Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 UDK - UDC 519.65:621.95 Izvirni znanstveni članek - Original scientific paper (1.01) Spremljanje obrabe vijačnega svedra (S390) z uporabo nevronskih mrež Using Neural Networks to Follow the Wear of a S390 Twist Drill Zdravko Krivokapič1 - Vukasin Zogovič1 - Obrad Spaič2 (1University of Montenegro, Montenegro; 2 Industrija alata, Bosnia and Herzegovina) Prispevek opisuje uporabo nevronskih mrež za zbiranje informacij ter postopkovnih parametrov odrezovalnega procesa (hitrost, podajanje in premer). Opazovan je vpliv dveh načinov ostrenja pri različnih časih obdelave na parametre obrabe. Material vijačnih svedrov (S390) je pridobljen s tehnologijo sintranja. Postopek za modeliranje pojava je učenje nevronske mreže z uporabo eksperimentalno pridobljenih podatkov. Nevronske mreže delujejo na načelu algoritma vzvratnega razširjanja napake. Nevronske mreže so učene s testnimi oblikami (posredno). Prikazani so dobljeni rezultati. © 2006 Strojniški vestnik. Vse pravice pridržane. (Ključne besede: mreže nevronalne, vrtanje, svedri vijačni, procesi obrabe) This paper deals with the use of neural networks for the integration of information as well as the parameters of the cutting process (speed, feed and diameter). Two sharpening methods and different working times related to the wear parameters are studied. The material used for the twist drill (S390) is obtained with power technology. Experimental results are used to train the neural networks, as one approach to the modeling of this process. The back-propagation algorithm is used as a model for neural networks. The neural networks with test shapes are trained (offline). The obtained results are presented. © 2006 Journal of Mechanical Engineering. All rights reserved. (Keywords: neural networks, drilling, twist drill, wear processes) 0 INTRODUCTION A great deal of attention has recently been paid by researchers to the application of neural networks in following the wear process of drilling ([1] to [5]). The application of neural networks for the information integration of the cutting-process parameters, the drill-sharpening method and the operation time in relation to the tool-wear parameters is treated in this paper. The experimental results obtained in the laboratory of the School of Mechanical Engineering were used for the neural-network training. The following train of effects was studied in particular: knowledge degree, the number of layers and the number of neurons in a discrete layer. The transformation function in f-number in the function of the neurons in a layer was analyzed as well. 1 NEURAL NETWORKS There are quite a number of neural network architectures. The back-propagation neural network model is among the most widely used in the state-following process. In this model, data are processed from the input to the output layer, whereas knowledge is performed by the algorithm of minimizing the square error by backward movement. The knowledge is controlled. It is the simplicity of the algorithm for this model that makes this architecture an attractive one, and the reason why it is analyzed in this paper. Fig. 1 shows the architecture of neural networks specific to the back-propagation model. The basic structure of this network is made up of three layers: input, secret and output. The data from the external environment are taken over by the input layer, whereas the output to the environment is generated by the output layer. It is the secret layer, one or more of them, that makes the transformation by extracting the input data, by means of the chosen transformed function (f). A procedure for determining the output values resulting from the observed neurons is shown in Fig. 2. 