The parameter of planar milling treatment that is based on artificial nerve network and genetic algorithm suits oneself optimize

9 document label piles up: A chapter number: Ruan XueyuAbstract of 　 of Li Congxin of 　 of Lin Jianping of 　 of 1001-2265(2000) 　 02-0005-04Adaptive Optimization Of Machining Parameters For Face Millingbased On Artificial Neural Network And Genetic AlgorithmsNi Qimin: The Keys To Adaptive Control With Optimization Of Machining Operations Are Modeling Of Cutting Process And Selecting Of Real-time Optimal Strategy.

In This Paper, we Propose To Moldel The Machining Process Using Artificial Neural Network And Realize On-line Optimization Using Genetic Algorithms.

A System For Adaptive Optimization Of Cutting Parameters Such As Feed Rate And Spindle Speed In A Face Milling Operation For Maximizing The Material Removal Rate Without Violating Machining Constraints Is Set Up.

Key Words: Cutting of metal of foreword of Face Milling;adaptive Optimization;artificial Neural Network;genetic Algorithms1 machines the choice of parameter to have main effect to machining productivity and economic benefits. In system of groovy computer CNC Machining, before treatment parameter is being machined manual of the experience according to process designing personnel, treatment or decide through leaving a line to optimize. However, in machining a process actually, machining operating mode is constant change, the treatment parameter that appoints beforehand when process designing cannot make machine a target to always be maintained best. A method that solves this problem is to give CNC to machine a system to offer to be able to optimize the best control system that processes parameter in real time according to machining conditional change. Such system is called from adaptive control system, can divide for the tie from adaptive control (ACC) and best from adaptive control (ACO) [1] . Tie from adaptive control it is through making cutting force or treatment power are maintained actually make productivity in limiting value the biggest change, its main drawback depends on lacking the feedback of pair of workpiece quality, the treatment parameter that considers onefold tie to get only may be restrained because of disobeying other and become unworkable [2] . Although people is right already best had much research from adaptive control [3 ～ 7] , but achieve economic system not much. Main difficulty depends on those who machine process model building what optimize strategy with real time to make. Be aimed at these two problems, the planar milling that the article put forward to be based on artificial nerve network to build model and genetic algorithm to optimize suits oneself optimize a method. Artificial nerve network establishs 2 treatment procedure the treatment model with appropriate and logical pattern is the premise that undertakes machining optimizing. Treatment model should parameter can be machined in the well and truly report inside wider operating mode range (rate of cutting speed, feed, cut wait greatly) output with treatment (quality of power of cutting force, cutting, surface) relation. In addition, treatment process model still must have pair of treatment environmental change, especially cutting tool wears away of the state get used to ability oneself. The complexity as a result of treatment process and nonlinear characteristic, accurate the treatment process model that builds an analytic form is special difficulty. The closest research shows [8, 9] , have suit oneself ability and nonlinear the artificial nerve network that handles ability, especially multilayer feel implement network, comfortable build a model at making a process, and normally other of be superior to builds modular technology. Be based on above analysis, this literary grace comes with artificial nerve network variable of map process input and the relation with process variable output. The cutting tool in machining a process to make process model can suit wears away the change of the state, outside treatment parameter is being divided in process input variable, still include cutting tool to wear away length is variable. When use, need undertakes to machining the cutting tool in the process to wear away above all online measure. 2.

