技术:小波包分析和经验模式分解。小波包分析是小波变换的延伸,以不同的尺度将动态信号正交地分解到相互独立的频带中,提供无冗余、不疏漏的独立频带分解信号的特征信息;经验模式分解方法基于信号的局部特征时间尺度,把动态信号分解为若干个本征模式分量,正交地给出分解信号的本征信息。所以二者分别从不同角度来分析信号,各具特色。为了提高机械故障诊断的准确性,结合小波包分解和经验模式分解方法在分析动态信号上的优势,特征评估方法在选取敏感特征方面的特点,以及径向基函数神经网络分类能力强的优点,提出了一种基于特征评估和神经网络智能故障诊断模型。该模型能够针对不同诊断问题选择其相应的敏感特征,克服了传统方法在特征选择上的盲目性。通过对滚动轴承局部损伤故障和烟气轮机转子轻微摩擦故障的诊断研究,应用结果表明:利用小波包分析和经验模式分解方法能从动态信号中精细地获得更多的故障特征信息;利用特征评估方法能够从原始特征集中评选出敏感特征,从而大大提高了径向基函数神经网络诊断的准确率。针对机械设备中早期故障和复合故障并发的复杂诊断问题,利用统计分析、经验模式分解、改进的距离评估技术、自适应神经模糊网络和遗传算法等技术,提出了一种综合多征兆域特征集和多个分类器组合的混合智能诊断模型。该模型运用多种信号预处理方法挖掘潜藏在动态信号中的故障信息,并综合利用从不同侧面表征机械设备运行状态的时域和频域统计特征,构成多元征兆域特征集来全面反映故障特性;利用基于不同输入特征集的多个自适应神经模糊网络之间的独立性和互补性,将其组合成混合智能模型。混合智能模型在机车轮对轴承的故障诊断中实现了轴承不同故障类型,不同故障程度,以及复合故障的可靠识别,获得了非常满意的诊断结果。同时,诊断结果也验证了提出的基于改进距离评估技术的特征选择方法的有效性。针对故障诊断中应用最多的无监督聚类算法——模糊C均值算法存在的问题,提出了一种新的混合智能聚类算法。该算法使用聚类评价指标自动确定聚类数;利用基于梯度下降的3层前馈神经网络通过无监督训练来自适应学习特征权值;运用基于点密度函数的算法计算样本权值。赋予特征和样本以相应的权重,强调敏感特征和典型样本对聚类的贡献,削弱无关特征和模棱两可样本对聚类的干扰,以提高聚类的性能。采用国际公认比较聚类算法性能的典型数据IRIS验证了混合智能聚类算法的有效性。在电力机车轮对轴承单一故障、早期故障和复合故障并发的诊断问题中,进一步验证了该算法不仅能正确地确定聚类数,而且聚类性能优于模糊C均值聚类算法,具有更好实用性和推广性能。阐述了基于网络的远程状态监测与故障诊断系统的必要性。介绍了“潜艇模型振动监测与分析系统”和“皮带输送机轴承状态检测与故障诊断系统”两个远程状态监测和故障诊断系统的总体框架;规划了两个不同系统的功能;提出了潜艇模型混合智能振动源辨识方法和皮带输送机滚筒轴承混合智能故障诊断方法;着重研究了混合智能技术在其中的应用。更多还原

【Abstract】 Faults of large-scale and complex mechanical equipments are characterised by complexity, uncertainty, syndrome, et al. If a single intelligent technique is utilized to diagnose these faults, it would be too difficult to obtain a satisfied diagnosis result. Generally, the diagnosis accuracy of the single intelligent technique is lower and generalization ability is weaker. Thus, it is urgent and necessary to present a novel idea and method to solve these practical engineering problems.According to the diversity and the complementarity between individual intelligent techniques, i.e. artificial neural network (ANN), fuzzy logic (FL), genetic algorithm (GA), et al., we may utilise their own merits and overcome their own shortcomings, and reinforce their advantages. By synthesising, integrating or fusing these individual intelligent techniques and different modern signal processing techniques and feature extraction methods via some means to propose hybrid intelligent diagnosis techniques, we can efficiently improve sensitivity, robustness and accuracy of a diagnosis system, reduce its uncertainty, ascertain the fault place exactly, and evaluate its severity. Therefore, it is quite worthy to investigate the hybrid intelligent technique and its applications in fault diagnosis for scientific theory studies and engineering applications. This dissertation just focuses on the extremely difficult but very attractive thesis. Aiming at incipient, slight and compound faults occurring in the mechanical equipments, the dissertation detailedly explores the fundamentals and engineering applications of the hybrid intelligent fault diagnosis techniques.The dissertation introduces basic conceptions and principles of fuzzy logic, neural network, clustering algorithm, genetic algorithm, et al., and provides an illustration for each technique to show its use and validity. Two advanced signal processing techniques suitable to nonstationary and nonlinear signals, wavelet packet analysis (WPA) and empirical mode decomposition (EMD), have been presented. WPA is an extended result of wavelet transform (WT). It orthogonally decomposes a dynamic signal into several independent frequency bands that link up mutually without redundant or omitted information. EMD method, which is based on the local characteristic time scales of a signal, adaptively decomposes the dynamic signal into a series of intrinsic mode functions (IMFs) and orthogonally presents intrinsic information of the signal. Thus, WPA and EMD have their own characteristics and could analysis the dynamic signal from different aspects, respectively.In order to improve the accuracy of fault diagnosis, we combine the superiority of WPA and EMD in processing dynamic signals, the advantage of feature evaluation method in selecting sensitive features and the strong classification ability of radial basis function neural network, and propose an intelligent