Be based on parameter of FNN intelligence grinding decision-making system

In The System, grinding Parameters Are Decided By Upper-layer Expert System, and Then Will Be Revised By Two FNN Units.

Wheve One Unit Uses The Actual Results Of Grinding To Repair These Grinding Parameters And Another Unit Uses The Current Status Of The Grinder To Repair Them.

So The Suitable Grinding Parameters Are Obtained.

By Simulating, it Has Been Proved That The Grinding Parameters Decision-making System Is Feasible And Has High Adaptive Ability.

Keywords: The production technology with modern introduction of Fuzzy Logic; Neural Network; Grinding System; Intelligent Decision-making1 forward flexible make (FMS) , computer is compositive make (CIM) and intelligence are made (the way such as IM) develops. Among them intelligence is made (IM) more the as the serious content of advanced production technology heat that makes research increasingly. It aims to introduce intelligent pilot theory and method production course, carry the thinking of imitate mankind expert and inferential activity, replace and the part in outspread production environment is mental, make production system can detect automatically thereby its run condition and environmental change, when getting the outside and inner drive, can make correct judgement and decision-making, in order to make sure the system achieves dovish treatment result. The precision control that grinding machines includes the many sided such as quality of appearance and dimension precision, position precision, surface. Compare with other machining photographs, the precision that grinding treatment asks is taller, grinding mechanism is more complex, behave especially go up in exterior quality. The factor that affects working surface quality in grinding process very much and affect each other between each elements, of index of exterior quality integrality online it is very difficult also to measure, this is built with respect to what be mathematical model it is difficult to caused huge. The evaluation of integrality of workpiece surface quality relies on factitious experience to decide again, subjective element holds very large proportion, caused very great difficulty to machine the evaluation of the result. In the light of traditional grinding treatment process precision controls a variety of difficulty that encounter, we put forward will traditional expert inference orgnaization to be the same as union of photograph of ambiguous nerve network to establish intelligence grinding parameter decision-making system. Expert system is used to be opposite in the system grinding parameter is preliminary and decision-making, when machining clearance and treatment end, undertake modulatory to machining parameter and undertake correction to the knowledge base, make the system was had very strong get used to ability and ability of self-study be used to oneself, the precision that raised grinding and grinding efficiency reduced the error that grinding processes. Can assure to achieve ideal treatment result from beginning to end when the machine tool is changed with environmental happening. 2 grinding parameter is decision-making of the system in making the grinding system in the tradition, speed parameter is original make choice of to measure from beginning to end, change and be adjusted what do not pledge as material of machine tool state, emery wheel state, workpiece. Be not as a result of grinding parameter best, cause grinding efficiency result of low, grinding is put in bigger error. But the mechanism that machines as a result of grinding is very complex, the parameter that affects grinding result is more, and be affected each other between each grinding parameter, the method that uses a convention cannot build a model to grinding process, the control method that renders a tradition is hard to come true. Artificial nerve network and ambiguous and logistic occurrence put forward new thinking for us, nerve network is had suit oneself, capability of self-study be used to is strong, do not have too much limitation to the input, quantity that exports parameter, three-layer network can approach random complex nonlinear the advantage such as the system. But have operation as a result of nerve network itself the amount is large, be immersed in easily local the smallest, after iteration reachs fixed number, convergent process becomes very slow wait for defect. Because this still cannot use its at the real time control of exact amount. A