壓縮包內(nèi)含有CAD圖紙和說明書,均可直接下載獲得文件，所見所得，電腦查看更方便。Q 197216396 或 11970985
Research on Fault Diagnosis of Fork Lift Truck Hydraulic System Based on Artificial Neural Network
Abstract—The structure and algorithm of BP neural net were described, therealization process of the fault diagnosis of hydraulic system based on BP neural net was discussed. According to the experiment and test of fault of fork lift truck hydraulic system, the BP net has better learning function, high net convergence rate and high stability of learning and memory. The diagnosis results indicate that the presented diagnosis method has high reliability and can attain the expected results, which can be applied to fault diagnosis of hydraulic system.
Keywords-Bp algorithm;Neural network;hydraulic system; fault diagnosis
I. INTRODUCTION
Because of the very complex structure of fork lift truck hydraulic system, once some faults happen in using process, it will have direct effect on operation efficiency. Therefore, the reliability and maintainability of the fork lift truck hydraulic system become increasingly high. At present, the traditional method of maintenance mainly depends on people’s experience, and it is very difficult to guarantee quality and efficiency of maintenance. Due to its self-organizing and nonlinearly adaptive nature, an artificial neural network potentially offers a new parallel processing paradigm that could be more robust and user-friendly than the traditional approaches. In fault diagnosis of hydraulic system, diagnosis information is acquired more easily by an artificial neural network than a single expert system based on regulation speculation. This paper describes application of BP neural network in fault diagnosis of the fork lift truck hydraulic system, and provides a newly solution methods.
II. A MODEL STRUCTURE OF BP NEURAL NETWORK AND TRAINING ALGORITHM
A. A model structure of BP neural network A typical structure of a three layer forward neural network is shown in figure 1. It includes input layer, hidden layer and output layer. In figure 1, circles represent neurons. Connecting line having weight between circles represents interaction strength between neurons, where is the connection weight between neuron i in the k-th layer and
neuron j in the k-1-th layer. is the threshold of neuron, (i=0~n) is the input of neurons, (j=0~m) is the output of neurons, and F(·) is a transfer function from the (k-1)-th layer to the k-th layer.
B. Learning algorithm of BP neural network
BP (Back propagation) neural network uses the error of the output layer to estimate the error of the direct precursor layer of the output layer, and then use the error to estimate the error of the preceding layer again and again. The estimates of error of the other layers again and again. The estimation of error of the other layers can be obtained. In this way, it may form the process that transmits the error of the output layer to the input layer of network along the transmission right about of the input signals. Thereby, the algorithm is called the Back Propagation algorithm. And the non-cycle network that uses the BP algorithm to learn is called BP network. Its course of learning is just the course of training. The training is to adjust the weights among neurons by certain manner when the samples vectors are put into neural network. The specific realizations of BP learning algorithm follow as:
? Initialize right aggregate wij, get the value of the lesser stochastic nonzero;
? Give many pairs of input and output samples (Xp, Dp), where p=1, 2, …, p, i is number of training mode pairs; Xp is input vectors, Dp is output expectation vectors.
? Calculate their actual output Yp=(y1p, y2p, …, ymp), in this course, many times of positive spread calculation is done in terms of the different number of network layer. Evaluate the objective function of the network, and the output error value can generally be denoted as:
? Judge whether the network satisfies the precision
? Where ε is the desired precise, the process of training will continue until the precision is attained.
? Adjusting the weights through dropping off one by one along the reverse according to grads can be computed by:
III. ESTABLISHING BP NEURAL NETWORK OF FAULT DIAGNOSIS OF HYDRAULIC SYSTEM
This paper is type of CPQ30 fork lift truck as a example. Fault rate of hydraulic system is
much higher, and many fault reasons also occur. Aimed at general fault of hydraulic system,
BP neural network verifies fault reason.
A. Analyzing fault mode and fault mechanism of
hydraulic system Analysis of fault mode and fault mechanism of hydraulic system is
shown in table 1.
B. Selecting the input and output vector of BP neural
network Units We consider fault mode x=(x1, x2, x3) as the input vector of neural network, and fault reason y=(y1, y2, y3, …, y7) as
the output vector of neural network. The nonlinear, mapping relation between fault mode and fault mechanism is established, Then we train neural network.. The input vector x1, x2, x3 indicate three kinds of faults. Namely, x1 indicates that temperature of pressure oil becomes more and more high; x2 indicates that lifting cylinder becomes powerlessness; x3 indicates that seals in the joint leak. The output vector y1, y2, y3, y4, y5, y6, y7 indicates seven kinds of faults. y1 indicates that pressure of main relief valve is too lower, large quantities of oil flows into relief valve; y2 indicates that Outer dirt of oil radiator is excess, and makes radiation efficiency low; y3 indicates that seals in piston damage; y4 expresses that directional valve damages; y5 expresses that non-return valve damages; y6 expresses that seal ring damages; y7 expresses that screws in joint aren’t tightened.
