外文文獻(xiàn)及翻譯--應(yīng)用高階神經(jīng)網(wǎng)絡(luò)對旋轉(zhuǎn)機(jī)器的多種故障進(jìn)行診斷

1、<p><b>　　畢業(yè)論文（設(shè)計(jì)）</b></p><p><b>　　英文翻譯</b></p><p>　　原文標(biāo)題 High Order Neural Networks for Simultaneous Diagnosis of Multiple Faults in Rotating Machines

2、 </p><p>　　譯文標(biāo)題 應(yīng)用高階神經(jīng)網(wǎng)絡(luò)對旋轉(zhuǎn)機(jī)器的多種故障進(jìn)行診斷 </p><p>　　學(xué)生姓名 學(xué)號 </

3、p><p>　　專業(yè)年級 </p><p>　　指導(dǎo)教師 </p><p>　　二Ｏ一二 年 五 月 二十八 日</p><p>　　應(yīng)用高階神經(jīng)網(wǎng)絡(luò)對旋轉(zhuǎn)機(jī)器的多

4、種故障進(jìn)行診斷</p><p>　　B. Zhong'，J. MacIntyre2, Y. He` and J. Tait2</p><p><b>　　摘要：</b></p><p>　　為了克服標(biāo)準(zhǔn)前向反饋神經(jīng)網(wǎng)絡(luò)的局限，以及由于高位神經(jīng)網(wǎng)絡(luò)的一定概念化和推測，所以高位神經(jīng)網(wǎng)絡(luò)對于故障診斷是非常有用的。本文描述了高位神經(jīng)網(wǎng)絡(luò)的理論和

5、結(jié)構(gòu)。它是一個(gè)初始化方法為hyperellipsoids和標(biāo)準(zhǔn)BP算法的訓(xùn)練算法，在這樣的網(wǎng)絡(luò)運(yùn)用一定概念和推測， 根據(jù)橢球狀單位網(wǎng)絡(luò)(HDANN)的等級制度診斷的人工神經(jīng)網(wǎng)絡(luò),在旋轉(zhuǎn)電機(jī)中常遇到多個(gè)缺點(diǎn)同時(shí)診斷的問題，包括幾個(gè)子網(wǎng)絡(luò)并且將一個(gè)大模式空間劃分成幾個(gè)更小的子空間時(shí)，子網(wǎng)絡(luò)可以在各自子空間中訓(xùn)練，并且整體網(wǎng)絡(luò)是有能力對多個(gè)故障上同時(shí)進(jìn)行診斷。最終，典型的故障數(shù)據(jù)從旋轉(zhuǎn)電機(jī)中被網(wǎng)絡(luò)測試出來，研究結(jié)果表示， HDANN可能得到更

6、加準(zhǔn)確和更加高效率的診斷結(jié)果，并且這種實(shí)時(shí)條件監(jiān)測和對旋轉(zhuǎn)電機(jī)的診斷是有用的。</p><p>　　關(guān)鍵詞：等級制度的診斷網(wǎng)絡(luò)；旋轉(zhuǎn)機(jī)械；故障診斷；神經(jīng)網(wǎng)絡(luò)</p><p><b>　　引言</b></p><p>　　近年來，越來越多故障診斷理論、方法和戰(zhàn)略，并且為大規(guī)模旋轉(zhuǎn)電機(jī)進(jìn)行定量方法的提議受到了關(guān)注。例如統(tǒng)計(jì)樣式分類方法，系統(tǒng)基于證明

7、的參量模型方法，等等。但這些方法總使模型和廣泛的演算復(fù)雜化。人工神經(jīng)網(wǎng)絡(luò)(ANN)技術(shù)由于它的并行處理的，聯(lián)想記憶，自已組織的，自我學(xué)習(xí)和非常強(qiáng)的非線性映射的能力，所以在故障診斷上很有潛力。特別地，神經(jīng)網(wǎng)絡(luò)的能力應(yīng)付高維度樣式分類和非線性樣式故障診斷分類方面是重要的。作者在這個(gè)主題發(fā)表了許多論文。作為樣式分類的方法，標(biāo)準(zhǔn)多層前向反饋神經(jīng)網(wǎng)絡(luò)決策空間以超平面和決定地區(qū)，形成的總是無邊際的，可能導(dǎo)致不夠精確的推測。雖然使用橢球狀單位要克服高