system that is based on ambiguous logic is to adopt a structure, pass subject degree of function and will be described accurately through language regulation. It does not need to learn training, but the performance that must understand beforehand and analyses a system, and the result of this one analysis is decided with regular form and the record comes down. What differ with nerve network is, nerve network is based on inhomogenous nerve yuan join structure, often completely opaque, and in ambiguous and logistic system, to all conditions, each separate measure can be mixed to comprehend by understanding. Nerve network has ability of self-study be used to and capacity of large-scale collateral processing, be good at quite on cognitive processing. And ambiguous system can make full use of the knowledge of course domain, can convey knowledge with less regular number, be good at quite on skill processing, special ambiguous on control problem logistic control has outstanding advantage 〔 2 〕 . Because nerve network and ambiguous logic have complementary sex, we combine nerve network and ambiguous and logistic look. the input of nerve network faintness of quantity, sendout is changed, use nerve network to realize ambiguous inference, compose builds ambiguous nerve network (FNN) . The difference that measures the input set to ask to machine a result together for treatment is worth, output is worth set to be used at undertaking correction to machining parameter for ambiguous increment. Be based on FNN we put forward grinding parameter decision-making system (1) seeing a picture, this system is the decision-making system that imitates expert thinking way. Systematic composition includes: Expert system (knowledge base, information inference machine, regular library) with treatment parameter increment decision-making FNN. Above all, we get from expert or the person place that have experience experience is worth compositive knowledge base. The initiative state when user general machine tool (workpiece material is qualitative, emery wheel material is qualitative, cooling fluid sort, emery wheel wears away state) ask with treatment (treatment precision, workpiece surface burn is spent, exterior micro-crack degree, exterior surface roughness) input system, inferential opportunity arises the properest grinding parameter according to inputting parameter to be extracted from inside the knowledge base (workpiece slewing speed, emery wheel slewing speed, speed of emery wheel feed) as initiative grinding ginseng data transmission gives grinding the course. Graph parameter of 1 intelligence grinding is decision-making the effect of grinding of whether reasonable will immediate impact that the system chooses as a result of grinding parameter, want to undertake examining to its so, trying to work of sampling observation part needs when grinding level and grinding process end and give correction net FNN1 feedback of sampling observation result namely. The input parameter that the poor value after treatment asks to be compared with treatment as a result is regarded as FNN1, and the correction that the output of FNN1 is regarded as grinding parameter is decision-making to inferential machine the initiative grinding parameter that go out undertakes correction. If correction is bigger (the error with the output larger presence of inferential machine) the sendout that explains the experience value in the knowledge base has an error to need to use FNN1 is amended to its. In grinding process as a result of emery wheel wear away ceaselessly the fall that can cause grinding efficiency, if impropriety amends grinding parameter and the wear extent to emery wheel to undertake compensating meeting,bring about grinding result error to increase, affect the performance characteristics of workpiece. The change that traditional grinding system is pair of emery wheel dimension only undertakes compensating, do not amend grinding parameter however. In this system FNN2 is used at wearing away actually according to emery wheel the state undertakes correction to grinding parameter. After grinding machines an end, emery wheel wears away actually the state is given FNN2 to be the same as emery wheel by feedback initiative state photograph is compared, poor value is delivered as input quantity FNN2. The output value of FNN2 is used at undertaking correction to grinding parameter for the correction of grinding parameter, correction result is passed as actual grinding parameter next time grinding