Neural Network Applications In Machining: A Review

3,897 views

Published on

This document was accepted for presentation in the proceedings of "International Conference on Advances in Manufacturing Technology '08".

1 Comment
1 Like
Statistics
Notes
No Downloads
Views
Total views
3,897
On SlideShare
0
From Embeds
0
Number of Embeds
4
Actions
Shares
0
Downloads
284
Comments
1
Likes
1
Embeds 0
No embeds

No notes for slide

Neural Network Applications In Machining: A Review

  1. 1. Neural network applications in machining: A review Ashish Khetan1 and Sankha Deb2 Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati-781039, Assam. a.khetan@iitg.ernet.in 1, sankha.deb@iitg.ernet.in 2 Abstract: For modeling of machining processes, the physics based techniques are not always successful due to the complicated physics involved and associated uncertainties. In such cases, data based modeling techniques like neural networks can be useful. The present paper reviews some of the applications of neural network in different areas of machining like turning, milling, drilling, grinding etc. A brief background of neural networks is also provided. The neural network applications in machining have been classified in following four categories: (1) Prediction of surface roughness (2) Prediction of tool wear (3) Prediction of cutting forces (4) Optimization of machining processes. The relative merits and demerits are given for choosing the type of neural networks, their architecture and the parameters of network. Keywords: Artificial neural networks, surface roughness, tool wear, cutting forces, optimization 1. Introduction The dream of making artificially intelligent manufacturing systems has motivated researchers to develop efficient computational techniques for modeling, monitoring, and control of manufacturing processes. The physics based modeling techniques alone, however, are not always succesful in modeling the manufacturing processes like machining due to the complicated physics involved and the associated uncertainties. In such cases, modeling based on data can be useful if there is sufficient data available for describing the behaviour of the process. The modeling based on data can be accomplished by soft computing tools like artificial neural networks (ANNs). Since the begining of 1990s, a number of researchers started applying ANNs for modeling of machining processes to predict the surface roughness, cutting forces, tool wear, and also for optimization of the processes. We have reviewed some of the recent research works, and based on the literature review, the merits and demerits of the different neural network architectures, mainly the multi-layer perceptron (MLP) neural networks and the radial basis function neural networks have been highlighted. The effects of choosing different network parameters on modeling accuracy is also given. The rest of the paper is organised as follows. A brief background of neural networks is given in section 2 and a review of neural network applications in machining is given in section 3.
  2. 2. 2. Background of neural networks Neural networks are basically connectionist systems, in which various nodes (called “neurons”) are interconnected. A typical neuron receives one or more input signals. The output signal provided by the neuron depends on its processing function. This output is transferred to connecting neurons in varying intensities. The neurons in the input layer receive input signals from the user and provide the output through the neurons in the output layers. Only the neurons in the input and output layers interact with the outside world/user; the rest are hidden. After choosing the network architecture, the network is trained. In the supervised training of the network, the network is presented with training pairs, each consisting of a vector from an input space and a desired network output. Through a defined learning algorithm, the network performs the adjustment of its parameters so that the error between the actual and desired output is minimized. Once trained, the network can be used for predicting the output for any input vector from the input space. This is called the “generalization property” of the network. The feed forward back propagation neural network (BPNN) and the radial basis function (RBF) neural network are the two most widely used neural networks. The back propagation algorithm iteratively adjusts the network weights to minimize the squares objective function, the sum of squared residuals (difference between the desired and estimated output). A RBF neural network comprises three layers- an input layer, a single layer of non-linear processing neurons, and an output layer. The architecture of a typical RBF network to calculate surface finish is shown in fig 1. Fig 1: A typical radial basis function neural network architecture. The weights Wk are adjusted by a multiple linear regression procedure so that sum of squared residuals is minimal. 3. Applications of neural network
  3. 3. The neural network applications in machining can be broadly classified into four major cateogries: prediction of surface roughness, prediction of tool wear, prediction of cutting forces, and optimization of machining opeartions. Subsequent subsections provide a review of the recent research done in these fields. 3.1 Prediction of surface roughness Neural networks have found applications for surface roughness prediction in turning and use the process parameters like feed, depth of cut, cutting speed, and cutting forces for modeling. The four parameters – feed, depth of cut, radial and axial cutting forces were employed for surface roughness prediction by Azouzi and Guillot [1], using ANN model with sensor based fusion technology. Risbood et al. [2] used the feed, the depth of cut, the cutting speed and an additional input parameter, the acceleration of radial vibration of the tool holder in their ANN model. They also made an observation that surface finish improves with increasing feed rate upto some value and then it starts deteriorating further. Kohli and Dixit [3] predicted the most likely values of surface roughness along with the upper and lower estimates. They also suggested a systematic procedure to select initial training and testing