437 Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 Input fro Input from layer X1 X2 XN Secret layer Output layer Output Carrying form form Fig. 1. Neural network architecture X1 X2 Xi XN. Neuron j in a secret knot Three basic transfer functions Name Sig Tgh Sin Function transfer f(z) f(z) f(z)= 1^ e f(z)= f(z)A1 ZTz f(z)=sin(z) Fig. 2. Neural network functioning algorithm 2 EXPERIMENT PROJECTION The experiment was carried out according to the Box– Wilson method of complex processes modeling using a complete multifactorial first-order plan with repeating in the central point of the multifactorial orthogonal plan. The characteristics employed in the experiment are listed in Table 1. The scheme for carrying out the experiment is illustrated in Figure 3. On the basis of the relation between the flank wear and the drilling duration (working time - stability) of the twist drill, for different cutting regimes, as illustrated in Figure 4 and Figure 5, it can be seen that it is very difficult to establish a relationship between the wearing belt and the drilling duration, so this paper presents an approach based on the employment of neural networks, as one of the possible approaches for establishing this relationship. Table 1. 3 ANALYSIS OF THE RESULTS Five parameters were used to train the neural network, as follows: three parameters for the cutting process (nominal diameter, r. p. m., and feedrate), the sharpening method, and the drilling length, while the value of the flank wear (ha) served as an output value, as illustrated in Figure 6. Fig. 3. The scheme for carrying out the experiment Characteristic Value/type Twist drill Standard DIN 338 Dimension 06; 07,75; 010 Material S –390 MICROCLEAN Point angle 135° Spiral grad. angle 15° Way of sharpening KRI / KO Material for examination Standard Č.4237 Thermal treatment Hardening/loosening Hardness HB/HRc 410 to 425 / 43 to 45 Tension hardness [dN/mm2] 1400-1450 Cutting regimes Drilling depth [mm] l = 3xd Number of revolutions [r.p.m.] 250; 355; 500 Feedrate [mm/r] 0.027, 0.053, 0.107 Cooling medium Yes Machine for exami nation FGU-32 Measuring device o f flank wearing DORMER 438 Krivokapič Z. - Zogovič V. - Spaie O. X O N -1 O N L L Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 Fig. 4. Wearing curves during drilling with the 06 twist drill, sharpened with a corrected cutting edge (KRI) In order to perform the training of a neural network, 90 data were chosen from the set of experimental results, as follows: 40 06 data, 40 010 data, and 10 07.75 data, as presented in Table 2. Table 2. Data for training the neural network drilling lenght [mm] Fig. 5 Wearing curves during drilling with 06 crossly sharpened twist drill (KO) Neural network of the feed-forward back-propagation type was used during training, where the following parameters were changed within the available software: INPUT d [mm] N n [r.p.m.] s [mm/r] l [mm] OUTPUT ha [mm] 6.00 1 500 0.027 1350 0.260 2 10.00 1 250 3 10.00 1 4 500 0.027 375 0.245 0.027 2400 0.230 6.00 5 10.00 1 250 500 0.107 0.000 . . 0.107 0 0.000 42 6.00 250 0.027 540 0.190 43 6.00 1 500 0.027 460 0.175 44 7.75 1 355 0.053 580 0.160 45 6.00 2 500 0.027 1000 0.150 46 10.00 1 500 47 10.00 2 500 0.107 60 0.095 48 6.00 1 500 0.027 750 0.130 . . 0.107 720 0.115 86 10.00 2 250 88 10.00 1 250 89 10.00 1 500 90 10.00 1 500 ___________ 85 7.75 1 355 0.053 230 0.086 0.107 1350 0.230 87 7.75 2 355 0.053 1860 0.220 0.107 1350 0.270 0.027 1230 0.195 0.027 3000 0.295 Sharpening 1- corrected cutting edge (KRI) method: 2 - crossly sharpening (KO) Spremljanje obrabe vijačnega svedra - Using Neural Networks to Follow the Wear 439 Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 Sharpening method Drill diameter. Number ofrevolul. Feed rate 1 parameter u .....' Drilling length NEURAL .NETWORK 3 parameters 1 parameter J Flank wear (ha) Fig. 6. Access scheme of neural-network training Table 3. Parameters of neural-network training Fig. 7. Trained function TRAINLM Adaptation of function learning LEARNGDM Function performance MSE Number of layers 3 Number of neurons I layer 18 II layer 15 III layer 1 Transfer function I layer TANSIG II layer III layer - Transformation function (trained function), - Adjustment of function learning, - Function performance, - Number of layers, - Number of neurons and the transmission function for all layers. The neural network converged with a performance of 5.00897x10-032 with 35 epochs (Fig. 7), wherein the parameters specified in Table 3 were employed. Table 4 presents the output values of flank wear (h), column A targeted values, column B values gained through training of the neural network, as well as the appropriate errors. The testing of the