【作者】 邵强；

【导师】 邵诚；

【作者基本信息】 大连理工大学 ， 运筹学与控制论， 2010， 博士

【摘要】 切削加工是使用切削工具,把坯料或工件上多余的材料层切去,使工件获得规定的几何形状、尺寸和表面质量的加工方法。在切削加工过程中所产生的颤振是影响工件质量的主要原因之一,颤振具有非线性、时变性和不确定性等特点,难以进行精确的测量和识别,多年来吸引了国际上众多学者持续对其进行研究,取得了一些重要的研究成果,但仍然存在一些问题有待于解决。本文综合了隐马尔可夫、支持向量机及核主元分析等理论和方法,对机床启动过程、刀具磨损状态等诱发颤振的因素进行诊断分析,并且对切削加工中的颤振监测和识别方法进行研究。论文完成的主要工作如下：(1)通过对切削加工时机床启动过程的故障特征分析,基于隐马尔可夫理论,建立了混合密度连续隐马尔可夫模型,对机床启动过程的故障进行识别。采用连续混合密度隐马尔可夫模型识别车床启动过程的运行状态,解决了隐马尔可夫模型的溢出问题,根据高斯密度函数特点,提取启动过程的工件松动、不平衡、不对中和正常启动等特征信息,依据信息特征进行故障诊断识别。该模型克服了传统的诊断方法容易丢失特征信息的弊端,方法简单、识别率高,适合应用于旋转机械的启动过程故障诊断。与隐马尔可夫模型进行比较分析,实验结果表明,该模型具有较好的识别效果。(2)通过刀具磨损量对颤振影响程度的分析,基于离散隐马尔可夫理论,建立了刀具磨损诊断模型。采用对切削加工中的动态切削力信号和刀柄振动信号进行快速傅立叶变换并提取特征量,将提取的特征谱矢量作归一化处理,然后利用自组织特征映射对归一化矢量进行预分类离散编码,编码量值作为观测序列引入到离散隐马尔可夫模型中进行机器学习,识别出刀具磨损程度,识别结果作为控制切削进给量大小的依据。该模型克服了传统识别方法的计算量大、算法复杂的缺点,识别速度高,具有良好的实时性,并通过与隐马尔可夫模型和分形理论比较分析,实验结果表明,该模型具有较好的识别效果,为正确识别切削颤振奠定基础。(3)针对切削力信号和工件振动信号的非线性、不确定性和时变性的特点,提取切削过程的大样本数据,建立了基于核主元分析与支持向量机结合的故障诊断模型(KPCA-SVM)。该模型通过KPCA方法提取非线性颤振数据中的线性主元信息,根据主元信息贡献率的大小,确定能够代表颤振特性的线性主元,然后,通过SVM的分类能力,对线性主元进行一对多方式分类,分类结果作为判定是否具有颤振趋势的依据,为控制任务提供数据基础。该方法弥补了传统识别方法难于充分描述颤振发展过程的缺陷,实验结果表明：对于能够描述切削过程的大样本数据,KPCA-SVM是一种新的有效的颤振趋势识别方法。与主元分析与支持向量机模型(PCA-SVM)的识别效果比较,具有一定的优越性。(4)针对切削加工过程中的颤振发生时的小样本数据,建立了基于支持向量机与隐马尔可夫模型(SVM-HMM)结合的诊断模型,辨识颤振发生的程度。该模型首先求取小样本数据在支持向量机下的最优比率,然后把最优比率转化成Sigmoid概率,作为观测序列输入到HMM模型,通过隐马尔可夫模型的良好的类内分类能力,对切削过程中能够表现颤振的振动信号和切削力信号做出有效训练和识别。实验结果表明,对于小样本数据,该方法对切削颤振具有较强的识别能力,识别效果优于支持向量机方法、隐马尔可夫方法,该方法克服了非颤振信息颤振化错误判断的弊端,是一种颤振诊断的新方法。更多还原

【Abstract】 Cutting is a machining way, by which superabundant material layer is cut by cutting tools from blank or work piece, in order to get specific figure, size and surface quality. Chatter in cutting process is one of the main causes which can affect the quality of work piece. It is difficult to accurately measure and identify chatter because of its nonlinear, time variation and uncertainty. For many years, lots of scholars from various countries always tried to study chatter and have achieved some important results. But there are still some problems that need to be solved. In this paper, the theories and methods of Markov, Support Vector Machine and Kernel Principal Component Analysis was synthesized, and was applied in diagnostic analysis of startup process and tool wear which are inducing factors for chatter, and the study about methods of monitor and identification of chatter in cutting process was performed. The main works in the research are as follows:(1) Through analysis of fault feature of startup process in cutting, based on Markov’s theory,Continue Hidden Markov Model with mixture probability Densities(CDHMM) was established in the research in order to identify fault condition during startup process. When it was used in identifying running state of startup process the overflow of Markov’s model was settled. Characteristic information including workpiece loosing, unbalance, asymmetry and normally starting was extracted based on Gaussian density function,, then the CDHMM did identify and diagnose fault according to these information.The new model was better than the traditional diagnostic method for it overcame the traditional model’s defect of losing information in feature information abstraction.The new method was simple, good in recognition rate, and suitable for diagnosis of startup fault about rotating machinery. The experimental results showed that the new model had better recognition effect than Hidden Markov model.(2) After analysis the effect of tool wear to chatter, a tool wear diagnosis model was established based on discrete hidden Markov model(DHMM). Dynamic cutting force signal and tool holder vibration signal was dealt by fast fourier transformation, and characteristic parameter was extracted and was normalized, then the parameter was presorted and coded by self-organizing feature maps. The coding value was introduced into the discrete hidden Markov models as observation sequence for machine learning, which could recognize condition of tool wear and feeding quantity was controlled according to the recognition results. The model overcame defect of the traditional identification means including large calculation and complex algorithm. The new model had higher recognition speed and better real-time character.The experimental results showed that the new method had better recognition results than the Hidden Markov Model which could establish basement for recognizing chatter correctly.(3) According to the nonlinearity, uncertainty and variability of signal of vibration and cutting force, a diagnostic model of fault was built in the research based on Kernel Principal Component Analysis and Support Vector Machine (KPCA-SVM) through extracting a large sample of cutting. This model extracted linear principal component information from nonlinear chatter value through KPCA, the linear principal component which could reflect character of chatter could be identified according to the contribution rate of principal component information. Linear principal component was classified by one-to-many mode and classification results could be used for judging chatter trend which provided data foundation for controlling duty. This method could fully describe development of chatter which is difficult for the traditional method. Experimental results showed that KPCA-SVM was a new effective method for recognizing chatter trend and better than PCA-SVM (Principal Component Analysis and Support Vector Machine) for large samples which could describe cutting process.(4) A diagnostic model was established based on Support Vector Machine and Hidden Markov Model(SVM-HMM) to recognize degree of chatter for small samples when chatter occurring in cutting process. This model got optimum rate of small sample in SVM, and the output results from SVM was transformed into Sigmoid probability, which then was transported into HMM model, where vibration signals and cutting force signals was effectively studied and recognized through HMM’s ideal class classification ability. In the research, the results showed that this method had better ability in recognition of cutting chatter and could get better recognition effect than SVM and HMM for small samples. This method overcame the disadvantage of recognizing non-chatter information as chatter and it was a new method for diagnosis of chatter.更多还原