fault diagnosis model based on feature evaluation and neural network. Aiming at various fault diagnosis problems, this model is able to automatically select the corresponding sensitive features and overcome blindness of traditional methods in selecting features. This model is applied to the local defects diagnosis of rolling element bearings and the slight rub fault diagnosis of a fume turbine rotor. The results demonstrate that more fault characteristic information can be precisely extracted by adopting WPA and EMD, the sensitive features can be easily selected from a large number of features with the feature evaluation method, and therefore the diagnosis accuracy has been greatly improved finally.Aiming at complex diagnosis problems of the intercurrent incipient fault and compound faults, a novel hybrid intelligent diagnosis model based on feature sets from multiple symptom domains and multiple classifier combination, is proposed, which combines statistics analysis, EMD, the improved distance evaluation technique, adaptive neuro-fuzzy inference system (ANFIS) and GA techniques. This model employs several signal preprocessing methods to mine the underlying fault information from dynamic signals. Time-domain and frequency-domain statistical features that reflect the equipment operation conditions from various aspects are synthesised to construct the multiple feature sets, which are able to completely present fault characteristics. Based on the independency and the complementarity of multiple ANFISs with the different input feature sets, we combine them and develop the hybrid intelligent diagnosis model. The practical application results of fault diagnosis of locomotive wheel pair bearings show the hybrid model is able to reliably recognise not only different fault categories and severities but also the compound faults. Thus, a desired diagnosis effect has been obtained via the hybrid model. Moreover, the application effect also validates the power of the proposed feature selection method based on the improved distance evaluation technique.Aiming at the existing shortcomings in the most popular unsupervised clustering algorithms used in the fault diagnosis field, fuzzy C-means (FCM) clustering algorithm, a novel hybrid intelligent clustering algorithm is developed. In this algorithm, the cluster number is automatically set by using the cluster validity index, feature weights are adaptively learned via a three-layer feed forward neural network with the gradient descent technique under the unsupervised mode of training, and sample weights are computed through the algorithm of distribution density function of data point. Then, the feature weights and the sample weights are assigned to the corresponding features and samples to emphasize the leading effect of sensitive features and typical samples, and weaken the interference of unrelated features and vague samples to improve the clustering performance. The test result of the benchmark data IRIS demonstrates the validity of the proposed algorithm. The algorithm is also employed to the single, incipient and compound fault diagnosis of locomotive wheel pair bearings. The results show that the hybrid intelligent clustering algorithm enables to automatically and correctly set cluster number, its clustering performance is superior to that of the FCM, and have a better practicability and generalisation.The necessity of developing remote condition monitoring and fault diagnosis systems is presented. The structures of two remote condition monitoring and fault diagnosis systems:“Monitoring and analysis system of vibration for the submarine model”and“Bearing condition monitoring and fault diagnosis system of strap transportation machines”, are introduced respectively. The different functions of the two systems are developed. The hybrid intelligent vibration source identification method for the submarine model and the hybrid intelligent fault diagnosis method for the roller bearings of the strap transportation machines are proposed. The application of the hybrid intelligent technique in the two systems is detailedly studied in the dissertation.更多还原