基于化学反应优化算法和支持向量机的滚动轴承故障诊断方法研究

发布时间：2018-05-01 12:17

本文选题：人工化学反应优化算法 + 化学反应优化　；参考：《湖南大学》2014年博士论文

【摘要】：滚动轴承广泛应用于机械行业,故对滚动轴承故障诊断技术的研究具有重要意义。滚动轴承的故障诊断实际上是一个模式识别的过程,其关键是故障的特征提取和状态识别。时频分析法是常用于振动信号的特征提取方法,其分为线性分析法和非线性分析法两种类型。线性时频分析方法包括短时傅立叶变换(STFT), Gabor变换和小波变换。非线性时频分析方法包括离散傅立叶变换(DFT), Wigner-Ville分布,Choi-Wiliam分布。STFT通过对信号的局部化时频分析建立时域和频域两者之间的关系,但信号的高频和低频分量又是由海森堡不确定性原理确定。Wigner-Ville分布是一种更加直观和合理的非线性时频分布,同时它还能够描述信号的瞬时功率谱,但其在卷积过程产生的虚假分量,造成信号时频特征的混乱。傅立叶变换将信号分解为一系列不同频率段的正弦波,而小波变换将信号分解成它的小波基。而滚动轴承的振动信号是非线性和非平稳的,因此以上方法在分析非平稳和非线性信号时都具有一定的局限性。希尔伯特—黄变换(HHT)是一种自适应时频分析方法,包括经验模态分解(EMD)和Hilbert谱分析(HAS)两个部分。EMD是依据信号自身的时间尺度特征将信号分解为有限个单分量信号,称之为内禀模态函数(IMF)。该方法在非线性非平稳数据的分析中具有较大的潜力,尤其是在分析典型的时频能量交涉信号中,但是,EMD的端点效应,模态混淆,过包络与欠包络,负频率,瞬时频率,以及缺乏一定的理论基础等一系列问题,仍需要进一步研究。最近Smith提出了一种新的自适应时频分析方法Local Mean Decomposition,局部均值分解(LMD)。LMD的原理是基于滑动平均处理将一个复杂多分量信号分解为若干乘积函数分量(PF),所使用的滑动步长(MA)对分解结果具有重要影响。此外,纯调频(FM)信号的标准也会对该方法的性能产生影响。与EMD方法相比,LMD具有更多的优势,如更少的迭代次数,更平稳的端点效应和更高的瞬时频率分辨率。由LMD分解得到的局部均值不受过包络和欠包络的影响,同时LMD局部幅值比EMD上、下包络线更具有振荡特性。因此,局部均值和局部幅度都基于信号的局部特征时间尺度。本文研究了另一种自适应时频分析方法Local Characteristic-Scale Decomposition即局部特征尺度分解(LCD)。通过LCD方法可以将信号分解为若干内禀尺度分量(ISC),每一个ISC分量都包含有信号的局部特征,因此通过该方法能够提取到更精确和有效的原始信号特征信息。LCD在减少端点效应和迭代时间和瞬时特性的精度等方面的表现都要优于EMD。此外LCD能够改善EMD分解信号产生的边际效应。根据上述优点,本文提出了基于LMD和LCD的滚动轴承故障诊断方法。滚动轴承故障诊断实际上是一个状态识别过程。状态识别采用统计学习理论,监督与非监督分类。状态识别常用的状态识别方法有人工神经网络(Artificial neural network, ANN^和支持向量机(Support Vector Machine, SVM)。 ANN主要缺点是难以确定网络结构和参数,还需要大量样本,而在实际中很难获得大量的样本。另外,收敛速度慢也大大增加了计算时间。SVM是一个基于统计学习理论和结构风险最优原则的有效的模式识别方法,不仅能解决ANN中存在的过拟合、局部最优、收敛速度慢等问题,针对小样本还具有很好的泛化概括能力。SVM被广泛的应用于模式识别和其他领域。然而,SVM的参数选择对分类效果有很大的影响,而参数的选择实际上是一个优化过程,因此优化算法被应用于SVM的参数选择。遗传算法(Genetic algorithm,GA)和粒子群算法(Particle swarm optimization, PSO)都应用于SVM参数的优化。GA算法具有收敛速度慢,容易丢失局部最优解等问题。此外,GA并不能解决特定的优化问题以及变种问题。PSO算法在解决问题时有容易描述、容易实现、收敛速度快等优势,但是存在不能有效避免过早收敛的缺陷。近来,一种新的基于化学反应原理的优化算法被提出后,经验证在很多方面都优于其它优化算法。化学反应优化算法的思路来自化学反应的发生,模仿化学反应中分子的微观运动,通过利用化学反应生成物具有最低能量的现象而实现优化。基于化学反应原理的优化方法有两种算法：一种是化学反应优化(Chemical Reaction Optimization, CRO), CRO原理是基于系统的势能,当势能降低到最低限度时,反应系统会逐渐达到平衡状态,因此将势能作为最小化问题的目标函数可行的；另一种是人工反应优化算法(Artificial Chemical Reaction Optimization Algorithm, ACROA),焓和熵可以作为最小化和最大化问题的目标函数(状态函数)。