Journal of Scientific & Industrial Research 530 Vol. 68, June 2009, pp. 530-539 J SCI IND RES VOL 68 JUNE 2009 Metal cutting process parameters modeling: an artificial intelligence approach Dejan Tanikic1*, Miodrag Manic2, Goran Radenkovic2, Dragan Mancic3 1 2 Technical Faculty Bor, VJ 12, 19210 Bor, Serbia Faculty of Mechanical Engineering Niš, A. Medvedeva 14, 18000 Niš, Serbia 3 Faculty of Electronic Engineering Niš, A. Medvedeva 14, 18000 Niš, Serbia Received 22 October 2008; revised 19 March 2009; accepted 17 April 2009 This study presents metal cutting process’ parameters modeling (cutting temperature, cutting force, and quality of machined surface) using artificial neural networks, and hybrid, adaptive neuro-fuzzy systems. Proposed models can be used for metal cutting process optimization, increasing productivity and reducing manufacturing costs. Keywords: Artificial neural networks, Metal cutting process, Neuro-fuzzy system Introduction In metal cutting process, work used is consumed in the process of plastic deformation of cutting layer and in overcoming of friction that occurs in the contact area of cutting tool (cutting insert) and work material (workpiece)1. Heat generated in chip forming zone, as well as in tool and chip contact zone and tool and workpiece contact zone directly influences quality and accuracy of machined surface and other phenomenon, which occur in metal cutting process2,3. Amount of heat generated in metal cutting process is expressed through work done in this process and mechanical equivalent of heat4. Theoretically, three zones5 (cutting, tool-chip contact, and tool-workpiece contact) of heat generation can be identified during turning. One novel approach is to remove heat generated through a cooling cycle as in interrupted cutting6. In metal cutting process, penetration of cutting wedge of tool into material of workpiece is enabled by cutting force. Cutting force magnitude, or amplitude of dynamic component of cutting force, can be used as control signal in adaptive controlled machining system7. A large number of models and theories for calculating cutting force are based on orthogonal cutting model8,9, and also a model, which includes strains, strain rates and cutting temperature10. *Author for correspondence Tel: +381-30-424-555; Fax: +381-30-421-078 E-mail: [email protected] Indicators of machined surface11, which identify quality of machined surface, include physical-mechanical characteristics of surface layer (micro hardness, residual stresses) and geometrical characteristics of machined surface (roughness, corrugation and shape deviation). Surface roughness is irregularities of material resulted from various machining operations12. This study presents metal cutting process’ parameters modeling (cutting temperature, cutting force, and quality of machined surface) using artificial neural networks, and hybrid, adaptive neuro-fuzzy systems. Main goals are qualitative analysis of metal cutting process, identifying and resolving most frequent problems, and improving manufacturing productivity. Experimental Details Proposed model (Fig. 1) shows way of component relations and information flow of material handling system, and linked information system that processes obtained data. Piezoelectric dynamometer with amplifier, for measuring cutting force, is positioned on lathe. Temperature changing of cutting tool and workpiece material is monitored by infrared camera. After machining, height of roughness of machined surface is measured by using an appropriate tool. All measured data is passed to data acquisition system. In data modelling system, data is modelled in two ways by artificial neural networks (ANN modelling), and hybrid, adaptive neuro-fuzzy system (NF modelling). After data TANIKIC et al: METAL CUTTING PROCESS PARAMETERS MODELING USING ANN 531 Fig. 1—Component relations and information flow of material handling system Table 1—Measuring equipment and accessories No. Measuring equipment and accessories Name Manufacturer Type 1 2 Middle lathe Cutting tool 3 4 5 6 7 8 Dynamometer Amplifier Computer system Computer Infrared camera Tool for measuring surface roughness “Potisje”-Ada Tool holder – “Sandvik” Cutting insert – “Sandvik” “Kistler” “Kistler” “Hewlett Packard” PA-C-30 PCLNR CNMG 9265A3 5007A HP 9000/300 Pentium II Varioscan 3021-ST Surftest SJ-301 “Jenoptik” “Mitutoyo” modeling, it is necessary to