process. Wear away when emery wheel serious when, FNN2 builds announcement system grind emery wheel. Repair in emery wheel grind after finishing, emery wheel will return initiative condition to begin round of new circulation again. Of correction annulus of network of 3 ambiguous nerve build 　 (the structure of network of 1) ambiguous nerve is in grinding process, the grinding parameter that basically affects workpiece surface quality and precision is rate value, accordingly, the rationality that speed parameter chooses is very crucial, because the parameter of initiative and effective rate of systematic output is the experience that extracts from inside the knowledge base by superstratum inference machine,be worth, this experience value is reasonable below current treatment condition it is sealed. So, need to undertake examining to each speed parameter in grinding process, according to examining the result undertakes modulatory to its. In the system FNN1 and FNN2 amend a network for grinding parameter. FNN1 structure sees a picture 2, the structure of FNN2 sees a picture 3. The structure of FNN1 and FNN2 divides it is 5: Layer of modular gelatinization layer, input layer, the concealed intermediate, output layer and turn over faintness to change a layer. Because their input, sendout is different, the effect that has severally is distinct also. Among them, the input quantity of FNN1 is: Exterior burn spends evaluation and evaluation of exterior surface roughness, output is: Correction of correction of correction of workpiece rotate speed, emery wheel rotate speed, speed of emery wheel feed. The input quantity of FNN2 is: Emery wheel wears away state and emery wheel wear away metabolic quantity. Sendout is: Emery wheel of speed of feed of correction of correction of workpiece rotate speed, emery wheel rotate speed, emery wheel is repaired grind correction, emery wheel to repair grind a quantity () of 2 values parameter. Their concealed layer node is counted by set etiquette, the correspondence of node of every concealed layer that we let FNN is medium blurs regular conditional statement, namely: Two in inputting a quantity when node of every concealed layer differs together only subject degree be linked together. Aim to make sure FNN can remember all regulation. Graph block diagram of structure of 2 　 FNN1 pursues 　 of block diagram of structure of 3 　 FNN2 (of parameter of 2) input, output choose the poor value that reachs faintness to differentiate to machine a result actually to machine a requirement together to be given FNN1 by the input, via FNN1 is being served as after faintness is changed faintness changes the input of the layer to measure. Include among them: Workpiece surface burn spends evaluation of evaluation, exterior surface roughness, micro-crack degree evaluation. The substance that produces as a result of grinding burn is: Grind when grinding is machined the cutting since bead, score and delimit brush action. Great majority wears bead is the horn before losing undertakes cutting, fall in taller grinding speed condition, make exterior layer has very high temperature, produce grinding burn. The principle of grinding burn is grinding temperature tall, burn and temperature have very close relationship. Accordingly everything affects the factor of temperature, right on certain level burn is influential: (Vf of speed of feed of 1) emery wheel. Increase when Vf, burn rate increases. (Vs of speed of 2) emery wheel. Vs increases, burn rate increases. (Vw of 3) workpiece slewing speed. Vw increases, burn degree reduces 〔 1 〕 . The account that grinding crackle produces also is hot, and the factor that affects grinding crackle has: (V of speed of feed of 1) emery wheel. V adds a general to bring about crackle to increase; (2) emery wheel answers rotate speed Vs. Vs decreases young general to bring about crackle to decrease; (Vw of 3) workpiece slewing speed. Vw increases will bring about micro-crack to reduce 〔 1 〕 . Because burn and the mechanism that micro-crack produces and influencing factor are same, because this FNN1 inputs parameter,micro-crack degree incorporates, spend by burn replace. Differentiate its faintness 4 grade; L(is serious) , M(is medium) , S(is slight) , ZE(does not have burn) . Grade of exterior surface roughness is surely: PL(big) , in the middle of PM() , PS(small) , ZE(0) , NS(is lost small) , NM(is lost in) , NL(loses big) . The sendout of FNN1 is differentiated by faintness 7 class: PL, PM, PS, ZE, NS, NM, NL. FNN2 