焓取决于物质的化学性质,温度与压力的状态,而熵用来测量化学系统组件的随机性或病症。ACROA具有一个参数,初始反应物,因此这种方法很容易使用。论文分别将CRO和ACROA应用于SVM的参数优化。结果表明：基于化学反应优化的支持向量机(Chemical Reaction Optimization-Support Vector Machine, CRO-SVM)和基于人工反应优化算法的支持向量机(Artificial Chemical Reaction Optimization Algorithm-Support Vector Machine, ACROA-SVM)在解决分类问题上都优于基于遗传算法的支持向量机(Genetic algorithm-Support vectormachine, GA-SVM)和基于粒子群算法的支持向量机(Particle swarm optimization-Support vector machine, GA-SVM)。在此基础上,论文分别将局部均值分解(Local mean decomposition, LMD)和局部特征尺度分解(Local characteristic-scale decomposition, LCD)与ACROA-SVM、CRO-SVM相结合应用于滚动轴承故障诊断。论文主要工作和创新点如下： 1.对LMD和LCD两种时频分析方法进行了研究,分别将LMD、LCD与经验模态分解(Empirical mode decomposition, EMD)方法进行了对比分析,仿真信号和滚动轴承故障实验信号的分析结果表明,相对于EMD方法,LMD和LCD在计算效率的端点效应等方面具有优越性。 2.对化学反应算法进行了理论研究,分析了GA和PSO这两种启发式算法的局限性,阐述了CRO和ACROA算法的原理及其化学反应过程,并提出了CRO和ACROA算法的参数。将CRO和ACROA和上述启发式算法进行了对比分析,总结了CRO和ACROA算法的优缺点。 3.提出了基于CRO, ACROA的SVM参数优化方法。在支持向量机中,泛化能力以及最小训练误差和最小模型复杂性之间的权衡是由内核参数和正则常数C决定,核函数参数定义从输入空间到输出空间之间的非线性映射.如果这些参数没有正确选择,SVM的性能就会减弱。本文采用CRO, ACROA对核参数和正则常数C进行优化,结果表明,CRO, ACROA相较于GA和PSO来说在训练速度和分类率方面有更好的性能,能获得最佳优化效果。 4.将CRO, ACROA-SVM和LMD、LCD等方法相结合对滚动轴承进行故障诊断。 (1)提出了一种基于LCD能量熵,ACROA算法设计的支持向量机简称LCD-ACROA-SVM)的滚动轴承故障诊断。首先,将振动加速度信号分解成若干个内禀尺度分量,然后,提出LCD能量熵的概念。其次,从包含主要故障信息的内禀尺度分量中提取能量特征作为支持向量机分类器的输入向量。最后,提出ACROA-SVM分类器用于识别滚动轴承故障模式。对内圈故障和外圈故障的滚动轴承进行分析,结果表明：基于ACROA-SVM的诊断方法和采用LCD方法提取不同频带能量水平能够准确有效地识别滚动轴承故障模式,提出的方法要明显优于经验模态分解方法,而且更加节省时间。 (2)提出了一种基于LMD和ACROA-SVM的滚动轴承故障诊断(简称LMD-ACROA-SVM)首先,采用局部均值分解方法将从滚动轴承中提取的原调制振动信号分解成若干个PF分量,其次,在包含主要故障信息的PF分量的包络谱中,不同故障特征频率处振幅的比值被定义为特征振幅比。最后,将特征振幅比作为ACROA-SVM分类器的输入并对滚动轴承的故障模式进行识别。结果表明：与LMD方法相结合的ACROA-SVM分类器可以有效地提高故障诊断的准确率,并且耗时少。 (3)提出了一种新的基于LCD和CRO-SVM的滚动轴承故障诊断,简称LCD-CRO-SVM。首先,采用LCD方法将滚动轴承原始振动信号分解成若干个内禀尺度分量之和。其次,在一系列内禀尺度分量的包络谱中,计算不同故障特征频率处的振幅比。最后,将这些振幅比作为CRO-SVM分类器的输入。实验结果表明：相比于其他方法,与LCD方法相结合的CRO-SVM分类器能够获得更高的分类精度和需要更少的时间。
[Abstract]:Rolling bearings are widely used in the machinery industry, so it is of great significance to the research of fault diagnosis technology of rolling bearings. The fault diagnosis of rolling bearings is actually a process of pattern recognition, and the key is the feature extraction and state recognition of the fault.