perform verification of models (model testing), and finally comparison of obtained data. Lathe, used for examining and measuring, is located in the Laboratory for Production Engineering, at the Faculty of Mechanical Engineering in Niš. Workpiece material (cylindrical bars, dimensions φ45x250 mm) was not previously thermally treated steel, with AISI designation 4140. This steel belongs to the group of doped, decent, cold drown steels, with strength of Rm=1050 N/mm2, and measured hardness of 205 HB. Measuring equipment with accessories, needed for determining cutting temperature, cutting forces and surface roughness during turning process is denominated in Table 1. Metal cutting process is carried out without using any coolant and lubrication fluids, which can obstruct directly recordings of cutting process. Pausing between two experiments enables system to cool down to room temperature. According to recommendations of manufacturer and experience of machinist, SANDVIK Coromant cutting tool [tool holder PCLNR 32 25 P12, and cutting insert CNMG 12 04 08 (grade 235)] has been chosen. Results and Discussion Cutting Temperature Because of stabilization of process temperature, measuring should be done shortly after beginning of cutting process 13 (30 s is enough for temperature stabilization). Room temperature before start of experiments was 28°C. Following parameters of cutting process were adopted: cutting speed V (80, 95, 110, 125 and 140 m/min), feed rate f (0.071, 0.098, 0.196 and 0.321 mm/rev), and depth of cut a (0.5, 1, 1.5 and 2 mm). Minimal measured cutting temperature 532 J SCI IND RES VOL 68 JUNE 2009 (214.82°C) occurs at: V, 80 m/min; f, 0.071 mm/rev; and a, 0.5 mm. Maximal measured cutting temperature (594.02°C) was measured at: V, 140 m/min; f, 0.321 mm/ rev; and a, 2 mm. Increasing feed rate causes increase in temperature of cutting; particularly at low cutting speeds. For example, during machining (V, 80 m/min) for change in f from 0.071 mm/rev to 0.321 mm/rev, temperature rises about 89%, while with same conditions, during machining (V, 140 m/min), temperature rises about 29% (a=0.5 mm). Larger values of a cause an increase in cutting forces, thereby resulting in increase in cutting temperature [increasing of cutting temperature between depth of cut (0.5-2 mm) is relatively monotonous, and takes values of 31-38%]. Cutting temperature is closely connected to cutting speed. For example, while depth of cut is constant (a=0.5 mm), increasing of V (80-140 m/min) causes temperature increase of 70% (f, 0.071 mm/rev) and only 16% (f, 0.321 mm/rev). Cutting Forces Largest component of cutting force in all experiment setups is main cutting force (F1), while radial force (F2) and axial force (F3) are quite minor. Resulting cutting force has minimal value of 5.234 kN (V, 80 m/min; f, 0.071 mm/rev; and a, 0.5 mm) and maximal value of 68.446 kN (V, 140 m/min; f, 0.321 mm/rev; and a, 2 mm). Monitoring ratio of components of cutting force, rather than monitoring magnitude of individual cutting force component, can successfully fulfill state of cutting tool monitoring14. During various cutting conditions, F1/F2 = 1.26-2.59, F1/F3 = 1.417-4.283, and F2/F3 = 0.603-2.355. While cutting with constant parameters (f, 0.071 mm/ rev; V, 140 m/min; a, <1.25 mm), magnitude of F2 is larger than magnitude of F3, while cutting with a >1.25 mm, magnitude of F2 is smaller than magnitude of F3. From this consideration, it can be concluded that general relation, which exactly defines ratio between F1, F2 and F3, does not exist. This ratio varies from case to case, and depends on general cutting conditions as well as adopted cutting regimes. Components of cutting force and resulting cutting force are in narrow relationship with feed rate. Increase of resulting cutting force (at changing feed rate, 0.071-0.321 mm/rev) is 179-289%. Cutting forces directly depends on depth of cut. Increasing depth of cut (0.5-2 mm) increases cutting force by 180-340%; at larger values of feed rates change is more intensive. Finally, cutting forces are least sensitive to changes in cutting speed. In the range of speed changes from 80-140 m/min, increase of resulting cutting force is 1-40%. Quality of Machined surface For arithmetic mean deviation (AMD, Ra) modeling, same input parameters as in temperature and force modeling are adopted but cutting speed takes values of 80, 110 and 