The Research On The Application Of Fuzzy Neural Network In Choosing Machining Parameters

On the background of Jilin Province science and technology developmentproject-research on fuzzy-neural network in machining (number: 20040333),inference and learning ability of fuzzy neural network were utilized to solve errorreflection problem and realize automatic choosing of machining parameters inCAM.As the bridge between CAD and CAM in process of automatic machining,development of CAPP is rather slow for the complexity, uncertainty and dynamicof its problems. Many non-leaner problems in CAPP can only be solved byartificial intelligence to simulate the thinking process of human experts. Forexample, the error reflection phenomenon in machining is now solved by workers'experience to machine several times and select proper machining quantity. But thismethod is rather subjective and cannot eliminate affect of the error reflectionphenomenon which has brought difficulties for computer aided machining. By farno effective method was found yet.The fuzzy neural network is the outcome that the neural network techniquecombines the fuzzy logic control techniques, which is a fuzzy control method basedon the neural network. A kind of fuzzy neural network was put forward to resolvethe error reflection phenomenon in machining. Although the fuzzy logic and theneural network have obvious different concept and contents, they are all used forsolving difficult problems on system control caused by the uncertainty andimprecision. The fuzzy logic imitates the logic thinking of person's brain and isused to deal with the imprecision in control. The neural network imitates thefunction of the brain nerves and can approach any non-linear function and reflectthe relation of input and output. Using the combination of them to solve the errorreflection problem has certain possibility.Optimization of machining parameters was attempted to be solved by fuzzyneural network. The standard back propagation method was modified and modelingof the fuzzy neural network was realized by Fuzzy C-mean clustering algorithm.The fuzzy neural network was designed too. This method has clear theories andperspicuity structure and its exactness and feasibility has been validated throughtraining and testing of the established network model. Interface program betweenfuzzy neural network and numerical control system was designed. In this waynumerical control machining program can be generated automatically. Machiningresults were satisfactory on Vturn-20 numerical control lathe. The main contentof the thesis includes following:1. The analysis and research of the error reflection phenomenon.Combining the error reflection theories and the machining circumstance analysisthe factors influence the error reflection coefficient, then educes the complicatednon-line relation of it and the semi-finished product error(Efront), the craft systemhardness(K),machining quantity(f),the work piece rigidity(HBS) and machiningtimes(Z).According to experience, along with the increment of machining times,the error can be minished , but in physically machining, the number of timesgenerally isn't more than 3 times. So we can suppose the machining times is 3,themachining proportion of each time is P1, P2, P3,The machining process can bedetermined as f, P1 and P2 were known. A fuzzy neural network with four input andthree output was formed as K,HBS, Efront, Eafter were input to the network while f,P1 and P2 can be got as output.2. Collect the experiment data. Use a easy eccentric clamp to machine on alather for many times with different hardness material( including iron, aluminum,20# steel, 45# steel), record the results( including K, Efront, Eafter, f, P1, P2) as thedata for network training and testing;3. Modeling of the fuzzy neural network based on fuzzy c-meanclustering arithmetic. Sampling input and output data were standardized with datastandardization arithmetic. Then they were clustered with fuzzy c-mean clusteringarithmetic to realize division of fuzzy subspace. Parameters of member functionswere determined at the same time.4.The structure optimization of the network model. Because the studyarithmetic of I FNN is based on the BP network 's arithmetic and it exists suchshortcomings as large expense, slow constringency speed, easy getting into localtittles, etc., usually faces the problem that can't be trained completely, it needs tobe bettered. The main resolving methods include choosing the appropriate originalauthority and using less study velocity. Because initial authorities can't beestimated accurately at the beginning and too small study velocity will cause toolong training time, testing collect method is used in the general neural networkcontroller's design to trial and error to confirm appropriate initial authorities andstudy velocity. The controller designed in the text is FNN controller, choose ofinitial authorities has advantage compared with generic NN because the initialauthorities have practical meaning. Self-adjustive study speed method was used toadjust study speed and append momentum method to settle the problem of easygetting into local minimum.5. Training and testing of network as well as experiments.The fuzzy neural network was trained with treated experiment data. Aftertraining, it was tested to find deficiencies. Numerical control machining programcan be generated automatically from output of the network. Machining experimentswere carried out on Vturn-20 numerical control lathe and the results weresatisfactory.Both analyses in theory and experiments results lead to the conclusion thatusing fuzzy neural to realize automatic choosing of machining parameters andautomatic generation of numerical control program is feasible.