Active pilot studies network of nerve of vibration of silk of machine of cut of electric spark line
- Time:
- Click:42
The Control System Is Aiming To The Active Vibration Control Of Axially Moving String Of Civil Wire-cutting Electrical Discharge Machine.
The Experiments On CivilDK3220B Wire-cutting Electrical Discharge Machine Have Been Executed.
The Results Show That The Control System Is Feasible And Can Improve The Cutting Accuracy Efficiently.
1 preface of ▲ the typical case that vibration of silk of machine of cut of electric spark line is problem of vibration of axial motion bowstring, the axial motion bowstring that two end prop up (the basic and mechanical model that graph 1) is afore-mentioned systems. The axial motion of bowstring makes the system has apparent top effect (Groscopic Effect) , incorporate is athletic bowstring each rank the weaveform that mode forms a stability no longer, mode does not have symmetry, what transmission wave has a space to become independent is phasic wait [1, 2] . Top effect disappears for zero hour in the athletic speed of bowstring only. As a result of the top effect of athletic bowstring and distributinging parameter character, stable control is reasonable theory go up and is difficulty on implementation, accordingly, inchoate research with a view to increases systematic damp, raise systematic stiffness, bring to bear on the passive control means such as rigid tie. To suit more complex environment is mixed the condition changes when contented system, raise control precision, use active pilot method to be in the way of a key that makes research gradually in recent years. Graph the characteristic that 1 two end prop up axial motion to be aimed at vibration of axial motion bowstring, if change when the top effect, distributinging parameter, difficulty that build a model, parameter,wait, using artificial nerve network to control is a kind of good method. Because artificial nerve network uses means of distributed information processing, have fault tolerance sex, get used to online study to mix oneself wait for a characteristic without the model, already caused wide attention in dominating a field. The article is used change measure conjugate gradient is algorithmic (SCG) and forward the network serves as supervisory layer, constructed statified control system, the electrode silk vibration that takes chance of silk thread cut to homebred high speed undertook be controllinged actively, achieved favorable result. Network of nerve of 2 SCG is statified control a system 2.
1 statified control construction of system is statified the basic idea that controls a system is: The self-induced controller that comprises with imitate circuit serves as rock-bottom controller, it is the executive part of whole control segment, with SCG forward network composition supervises a controller, the parameter that the system changes when dogging changes, adjust the gain of rock-bottom controller. Block diagram of whole control system is shown 2 times like the graph. Graph vibration of bowstring of motion of supervisory layer of network of nerve of 2 SCG controls network of nerve of systematic block diagram actively to supervise a layer by two develop Chinese hackberrya the SCG with identical structure is forward network NN1 and NN2 composition. Among them NN1 is used at building to be accused the system differentiates with what survey a system model, NN2 uses as controller, its output the gain of controller of G(k) adjustable ground floor. Self-induction is made move implement use the pottery and porcelain that suppress report (sensor of PZT) self-induced parity / make move implement. It actually by a flexibility radical piece the layer suppressing report that adds composition of two pottery and porcelain that suppress report shuts bridge, the sensor that finish is mixed move implement function. 2.