8、位神經(jīng)網(wǎng)絡(luò)這個(gè)局限，但是對于故障診斷應(yīng)用方面是有用。所以在本文橢球狀單位網(wǎng)絡(luò)被描述，初始化方法為hyperellipsoids，并且訓(xùn)練算法也被描述。一個(gè)等級制度的人工神經(jīng)網(wǎng)絡(luò)(HDANN)為旋轉(zhuǎn)電機(jī)的眾多個(gè)故障進(jìn)行診斷和測試結(jié)果。</p><p><b>　　1.1 網(wǎng)絡(luò)結(jié)構(gòu)</b></p><p>　　EQ是定義的橢球狀單位。一個(gè)單位可能達(dá)到令人滿意的估計(jì)，并且網(wǎng)

9、絡(luò)包括線性輸入裝置和橢球狀輸出裝置二層數(shù)。每個(gè)暗藏的單位只連接到一套輸出裝置，并且每套輸出裝置有緊密的暗藏的單位。就特點(diǎn)傳染媒介而論，是從原始數(shù)據(jù)得到的特征，而提取的數(shù)據(jù)是從旋轉(zhuǎn)電機(jī)中得到的，其網(wǎng)絡(luò)圖形如圖 1表示。在這網(wǎng)絡(luò)中每個(gè)輸入單位一起連接到每個(gè)隱藏的單位二個(gè)壓重上。 分別地, 在相應(yīng)的尺寸上決定了橢圓體的主要軸的中心共縱線和長度。</p><p><b>　　圖1 網(wǎng)絡(luò)結(jié)構(gòu)</b>&

10、lt;/p><p>　　1.2 HDANN在多個(gè)故障診斷中同時(shí)應(yīng)用</p><p>　　在旋轉(zhuǎn)電機(jī)典型的眾多故障中，例如失配、摩擦、軸裂縫、不同心、油旋轉(zhuǎn)、摩擦旋轉(zhuǎn)、軸承不精確性和他們的雙重和三倍缺點(diǎn)等，均在這個(gè)部分被考慮進(jìn)去。</p><p><b>　　1.3 故障數(shù)據(jù)</b></p><p>　　在測試裝備電動子中得到

11、的各種各樣的錯(cuò)誤數(shù)據(jù)來說明其物理缺點(diǎn)?；瑒虞S承支持電動子，并且數(shù)據(jù)在軸承中使用的位移傳感器，在水平和垂直的方向進(jìn)行收集。電動子以3000轉(zhuǎn)每分鐘的速度運(yùn)轉(zhuǎn)。在采樣期間，以1.6千赫的取樣頻率收集數(shù)據(jù)，并且在每1024點(diǎn)時(shí)候進(jìn)行抽樣收集。研究表明，如果輸入空間是高幅員，特別當(dāng)少量訓(xùn)練數(shù)據(jù)是可利用的時(shí)候，這樣會嚴(yán)重削弱網(wǎng)絡(luò)的推斷能力。然而，在條件監(jiān)測應(yīng)用中，用大量故障數(shù)據(jù)是極端罕見的。這里用于應(yīng)付這個(gè)問題的方法是假設(shè)增加隨機(jī)噪聲的訓(xùn)練樣品，