140 m/min. AMD directly depends on feed rate. So, increasing feed rate increases AMD; more intensive at higher cutting speeds. For example, for machining (a, 0.5 mm), on increasing f from 0.071 mm/ rev to 0.321 mm/rev at V= 140 m/min, Ra increases by 211%, while for same cutting conditions and at V= 80 m/min, Ra increases only by 42%. Increasing depth of cut increases Ra,. In machining (f, 0.071 mm/rev), at a increasing from 0.5 mm to 2 mm, Ra increases by 173%, while at machining with f= 0.321 mm/rev, Ra increases by only 4% (V, 140 m/min). Increasing V also influences Ra. Increasing V from 80 m/min to 140 m/min (a, 0.5 mm), Ra increases 138%, while Ra increases only 55% when a = 2 mm. Big influence on height of roughness has built up edge (BUE), as well as constant presence of frictional forces, which cause interrupted character of cutting process. At higher cutting speeds, this effect is almost completely eliminated. So, real value of roughness approaches its theoretical value. Real value of roughness depends on geometric factors, as well as cutting speed, workpiece characteristics, cutting tool angles, coolant and lubricant fluids, elastic properties of workpiece, state of tool (sharp or worn) etc. Data Modeling using Artificial Neural Networks (ANN) In ANN structure15,16 (Fig. 2), number of hidden layers was set to 1 (with 2 neurons), after that number of hidden layers and neurons in them was increased, with tracking errors and possibility of generalization of appropriate network architecture. Number of training cycles was varied, too. Testing was done on ANNs (Fig. 3) with one hidden layer of neurons with 2, 3 and 5 neurons (Ns=2, Ns=3, Ns=5), and with two hidden layers of neurons with 2+2 and 3+2 hidden neurons respectively (Ns1=2 Ns2=2; Ns1=3 Ns2=2). ANN with two hidden layers (with 3 and 2 neurons) has been found most accurate, and data generalization has been successfully performed. So, adopted ANN architecture is 3-3-2-1. Matlab software was used for ANN modeling. Adopted learning algorithm (Levenberg-Marquardt) provides best convergence in approximation of an unknown function (function prediction). TANIKIC et al: METAL CUTTING PROCESS PARAMETERS MODELING USING ANN 533 Fig. 2—Schematic structure of artificial neural network (ANN) Fig. 3—Error of different networks architectures subjected to number of training cycles Three independent systems were created: I) ANN for cutting temperature prediction (ANNTEMP); ii) ANN for cutting force prediction (ANNFORCE); and iii) ANN for arithmetic mean deviation of machined surface prediction (ANNSURF). ANNs for cutting temperature and cutting speed prediction were trained with training set of 80 experimentally obtained data. Rest 31 data set was used for verification (testing) of ANN. ANNSURF was trained with training set of 48 experimentally obtained data, and, after that tested with remaining 25 data set. 534 J SCI IND RES VOL 68 JUNE 2009 Fig. 4—Architecture of characteristic adaptive neuro-fuzzy system Data Modelling using Adaptive Neuro-Fuzzy Systems (ANFIS) ANFIS17,18 contains a few number of layers with nodes (neurons), each of which performs a particular function (node function) on the incoming signals as well as a set of parameters pertaining to this node. The type of the function, which the node performs, may vary from node to node, and a choice of node function depends on overall input-output function that network simulates19. In general, this system represents the way for adjusting existential base of rules, using learning algorithm, which is based on assembly of input-output pairs, used for training. In architecture of ANFIS (Fig. 4), it is supposed that system has just two input values x and y (Level 1), and one output value z (Level 5). In case, rule base consists just two fuzzy IF-THEN rules (Takagi-Sugeno type), as shown in Level 2: Rule 1: IF x is A1 AND y is B1, THEN f1 = p1x + q1y + r1 Rule 2: IF x is A2 AND y is B2, THEN f2 = p2x + q2y + r2 First part of fuzzy rule (after IF part of rule) is called premise, while second part of the rule (after THEN part of rule) is called consequent. From ANFIS system architecture, it is obvious that for given values of premise parameters, output value can be presented as linear combination of consequent parameters20 as f = w1 w2 f1 + f = w1 f1 + w 2 f 2 = w1 + w 2 w1 + w 2 2 ( ) ( ) ( ) ( ) ( ) ( ) = w1 x p1 + w1 y q1 + w1 r1 + w 2 x p 2 + w 2 y q 2 + w 2 r2 …(1) Operation of hybrid algorithm is composed from one forward and one backward pass. In forward pass signals are running until Level 4, and consequent parameters are identified by least squares estimate. In backward pass, error rates propagate backward and premise parameters are updated by gradient descent. On data modeling computer system, Matlab software is installed (toolbox for modeling data using hybrid, adaptive neuro-fuzzy systems). Based on this system, appropriate application is created (called NeuroFuzzy), which enables post processing of data about cutting temperature, resulting cutting force and arithmetic mean deviation of machined surface using experimentally obtained data. Three neuro-fuzzy systems (NF) were created: First one for NFTEMP, second for NFFORCE and third for NFSURF. Just like with modeling using ANNs, input parameters adopted are V, f and a, while output values are NFTEMP, NFFORCE and NFSURF. After analytical consideration of different architectures of ANFIS behaviors, TANIKIC et al: METAL CUTTING PROCESS PARAMETERS MODELING USING ANN 535 a) b) c) Fig. 5—Cutting temperature in relation to: a) cutting speed; b) depth of cut; and c) feed rate parameters adopted were number of membership function of each input value (2), type of membership function (gbellmf), type of output membership functions (constant), method of optimization used for modeling (hybrid), and number of training epochs (60 for NFTEMP and NFFORCE, and 30 for NFSURF). Comparative Analysis Fig. 5 gives comparative results of exploitation of ANNTEMP and NFTEMP, as well as measured values of cutting temperatures. Measured values of resulting cutting force, and predicted values obtained by ANNFORCE and NFFORCE are shown in Fig. 6. Comparative results of exploitation of ANNSURF and NFSURF, and measured values of arithmetic mean deviation of machined surface are shown in Fig. 7. From Figs 5-7, it could be observed that all measured and predicted values match. Besides, measured and modeled values, trendlines (a graphic representation of trends in data sets) are shown in following figures. In this particular 536 J SCI IND RES VOL 68 JUNE 2009 a) b) c) Fig. 6—Resulting cutting force in relation to: a) cutting speed; b) depth of cut; and c) feed rate case, polynomial trendline was used (this type of trendline uses following equation to calculate least squares fit through y = b + c1 x + c2 x 2 , where b, c 1 and c2 are constants). Analysis of maximum as well as average mean errors for every single case (Table 2) indicates that ANFIS gives a bit more accurate values than ANN. In almost all situations (accept maximum error in cutting force prediction), error was minor in case of using ANFIS. System for Tool State Prediction Main constraint factors, which occur during analysis of tool state prediction, are related to: i) Providing required quality of machined surface; and, ii) Maximum exploitation of used tool. So, successful system must take into consideration all relevant factors, as well as their overall influence on current state of the tool. Systems for cutting temperature, cutting force and surface quality prediction, based on ANNs and ANFIS technologies, represent subsystems of this global system (Fig. 8). ANFIS, showing 537 TANIKIC et al: METAL CUTTING PROCESS PARAMETERS MODELING USING ANN a b) Fig. 7—Arithmetic mean deviation in relation to: a) cutting speed; b) feed rate better characteristics, were adopted for sub-systems of main system. First variant represents “parallel” connection of this sub-systems where input parameters of all models are cutting regimes, while output values, together with other relevant parameters goes in artificial intelligence based system for prediction of cutting tool. In second variant “serial” outputs from foregoing subsystem together with cutting regimes, represents inputs in subsequent sub-system (or global system). At least, output from last sub-system, together with other relevant factors, goes into global (identical to previous case) system, which gives prediction of cutting tool state. Principally, second variant is a little bit complex, but accuracy of such a system would be better (because of more precise definition of input parameters of each sub-system). Global (main) system can be based on ANNs, ANFIS technologies, or some other artificial intelligence based system. The output from this system is evaluation of cutting tool state. This signal could be a significant and Table 2 — Maximal and mean errors of used systems Model Max. error, % Mean error, % Cutting temperature ANNTEMP NFTEMP 14.057 8.773 3.496 3.396 Cutting force ANNFORCE NFFORCE ANNSURF NFSURF 13.637 17.843 4.388 3.876 Arithmetic mean deviation 12.628 6.932 11.321 6.310 538 J SCI IND RES VOL 68 JUNE 2009 Fig. 8—Possible variant solutions of tool state prediction system objective cutting tool state identifier, and could be further used for projection of intelligent manufacturing. 