【作者】 关山；

【导师】 王龙山；

【作者基本信息】 吉林大学 ， 机械制造及其自动化， 2011， 博士

【摘要】 随着制造装备自动化与制造模式集成化的发展,刀具切削状态的实时监测成为实现加工过程自动化的一项关键技术,也是目前尚没解决的重要难题。因此,研究加工过程中的刀具状态监控技术对实现加工过程自动化、提升产品加工质量、提高生产率是至关重要的也是非常迫切的。本文针对变切削条件下刀具磨损的分类及磨损量预测这一课题,通过大量的切削实验,在实验数据基础上,将现代信号处理方法,如:经验模态分解,高阶谱分析和小波变换等,引入刀具磨损特征的提取中,提出了基于最小二乘支持向量机的声发射信号多特征分析与融合的刀具磨损分类与磨损量预测方法。主要研究内容包括以下几个方面:1、以往的刀具状态监测中,多数都采用传统的信号处理方法,如:时域统计分析、傅里叶变换、功率谱估计等,提取对应刀具不同磨损状态的信号特征。由于刀具在磨损过程中发出的各种信号的非平稳、非线性及非高斯特性,传统的信号处理方法,已经远远的不能满足特征提取的要求。本文将更适于非平稳、非线性、非高斯信号处理的现代信号处理方法,如:经验模态分解、高阶谱分析和小波包分析,引入刀具磨损信号的特征提取。首先,提出了基于EMD与AR模型相结合的刀具磨损特征提取方法。通过EMD分解将非平稳的声发射信号分解为有限个平稳的IMF分量之和,针对不同磨损阶段信号EMD分解后得到IMF分量个数不同的问题,提出了基于相关系数法的IMF分量选择方法。建立各IMF分量的AR模型,提取模型系数构造特征向量。结合声发射信号实例,讨论了AR模型的定阶问题,利用所建模型对各IMF分量进行预测,预测误差分析结果证明了建立4阶AR模型的准确性,从而间接的证明了AR模型系数作为表征刀具磨损特征向量的有效性。其次,将双谱分析理论引入刀具磨损状态监测信号的特征提取中,提出了基于双谱奇异值分解的刀具磨损特征提取方法。在对不同切削条件、不同磨损阶段的实验数据去均值及归一化处理的基础上,进行双谱分析,构造基于双谱的特征向量矩阵,然后对特征向量矩阵进行奇异值分解,选取奇异谱构造特征向量。实验结果表明:该方法可以有效的减少切削条件的影响,更适于变切削条件下刀具磨损特征的提取。第三,提出了建立在小波包最优基基础之上的小波包能量及时域统计特征的信号特征提取方法。结合声发射信号实例,探讨了小波包分解层次的确定方法和小波包最优基的选择方法。在此基础上,提取最优小波包能量和时域统计特征构造特征向量,通过引入基于类内、类间距离评估因子的特征评估方法对所得的高维时域统计特征进行评估,将时域统计特征从72维降至13维,有效的减少了特征的维数,提高了特征分类的性能,避免了由于高维输入特征而产生的“维数灾难”。2、提出了信号多特征分析与融合的刀具磨损特征提取方法。在传统的刀具磨损监测系统中,为了提高监测的准确率,许多学者提出了基于传感器融合的监测方法,但是这种方法用于实际的监测,却存在一定的弊端:首先,多传感器融合大大增加了监测系统的成本,其次,基于多传感器融合的监测系统在安装上可能会干涉加工人员的操作,甚至影响到加工机床的性能。鉴于此,本文采用上述提及的现代信号处理技术,提取同一信号在不同域内的特征,从不同的角度反映刀具的磨损状态,构造联合多特征向量。然后利用核主元分析法对联合多特征向量进行融合降维处理,通过提取累积贡献率大于85%的主元,生成对应刀具磨损的融合特征,有效剔除了联合多特征中与刀磨损相关性较小或冗余的特征。降维后融合特征的散度图表明:所保留的融合特征具有更好的聚类性。3、将最小二乘支持向量机引入刀具磨损的分类与磨损量预测。针对人工神经网络训练需要大量样本,学习算法收敛速度慢,且训练过程中易陷入局部极小值等缺点,将最小二乘支持向量机引入刀具磨损的分类,实例分析的结果证明,在正确选定最小二乘支持向量机核函数参数的前提下,融合特征对刀具磨损的识别率要高于联合特征及单一特征的识别率,基于最小二乘支持向量机的分类模型要优于基于神经网络的分类模型。采用最小二乘支持向量机回归算法,通过构造并联的双回归支持向量机,有效的实现了刀具磨损量提前10s预测。4、开发了基于数字信号处理器的刀具磨损监测系统。该系统运算速度快,完全满足监测系统实时性的要求;系统柔性强,所有分类与预测算法均通过软件实现;提供通用的传感器信号接口,传感器可根据实际监测信号的变化方便更换;对于不同的加工方法,在未知信号特征及监测算法的情况下,本系统可以做为一个数据采集器,通过系统提供的高速USB接口,可将现场数据及时上传计算机,经技术人员分析与处理,编写新的特征提取、分类和预测算法程序,再回传给系统,实现不同加工方法下刀具的状态监测。因此系统通用性强,在理论上适用于任何加工方法刀具状态监测。更多还原