Time frequency analysis is a feature extraction method commonly used for vibration signals. It is divided into two types: linear analysis and nonlinear analysis. Linear time-frequency analysis methods include short time Fu Liye transform (STFT), Gabor transform and wavelet transform. The nonlinear time-frequency analysis method includes discrete Fourier transform (DFT), Wigner-Ville distribution, and Choi-Wiliam points. .STFT establishes the relationship between the time domain and the frequency domain by the localization time frequency analysis of the signal, but the high frequency and low frequency components of the signal are also determined by the Heisenberg uncertainty principle that the.Wigner-Ville distribution is a more intuitive and reasonable nonlinear time-frequency distribution, and it can also describe the instantaneous power spectrum of the signal, but it is in the case of the instantaneous power spectrum of the signal. The false component produced by the convolution process causes the chaotic time frequency characteristics of the signal. The Fu Liye transform decomposes the signal into a series of sinusoidal waves of different frequency segments, and the wavelet transform decomposes the signal into its small wave basis. The vibration signal of the rolling bearing is nonlinear and non-stationary, and the above method is used to analyze the non-stationary and nonlinear signals. Hilbert Huang Bianhuan (HHT) is an adaptive time-frequency analysis method, including two parts of empirical mode decomposition (EMD) and Hilbert spectrum analysis (HAS)..EMD is decomposed into a finite single component signal based on the time scale characteristics of the signal itself, which is called the intrinsic mode function (IMF). This method is not the same as the intrinsic mode function (IMF). The analysis of linear nonstationary data has great potential, especially in the analysis of typical time-frequency energy negotiation signals. However, the endpoint effect of EMD, modal confusion, over envelope and under envelope, negative frequency, instantaneous frequency, and lack of a certain theoretical basis, still need further study.