data sets. Pal and Chakraborty [4] used the input parameters– cutting speed, feed, depth of cut and cutting forces in BPNN and reported the convergence of the mean square error both in training and testing data sets. For the prediction of surface roughness in hard turning, Özel and karpat [5] added the tool edge geometry, the Rockwell- C hardness of the work piece and the cutting length to the set of input parameters in their work. Ho et al. [6] developed an adaptive neuro-fuzzy inference system for the prediction of surface roughness. They used the grey level of the surface image as a new input parameter instead of using cutting forces. Jiao et al. [7] used the fuzzy adaptive networks with only three parameters cutting speed, feed and depth of cut. The method uses newly available operating data to regularly improve the network. Abburi and Dixit [8] used the fuzzy inference system to predict the surface roughness for the given process parameters and also to predict the inverse relation. RBF neural network which involves comparatively less complexity in modeling was tested by Sonar et al. [9] and found to be slightly inferior in predicting the surface roughness, compared to a MLP neural network. Prediction of surface roughness has also been done for milling processes. Tsai et al. [10] employed the neural network for real time prediction of surface roughness in end milling. They used the vibration intensity per revolution as an additional parameter along with the process parameters, spindle speed, feed, and depth of cut. Machining tolerance was used as a new parameter in the work done by Oktem et al. [11] and it was found that the surface roughness becomes particularly high for lower values of machining tolerance. For the surface roughness prediction in face milling, Bernardos and Vosniakos [12] used a total of nine factors for ANN modeling, out of which they found that the feed rate per tooth, longitudinal feed component of the cutting force, the depth of cut, the engagement
  4. 4. of cutting tool and the cutting fluid are the most influential ones. They employed Levenberg-Marquardt algorithm for training the ANN, orthogonal array for choosing the factor levels and Taguchi’s design for designing the experiment. Surface roughness has also been predicted in surface grinding operation using the neural networks. Aguiar et al. [13] used three input parameters– depth of cut, acoustic emission and electric power signals, digitally processed through known statistics, for the surface roughness prediction in grinding. 3.2 Prediction of tool wear Many researchers have attempted to predict the flank wear for tool condition monitoring in various machining operations like turning, milling, drilling etc. Tool condition monitoring may be done in two ways: off-line and on-line, applying direct or indirect methods. Direct methods rely on sensing technology that measures the wear using the optical, radioactive, and proximity sensors, and electrical resistance techniques. However, the direct methods are not easily achievable because of the complexity of measuring the above signals during the process. Indirect methods measure other factors that are responsible for tool wear such as cutting forces, acoustic emission, temperature, vibration, spindle motor current, cutting conditions, torque, strain, and also factor like snapshot images of the cutting tool. Liu and Altintas [14] proposed a two step method for tool condition monitoring in turning operation. They derived an expression based on physics, to calculate the flank wear, first in terms of cutting force ratio and other cutting parameters, and then used the calculated flank wear as an input to the neural network, along with other inputs like cutting force ratio, feed rate and cutting speed, to predict the flank wear. Li et al. [15] integrated the neural network with an analytical model, Oxley’s theory, and formed a new hybrid machining model for the prediction of the flank wear. Scheffer et al. [16] proposed a combination of static and dynamic neural networks, trained on-line and off-line, to predict the flank wear. Recurrent neural network design was employed for the continuous tool wear monitoring in [17]. Yao et al. [18] presented a new method of tool wear detection by fusing the two tool wear estimations, based on the cutting conditions and the detected signals. The detected signals include the spindle motor current, feed motor current and acoustic emission. Quan et al. [19] used a sensor integration strategy for prediction of the flank wear by combining the information obtained from acoustic emission sensor and power sensor with the machining parameters. Choudhary and Bartarya [20] used the temperature at the cutting zone and the surface finish for the prediction of the tool wear. Lee et al. [21] in their work concluded that by using an appropriate cutting force ratio flank wear can be estimated very accurately.
  5. 5. For the tool wear monitoring during face milling, Ko and Cho [22] combined the neural network classifier with an adaptive signal processing scheme. The method updated the model parameters adaptively at each sampling instant. Chen and Jen [23] analyzed the training efficiency and the test performance of the different data fusion methods and found that the two methods, index multiplication group (IMG) and the indices multiplication and division group (IMDG), are the most suitable ones. They also concluded that the performance of the monitoring system can be significantly improved with suitable selection of data fusion method. Sanjay et al. [24] estimated the flank wear in drilling using different structures of ANN with input parameters – drill diameter, feed, cutting speed, time, force and torque. Their work demonstrated the dependence of tool wear on these variables and they found the three layered neural networks, one with 2 and the other with 10 neurons in the hidden layer, as the best structures. Lin and Ting [25] used the BPNN for drill wear monitoring and found that the neural networks with two hidden layers learn faster and can more accurately estimate tool wear than the networks with one hidden layer. Mahfouz [26] compared several architectures of feed forward BPNN for tool condition monitoring of twist drill