Network of nerve of tatty of 1 cutting tool measures cutting tool to wear away the state is right cutting force and workpiece surface quality have main effect, accordingly, for accurate forecast cutting force and treatment quality, must measure cutting tool to wear away above all length. Measure cutting tool to wear away the method of the state has direct way and indirect method, direct way can be divided again measure law and blame contact to measure a way for the contact. The contact measures a law to be able to be used at after the event to measure only commonly, realize online estimation hard. Blame contact measures a law to use laser and image processing technology more directly, because be handled the limitation of speed or precision and cost relative to taller reason, use very hard also at be being measured in real time. Use actually at present it is indirect method of measurement mostly, if cutting force detects,law, vibration detects law, AE law. Research shows [10, 11] , because machine the complexity of the process, only basis is onefold of sensor detect information reckons cutting tool wears away state, can appear very tall measure by accident, sign up for by accident, the method that uses confluence of much sensor information commonly consequently undertakes detecting, but computational amount is larger. The article is based on the ability of large-scale collateral computation of nerve network and information confluence ability, put forward cutting tool to wear away the measurement technique of artificial nerve network of length. Network type chooses to be multilayer perception implement, the information of place confluence (namely the network is inputted) include rate of cutting speed, feed and cutting power can the 4 features parameter of chart. The network outputs node to have only, namely cutting tool wears away the estimation value of length. Multilayer feel implement the division check the number of the amount of concealed layer and each concealed layer has very big effect to the function of the network. Undertake trying to a variety of network structures, compare its to learn speed and precision, structure of affirmatory finally network is 6-4-1. 2.

2 processes are forecasted implement the process is forecasted implement nerve network wears away according to the cutting tool that already estimated length and treatment parameter will forecast a process to output variable. Network type also chooses to be multilayer perception implement. Input node in all 5, wear away for cutting tool respectively rate of speed of length, cutting, feed, cut mix greatly cut wide; Output node is 3, part P of power of F of corresponding cutting force, cutting and exterior surface roughness Ra. Concealed layer structure uses trial and error method to decide, use two concealed layer, check the number of concealed layer division mixes 7 for 15 respectively. 2.

Before the network is using two afore-mentioned nerve, 3 training algorithm must undertake the line trains leaving. Multilayer perception implement model the training algorithm with nerve the most commonly used network is BP algorithm, but this algorithm is had be immersed in easily local the least weakness. And imitate anneal is algorithmic [12] can jump out local and best trap, find overall situation (or approximate) optimum solution. two kinds algorithmic union rises, can produce both advantage adequately. Its train measure to be as follows: ① produces network parameter randomly initiative condition R, algorithm of BP of basis of ② of 　 of 　 of T0 of your T ＝ produces R fall to be awaited choose ＝ of your E of ③ of 　 of 　 of condition R ′ [Ep (R ′ ) - Ep (R) ] ／ Ep (R) , the ④ of 　 of function of sum of squared error that Ep is training example is like E ≤ 0, make ′ of R ＝ R; Otherwise with probability EXP(- E ／ KBT) accept ⑤ of 　 of 　 of ′ of R ＝ R to repeat ⑥ of second 　 of M of ② ～ ④ to reduce ⑦ of 　 of temperature T 　 to repeat ② ～ ⑥ till temperature very low or the nerve after already achieving precision to ask the course trains. 3 genetic algorithm are used at treatment to optimize 3.

1 treatment optimizes a model from raise treatment efficiency to set out, it is with rate of purify of the oldest stuff optimize a target. To planar milling, MRR of material purify rate can express to lead F for feed, cut deep Ap and the product that cut wide Aw. In considering effective treatment, cut mix greatly it is more difficult to cut widely online adjustment, online control parameter chooses to be N of rotate speed of easy pilot main shaft and feed to lead F(to cut deep, cut wide can be in according to working procedure of workpiece geometry and treatment when be being optimized from the line preliminary and affirmatory) . The function that machines a tie to basically have machine tool and cutting tool itself is restricted and be asked by the quality of treatment workpiece, tie of tie of force of the the limits restriction that includes rotate speed and feed rate, biggest cutting, power of the biggest cutting and tie of workpiece surface surface roughness. The place on put together is narrated, optimize a model to be as follows: Optimize a target: 　 of MaxMRR ＝ Fapaw designs variable: X ＝ [N 　 F] T 　 restrains a condition: The input restrains Nmax of ＜ of Nmin ＜ N, output of 　 of 　 of Fmax of ＜ of Fmin ＜ F restrains F ＜ Fmax, p ＜ Pmax, 　 of Ra ＜ Ramax this is the tie optimizes a problem, can use punish function law its translate into optimizes a problem without the tie: 　 of 　 of 　 of 　 of 　 of MaxH ＝ Fapaw - R1 (＜ Fmin - F ＞ Nmin of 2 ＋ ＜ - 2 　 of N ＞ F of ＜ of ＋ of 　 of 　 　 　 - 2 ＋ of Fmax ＞ ＜ N - Nmax ＞ 2) 　 of 　 of 　 　 　 - R2 (＜ , r1, R2 is the castigatory factor that when disobeying input, output to restrain, brings to bear on respectively, definition of ＜ ＞ function is: 3.