C. Selecting the structure of BP neural network and training sample
1) Determining the structure parameter of BP neural network G.Cbenko, Licheng Jiao thought a three-layer straight feed-forward network with enough nodes of hidden layer approaches to any continuous mapping with arbitrary accuracy. According to mapping existing theory, a three-layer network with a hidden layer can obtain expectation accuracy. If determining concrete problems, once the number of hidden nodes is determined, its structure can be determined. The structure of BP neural network which applies fault diagnosis to hydraulic system of fork lift truck includes three layer, namely, the number m of hidden layer nodes is 4, the number n of output layer nodes is 7, and according to empirical formula where h equals h=（） , the number h of hidden layer nodes is 5.
The transfer function from the input layer to the output layer in network training is very often a sigmoid function, namely, ，the transfer function from the hidden layer to the output layer is linear function.
2) Selecting training samples Supposing F represents learning sample in network, from the above determined network structure, some parameters are obtained as following: F equals (X, Y), X equals (x1, x2, x3), and Y equals (y1, y2, y3, … ,y7). Where, X expresses the input sample, and Y expresses the output sample. Training samples in network is shown in table 2
IV. NETWORK TRAINING AND NETWORK TESTING
Training samples from the above table 2 are applied to compiled MATLAB program, where, network training error index is 0.00001, and test results is shown in table 3. In view of table 3, if network training error index is 0.00001, the output result of sample 1 equals [0.9956, 0.9989, 0.0008, 0.0012, 0.0023, 0.0011, 0.0015], however, the expectation value of the output is [1, 1, 0, 0, 0, 0, 0]. This shows that the diagnosis result of BP neural network is consistent with the actual result. The value of fault nodes is close to one, the value of non-fault nodes is close to zero, and it is proved that the approach has relatively accuracy and reliability.
V. CONCLUSION
This paper proposes a newly approach of fault diagnosis of fork lift truck hydraulic system based on BP neural network. Through experiment showed, this approach is feasible, has characteristics of fault tolerance, conjecture, memory, adaptation, parallel process etc, and has certain practical value.
REFERENCES
[1] Tianjue Lei. Hydraulic Engineering Hand Book. Machine Industry Press, Beijing, 1998.
[2] Heqing Li, Qing Tan. “Fault Tree Analysis on the Powerlessness of Lifting Cylinder in Hydraulic System of the Fork Lift Truck”,Chinese Journal of Machine Tool & Hydraulics,Vol.36 No.2:199-201, 2008.
[3] Sanquan He, Yaozhu Mo. Comprehensive Transport & Assembly Machinery. People Traffic Press, Beijing, 2002.
[4] Zhong Li, Guoping Yang. Hydraulic Transmission Theory, Fault diagnosis and Elimination of Construct Machinery. Machine Industry Press, Beijing, 2005.
[5] Kaili ZHou, Yaohong KANG, Model of the Neural Network & Simulation Program Design for Matlab. Tsinghua University Press, Beijing, 2005.
[6] Qingsheng Xie, Jian Yi, Method of the Neural Network in Mechanical Engineering, China Machine Press, Beijing, 2003.
