【作者】 李威霖；

【导师】 傅攀；

【作者基本信息】 西南交通大学 ， 机械制造及其自动化， 2013， 博士

【摘要】 刀具状态监测是先进制造技术的关键技术之一,是保证不间断生产、实现加工自动化的关键,对提高产品加工质量与效率、保护加工设备、提高制造业水平具有重要意义。论文在深入探讨国内外刀具状态监测研究现状的基础上,以车削和铣削加工刀具为监测对象,切削力和切削振动为刀具状态监测信号,对目前该研究领域存在的主要技术难题进行创新性研究。具体研究内容如下：(1)讨论了刀具磨损的特点、形式及影响因素,针对切削过程的复杂性,确定了以切削力和切削振动为监测信号的间接在线监测方式,采用均匀试验设计方法进行了试验设计,获得了刀具不同磨损状态下的实验数据,为后续研究提供数据支持。(2)研究刀具磨损状态监测特征提取技术。首先采用小波阈值降噪技术消除监测信号中的高频干扰噪声,然后再对监测信号进行时域和频域统计特征分析的基础上,引入小波包分析技术,提取信号的小波包频带能量及小波熵特征,得到从不同角度反映刀具状态的特征集。通过分析这些特征参数与刀具磨损的相关性可知：特征参数与磨损量呈非线性关系,不能通过单一特征判断刀具磨损状态,存在诸多不相关特征。(3)研究刀具磨损状态监测特征选择技术。原始特征集中的不相关和冗余特征将使得识别模型的学习样本数及计算量都成倍增加,从而降低整个系统的运行效率和精度,因此对原始特征进行合理的选择尤为重要。针对刀具状态监测特征的维数高、样本少及类别多的特点,提出基于“一对一”多分类支持向量机递归特征消去的刀具状态监测特征选择算法。在递归特征消去算法基础上,将多分类支持向量机作为特征评价分类器,逐个消去不相关和冗余特征。实验表明该算法能有效剔除与刀磨损相关性小或冗余的特征,提高刀具磨损状态监测系统的学习效率和识别精度。(4)不完备先验知识下的刀具磨损状态评估技术研究。针对实际加工生产中,各工况下刀具全寿命先验样本获取困难,导致传统监测方法适用性差的问题,在隐马尔科夫模型的基础上提出基于因子隐马尔科夫模型(FHMM)的刀具状态评估技术。利用新刀和钝刀状态下的先验观测序列建立FHMM,根据刀具磨损过程中观察序列与模型的对数相似度获得其性能指标来评估刀具磨损状态,通过设置适当的阈值对磨损状态进行报警。同时,为监测系统引入学习能力,使其可在使用过程中不断完善自身知识库,提高系统可靠性。实验结果表明：该策略能在只具有新刀和钝刀先验知识的情况下,实现刀具磨损状态的初步估计,避免刀具的过度磨损；并且系统具备学习与完善能力,扩展性好。(5)研究刀具磨损量识别技术。在获取一定量的较为完备的先验数据后,可建立更加精确的磨损量识别模型,而传统的人工神经网络识别模型往往需要大量训练样本,且存在收敛速度慢、易陷入局部极小值、识别精度差等问题,本文提出基于最小二乘支持向量机(LS-SVM)的刀具磨损量识别方法。针对LS-SVM的正则化参数和核函数宽度对识别精度的影响大的问题,在标准粒子群优化算法基础上,提出自适应惯性权重粒子群优化算法,用于其参数的优化选择,并引入留一交叉验证法对寻优过程中的参数进行评价。实验结果表明：本文的自适应粒子群优化算法参数优化能力比标准粒子群优化算法强；基于粒子群优化LS-SVM模型的刀具磨损量识别精度高于传统神经网络,且所需先验样本更少。(6)研究刀具时序状态监测结果的优化与预测技术。针对刀具状态监测结果往往存在围绕真实状态值上下波动的系统误差,从而影响监测系统精度的问题,提出了基于卡尔曼滤波的时序监测结果优化技术,采用信号与噪声的状态空间模型,利用前一时刻的估计值和当前时刻的监测结果来估计当前时刻的优化值。经卡尔曼滤波优化后的刀具状态监测结果变化具有一定规律性,基于此提出基于自回归移动平均模型的监测结果预测技术,根据刀具历史监测数据预测刀具未来时刻的状态。实验证明：卡尔曼滤波算法能有效减小状态监测结果中的系统误差,优化的性能指标与刀具磨损量相关性更强,优化后的磨损量识别结果精度更高；自回归移动平均模型能较为精确地预测刀具未来时刻的状态。更多还原

