当前位置:首页>> >>


基于SVM的刀具状态识别系统设计(无源码).rar

收藏

资源目录
    文档预览:
    编号:20180914220233526    类型:共享资源    大小:5.02MB    格式:RAR    上传时间:2018-09-15
    尺寸:148x200像素    分辨率:72dpi   颜色:RGB    工具:   
    45
    金币
    关 键 词:
    基于 SVM 刀具 状态 识别 系统 设计 源码
    资源描述:
    NANCHANG UNIVERSITY外文资料原文及译文(2010—2014 年)题 目 基于 EMD 和 SVM 与 AE 传感器的刀具磨损识别学 院: 信息工程学院 系 自动化 专 业: 测控技术与仪器 班 级: 测控技术与仪器 131 学 号: 5801313044 学生姓名: 韩体壮 指导教师: 李鸣 起讫日期: 2017.3.10~2017.5.31 英文原文英文原文:Tool Wear Identifying based on EMD and SVM with AE SensorXu Tao Feng ZhigangDept. of Automation, Shenyang Institute of Aeronautical EngineeringShenyang, 110136, ChinaEmail: wyhxt2000@163.comAbstract – By the contrast with conventional methods, Acoustic Emission (AE) sensor possesses better performance for tool wear identifying. So, AE sensor is employed into identification of tool wear in this paper. Because of the diversity and time varying of AE, Empirical Mode Decomposition (EMD) and Support Vector Machine (SVM) are employed to analyze AE signal. EMD is suitable for analyzing non-stationary signal, and SVM possesses excellent classification capacity for small samples. According to these features, a method of identifying fault of tool wear based on EMD and SVM was presented. The characteristics of the tool under different conditions were extracted by EMD, and the tool wear was identified by SVM classifier. Experiment results show that the method based on EMD and SVM is suitable for identifying tool wear, and therate of successfully identifying is 95%.Keywords – Tool wear, AE sensor, empirical mode decomposition, support vector machine.英文原文I. INTRODUCTIONWith the increasing competition of market, modern manufacture should maximize quantity of productions and minimize cost meanwhile. The direction of modern manufacture is the implementation of automation and unmanned operation. But, tool wear is identified by the operator in conventional method. It is subjective and becomes bottleneck of the development of modern manufacture. So, on-line detection method for tool wear has been studied extensively. Conventional methods include current, optical fiber and image, which could detect tool wear directly. But, these methods did not achieve good results because of existence of cutting chip and thermal stress [1,2]. Cutting force and power, which could detect tool wear indirectly, did not achieve good results because of the faintness of feed rate change and cutting depth change [3-6].Acoustic emission reflects the change of internal lattice in mental material. So it includes much information about tool wear, and widely used in fault detection fields [7,8]. But, it is difficult to extract feature of tool wear because of the diversity and irreversibility of AE signal. Conventional method for feature extraction of AE signal was wavelet package decomposition, which extracted feature from the power of different frequencyband [7-10].EMD could decompose signal adaptively based local characteristic. In the process of decomposition, base function could be created adaptively. Then the Intrinsic Mode Functions (IMF), which are achieved after decomposition, include local feature of original signal to implement multi-resolution analysis adaptively [11]. Therefore, EMD is used to extract feature of tool wear, and SVM is used to identify tool wear mode in this paper.This paper will be structured as follows; In Section II we will introduce the theorem of EMD approach; In Section III we will introduce the theorem of SVM approach; In Section 4 we will describe our approach, explaining each step; In section 5 we will show how applying this method for tool wear identifying. Finally, we will present meaning conclusion.英文原文II. EMPIRICAL MODE DECOMPOSITIONIn 1998, Norden E. Huang, who worked in NASA Goddard space flight center, proposed Hilbert-Huang transform to process those signals with nonlinear and non-stationary characteristics [12]. The main feature of the transform is to employ Empirical Mode Decomposition technology, which decomposes data into a group of orthogonal IMFs with mode functions. Here, IMFs must satisfy next two conditions. The first is that the number of extremum equals the number of zero-crossing point. The second is the local symmetry of upper envelope line and lower envelope line [13]. In fact, AE signal maybe not satisfy these conditions because of its complexity in the process of tool wear. Yet, Norden E. Huang supposed that any signal was composed by some different intrinsic modes, which might be linear or nonlinear [14,15]. The number of extremum equaled the number of zero points in local group with the symmetry characteristics of the upper envelope line and the lower envelope line. At any time, a signal might include many intrinsic modes. If these modes overlapped with each other, they might format compound signal. Hence, any signal may be decomposed into infinite IMFs, which could be achieved by the method below.