# Those who be based on the axes contrail of nerve network is automatic identify

- Time:
- Click:403

Of 1BP algorithm roughly measure (1) initialization. Worth of advantageous position of random initialization network and node threshold values. (2) provide training example. The input learns example X0, ... , xn and pedagogic signal T0, ... , tm, and study rate η , error allows to be worth E. (3) ordinal computation each node output. (4) computation trains an error. (5) worth of correction network advantageous position and threshold values. (6) whether is overall error less than judgement network the error allows to be worth E, be less than stop, turn otherwise time (3) . 2.

The improvement of 2BP algorithm the processing that the key of BP algorithm depends on pair of study rate η , η is too big, oscillation is caused in can learning a process; η is too small, make study rate too slow. Consequently, better idea is η value is changed according to case trends in learning a process, for this, we are taken 2.

3 best and implicit the affirmatory nerve network of layer division check the number is implicit L of layer division check the number is a very important segment surely really. If L is too small, criterion the network cannot undertake learn and discriminating at all; If L is too big, criterion the ability of self-study be used to of the network is very poor, cannot identify the pattern that have not has seen before. Chose a best L to be worth only, ability makes the network has more powerful rash club quite gender, fight noise ability and ability of self-study be used to. General and best implicit L of layer division check the number can decide by the experience formula below: In type, m and N are the input of the network and output node number respectively; C is interpose at 1 ～ the constant of 10. Because middling of this experience formula counts C to decide hard, accordingly, in applying actually, it is used rarely. We put forward a kind here relatively effective affirmatory and best implicit the experience method of L of layer division check the number. Your integer is after the experimentation with much course, discovery is optimal and implicit L of layer division check the number not only outside be being concerned with M, N, the number that still reachs training example with the linear dependency between training example is concerned. Nevertheless, to system of average network of pattern recognition nerve, best and implicit L of layer division check the number can fall into interval commonly [2, k] or [K, 2K+3] (this besides see a table outside 1 medium test result, OK still from document 4 medium experimental result data try test and verify) . So, be in affirmatory and best implicit when L of layer division check the number, can take a cost from K place above all, next, can use " law of trial and error " from K two side take a cost respectively, compare the convergent rate of the network, decide the interval that L is in, find out optimal L value thereby. Watch 1 it is the condition that inputs node and 4 output node in 7 next using " law of trial and error " reach. by the watch 1 visible, to same the network, best and implicit L of layer division check the number is as change of training example numerary change, but the value of L is in roughly however interval [2, k] or [K, 2K+3] inside change. Consequently, to system of average network of pattern recognition nerve, be in affirmatory and best implicit when L of layer division check the number, this methodological can yet be regarded as a kind of effective referenced method. Those who need a specification is, best and implicit layer division check the number involves a lot of other factors surely really (like the linear dependency between example, the complexity of example) , it is a complex process, the method that still does not have a kind of effective at present consequently can be abided by. This kind of means that the author offers, remain formula of a kind of experience, still do not have versatility in applying actually, but have referenced value only. The identifying of 3 axes contrail we are right " banana form " , " elliptical " , " inside 8 " and " outside 8 " contrail of 4 kinds of axes undertakes discriminating, they part correspondence carries at nerve network output 1, 2, 3, 4 node. Nerve network is 3 structures, 7 input node and 4 output node, implicit layer division check the number is taken 12. 3.

The 1 measure that identifies axes contrail with nerve network (the graphical data of contrail of 1) countershaft heart presses type respectively (6) ～ (12) computation its fixity quadrature ～ . (2) inputs computation as a result go learning in nerve network. The ～ that gets consideration regards the input of nerve network as example, let its press type (13) ～ (16) undertake study. Some amounts are too small in the ～ that gets as a result of computation, to raise the area graduation between example, in order to facilitate the network learns, we undertook the amplification of certain multiple to its. (3) is treated measure a graph to undertake classified identifying. After network learning process is finished, can serve as classification implement undertake pattern recognition. When identifying, press measure first (1) treat identify axes contrail to beg ～ , regard nerve as the input of the network computation as a result next, let its identify. 3.

The identifying of 2 axes contrail passes the equation below, we can win needs axes contrail figure. And, the data middling that considers actual axes contrail is mixed have noise, for this, we are in when making axes contrail figure, factitious also ground joined a few noise. In type, w is horny frequency, a1, a2, a3 and B1, b2, b3 is X(t) and Y(t) respectively 1, 2, heft of 3 times frequency; α 1, α 2, α 3 with β 1, β 2, β 3 it is corresponding starting phase respectively. Through changing these 12 parameter, can win the axes contrail figure that we need. The figure of axes contrail example that article place uses and wait for identify a graph to have banana look, elliptic, inside 8 glyphs, outside 8 glyphs (graph summary) . Watch 2 in identify a result to wait for those who identify a figure 10 kinds. The output result that 4 output that expresses 2 medium data to identify network of the nerve in the process for each carry, be boundless key link. The numerical value that a certain output in identifying a process some carries is the greatest, identify the figure that is delegate of this port place namely as a result this. With (A) group for exemple, the numerical value of the 1st port in the group is the greatest, so this second identifies the figure that eventuate port represents 1 times. As a result of port the graph of a delegate is " banana form " , so final identifying is as a result " banana form " . By the watch 2 visible, identify correct rate for 100% . When the form that treats knowledge graph and example are adjacent, the degree of confidence that identifies a result is bigger, when the noise in awaiting knowledge graph data increases, degree of confidence drops subsequently. Nevertheless, as a whole, identifying a result still is satisfactory, particularly right ellipse and banana form, ellipse and inside the graph with these two groups of 8 similar older rate, nerve network showed good identifying ability. This kind of stronger rash club sex of nerve network, it is one of main reasons that its use extensively at pattern recognition. 4 problems of discuss (1) is in the process of computational ～ , as a result of certain a numerical value is very little, identify to facilitate network study is mixed, multiply its with a proper coefficient. The fact makes clear, reasonable choice coefficient, can more effective raise the area graduation between example. The principle that chooses coefficient is, the gap that reduces example data to go up in quantitative class as far as possible, to ～ medium lesser undertake enlarge, make these 7 numerical value are in as far as possible same a quantitative class limits inside. (The article is in 2) when all sorts of typical figure that produce axes contrail, included only 1, 2, heft of 3 times frequency, and the bases in the graphical data of actual axes contrail besides these 3 kinds of heft, still include a few other part, if assign times frequency,wait. Accordingly, when applying actually, but according to the circumstance, a few other weight are added when all sorts of typical form that synthesize axes contrail. (3) nerve network has stronger fault tolerance capacity, should mix in the graphical data of axes contrail have a few when noise, not influence identifying result, become when noise is heavier, identify a result with respect to can serious effect, for this, can undertake to real data filter wave is handled first, hand in again next identify by nerve network. 5 epilogue are had as a result of ～ translation, scale and rotate fixity, can the diagnostic parameter of the graphical figure of token axes contrail, and data less, consequently special agree with to the compose of nerve network is built and learn, can develop its adequately ability of stronger fault tolerance, get used to the advantage such as ability of ability, self-study be used to and collateral processing capability oneself. The author used nerve network to undertake emulation discriminating to the graphical figure of 4 kinds of relatively typical axes contrail, the result makes clear, this method is effective, feasible. CNC Milling