【Abstract】 With the development of automation of manufacturing equipment and integration of manufacturing mode, real-time tool condition monitoring has become key technology in achieving automatic machining process and is an important problem not solved currently. The research on tools condition monitoring in machining process is of great importance and urgent to the realization of automatic machining process, improvement of product quality, and high productivity.To classify tools wearing and forecast wearing capacity under alternating cutting conditions, a multi-feature fusion method based on least squares support vector machines (LS-SVM) was established through the analysis of large quantity of experimental data. Modern digital signal processing (DSP) methods such as empirical mode decomposition, high-order spectral analysis and wavelet transform were used in the characteristics extraction of tools wearing. The main accomplishments are summarized as follows:1、In the past, traditional signal processing methods are mostly used in characteristics extraction in monitoring tools wearing. Methods such as statistical analysis in time domain, Fourier transform and PSD estimation cannot meet the signal processing requirements because signals from tools wearing are non-stationary, non-linear and non-Gaussian. In this paper, Modern signal processing methods such as empirical mode decomposition, high-order spectral analysis and wavelet transform are introduced tool wear signal extraction.First, the method of characteristics extraction of tools wearing based on the combination of Empirical mode decomposition (EMD) and autoregressive (AR) model is proposed. EMD can decompose non-stationary emission signals into finite intrinsic mode function (IMF) components. The method of IMF selection based on correlation coefficient method was proposed with the aim of addressing the issue that EMD of acoustic emission signals at different wearing stages produce different numbers of IMF. AR model coefficient contains important information of system and is most sensitive to status changes, so AR model was established for each IMF component after constructing feature vectors via extracting model coefficient. The order of AR model directly influences the accuracy of model and system identification. The order determination of AR model was addressed using sample acoustic emission signals of tools wearing. Forth-order AR model proved to be accurate by prediction error analysis. It is indirectly proved that the AR model coefficients as feature vectors characterize the effectiveness of the tool wear.Second, the theory of bispectrum analysis was successfully applied to the characteristics extraction of tools wearing. As a result, a method of feature extraction based on bispectrum singular value decomposition was proposed. After the removal of mean values and the normalization process of experimental data from different cutting conditions and different wearing stages, an eigenvector matrix was constructed on the basis of the bispectrum analysis of experimental data. The eigenvector was subsequently constructed by extracting singular spectrum through singular value decomposition of eigenvector matrix. By analyzing sample acoustic emission signals of tools wearing, this method proved to be effective in reducing the influence of cutting conditions and suitable for feature extraction of tools wearing under alternating cutting conditions.Next, The energy characteristics extraction of wavelet package and statistical feature of time domain, which is established on the basis of the best bases of wavelet package, was proposed. Determination of decomposing levels and selection of best bases of wavelet package were investigated using sample acoustic emission signals. High dimensional time domain statistical features was evaluated on the distance ratio of within/between class. As a result, the dimension of time domain was reduced from 72 to 13. This improved the performance of feature classification and avoids the Dimensional disaster.2、A characteristics extraction method of tools wearing based on multi-feature fusion was proposed. In the traditional monitoring system of tools wearing, the sensor fusion based monitoring method is used by many researchers to improve the accuracy of monitoring. But this traditional method has some limitations in its practical applications. First, it largely increases the cost of the monitoring system. Second, installing of monitoring system may interfere the operator, even impede the performance of machine tools. In view of this, multiple feature joint base on modern signal processing methods was proposed. The characteristics of the same signal in different domains was extracted by modern signal processing technology, then the joint multi-feature vector was contructed. By using the Kernel principal component analysis method, the joint multi-feature vector was fused. Select the principal components whose Cumulative contribution rate is greater than 85% and generate fusion feature vector corresponding to the tool wear, which can effectively exclude features that is redundant or has little relevance to tool wear. Divergence diagram of the fusion feature after descending dimension shows that: the fusion features retained have better character of clustering.3、Introduced the least squares support vector machines to tool wear classification and wear prediction. To overcome shortcomings such as large quantities of samples required in artificial neural network training, low convergence rate of learning algorithm and trapped in local minimum in training, least squares support vector machines was applied to classifying and forecasting tools wearing. The analysis of sample data proved that the identification rate of fusion characteristic is higher than that of single one, and the model of classification and forecasting based on least squares support vector machines is superior to the one based on artificial neural network. Using least squares support vector machine regression algorithm, by constructing a parallel dual regression support vector machine, tool wear prediction 10s ahead is effectively realized.4、A monitoring system of tools wearing based on digital signal processor was developed. The features of this system are: fast computing speed to meet the requirement of real-time monitoring; high flexibility because all algorithms for classification and forecasting are realized using software only; and universal interface for sensor signals so that sensors can be replaced at any time as required by real-time signals. This system can be used as a digital collector even the signal characteristics and algorithms are unknown because the high speed USB interface can transfer data to computer for prompt processing. After being analyzed and reprogrammed for classification and forecasting by the users, the data are fed back to the system. As a result, monitoring of cutting conditions by different processing methods can be realized. Theoretically, this monitoring system can be applied to any processing method.更多还原