(1) Determine all extremum points of in local)(txgroup, then connect these maximal points to format upper envelope line by cubic spline lines. Meanwhile, connect those minimal points to format lower envelope line by cubic spline lines. These two line envelope all data of the signal.(2)The mean value of the two envelope lines is denoted by 1 , then calculate(1) 11)(uxtyt(3)Judge whether y1(t) is IMF. If y1(t) doesn’t satisfy the IMF’s condition, repeat step (1) an (2) while y1(t) as the original data until y1(t) satisfies the IMF’s condition. Now, let y1(t)=c1(t) . So, the first IMF of x1(t) is c1(t) , which denotes high-frequency component of x1(t)(4) Extract c1(t) from x1(t) , then achieve signal r1(t) without high-frequency component(2))()(1tctrAs the original data, r1(t) is processed according to the steps (1), (2) and (3) to achieve the second IMF, denoted by c2 (t) . Repetition of procession, n IMFs are achieved英文原文(3)nnrcr12When cn (t) or rn (t) satisfies termination condition that rn (t) becomes a monotonic function, repetition is terminated. From equations (2) and (3), we may have(4))()(1trctxninHere, rn (t) is the residual function that denotes the trend of signal. Yet, each IMF c1(t),c2 (t), ,cn (t)denotes different frequency band component from high to low respectively.EMD could decompose the signal into waveform and trend as different scales, then a series of IMFs with different feature scale are achieved. Each IMF focuses on local characteristic of the signal. Then, features extraction could be more accurate and effective. So, EMD is employed to extract feature from AE signal when tool wear.英文原文III. SUPPORT VECTOR MACHINEIn 1995, Vapnik proposed SVM method, which is regarded as the replacement of classification methods, especially when less samples and nonlinear situation [16,17]. It possesses better performance for generalization.Suppose that two class of training sample available.(1)niyxi ,,),( .1)1,(yHere, l is the number of samples, n is the number of dimensions and y is the identifier of classes. Nonlinear problem could be converted into linear problem in higher-dimension space, in which the optimal classifier face is achieved. Base on theorem of functional, only when a sum-function satisfies Mercer condition, it corresponds to the inner product of a transformation space. So, an appropriate inner production function in the optimal classifier face could implement linear classifier after nonlinear transform. Radial Basis Function (RBF) is chosen in this paper.{ 2} (2)exp),(iKiXGeneral formation of the classifier face equation is below.(3)0bwxHere, w is weight vector and b is classifier threshold. If all samples can be classified by this equation, it must satisfy the equation below(4)iiixy1])[( li,.21xi is relaxation factor.Now, the distance between two classes equals ,The classifier face which maximizes w/the distance is called optimal classifier class. To seek the optimal classifier hyper-face, quadratic programming problem with constraints (4) should be solved(5)lic12minHere, ci is large than 0 and constant as specified, which determines punishment to the samples, which have been classified incorrectly. The problem above can be converted into dual formation as below.英文原文(6))(21max, jijiljili xKyaHere, the constraint condition is (7)01liiyaliCai.2,1Based on KKT condition, optimal paramete satisfiesi(8)ibiixwy,So, optimal parameters corresponding to most samples will become 0. Only few of them will be not 0. Those corresponding samples are support vector. After the optimal parameter have been solved as equations (10), discriminant function of SVM classifier is(9)bxKyayii)(sgnFinally, the classifier can be determined by the signal of the discriminant function.