【Abstract】 Tool condition monitoring is one of the key technologies of the advanced manufacturing technology, key to enable uninterrupted production, realize automatic and unmanned machining process. It has great significance for ensuring machining efficiency and quality, protecting equipment, improving manufacturing standards and so on. Taking turning and milling tools as monitoring objects, cutting force and vibration signal as monitoring information, this paper takes some innovative studies on the key problems of tool condition monitoring field on the basis of discussing the review of developments in tool condition monitoring. The main research contents are as follows:(1) The wear feature, mechanism and influencing factor of cutting tool are briefly discussed. An indirect online monitor system based on cutting force and vibration signals was build, and a uniform experimental design method is used to arrange the experiments to collect tool condition monitoring data.(2) Research on feature extraction technology. First, using wavelet threshold denoising technology to eliminate high-frequency interference noise in monitoring signals, and then based on the statistical characteristics in time domain and frequency domain, wavelet packet analysis technology was studied, the wavelet packet energy and wavelet entropy features of cutting force and vibration signals were extracted. After getting the original feature set which reflect tool wear states from different perspectives, the correlation between tool wear and these eigenvalues was analyzed. Experimental results have shown:there is strong nonlinear relationship between tool wear and eigenvalues, tool wear condition can’t be got through a single feature.(3) Research on feature selection technology. There are many irrelevance and redundancy eigenvalues in original features set. It will be lead to sharp increase in learning samples and computational complexity of identification model if all the original features are used to monitor tool condition, which will reduce the speed and efficiency of monitoring system. Traditional feature selection algorithm of tool condition monitoring ignores the relationship of classifier, and need a large number of learning samples, is not suitable for feature selection under small samples. This paper proposed a tool condition monitoring feature selection technology based on one-versus-one support vector machine recursive feature elimination (SVM-RFE). The contribution of eigenvalues to classifier is used as the selection criteria to eliminate the irrelevance and redundancy features. Experiments show that the algorithm can effectively remove the irrelevant or redundant features, improve learning efficiency and identification accuracy of tool condition monitoring system.(4) Research on tool wear status assessment technology under incomplete prior knowledge. Aiming at lacking priori samples and difficult knowledge acquisition in the actual production process, which will make the applicability of monitoring system poor, a tool wear status assessment method based on factorial hidden markov model (FHMM) was proposed. First establish FHMM monitoring model using the priori knowledge of blunt and new tool. The performance indicators that can reflect tool wear state can be got from the similarity between monitoring sequence and trained FHMM. Meanwhile, self-improvement ability was introduced to the monitoring system so that the knowledge base of FHMM can continually improve in monitoring processes. The experimental results showed that:this method can achieve a preliminary estimate of tool wear using the priori knowledge of blunt and new tool, and this system has the ability of self-learning and self-improvement.(5) Research on tool wears recognition technology. There is a lack of the priori samples used to train recognition model, while the tool wear identification model based on traditional artificial neural network often requires a large number of training samples, and have some disadvantages such as running into local minimum value easily, slow convergence rate and so on. Against these issues, this paper proposed a tool wear recognition method based on least squares support vector machines (LS-S VM). Meanwhile, a self-adaptive weight particle swarm optimization (PSO) algorithm was proposed to search optimum value of the kernel function parameter and error penalty factor which affect the precision of the LS-SVM significantly. Experiments show that:PSO algorithm can efficiently search the optimal parameter of LS-SVM model under small sample; tool wear identification accuracy based on PSO-LS-SVM model is better than neural networks.(6) Research on optimization and prediction technology of tool condition monitoring results. By analyzing and testing results show that the system accuracy becomes poor because the interference signals existed in the intelligent condition monitoring results. Against this issue, this paper proposed an optimization technique for time sequence monitoring results based on kalman filter. According to the state space model of signal and noise, the present value of tool conditions can be estimated using the previous estimated value and the present observed value. There is certain regularity in the kalman filter optimized tool condition monitoring results. Based on history monitoring data, the future tool status can be predicted by certain prediction model, In this paper, a monitoring results prediction model based on autoregressive moving average (ARMA) method was proposed. Experimental results show:kalman filtering optimization algorithm can effectively eliminate interference noise in monitoring results, the relationship between status indicators and tool wear is stronger, the identification accuracy is higher, the ARMA method can effectively predict tool status.更多还原