2 SCG is forward nerve network control is politic and forward the method of power worth study of the network uses BP algorithm normally, its are main defect is convergent speed slower. In its numerous improvement in the method, change measure conjugate gradient is algorithmic (SCG) is a more outstanding kind. Those who use SCG algorithm is forward the network calls SCG network. SCG algorithm is when K of long α of computational search pace, used extraordinary skill, it not only can handy and accurately computational α K, still consider Hessian matrix qualitative. Graph 3 for SCG network sketch map of L layer. Set L layer to have Nl nerve yuan, o(l)j(j=1... Nl) is L layer the output of J unit, i(l)j is L layer the input of J unit increases authority cost, w(l-1)ij is L-1 layer I is unit to L layer the connection of J unit counterpoises, criterion: (Pattern of study of 1) is right input X=(x0, x1, ... Xi, ... Xn) and expectation output Yd=(yd0, yd1, ... Ydi, ... Ydm) composition, corresponding X of a certain input, the network is answered for Y=(y0, y1, ... Yi, ... Ym) . Mode of a certain study is opposite correspondence, error function is: ( of 2) (3) basis slants the operation regulation of derivative (4) pursues right of the type on sketch map of network of layer of 3 L is the 2nd times, the first definition is generalization error τ (L)j: (5) has: (The type on 6) makes clear, the error gradient element of corresponding W(l-1)ij is equal to L layer the generalization error of J unit multiplies the output with an unit before I. The 2nd of generalization error can get directly: (7) divides kind of case the first times: (1) to outputting a layer: (8) inputs the generalization error of the layer to be: of of of τ J=-(ydj-yj)yj(1-yj) (9) (2) to concealed layer: ( of of 10) (the generalization error of layer of 11) concealed is: (12) is visible, of generalization error beg taking a method is to the input from output layer layer recurrence is gotten. After the error gradient that gets each mode correspondence, undertake to them cumulative, can get the global error gradient of the network: (The initialization of algorithm of SCG of of 13) works, besides to each network authority is worth an ode with (- 1, outside the random number between 1) , still two parameter need initialization: σ and measure factor λ . λ has an elevator to make in algorithm, reason is initiative the particular concern extraction a cost of the value is not big, should be less than those who be equal to 10-6 to only real can. σ once after initialization, keep changeless. Theoretic tell, the initiative value of σ will affect algorithmic performance, from algorithmic precision consideration, σ should Yue Xiaoyue is good. Numeric experiment makes clear, σ wants only enough small (σ ≤ 10-4) , small to the influence of algorithmic function Hu Jiwei. This also reflected the good point with good stability of SCG algorithmic numerical value. According to afore-mentioned theoretical analysises and derivation process, used Borland C++ to weave control software. Length of be confined to, do not make farther introduction. Experiment of cut of 3 octagonal workpiece the SCG network that to examine the article offers is active pilot effect, actual cutting test had to octagonal workpiece on machine of DK3220B line cut. This machine tool has canister of silk of a lay aside only, rate walking along silk is not adjustable. Experimental system is shown 4 times like the graph. Graph 4 brace up laboratory of system of active control test is with electrode silk diameter 0.
12mm, workbench feed speed is 2mm/min, material of treatment spare parts is T8 tool steel character, the surface quenchs processing, test specimen ply is 6mm, the tagging of outline dimension is 10mm × 10mm. Control brought to bear on in X direction only in the experiment, converter of A/D, D/A all uses 12. Get the following result through the experiment: (After 1) carries out active control, cut seam width the biggest dropped 0.
155mm, dropped on average 0.
074mm, drop extent is 20% . The vibration that makes clear electrode silk got effective inhibition; (After 2) carries out active control, average size deviation by 0 original.
179mm is reduced for 0.
142mm, decreased 0.
037mm, the dimension precision that cut machines also has different rate rise; (After 3) carries out active control, cut surface surface roughness is the biggest dropped 1.
28 μ M, drop on average for 0.
68 μ M, cut surface surface roughness got reducing; (4) is in bring to bear on on pilot X direction, cut seam width and exterior surface roughness to drop the biggest, visible control effect is more apparent, be in at the same time mix two without pilot Y direction inclined the control effect that has different rate up, because machine the oscillatory contrail of silk of the electrode in the process to show irregular elliptic reason,this is. The article put forward 4 conclusion to be based on the strategy of statified and oscillatory active control of nerve network, comprise rock-bottom self-induced controller by imitate circuit namely, by SCG nerve network composition supervises a controller, those who pass supervisory layer controller is online monitor the work that will adjust rock-bottom self-induced controller. Such control system can be answered quickly already need in order to satisfy real time pilot, have again get used to ability oneself. Further, the axial motion spring that uses the article controls a technology actively to take silk thread cut in homebred high speed silk undertook on machine brace up research of active control experiment, the experiment makes clear, active control reduces the vibration of silk of the electrode in line cut treatment effectively, raise line cut to machine quality, have stronger applied perspective. CNC Milling