Smith recently proposed a new adaptive time-frequency analysis method, Local Mean Decomposition. The principle of local mean mean decomposition (LMD).LMD is based on sliding average processing to decompose a complex multicomponent signal into a number of product function components (PF). The sliding step length (MA) used is important for the decomposition results. In addition, pure frequency modulation (FM). The standard of the signal will also affect the performance of the method. Compared with the EMD method, LMD has more advantages, such as fewer iterations, more stable endpoint effects and higher instantaneous frequency resolution. The local mean of the LMD decomposition is not enveloped and under enveloping, and the local amplitude of LMD is more than that of the EMD and the lower envelope. The local mean and local amplitude are all based on the local characteristic time scale of the signal.
Another adaptive time-frequency analysis method, Local Characteristic-Scale Decomposition, local feature scale decomposition (LCD), is studied in this paper. The signal can be decomposed into a number of intrinsic scale components (ISC) by LCD, and each ISC component contains local characteristics of the signal, so it can be extracted more accurately and effectively by this method. The performance of the original signal feature information.LCD is superior to EMD. in reducing the endpoint effect, the iteration time and the accuracy of the instantaneous characteristics. In addition, LCD can improve the marginal effect of the EMD decomposition signal. Based on the above advantages, this paper presents a fault diagnosis method for rolling bearings based on LMD and LCD.
The fault diagnosis of rolling bearing is actually a state recognition process. State recognition uses statistical learning theory, supervised and unsupervised classification. State recognition methods commonly used for state recognition are Artificial neural network, ANN^ and support vector machine (Support Vector Machine, SVM). The main disadvantage of ANN is that it is difficult to determine the network The structure and parameters of the collaterals need a large number of samples, and it is difficult to obtain a large number of samples in practice. In addition, the slow convergence speed also greatly increases the calculation time.SVM is an effective pattern recognition method based on the statistical learning theory and the structural risk optimal principle. It can not only solve the overfitting, local optimal, and slow convergence speed in the ANN. .SVM is widely used in pattern recognition and other fields for small samples. However, the parameter selection of SVM has a great influence on the classification effect, and the selection of parameters is actually an optimization process. Therefore, the optimization algorithm should be used for the parameter selection of SVM. Genetic algorithm (Genetic algorit) HM, GA) and particle swarm optimization (Particle swarm optimization, PSO) are applied to the optimization of SVM parameters. The.GA algorithm has the problems of slow convergence speed and easy to lose local optimal solution. In addition, GA can not solve specific optimization problems and the.PSO algorithm is easy to describe, easy to realize, fast convergence speed and so on. Potential, but there is a defect that can not effectively avoid premature convergence.
Recently, a new optimization algorithm based on the principle of chemical reaction has been proposed and proved to be superior to other optimization algorithms in many aspects. The idea of chemical reaction optimization algorithm comes from the occurrence of chemical reactions, imitating the microscopic movement of molecules in chemical reactions, and achieving the advantage of using the phenomenon of the lowest energy of the chemical reaction generation. There are two algorithms based on the chemical reaction principle: one is Chemical Reaction Optimization (CRO), and the CRO principle is based on the potential energy of the system. When the potential energy is reduced to a minimum, the reaction system will gradually reach the equilibrium state. Therefore, the potential energy is feasible as the objective function of the minimization problem; The other is the artificial reaction optimization algorithm (Artificial Chemical Reaction Optimization Algorithm, ACROA). Enthalpy and entropy can be used as the objective function (state function) for minimizing and maximizing the problem. The enthalpy depends on the chemical properties of the matter, the state of temperature and pressure, and the entropy is used to measure the randomness of the chemical system components or the disease.ACROA There is a parameter, initial reactivity, so this method is easy to use. CRO and ACROA are applied to SVM parameters optimization. The results show that the support vector machine based on chemical reaction optimization (Chemical Reaction Optimization-Support Vector Machine, CRO-SVM) and support vector machine based on artificial reaction optimization algorithm (Art) Ificial Chemical Reaction Optimization Algorithm-Support Vector Machine, ACROA-SVM) is superior to genetic algorithm based support vector machines (Genetic algorithm-Support vectormachine, GA-SVM) and support vector machines based on Particle Swarm Optimization in solving classification problems. SVM). On this basis, the paper applies local mean decomposition (Local mean decomposition, LMD) and local feature scale decomposition (Local characteristic-scale decomposition, LCD) to ACROA-SVM and CRO-SVM, in the fault diagnosis of rolling bearings.