2 real time optimize algorithm efficient, reliable optimizing algorithm is implementation in the requirement that the line optimizes. The heredity that gets edificatory from inside biology evolution is algorithmic [13] , use heuristic search technology to seek optimum solution, optimize methodological photograph to compare with the tradition, have rash club sex efficiency of good, search is tall, right target function limitation is little, use collateral machine to make the good point such as operation of collateral high speed easily, agree with consequently afore-mentioned those who optimize a model is online seek solution. Algorithmic and main process is as follows: ① encode represents all sorts of extraction of one each variable to be worth with the binary code of certain length, of will respective variable 2 become a string repeatedly into the code, get code of a binary system is strung together, call a chromosome. The length of chromosome depends on the binary number that is used at rotate speed and feed rate replacement in numerical control system. For example, digit of rotate speed replacement is 3, digit of feed rate replacement is 4, criterion the length of chromosome is 7. ② produces initiative group to press random method by the computer, generation a series of chromosome, every chromosome represents each put oneself in another's position, a certain quantity of individual composition is primitive group. ③ computation grade of fit presses encode regulation, each individual chromosome the independent variable of place correspondence takes a cost (Ni, fi) era is entered (* ) type, the value of earnings target function regards his as grade of fit. ④ seed selection regards a choice as pressure with grade of fit, choose a pair randomly individual, as the parents of raise up seed. ⑤ cross has the father and mother of random pitch on cross. ⑥ mutation selects a group with certain probability in a certain number of individual, to what already chose every are individual, choose some randomly, retroflexion this number. Till group grade of fit tends,⑦ repeats ③ ～ ⑥ stable, rise no longer. Treatment of 4 planar milling suits oneself optimize systematic machining best depend on getting used to treatment model oneself from the key of adaptive control build and optimize strategy in real time make. This literary grace built the process model that planar milling machines with network of two artificial nerve, came true with genetic algorithm online optimize, be based on above algorithm, the planar milling that built next plan institute to show suits oneself optimize a system. Planar milling CNC Machining suits oneself optimize a system to pursue this system basically comprises by 3 parts: Measure nerve network to be opposite of the cutting tool in machining a process wear away the state undertakes real time estimates; Network of process model nerve builds foregone cutting tool to wear away parameter of the treatment below length and the relation with process variable output, the computation that is the grade of fit in optimizing a program provides partial data; Online optimize algorithm to assure to offer best treatment parameter for numerical control system from beginning to end () of rotate speed of feed rate, main shaft, make machining environmental happening to change (cutting tool wears away, cut because of what workpiece geometry changes and cause deep, cut broad change to wait) below the circumstance, treatment system still can work below best condition, fall in the premise that does not violate treatment obligation namely, obtain as far as possible old material purify rate. 5 conclusion 　 (the artificial nerve network that 1) has capability of large-scale collateral processing and information confluence capacity, usable Yu Dao provides tatty online measure. (2) is multilayer perception implement the network is comfortable build a model at making a process. (The genetic algorithm with the good, tall efficiency that seek solution can use sex of 3) rash club at best the real time that gets used to a system oneself is optimized. (The planar milling of 4) article construction optimizes a system to because cutting tool wears away,can suit, the treatment environment that workpiece geometry causes changes, but inhomogenous to material wait for not certain factor to still cannot be compensated. In addition, the process model of the article gets through training from the line, when workpiece material or cutting tool produce change, must undertake training afresh. Further research can be discussed online get used to the theory that builds a model and method oneself. CNC Milling