中文翻譯：
對基于仿真神經(jīng)網(wǎng)絡(luò)的叉車液壓系統(tǒng)故障診斷研究
摘要：
闡述了BP神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu)和算法，對基于BP神經(jīng)網(wǎng)絡(luò)的液壓系統(tǒng)故障診斷產(chǎn)生過程進行了討論。根據(jù)實驗和測試叉車液壓系統(tǒng)的故障，BP網(wǎng)絡(luò)有更好的學習功能，高網(wǎng)絡(luò)歸納整理速度，和高穩(wěn)定性地學習和記憶能力。診斷結(jié)果表明，本文所提出的診斷方法具有較高的可靠性，可達到預(yù)期的結(jié)果，可用于液壓系統(tǒng)故障診斷中。
關(guān)鍵詞-BP算法；神經(jīng)網(wǎng)絡(luò)；液壓系統(tǒng)； 故障診斷
1.簡介
因為叉車液壓系統(tǒng)結(jié)構(gòu)非常復(fù)雜,一旦在使用過程中發(fā)生某一故障,它將直接影響運行效率，因此叉車液壓系統(tǒng)的可靠性和可維護性變的越來越重要。目前傳統(tǒng)的維護方法主要取決于人們的經(jīng)驗，很難保證維護的質(zhì)量和效率，由于其自組織性和非線性自適應(yīng)性質(zhì)，仿真神經(jīng)網(wǎng)絡(luò)可能提供一種新的可以并行處理模式比傳統(tǒng)的方法更強大更適用。
在液壓系統(tǒng)故障診斷中，診斷信息通過仿真神經(jīng)網(wǎng)絡(luò)比一般基于規(guī)律預(yù)測的專家系統(tǒng)更容易獲得。本篇介紹了BP神經(jīng)網(wǎng)絡(luò)在叉車液壓系統(tǒng)故障診斷中的應(yīng)用，提供了一種新的解決方法。
2. 一種BP神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu)的訓(xùn)練算法
A．一種BP神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu)
典型的三層神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)如圖1所示它包括輸入層、隱層和輸出層。在圖1中,圓代表神經(jīng)元。
圓之間連接線具有權(quán)重代表神經(jīng)元之間相互作用的力，是連接在k-th層中神經(jīng)元i和在k-1-th層中神經(jīng)元j之間的權(quán)重。是神經(jīng)元的輸入端口（），是輸出神經(jīng)元，F(xiàn)(·)是從(k-1)k-th層到k-th層的傳遞函數(shù)。
B.學習BP網(wǎng)絡(luò)神經(jīng)算法
BP（反向傳播）神經(jīng)網(wǎng)絡(luò)用輸出層的誤差估算它的前導(dǎo)層誤差，然后再用這個誤差判斷它的前一層誤差依次下去。一層又一層的估算其它層地誤差，就可以得出其它層的誤差 ，這樣，可以形成傳遞輸出層對輸入層網(wǎng)絡(luò)誤差并影響輸入層信號的過程。因此該算法叫反向傳輸算法，non-cycle網(wǎng)絡(luò)使用BP算法演示叫做BP網(wǎng)絡(luò)，他的訓(xùn)練過程就是學習過程。演算就是當樣本矢量輸入神經(jīng)網(wǎng)絡(luò)時用某種方式調(diào)整神經(jīng)元之間的重量。具體的BP演示算法如下：
l 初始化得到正確的值，獲得更小的隨機非零值。
l 給出幾組輸出輸入樣本值（），p=1，2，…，p，i是演算模式組的順序號是輸入量，輸出的期望值。
l 計算它們真實的輸出值=(，，…)，在這個過程中，無數(shù)次的傳輸計算在不同數(shù)量的網(wǎng)絡(luò)層中完成。
評估的目標函數(shù)網(wǎng)絡(luò)和輸出的誤差值一般可以表示如下式：
l 判斷網(wǎng)絡(luò)是否滿足精確度要求：
l 是理想的精確值這個演算過程一直進行到得到滿足精確度值時。
l 通過依據(jù)梯度減掉一個又一個逆向值來調(diào)整權(quán)重，可根據(jù)下式計算：
3.建立液壓系統(tǒng)故障診斷的BP神經(jīng)網(wǎng)絡(luò)
本文是以CPQ30型叉車作為一個例子。液壓系統(tǒng)的故障率非常的高，而且許多操作失誤也存在。針對一般的液壓系統(tǒng)故障，BP神經(jīng)網(wǎng)絡(luò)驗證了其故障的原因。
A. 液壓系統(tǒng)故障形式及故障機理分析
液壓系統(tǒng)的故障形式及故障機理分析見表1。
B. 選擇BP神經(jīng)網(wǎng)絡(luò)單元的輸入輸出矢量
我們定義故障形式x=（，，）為神經(jīng)網(wǎng)絡(luò)的輸入量，故障原因y=（，，，…，）作為神經(jīng)網(wǎng)絡(luò)的輸出量，故障形式和故障機理之間的非線性映射關(guān)系就確立了，我們繪制神經(jīng)網(wǎng)絡(luò)。輸入三個變量，，表示三個不同的故障點，表示壓力油的溫度越來越高，表示舉升液壓缸壓力不夠，x3表示連接處油泄漏，輸出矢量分別為，，，，，，，表示七種不同的故障點。表示主安全閥壓力偏低了，大量的油流入安全閥；表明外層散熱的污垢油太多，使散熱效率變低；
表示密封活塞損壞；表示換向閥處于損壞；表示單向閥處于損壞；表示密封圈損壞；表示螺釘沒有擰緊。
C. 選擇BP神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu)和訓(xùn)練演示樣本
1）.確定BP神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu)參數(shù)
G.Cbenko, Licheng Jiao 想到了一個三層前饋網(wǎng)絡(luò)，隱層有足夠結(jié)點近似于具有任意精度任意連續(xù)性地映射的方法。根據(jù)映射現(xiàn)有的理論，建立的含有隱層的三層網(wǎng)絡(luò)能獲得期望的精度。一旦隱層節(jié)點數(shù)確定了，它的結(jié)構(gòu)就確定了。
提供給叉車液壓系統(tǒng)診斷的BP神經(jīng)網(wǎng)絡(luò)包括三層，也就是說隱層節(jié)點數(shù)m的值為4，輸出層節(jié)點數(shù)n值為7，并根據(jù)經(jīng)驗公式h=（），隱層節(jié)點數(shù)h等于5.