The main work and innovation of this paper are as follows:
1. the two time frequency analysis methods of LMD and LCD are studied. The comparison of LMD, LCD and empirical mode decomposition (Empirical mode decomposition, EMD) method is carried out respectively. The analysis results of the simulation signal and the rolling bearing fault experimental signal show that the LMD and LCD are superior to the EMD method in calculating the efficiency of the end effect. The more sex.
2. the theoretical study of chemical reaction algorithm is carried out. The limitations of the two heuristic algorithms of GA and PSO are analyzed. The principle of CRO and ACROA and the chemical reaction process are expounded. The parameters of the CRO and ACROA algorithms are proposed. The comparison analysis of CRO and ACROA and the above heuristic algorithms is made, and the advantages and disadvantages of the CRO and ACROA algorithms are summarized.
3. a SVM parameter optimization method based on CRO and ACROA is proposed. In support vector machines, the tradeoff between generalization ability and minimum training error and minimum model complexity is determined by kernel parameters and regular constant C. The kernel function parameters define the nonlinear mapping between the input space and the output space. If these parameters are not selected correctly, the parameters are not selected correctly. The performance of SVM will be weakened. In this paper, CRO, ACROA is used to optimize the kernel parameters and regular constant C. The results show that CRO and ACROA have better performance in training speed and classification rate than GA and PSO, and can obtain optimal optimization results.
4. fault diagnosis of rolling bearing is carried out by combining CRO, ACROA-SVM and LMD, LCD and other methods.
(1) a fault diagnosis of rolling bearings is proposed based on the LCD energy entropy and the support vector machine (LCD-ACROA-SVM) designed by ACROA algorithm. Firstly, the vibration acceleration signal is decomposed into several intrinsic scale components, and then the concept of LCD energy entropy is proposed. Secondly, the energy characteristics are extracted from the intrinsic scale components including the main fault information. As the input vector of the support vector machine classifier, the ACROA-SVM classifier is proposed to identify the rolling bearing fault mode. The rolling bearing of the inner ring fault and the outer ring fault is analyzed. The results show that the ACROA-SVM based diagnosis method and the LCD method can be used to extract the different frequency band energy levels accurately and effectively. Bearing failure mode, the proposed method is much better than the empirical mode decomposition method, and saves time more.
(2) a rolling bearing fault diagnosis based on LMD and ACROA-SVM is proposed. First, the local mean decomposition method is used to decompose the original modulation vibration signals extracted from the rolling bearings into several PF components. Secondly, in the envelope spectrum of the PF components containing the main fault information, the amplitude of the amplitude of the different fault characteristics is at the frequency. The ratio is defined as the characteristic amplitude ratio. Finally, the characteristic amplitude ratio is used as the input of the ACROA-SVM classifier and the fault mode of the rolling bearing is identified. The result shows that the ACROA-SVM classifier combined with the LMD method can effectively improve the accuracy of fault diagnosis and consume less time.
(3) a new fault diagnosis of rolling bearing based on LCD and CRO-SVM is proposed, for short, LCD-CRO-SVM. first, using LCD method to decompose the original vibration signal of rolling bearing into the sum of several intrinsic scale components. Secondly, the amplitude ratio of different fault characteristic frequencies is calculated in the envelope spectrum of a series of intrinsic scale components. Some amplitude ratios are used as the input of the CRO-SVM classifier. The experimental results show that the CRO-SVM classifier combined with the LCD method can obtain higher classification accuracy and less time than the other methods.

【学位授予单位】：湖南大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TH165.3

【参考文献】