Summary: The milling cutter that puts forward a kind to be based on network of nerve of radial radical function wears away monitoring method. The output of network of nerve of radial radical function is cutting tool tatty particular value, such is helpful for that wear away to cutting tool undertake all sorts of real time are compensated. The experiment makes clear, use network of nerve of radial radical function to undertake condition identifying can is opposite small-sized establish milling cutter wear away undertake monitoring, can obtain favorable result. The network of BP of training rate excel that proves RBF network at the same time. Nerve network regards supervisory system as a kind of method of condition identifying, have the very strong collateral operation capacity that gets used to sex of ability, fault tolerance ability, rash club and height oneself. These characteristics make this method is handled in signal (include pattern recognition and intelligent control) the domain received general application, obtained favorable control result. Radial radical function is used in research (RBF) nerve network makes condition identifying. Radial radical function uses gauss function, the center of gauss function and variance get together through Fuzzy-C faintness kind will obtain. 1 feature generates the signal that collects through be opposite to undertake study to discover, peak value of the variance, biggest music score (200 ～ 600kHz) , one pace wears away to cutting tool from population and mean effective value the most sensitive. Variance S2=(S(Xi-X)(Xi-X))/(N-1) is the biggest chart peak value 200 ～ 600kHz one pace from population P=R1/R0, r1=SXIXI-1/N, XRMS=(SXI2/N)½2 of mean effective value of R0= auto-correlation coefficient is radial base network of radial radical function has the designs network of nerve of radial radical function structure of function nerve network only layer of a concealed, exporting unit is unit of linear sue for peace, output be output of each concealed unit to add namely authority and. The action function of concealed unit uses radial radical function (RBF) , the power worth that inputs concealed unit is secured for 1, only concealed unit is worth Wj=(j=1 to the authority between output unit, 2, ... , n) is adjustable. The most commonly used radial base function is gauss function, its adjustable parameter has two, namely central position and variance (the width parameter) of function, when using this kind of function of whole network adjustable parameter has 3 groups, it is respectively: Each base the central position of function, variance and the power worth that export unit. Graph 1 radial base function (RBF) network structure is radial base the designs radial radical function network structure of function nerve network uses input layer 4 node, concealed layer 10 node, output a network structure of a node. Use collect data (set in all N example) 4 eigenvalue (the input that 200 ～ 600kHz) regards coefficient of auto-correlation of mean effective value, variance, one pace and peak value of the biggest music score as radial radical function, output what layer node is milling cutter to wear away value. Get together with faintness kind of method that Fuzzy C- differentiates become example constituent M group, with every groups get together kind of center as individually base function center, base function uses gauss function, again with formulary S=d/2M cipher out the variance of each center. Through cipher out of least square method each authority are worth finally. 3 tests result adopts requirement of the following treatment in the experiment: Milling cutter of ① Ø4 high-speed steel, 1120r/min of main shaft rotate speed, feed speed 15.
68mm/min, cut deep 5mm, workpiece uses nodular cast iron (24HRC) : 03 high-speed steel of ～ milling cutter, 900r/min of main shaft rotate speed, speed of tool carrier feed 12.
6mm/min, cut deep 3mm, workpiece uses 45# to move qualitative steel (22HRC) . Cutting tool tatty measures a tool to be microscope of JGX-1 small-sized tool, precision is 0.
01mm. Use 72 groups of collection data (each collect 36 groups below condition of two kinds of cutting) 4 eigenvalue (coefficient of auto-correlation of mean effective value, variance, one pace and peak value of the biggest music score (600kHz)) regards 200 ～ as training example, according to in front narrates radial base the design method of function nerve network, the radial that gets needing base function nerve network: Next with additionally 20 groups of data (obtain below cutting condition ① ) regard a test as example. Graph the 2 wear curve that are 20 groups of data that are used at the test (the time-interval that checks sample is 10s) . The curve in the graph is the cutting tool that measures actually wears away value, Λ curve is the cost that passes network of nerve of radial radical function to get. Graph dot of data of ○ of curve of test of 2 nerve network: -- A=T: ... optimal approach graph network of RBF of chart of regression of 3 lines form and BP network function train methodological network type to train time quite (Fuzzy-C of number of S) training pace gets together kind of algorithmic RBF network 1.
207 improve network of BP algorithmic BP 130.
422016 graphs 3 be output of network of radial radical function and cutting tool is actual wear away the form that be worth a line returns to a graph. The A in this graph is cutting tool is actual wear away value: T is the output of RBF network. By right watch knowable, the training rate far outclass of RBF network improves algorithmic BP network. 4 conclusion use coefficient of auto-correlation of the mean effective value that collects signal, variance, one pace and peak value of the biggest music score (200 ～ 600kHz) is measured for the feature, effectively estimation gives cutting tool to wear away value. In respect of real time sex, be in advocate on the personal computer that frequency is 166MHz, arrive from data processing the condition of network of nerve of radial radical function identifies (35ms) , place is 200ms about in all with time, and in applying actually, the time-interval that we collect data every time can be set wait for 1s, 2s, accordingly, this research can satisfy requirement of real time sex. CNC Milling