12、從而訓(xùn)練出比較精確的結(jié)果，并且使用這種方法，200個(gè)小組樣本分為七個(gè)類型的故障來生成，訓(xùn)練和測試網(wǎng)絡(luò)結(jié)構(gòu)。</p><p><b>　　1.4 網(wǎng)絡(luò)訓(xùn)練</b></p><p>　　所有子網(wǎng)絡(luò)為橢球狀單位網(wǎng)絡(luò)和八套輸入裝置，各種各樣的子網(wǎng)絡(luò)輸出裝置的數(shù)量是根據(jù)故障的進(jìn)行分類的，這樣是正確的。每個(gè)子網(wǎng)絡(luò)獨(dú)立地，從200個(gè)樣式任意地選擇的100個(gè)小組樣式，使用了以上學(xué)習(xí)算

13、法進(jìn)行訓(xùn)練。</p><p>　　1.5 網(wǎng)絡(luò)測試結(jié)果</p><p>　　測試的過程中，在訓(xùn)練期間沒使用的斷層類型而被使用了。因此樣式的總數(shù)為測試的7個(gè)100唯一斷層類型、21個(gè)100雙重?cái)鄬宇愋秃?5個(gè)100三倍斷層類型?？赡艿墓收峡倲?shù)是63。所以，測試結(jié)果如下所示：</p><p>　　1．作為唯一斷層類型，最后的診斷結(jié)果1。 正確診斷的百分比是99.6%。&

14、lt;/p><p>　　2．作為雙重?cái)鄬宇愋?，最后的診斷結(jié)果2。正確的百分比是98.9%。 </p><p>　　3．為三倍斷層類型，最后的診斷結(jié)果3。 正確診斷的百分比是96.4%.</p><p>　　分析以上不正確診斷的發(fā)生原因，我們發(fā)現(xiàn)三倍缺點(diǎn)類的數(shù)量是太大，并且有些類群很接近對應(yīng)的群，而形成了hyperellipsoids重疊。如果測試圖形卡在重疊的區(qū)域， H

15、DANN可以給不正確結(jié)果。在實(shí)際應(yīng)用，只要知道了斷層類型就可以從真正的情況中增加再訓(xùn)練的子網(wǎng)絡(luò)，并且能達(dá)到更高的診斷準(zhǔn)確性。</p><p><b>　　概要</b></p><p>　　橢球狀單位網(wǎng)絡(luò)劃分輸入空間與hyperellipsoids形成一定決定地區(qū)。這對于故障診斷的應(yīng)用是一個(gè)適當(dāng)?shù)倪x擇。對于這樣的網(wǎng)絡(luò)，一個(gè)等級制度的診斷方法能同時(shí)為多個(gè)故障診斷是可實(shí)行的

16、。</p><p><b>　　參考文獻(xiàn)</b></p><p>　　[1] Duda RO, Hart PE. Pattern Classification and Scene Analysis, Wiley, New York, 1973</p><p>　　[2] Watanabe K, Himmelblau DM. Fault di

17、agnosis in nonlinear chemical process: Theory. AIChE J 1983; 29: 243–250</p><p>　　[3] Venkatasubramanian V, Chen K. A neural network methodology for process fault diagnosis. AIChE J 1989; 35(12): 1993–2002&

18、lt;/p><p>　　[4] Yan Tinghu, Zhong Binglin, Huang Ren. Neural network technique and its application in fault diagnosis for rotating machines. J Vibration Eng 1993; 6(3): 205–211</p><p>　　[5] MacIn

19、tyre J,Jennings I.Condition monitoring software that thinks for itself.Proc MAINTEC 97 International Maintenance Conference,Birmingham,UK,1997</p><p>　　[6] MacIntyre J.Neural networks in condition monitorin

20、g just another fad?Proc 5th International Conference on Profitable Condition Monitoring,Harrogate,UK,1996</p><p>　　[7] Leonard JA,Kramer MA.Limitations of the back propagation approach to fault diagnosis an

21、d improvement with basis functions.AIChE Annual Meeting,Chicago, IL,1990</p><p>　　[8] Kavuri SN,Venkatasubramanian V.Representing bounded fault classes using neural networks with ellipsoidal activation funct

22、ions.Computers in Chemical Eng 1993;17(2):139–163.</p><p>　　[9] Kavuri SN,Venkatasubramanian V.Using fuzzy clustering with ellipsoidal units in neural networks for robust fault classification.Computers in C

