【作者】 魏秀业；

【导师】 潘宏侠；

【作者基本信息】 中北大学 ， 火炮、自动武器与弹药工程， 2009， 博士

【摘要】 对复杂机械系统进行状态监测与故障诊断是人们普遍重视和关注的课题,本文在深入系统地研究粒子群优化算法理论以及算法参数、性能的基础上,提出了基于参数策略的粒子群改进算法,以齿轮箱为研究对象,研究基于粒子群优化的齿轮箱智能故障诊断理论与方法,主要研究内容及结论如下:1.阐述了粒子群优化算法的基本原理,通过对其粒子速度进化方程的分析,研究粒子本身的行为和社会行为以及主要控制参数对粒子群优化算法性能的影响。2.为了改善了基本粒子群优化算法收敛性能,提出了两种基于参数改进策略的粒子群优化算法,即动态加速常数的粒子群优化算法和速度自适应粒子群优化算法,改进算法是在基本算法中分别将加速常数和最大限制速度设置为随进化代数变化的函数,并在测试函数和神经网络中进行了仿真研究,结果表明:改进算法加快了基本算法的收敛速度,并给出了改进算法参数的合理取值范围。进一步研究了加速常数、最大限制速度与惯性权重之间互相协同对粒子群优化算法的控制的优劣问题。结论是采用动态加速常数协同惯性权重的粒子群优化算法,可以克服标准粒子群算法搜索后期陷入局部搜索的不利之处,算法具有良好的性能。3.针对核主元分析方法在核函数参数选择上的盲目性,提出并实现了基于粒子群优化的核主元分析的故障特征选择方法。建立以Fisher判别函数为优化目标的适应度,利用粒子群算法中多个随机粒子实现核函数参数的优化,改善了核主元分析方法的性能。通过IRIS数据仿真分析,验证了该方法用于特征向量的选择的正确性和有效性。将优化的核主元分析方法应用于齿轮箱典型故障的特征提取中,结果表明:参数优化的核主元分析能有效地降低齿轮箱特征向量的维数,较线性主元分析取得更好故障识别效果。该方法在机械故障信号的非线性特征提取中具有优势。4.提出了基于粒子群优化的齿轮箱传感器优化配置方法,解决多测点传感器的布置和定位问题。建立了基于模态置信准则的适应度,根据齿轮箱有限元模态计算结果,用改进的粒子群优化算法寻求满足适应度要求的传感器布置方案,齿轮箱试验模态分析和频响特性分析的结果验证了所提出的方法的合理性。5.提出了基于动态加速常数和速度自适应粒子群优化的智能故障诊断方法。以齿轮箱振动信号的时、频域特征为神经网络输入,以齿轮箱的主要故障形式为输出,建立基于粒子群优化算法神经网络故障诊断系统。在训练和诊断过程中,粒子群优化算法作为一种粗优化或离线学习过程,调节和优化具有全局性的网络参数,如权值和阈值等;而用神经网络学习作为一种细优化或在线学习过程方法,优化具有局部性的参数。诊断结果表明:所提出的智能诊断方法提高了齿轮箱故障诊断性能,为非线性复杂系统的故障诊断效率和精度以及诊断自动化的提高提供一种通用的解决方案。更多还原

【Abstract】 Condition monitoring and fault diagnosis on complicated machine is a popular subject taken importantly and cared about by people.This thesis has proposed a modified algorithm for PSO based on systematical and deep research on particle swarm optimization (PSO) algorithm and its parameters, performance. It takes gearbox as a researched object, has studied on the theory and method of gearbox intelligent fault diagnosis with PSO. Its contribution as follows:1. It introduces the basic principle of PSO algorithm, and studies on the particle’s itself behavior and social behavior of the algorithm by analysis of the particle’s velocity evolutionary equation, then analyses the influences on performance of PSO by the main control parameters.2. It puts forward two modified PSO algorithms based on the strategy of parameters change to improve the convergence performance of the basic PSO algorithm, which are the PSO with dynamic accelerating constants (CPSO) and the PSO with adaptive velocity (VPSO). The accelerating constants and maximum limited velocity are set dynamic alternation with iteration in modified PSO. They are carried out simulated research by the test function and artificial neural networks, and the results show that modified PSO speed-ups the convergent velocity of basic PSO algorithm, while the rational parameter range of the modified PSO has been suggested. In addition, it studies on the priority problem of algorithm controlled by accelerating constants, the maximum limited velocity, and their coordinate with inertia weight. The conclusion is that the PSO with dynamic accelerating constant and coordinating with inertia weight (WCPSO) may overcome the disadvantage of the standard PSO which easily converges in local extreme point in the final period of searching process, and it has better performance.3. Aimed at the blind setting of parameter in kernel principal component analysis (KPCA), it proposes and realizes feature extraction based on kernel principal component analysis optimized by PSO algorithm. Firstly, it constructs a fitness function which Fisher discriminate function is optimized object, then WCPSO is used to optimize it by its many random particles to improve the performance of KPCA. Iris data are researched in simulation, which testify KPCA validity in feature extraction. The optimized KPCA is applied to feature extraction of gearbox typical faults.The results indicate that KPCA after parameter optimized can effectively reduce the dimensions of feature vector of gearbox, and it has a better fault classification performance than linear principal component analysis (PCA). This method has an advantage in nonlinear feature extraction of mechanical failure signal.4. It presents a method of optimum placement of sensors in gearbox based PSO algorithm to solve the problem of sensors layout and localization. First, it establishes the fitness function based on mode assurance criterion;then according to the results of gearbox finite element mode analysis, WPSO is applied to optimize it to find the optimal strategy of optimum placement of sensors in gearbox. The results of the test modal analysis and frequency response character analysis for gearbox testify the rationality of the method supposed in the thesis.5. It sets up intelligent fault diagnosis based on PSO with dynamic accelerating constants and adaptive velocity. It establishes the neural network (NN) optimized by PSO for fault diagnosis, in which the features in time domain and frequency-domain from gearbox vibration signal are taken as input vector, while its main fault types as output vector of NN. In training and detecting process, PSO algorithm regulates and optimizes the global parameters such as weights and the thresholds of NN, which acts as a roughly optimizing or off-line studying process; while neural network does local ones, which acts as a fine optimizing or on-line studying process. The diagnostic results show that the method of intelligent fault diagnosis has improved the performance of gearbox fault diagnosis, and provided a universal solving programme for the nonlinear and complicate system to improve the efficient and accuracy of diagnosis and its automation.更多还原