Research on Milling Cutter Wear Monitoring Technology Based on Radial Basis Function Neural Network

Neural network as a method to monitor the status of the system, has a strong ability of self-adaptation, fault tolerance, robustness and a high degree of parallel computing capabilities. These features make the method widely used in signal processing (including pattern recognition and intelligent control) and have achieved good control effects. Radial basis function (RBF) neural network was used for state recognition in the study. Radial basis functions use Gaussian functions, and the center and variance of Gaussian functions are obtained by fuzzy-C fuzzy clustering. 1 Feature generation After studying the collected signals, it was found that the variance, maximum spectral peak (200-600 kHz), one-step autoregression coefficient and rms value are the most sensitive to tool wear. Variance s2=(S(Xi-X)(Xi-X))/(N-1) Maximum spectral peak 200-600 kHz One-step auto regression coefficient P=R1/R0, R1=SXIXI-1/N, R0=Autocorrelation The root mean square of the coefficient XRMS = (SXI2/N) 1⁄2 2 Radial basis function neural network design Radial basis function neural network structure The radial basis function network has only one hidden layer, the output unit is the linear summation unit, ie the output Is the weighted sum of the output of each hidden unit. The function of the hidden unit uses the radial basis function (RBF). The weight input to the hidden unit is fixed to 1, and only the weight wj=(j=1,2,...,n) between the hidden unit and the output unit is adjustable. . The most commonly used radial basis function is a Gaussian function. There are two adjustable parameters: the center position and the variance (function width parameter). When using these functions, there are three groups of adjustable parameters for the entire network: Base function's center position, variance, and weight of the output unit.

2
Fig. 1 Radial basis function (RBF) network structure

Design of Radial Basis Function Neural Network The network structure of Radial Basis Function uses the network structure of 4 nodes of the input layer, 10 nodes of the hidden layer, and 1 node of the output layer. Using the four eigenvalues ​​(root mean square, variance, one-step autocorrelation coefficient, and maximum spectral peak value of 200-600 kHz) of the collected data (for a total of N samples) as the radial basis function input, the output layer node is a milling cutter The wear value. The fuzzy clustering fuzzy C-division method is used to transform the sample components into M groups. The cluster center of each group is used as the center of each basis function, and the basis function is Gaussian function. Then the variance of each center is calculated by the formula s=d/2M. Finally, each weight is calculated by the least squares method. 3 Experimental results The following processing conditions were used in the experiment: 1Ø4 high speed steel milling cutter, spindle speed 1120r/min, feed rate 15.68mm/min, depth of cut 5mm, workpiece using nodular cast iron (24HRC): ~ 03 high speed steel milling cutter, Spindle speed 900r/min, tool feed speed 12.6mm/min, depth of cut 3mm, workpiece using 45 # quenched and tempered steel (22HRC). The tool wear measuring tool is JGX-1 miniature tool microscope with an accuracy of 0.01mm. The four eigenvalues ​​(root mean square value, variance, one-step autocorrelation coefficient, and maximum peak value (200-600 kHz)) of the collected 72 sets of data (36 sets for each cutting condition) were used as training samples. The design method of the radial basis function neural network obtains the required radial basis function neural network: then another set of 20 data (obtained under cutting condition 1) is used as a test sample. Figure 2 is a wear curve for 20 sets of data used for testing (10s for the test sample interval). In the picture The curve is the actually measured tool wear value, and the Λ curve is the value obtained by the radial basis function neural network.

1
Figure 2 Neural Network Test Curve

2
â—‹ Data Points: - A=T:... Best Approximation Curve Figure 3 Linear Regression

Table RBF network and BP network performance comparison training method Network type training time (s) Number of training steps Fuzzy-C clustering algorithm RBF network 1.20 7 Improved BP algorithm BP network 130.42 2016 Figure 3 is the radial basis function network output and the actual tool Wear value linear regression. In the figure, A is the actual wear value of the tool: T is the output of the RBF network. From the right table, the RBF network training speed is much higher than the BP network of the improved algorithm. 4 Conclusions Using the rms value, variance, one-step autocorrelation coefficient and maximum peak value (200-600kHz) of the acquired signal as feature quantities, the tool wear value can be estimated effectively. In real-time, on the microcomputer with a main frequency of 166MHz, the state recognition (35ms) from the data processing to the radial basis function neural network takes about 200ms in total. In practical applications, we collect data every time. The time interval can be set to 1s, 2s, etc. Therefore, this study can meet the real-time requirements.