trained network was performed for 06 - KO with the drilling parameters, n=500 r.p.m., s=0.027 mm/r, and the drilling lengths that were not trained. The results of the work (simulation) are represented by a wearing curve in relation to the wearing curve derived based on the experimental data (Figure 8). Figure 9 presents a wearing curve for BV 06.00 - KO, which was obtained by the simulation of a trained network for the cutting parameters, n=500r.p.m. and s=0.053mm/r, in relation to the wearing curve for cutting parameters, n=500r.p.m. and s=0.027mm/r, as well as n=500r.p.m. and s=0.107mm/r, obtained on the basis of experimental data. 4 CONCLUSION The modern conditions of processing by cutting and the growing requirements concerning the quality of cutting materials make the study of cutting regimes very important. The incorrect choice of cutting-regime values – in addition to the fact that a tool has all the other qualities – brings about rapid wear and a decrease of durability, and even the breakage of certain parts. Traditional methods of experimental work demand a significant waste of time and resources, because of the subject of reference, i.e., the influence of every factor is being examined separately, with a fixation on other factors’ meaning. In modern conditions of production, where there are no algorithmic solutions and there are no completely defined theoretical solutions, or where 440 Krivokapič Z. - Zogovič V. - Spaie O. Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 Table 4. Record number 2 0.245 4 0.000 2.2204e-016 4.4409e-016 . . . 42 0.190 43 . . . A B 0.260 0.230 0.000 0.175 44 0.160 45 0.150 46 0.095 47 0.130 48 0.115 85 0.086 86 0.230 87 0.220 88 0.270 89 0.195 90 0.295 0.26 0.245 0.23 -4.4409e-016 0.19 2.7756e-016 0.175 Error 2.2204e-016 -6.3838e-016 -2.2204e-016 1.6653e-016 -5.5511e-017 0.16 3.0531e-016 0.15 -1.3878e-016 0.095 -4.1633e-016 0.13 1.1102e-016 0.115 1.3878e-017 0.086 1.3878e-016 0.23 4.7184e-016 0.22 2.7756e-017 0.27 0 0.195 1.6653e-016 0.295 5.5511e-017 Fig. 8. there is a theory but it is not possible in practice to process all theoretical cases using an algorithm in a satisfactorily time interval, expert systems are used. This paper presents a performed training of a neural network based on experimental data with five input parameters and one output parameter. The inputs were the cutting process parameters, the method of sharpening and the drilling length, and the output (targeted) dimension was the value of Fig. 9. the wearing parameter. The trained network was tested on experimental data, and then employed in the determination of a wearing belt under regimes that were not used in the experiment. These initialized studies should serve as the initial starting point for the tracking of tool conditions using artificial intelligence, i.e., the intelligent tracking of tool conditions. Spremljanje obrabe vijaenega svedra - Using Neural Networks to Follow the Wear 441 Strojniški vestnik - Journal of Mechanical Engineering 52(2006)7-8, 437-442 5 REFERENCES [1] Purushothaman, S., Srinivasa, G.Y. (1994) A back-propagation algorithm applied to tool wear monitoring, Int. J. Mach. Manufact., 1994, Vol. 34, No. 5, pp. 625-631. [2] Byrne, G., et al. (1995) Tool condition monitoring - The status of research and industrial application, Annals of CIRP, Vol. 44/2, 1995, pp. 541-567. [3] Rangawala, S., Dornfeld, D. (1990) Sensor integration using neural network for intelligent tool condition monitoring, Journal of Engineering for Industry, August 1990, Vol. 112. [4] Lin, C.S., Ting, J.C. (1996) Drill wear monitoring using neural network, Int. J. Mach.Manufact., 1996, Vol. 36, No. 4, pp. 465-475. [5] Krivokapic, Z., Zogovic, V. (1998) Following the twist drill wear using the neural networks, 24. JUPITER Conference with foreign participants, Beograd. Authors’ Address: Prof. Dr. Zdravko Krivokapič Vukasin Zogovič University of Montenegro Faculty of Mechanical Eng. Cetinski put bb Podgorica, Montenegro Obrad Spaič Faculty of Production and Management Trg palih boraca 1 89101 Trebinje, BiH Prejeto: Sprejeto: Odprto za diskusijo: 1 leto 2.11.2005 22.6.2006 Received: Accepted: Open for discussion: 1 year 442 Krivokapič Z. - Zogovič V. - Spaič O.