英文原文IV. FAULT IDENTIFYING ALGORITHM BASED ON EMD AND SVMFirstly, EMD was used to analyze the AE signal to extract feature vector under different conditions, including normal tool and tool wear. The process for feature extraction is below.(1) Decompose the AE signal with EMD, then choose IMFs containing main information.(2)Calculate the total energy of each IMF2 -)(tiicEd )( Ni.2,1(1)(3) Construct feature vector T based on energy above.(2)],.[21nEFor the convenient analysis and process, vector T need to be normalized. Then, the feature vector T ' is achieved after normalization. After SVM has been trained by learning samples T ' , SVM possesses the capability for identifying tool wear. Finally, SVM can be tested to identify tool wear by the testing data.英文原文V. EXPERIMENTData processed in this paper is origin from working process of a CNC lathe, where workpiece material is GH648 and cutting tool material is KC5010. Data acquisition system is PXWAE AE detection system from Beijing Pengxiang company. AE sensor is PXR30. Normal data and tool wear data are illustrated by figure 1, 2 respectively.Fig. 1 AE data for normal toolFig. 2 AE data when tool wearFirstly, AE signal is decomposed by EMD to achieve every IMF. Analysis results of Normal data and tool wear data are illustrated by figure 3, 4 respectively. The number of IMFs is not constant. After much analysis, we know that the number is between 7 and 10. Because the former 7 IMFs contains most information of AE signal, we define the length of feature vector is 7. If the number of IMFs exceeds 7, those components are overlapped onto the 7th component.Fig. 3 EMD results of AE data for normal tool Fig. 4 EMD results of AE data when tool wear英文原文Feature vector can be calculated based on IMFs. Some of those vectors are as shown in table 1.Tab. 1 Learning samples of eigenvectorNo. Feature Vector description1 0.0228 0.0726 0.2714 0.5233 0.0834 0.0057 0.0209 Normal2 0.0316 0.0804 0.3093 0.4182 0.1122 0.0431 0.0052 Normal3 0.1017 0.3642 0.3249 0.1328 0.0218 0.0240 0.0306 Normal4 0.0857 0.3142 0.1885 0.2359 0.1133 0.0186 0.0437 Normal5 0.0641 0.3478 0.3398 0.1467 0.0389 0.0061 0.0566 Normal6 0.0674 0.1795 0.1922 0.2507 0.1203 0.0152 0.1747 wear7 0.0635 0.1479 0.1413 0.3217 0.2275 0.0162 0.0819 wear8 0.0583 0.0747 0.0985 0.0471 0.6061 0.0548 0.0604 wear9 0.0565 0.1528 0.1315 0.5153 0.0023 0.0019 0.1398 wear10 0.0642 0.0751 0.1345 0.3711 0.2727 0.0185 0.0638 wear200 samples are chosen as the training samples of SVM from normal data and tool wear data. Here, there are 100 samples from normal and tool wear data respectively. Target samples of normal data and tool wear data are +1, -1 respectively. After SVM has been trained by the 200 samples, it possesses the capability for identifying tool wear. Another 120 testing samples are chosen to test SVM. Some testing results are illustrated in table 2.Tab. 2 Testing samples of eigenvectorNo. Actual State Feature Vector SVM Output Identifying Result1 Normal 0.0674 0.1795 0.1922 0.2507 0.1203 0.0152 0.1747 8.0002 Normal2 Normal 0.0635 0.1479 0.1413 0.3217 0.2275 0.0162 0.0819 2.1540 Normal3 Normal 0.0583 0.0747 0.0985 0.0471 0.6061 0.0548 0.0604 1.7056 Normal4 Normal 0.0565 0.1528 0.1315 0.5153 0.0023 0.0019 0.1398 2.9264 Normal5 Normal 0.0642 0.0751 0.1345 0.3711 0.2727 0.0185 0.0638 3.3842 Normal6 Wear 0.1072 0.1740 0.1853 0.3892 0.0809 0.0059 0.0574 -1.5044 Wear7 l Wear 0.0706 0.2255 0.1824 0.1825 0.1158 0.0040 0.2193 -3.8831 Wear8 Wear 0.0711 0.1234 0.2654 0.1749 0.1518 0.0159 0.1975 -2.4048 Wear9 Wear 0.0408 0.0913 0.2163 0.1185 0.3594 0.0127 0.1611 -1.3708 Wear10* Wear 0.0896 0.2230 0.1888 0.3363 0.0845 0.0418 0.0360 0.5745* Normal*Those large than 0 is normal data and less than 0 is tool wear data by the sign of output. Testing results show that 3 normal samples are identified as faulty state and 3 tool wear samples are identified as normal state. As a result, the rate of successfully identifying is 95%.
    展开阅读全文
    1
      金牌文库所有资源均是用户自行上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作他用。
    0条评论

    还可以输入200字符

    暂无评论,赶快抢占沙发吧。

    关于本文
    本文标题:基于SVM的刀具状态识别系统设计(无源码).rar
    链接地址:http://www.gold-doc.com/p-158405.html

    当前资源信息

    4.0
     
    (2人评价)
    浏览:45次
    bysj上传于2018-09-15
    1
    关于我们 - 网站声明 - 网站地图 - 资源地图 - 友情链接 - 网站客服客服 - 联系我们
    copyright@ 2014-2018 金牌文库网站版权所有
    经营许可证编号:浙ICP备15046084号-3
    收起
    展开