23、hemical Eng 1993; 17(2):139–163</p><p>　　[10] Pao Y.Adaptive Pattern Recognition and Neural Networks,Addison-Wesley,New York,1989</p><p>　　[11] Hrycej T.Modular Learning Neural Networks.Wiley,N

24、ew York,1992</p><p>　　[12] Holmstrom L,Koistinen P.Using additive noise in back propagation training.IEEE Trans Neural Networks 1992;3(1),24–37</p><p>　　[13] Bishop C.Neural Networks for Patte

25、rn Recognition,Oxford University Press,1995</p><p>　　High Order Neural Networks for Simultaneous Diagnosis of Multiple Faults in Rotating Machines</p><p>　　B. Zhong'，J. MacIntyre2, Y. He` an

26、d J. Tait2</p><p><b>　　Abstract:</b></p><p>　　To overcome the limitations of the standard feedforward neural networks,high-order neural networks(i.e.ellipsoidal unit networks),which

27、are very useful for fault diagnosis applications due to their bounded generalisation and extrapolation,are proposed.This paper describes the theory and structure of such networks.A method for initialising hyperellipsoids

28、 and a training algorithm based on the standard backpropagation algorithm are proposed.Utilising the properties of bounded generalisation and ex</p><p>　　Keywords: Ellipsoidal unit networks;Hierarchical diag

29、nostic networks;Rotating machinery;Simul- taneous diagnosis of multiple faults;netural network</p><p>　　Introduction</p><p>　　In recent years, more and more attention has been given to the theor

30、y, method and strategy of fault diagnosis for large-scale rotating machines, and quantitative methods have been proposed, such as statistical pattern classification methods,system identification-based parameter model met

31、hods,and so on. These methods always need complicate models and extensive calculation. The Artificial Neural Network(ANN)technique offers potential for fault diagnosis due to its parallel processing, associative </p&g

32、t;<p>　　1.1 Network Architecture</p><p>　　The ellipsoidal unit defined by Eq. (3)is capable of approximating the Gaussian distribution. For the symmetric distribution, one unit can achieve a satisfact

33、ory approximation, and the network consists of two layers of linear input units and ellipsoidal output units. Each hidden unit is connected to only one output unit and each output unit has a dedicated set of hidden units

34、 for some distribution. Considering the distribution characteristics of feature vectors abstracted from raw data from rota</p><p>　　Fig1 network Architecture</p><p>　　1.2 Applications of HDANN i

35、n Simultaneous Diagnosis of Multiple Faults</p><p>　　Typical faults of rotating machines such as unbalance, rubbing, shaft crack, misalignment, oil whirl, rubbing whirl, bearing inaccuracy, and their compose

36、d forms of double and triple faults are considered in this section.</p><p>　　1.3 Fault Data</p><p>　　The various fault data were obtained by introducing physical faults on an test rig rotor syst

37、em driven by an electric motor. The rotor is supported by journal bearings, and data are collected at the bearings in both the horizontal and vertical directions using displacement sensors. The rotor was held at a consta

38、nt speed of 3000 rpm during sampling, and data was collected at a sampling frequency of 1.6 KHz, with 1024 points being collected in each time sample.</p><p>　　1.4 Training and Testing Patterns for Neural Ne

39、tworks</p><p>　　Taking the energy distribution of eight different frequency intervals from vibration spectra as feature vectors to calibrate various fault classes, the standard training patterns for Net1 are

40、 shown in Table 1.</p><p>　　*F1–F7 denote the fault class of unbalance, rubbing, axis crack, misalignment, oil whirl, rubbing whirl and bearing inaccuracy, respectively. Much research has indicated that a ne

41、twork’s ability to generalise is seriously impaired if the input space is of high dimensionality, particularly when small amounts of training data are available. Much research has indicated that a network’s ability to ge