表一. 液壓系統(tǒng)的故障類型及故障機理分析
樣本點
符號
故障類型
故障原因
符號
1
壓力油溫度上升過快
主壓力閥壓力太低，大量的油流入安全閥；
外層散熱污垢油過多，降低散熱效率
2
舉升液壓缸壓力降低
活塞密封損壞；
換向閥損壞；
單向閥損壞；
3
聯(lián)接處密封泄露
密封圈損壞；
螺釘未擰緊；
表二. 對BP神經(jīng)網(wǎng)絡(luò)測試樣本
樣本點
樣本輸入 樣本輸出
1
1
0
0
1
1
0
0
0
0
0
2
0
1
0
0
0
1
1
1
0
0
3
0
0
1
0
0
0
0
0
1
1
表三. 故障診斷結(jié)果
樣本點
輸出結(jié)果
1
0.9956
0.9989
0.0008
0.0012
0.0023
0.0011
0.0015
2
0.0024
0.0020
0.9809
0.9982
0.9919
-0.0012
0.0015
從輸入層到輸出層的傳遞函數(shù)在網(wǎng)絡(luò)測試中常常是一個函數(shù)的功能，也就是
，從隱層到輸出層的線性函數(shù)。
2）.選擇測試樣本
假設(shè)F代表網(wǎng)絡(luò)中測試的樣本點，從上述已經(jīng)確定的網(wǎng)絡(luò)結(jié)構(gòu)中，可以獲得以下參數(shù)，F(xiàn)=（X，Y），X=（，，），Y=（，，，…）。X表示輸入信號，Y表示輸出信號。網(wǎng)絡(luò)測試點見表2中。
4.網(wǎng)絡(luò)訓(xùn)練與網(wǎng)絡(luò)測試：
表2中的訓(xùn)練樣本適用于編寫MATLAB程序，網(wǎng)絡(luò)訓(xùn)練誤差指數(shù)是0.0001，測試結(jié)果見表3.
針對表格3，如果網(wǎng)絡(luò)訓(xùn)練的誤差是0.00001，表1中樣本輸出結(jié)果等于[0.9956,0.9989,0.0008,0.0012,0.0023,0.0011,0.0015]，然而輸出的期望值是[1,1,0,0,0,0,]。這就說明BP神經(jīng)網(wǎng)絡(luò)診斷結(jié)果符號實際要求的結(jié)果。有故障節(jié)點的值接近于1，無故障節(jié)點的值趨向0證明了該方法具有一定的精度和可靠度。
5. 結(jié)論：
本文提出了基于BP神經(jīng)網(wǎng)絡(luò)的一種新的叉車液壓系統(tǒng)的故障診斷方法。通過實驗證明，這種方法是可行的，具有故障容差（容錯性）、預(yù)測、記憶、自適應(yīng)、并行處理等特性，具有一定的實用價值。
參考文獻
[1]雷天覺.液壓工程手冊。北京，機械工業(yè)出版社，1998
[2]李清何，譚青.叉車液壓系統(tǒng)起升液壓缸無力的故障樹分析，中國機床與液壓期刊No.2:199 -201,2008年
[3]何三全，莫耀祖.綜合運輸與裝卸機械，北京，人民交通出版社2002
[4]劉忠，楊國平.工程機械液壓傳動原理故障診斷與排除，北京，機械工業(yè)出版社2005
[5]周開利，康耀紅 .神經(jīng)網(wǎng)絡(luò)模型及其MATLAB仿真程序設(shè)計，北京，清華大學出版社，2005
[6]謝慶生.尹健.機械工程中的神經(jīng)網(wǎng)絡(luò)方法，北京，中國機械工業(yè)出版社，2003