2 Conclusions This work shows possibility of implementation of artificial intelligence based systems in metal cutting process. In first phase, for different values of input parameters (cutting speed, feed rate and depth of cut), data acquisition is performed for cutting temperature, cutting force and arithmetic mean deviation of machined surface. Modeling of cutting process was performed in next phase, using experimentally obtained data, and artificial intelligence based approach (ANNS and ANFIS systems). For some unknown values of input parameters, system can predict some output parameters of interest. For proposed experiment setup, ANFIS system showed slightly better performances, and have priority in the modeling of this type of data. Finally, global system for predicting state of the cutting tool was proposed (with sub-systems for cutting temperature, cutting force and arithmetic mean deviation prediction). 3 4 5 6 7 8 9 10 11 References 1 Da Silva M B & Wallbank J, Cutting temperature: prediction and measurement methods – a review, J Mater Process Technol, 88 (1999) 195-202. 12 Devillez A, Schneider F, Dominiak S, Dudzinski D & Larrouquere D, Cutting forces and wear in dry machining of Inconel 718 with coated carbide tools, Wear, 262 (2007) 931-942. Childs T, Maekawa K, Obikava T & Yamane Y, Metal Machining – Theory and Applications (Arnold, London NW1 3BH Great Britain) 2000. Arshinov V & Alekseev G, Metal Cutting Theory and Cutting Tool Design (Mir Publishers, Moscow) 1979. O’sullivan D & Cotterell M, Temperature measurement in single point turning, J Mater Process Technol, 118 (2001) 301-308. Dessoloy V, Melkote S N & Lescalier C, Modeling and verification of cutting tool temperatures in rotary tool turning of hardened steel, I J Mach Tools Manuf, 44 (2004) 1463-1470. Dimla Sr D E & Lister P M, On line metal cutting tool condition monitoring, I: Force and vibration analysis, I J Mach Tools Manuf, 40 (2000), 739-768. Merchant M E, Mechanics of the metal cutting process – Orthogonal cutting and the type of chip, J Appl Phys, 16 (1945) 267-274. Piispanen V, Theory of formation of metal chips, J Appl Phys, 19 (1948), 876-881. Oxley P L B, Mechanics of Machining, An Analitical Approach to Assessing Machinability (Ellis Horwood Limited, Chichester, UK) 1989. Radovanovic M, Tehnologija masinogradnje, obrada materijala rezanjem (Univerzitet u Nisu, Masinski fakultet, Nis, Srbija (in serbian)) 2002. Erzurumlu T & Oktem H, Comparison of response surface model with neural network in determining the surface quality of moulded parts, Mater Design, 28 (2007) 459-465. TANIKIC et al: METAL CUTTING PROCESS PARAMETERS MODELING USING ANN 13 14 15 16 Kwon P, Schiemann T & Kountanya R, An inverse estimation scheme to measure steady-state tool-chip interface temperatures using an infrared camera, I J Mach Tools Manuf, 41 (2001) 1015-1030. Das S, Chattopadhyay A B & Murthy A S R, Force parameters for on-line tool wear estimation: A neural network approach, Neural Networks, 9 (1996) 1639-1645. Santochi M & Dini G, Use of neural networks in automated selection of technological parameters of cutting tools, Compu Integrated Manuf Syst, 9 (1996) 137-148. Freeman J A & Skapura D M, Neural Networks: Algorithms, 17 18 19 20 539 Applications and Programming Techniques (Addison-Wesley Publishing Company, Inc., USA) 1991. Sick B, On-line and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research, Mech Syst Signal Process, 16 (2002) 487-546. Werbos P, Beyond regression: New tools for prediction and analysis in the behavioral sciencies, Ph D Thesis, Harvard University, 1974. Jang J S R, ANFIS: Adaptive-Network-Based Fuzzy Inference System, IEEE Trans Syst, Man Cybernetics, 23 (1993) 665-685. Jantzen J, Neurofuzzy modelling, Tech. report no. 98-H-874 (Tech Univ of Denmark, Denmark) 1998.
© Copyright 2026 Paperzz