42、neralise is seriously impaired if the input space is of high dimensionality, particularly when small </p><p>　　1.5 Network Training</p><p>　　In HDANN, all subnetworks are ellipsoidal unit networ

43、ks with eight input units, the number of output units of various subnetworks is determined according to assigned fault classed. Each subnetwork is trained independently with 100 groups of patterns selected randomly from

44、the 200 patterns for each fault type, using the learning algorithm described.</p><p>　　1.6 Network Testing Results</p><p>　　The testing process was carried out using the fault patterns not used

45、during training. The total number of patterns for testing were therefore 7×100 single fault patterns,21×100 double fault patterns, and 35×100 triple fault patterns. The total number of possible classes was

46、 therefore 63.The test results and discussion are as follows:</p><p>　　1. For single fault patterns, the final diagnostic result is given by Stage 1.The percentage of correct diagnosis is 99.6%.</p>&

47、lt;p>　　2. For double fault patterns, Stage 1 triggers Stage 2, and the final diagnostic result is given by Stage 2.The percentage of correct diagnosis is 98.9%.</p><p>　　3. For triple fault patterns, Stag

48、e 2 triggers Stage 3,and the final diagnostic result is given by the Stage3.The percentage of correct diagnosis is 96.4%.</p><p>　　Analysing the reason for incorrect diagnosis results, we find that the numbe

49、r of triple fault classes is too large, and some clusters of classes are so close that hyperellipsoids formed for corresponding clusters are overlapping. If a test pattern falls in the overlapping. If a test pattern fall

50、s in the overlapping region, HDANN may give incorrect results. In practical. applications, known fault patterns from real conditions can be added to retrain various subnetworks and higher diagnostic accur</p><

51、p><b>　　Summary</b></p><p>　　Ellipsoidal unit networks divide input space with hyperellipsoids and form bounded decision regions. This is an appropriate choice for fault diagnosis applications.

52、 Based on such networks, a hierarchical diagnostic strategy for simultaneous diagnosis of multiple faults is recommended and practicable.</p><p>　　References</p><p>　　1. Duda RO, Hart PE. Patter

53、n Classification and Scene Analysis, Wiley, New York, 1973</p><p>　　2. Watanabe K, Himmelblau DM. Fault diagnosis in nonlinear chemical process: Theory. AIChE J 1983; 29: 243–250</p><p>　　3. Ven

54、katasubramanian V, Chen K. A neural network methodology for process fault diagnosis. AIChE J 1989; 35(12): 1993–2002</p><p>　　4. Yan Tinghu, Zhong Binglin, Huang Ren. Neural network technique and its applica

55、tion in fault diagnosis for rotating machines. J Vibration Eng 1993; 6(3): 205–211</p><p>　　5. MacIntyre J,Jennings I.Condition monitoring software that thinks for itself.Proc MAINTEC 97 International Mainte

56、nance Conference,Birmingham,UK,1997</p><p>　　6. MacIntyre J.Neural networks in condition monitoring just another fad?Proc 5th International Conference on Profitable Condition Monitoring,Harrogate,UK,1996<

57、/p><p>　　7. Leonard JA,Kramer MA.Limitations of the back propagation approach to fault diagnosis and improvement with basis functions.AIChE Annual Meeting,Chicago, IL,1990</p><p>　　8. Kavuri SN,Ven

58、katasubramanian V.Representing bounded fault classes using neural networks with ellipsoidal activation functions.Computers in Chemical Eng 1993;17(2):139–163.</p><p>　　9. Kavuri SN,Venkatasubramanian V.Using

59、 fuzzy clustering with ellipsoidal units in neural networks for robust fault classification.Computers in Chemical Eng 1993; 17(2):139–163</p><p>　　10.Pao Y.Adaptive Pattern Recognition and Neural Networks,Ad

60、dison-Wesley,New York,1989</p><p>　　11.Hrycej T.Modular Learning Neural Networks.Wiley,New York,1992</p><p>　　12.Holmstrom L,Koistinen P.Using additive noise in back propagation training.IEEE Tr