main basis every time the emery wheel after grinding process ends wears away the state wears away namely change is decision-making the correction that gives grinding parameter. Wear away to output 2 values parameter to the network when certain level when emery wheel (emery wheel is repaired worry a situation) announcement system undertakes repairing grinding to emery wheel. Emery wheel is repaired grind after ending, the emery wheel in inputting a quantity wears away the state regains initiative position. Emery wheel wears away the state is divided to be 6 class by faintness: PL(is serious) , PSL(is more serious) , PM(is medium) , PLS(is slighter) , PS(is slight) , ZE(is not had wear away) . Emery wheel wears away metabolic quantity is differentiated to be 4 class by faintness: PL, PM, PS, ZE. The input parameter of FNN2 can be in actual grinding process through detecting the relation of radial feed speed and radial grinding force gets secondhand. Repair in the in outputting parameter emery wheel of FNN2 grind for binary -state variable (exist only 0 with two kinds of 1 circumstances) 0 delegates are not repaired grind emery wheel, 1 delegate is repaired grind emery wheel. The others sendout is differentiated to be 7 class by faintness: PL, PM, PS, ZE, NS, NM, NL. The modular gelatinization of parameter of the FNN in the system and turn over faintness to change choice triangle as subject spend function, the gelatinization returning a model that outputs parameter uses method of centre of gravity, the abscissa of result of will ambiguous inference and center of ambiguous subclass area serves as the sendout of corrective quantity. Its computational formula is: (Here: Xb of ≤ of Xa ≤ X, s(xs of centre of gravity, ys) to subject spend function to the left when bound and the circumstance in right bound, the semmetry that has its fiction is patulous (4) seeing a picture can get the controller inside whole width limits exports corrective quantity. Graph 　 of law of 4 patulous centre of gravity (the learns algorithmic 　 FNN nerve of network of 3) ambiguous nerve yuan function uses Sigmoid function: F(x)=1/(1+e-x) 　 (2) 　 algorithm is used the most mature now, the BP with the widest application is algorithmic. Adopted patulous Delta regulation, inspect the sum of squares of accident: (3) is learning the with reducing E quickly as far as possible means in the process to undertake. General it is relied on whether to search along gradient direction in power worth space. Additional, assuming power worth is adjusted is the condition that trains to be carried out to order centrally in training limits next undertaking. Accordingly, the error for whole training part is the smallest, due entire network is average E of accident sum of squares always the smallest. (A number with the right training that 4) (P centers to train example) the formula that adjusts to power worth is: (5) can roll out to outputting a layer: (6) to concealed layer: (7) 　 (8) 　 　 (I is limits of input layer suffix, j is limits of concealed layer suffix, k is limits of output layer suffix) 　 　 (the improvement with 4) algorithmic study -- change the basis that worth of the power in algorithm of BP of 　 of measure gradient algorithm adjusts is output error to be worth to authority slant guide (gradient) , but the way that gradient can decide to adjust only, and adjustment measure will decide convergent rate. In deciding measure BP network to train a process, often appear to error of the output when certain level in network astringent substantially the circumstance of concussion, make finally network loss of accuracy. Appear because the study condition of the network is chief,the reason of this kind of circumstance is it is constant, when the network authority is worth convergence to arrive when certain level, if press original study pace to grow,undertake adjustment be meetinged occurrence exceeding tone, bring about an error finally to increase. And be in what the network trains each authority in the process to be worth gradient drop speed is different, what if grow to undertake adjustment be meetinged to its with same situation,create a network is not stable, appear thereby concussion and not convergent circumstance. In the light of a variety of defect that decide a pace to grow BP algorithm we put forward to change measure gradient is algorithmic. Study condition is chief adjust formula to be: (=sign(Tp-1 of 9) λ , tp) 　 (10) among them: η grows to study a situation (η ＜ 1.

6) ; λ grows study factor for the pace; T is gradient value; Tp-1 is the gradient value when P-1 second iteration; Tp is the gradient value when P second iteration. (A) changes 　 of curve of measure network astringent (two kinds of 5 algorithm are in graph of astringent of network of B) calm measure the convergent curve when network application is changing in measure algorithm, study pace grows to be measured for adventitious. If two iteration are same variable gradient value explains study pace has grown instead big, had appeared to exceed tone, need to reduce its. If twice the gradient of iteration is the same as to, can increase study pace to grow to raise convergent rate thereby. And change the study pace in measure algorithm grows is according to the network the gradient of each weight undertakes be adjustmented alone, adjust in this kind the gradient that can make all network advantageous position is worth below means can achieve minimum to raise network precision thereby. Check the number of division of layer of number, concealed, output nods in same input red-letter day the from the network output error below the condition such as frequency of collect of example of worth of node number, initiative power, training, training and astringent curve (can see in 6) seeing a picture: When the output that measure network changes when training frequency is achieved 600 times the error is achieved 0.

00006, and the output error that decides measure network is 0.

02. And can see from inside the graph change the error curve of measure network is in convergent condition from beginning to end, and the scale when training a frequency to arrive 320 times spends BP algorithm the error curve of nerve network already appeared the condition that concussion and error add. Still can see be sent on a diplomatic mission use from inside the graph change error of the output when the ANN of measure algorithm is training already was achieved 0.

008 be equal to the output error when using scale to spend BP algorithm ANN to train 200 times to be worth. Graph the convergent curve that 6 　 FNN approachs ambiguous regulation 4 emulate reach test result to will change measure gradient algorithm is applied in intelligence grinding parameter in decision-making system, undertake to grinding parameter decision-making. The relation that inspects workpiece grinding result and grinding parameter is expressed as follows, the relation that is the same as grinding parameter as a result because of grinding treatment is quite complex, and the FNN in this system outputs a network more for little input, so corresponding and same group is inputted but correspondence is solved many groups. According to expert experience attainable ambiguous and regular form is like: (Br=PL Then Vw=NS And Vf=PL And Vs=NS of Δ of Ra=PL And of 1)if Δ . (35 regulation such as Br=PSL Then Vw=PS And Vs=PL And Vf=NS of Δ of Ra=PL And of 2)if Δ . Emery wheel wears away state and emery wheel wear away ambiguous and regular form is like metabolic quantity correspondence: (Emery wheel of E=PL Then Vs=PL And Vw=NL And Vf=PL And Rs(of Δ of 1)if E=PL And is repaired grind) =1. (Br=NS Then Vs=PS And Vw=PS And Vf=PS And Rs=0 of Δ of 2)if E=NS And 24. Express burn of Ra of surface roughness Δ to spend 　 of ↑ of ↑ of ↓ of ↑ of Vs of rotate speed of emery wheel of ↑ of ↑ of ↑ of ↑ of Vf of speed of feed of ↓ of ↓ of ↑ of ↑ of Vw of rotate speed of workpiece of Cr of Δ Br micro-crack Δ to use ambiguous regulation to get 100 groups of data, these 100 groups of data are set training example undertakes training to FNN network. The convergent curve that in iteration the system on 20000 foundation outputs error and FNN net sees a picture 7. In network training course, the output error that when iteration is achieved 20000 times only group input measures correspondence is achieved 0.

9% . In finite second output of the system on iteration frequency achieved expectant error demand, convergent in order, precision is relatively ideal. Good training decision-making system is in 3MZ, test and verify undertakes on 1310 bearing cylindrical grinder, use grinding clearance to be opposite next time grinding parameter undertakes decision-making, obtained better decision-making result. The grinding parameter that proves the article puts forward is decision-making the system is feasible. 5 conclusion 　 (parameter of the intelligence that 1) is based on double FNN grinding is decision-making system. Expert system is used to undertake to grinding parameter in decision-making system decision-making, it is the knowledge base set in expert system modifiable, decision-making system but as grinding process undertake the knowledge base undertakes abundance be been opposite ceaselessly. In actual grinding (try grind) actual grinding result can get through workpiece of fixed sampling observation in the process and undertake correction to the knowledge base by grinding parameter correction FNN. (2) realizes ambiguous reasoning process with ambiguous nerve network, use double FNN to undertake correction to grinding parameter. Among them, FNN1 basis machines a result to undertake correction to grinding parameter and knowledge base, FNN2 wears away actually according to emery wheel the state undertakes correction to grinding parameter in grinding process clearance, assure the rationality of decision-making parameter thereby. (3) changes measure gradient is algorithmic. This algorithm is used detect the directional change of variable gradient undertakes be adjustmented accordingly to studying measure, make network astringent speed is accelerated apparently, convergent precision increases, reduced the occurrence of concussion and circumstance of disperse the internal heat with sudorifics. Get applied in decision-making system and obtained relatively ideal result. This system is in theoretic have feasibility, aborning has higher economic value, the intelligence that makes a process in domain of advanced production technology especially is decision-making believe to be able to have positive effect with control. CNC Milling CNC Machining