wear. Fully connected neural networks were found to be better than decoupled networks. The algorithm utilizes the analysis of vibration signature as the main and only source of information from the machining process. Neural networks have also been used in estimation of tool usage (life or wear) for micro- machining operations. Tansel et al. [28] made excellent tool usage estimation in micro end milling by using two encoding techniques, force-variation based encoding (FVBE) and segmental- average-based encoding (SABE), along with BPNN. Tansel et al. [29] proposed an off-line tool wear estimation method primarily for the micro-machining of non-metals. They introduced the neural network based periodic tool inspector (N²PTI) to evaluate the cutting force signals recorded at identical cutting conditions and to estimate the usage, they found the wavelet-transformation based encoding (WTBE) and back propgation neural network, to be the best combination for N²PTI. 3.3 Prediction of cutting forces Neural networks have also been used for the prediction of the cutting forces during various machining operations. Szecsi [30] modeled the cutting forces using the three layer feed forward neural network. He used a set of twelve input parameters to get the three components of cutting force, with 7-8 neurons in the hidden layer of the network, within an accuracy of 3.5%. Supervised neural networks have also been used by Zuperl and Cus [31] to successfully estimate the cutting forces developed during a ball-end milling process. The author used the BPNN and predicted the three components of force in a oblique cutting to an accuracy of ±4%. Zuperl et al. [32] used the BPNN to evolve a generalized model for the prediction of the cutting forces during a ball-end milling
  6. 6. process based on a set of ten input parameters. They estimated feed cutting force within an error of 4% and also reported the superiority of the radial basis function neural network over the BPNN in predicting the forces. Kadirgama and Abou-El-Hossein [33] employed the neural network to predict the cutting forces in milling. Hao et al. [34] introduced the ANNs for the prediction of the cutting forces of a self-propelled rotary tool. To overcome the problem of convergence to local minima in the error space, they used the hybrid of genetic algorithm (GA) and back propagation algorithm. Aykut et al. [34] used the ANNs for modeling the effects of machinability on chip removal cutting parameters for face milling. They predicted cutting forces by changing the cutting speed, the feed rate, and the depth of cut within an average error of 2% and 10% for training and testing respectively, under dry conditions. Alajmi and Alfares [35] developed a model, using BPNN with an enhancement by differential evolution (DE) algorithm, to predict the cutting forces in turning. Their results showed an improvement in the reliability of predicting the cutting forces over the previous works using simple back propagation network. 3.4 Optimization of machining processes Neural networks have also been used for optimization of machining processes. Cus and Zuperl [36] used neural networks for the optimization of the cutting parameters. They compared feed forward neural network with RBF neural network and found that the former gives more accurate results but requires more time for training and testing. The procedure is more suitable for fast and approximate determination of cutting conditions, when there is not enough time for thorough analysis. Tansel et al. [37] optimized the cutting coinditions from experimental data using genetically optimized BPNN system (GONNS). They tested performance of the GONNS for the two cases, first keeping the cutting forces in the desired range, while maximizing metal removal rate in micro-end- milling, and second while obtaining the best possible compromise between the roughness of the machined mold surfaces and the duration of finshing cut. Zuperl et al. [38] used the ANNs with analytical module OPTIS to optimize the cutting conditions under the constraints of low machining costs and high productivity, taking into account the given limitaions of the cutting process. Chiang et al. [39] proposed an architecture with two different kinds of neural networks for on-line determination of optimal cutting conditions in an end milling proceess: a BPNN to model the cutting processes and the second network, which parallelizes the augmented Lagrange multiplier algorithm, to determine the corresponding optimal cutting parameters by maximizing the material removal rate according to appropriate operating constraints . Tandon et al. [40] used the ANNs with a newly developed particle swarm optimization (PSO) technique to optimize the cutting conditions in NC end
  7. 7. milling operation. EL-Mounayri et al. [41] optimized the process parameters of flat and ball-end milling, using a neural network based model. They implemented RBF neural network and reported its several advantages over the traditional BPNN, including high efficiency, easy definition, accurate results and very fast convergence. Zuperl et al. [42] proposed an architecture with two different kinds of neural networks for on-line optimal control of a milling process. They selected feedrate as the variable to be optimised, and estimated the milling state by the measured cutting forces. 4. Conclusions The paper presents a review of some of the applications of neural network in machining operations: prediction of surface roughness, prediction of tool wear, prediction of cutting forces and optimisations of machining operations. The study reveals that different types of neural networks including feed forward BPNN, RBF neural network and adaptive fuzzy networks have been used in modeling and prediction. The accuracy, reliability and effectiveness of the neural network depends on a number of factors among others like the input parameters choosen, number of hidden layers, the number of neurons in the hidden layers, and its type and architecture. The literature shows that neural networks have a lot of potential to offer for application in modeling the machining processes. However, in our opinion physics of the process should be well understood to complement the modeling successfully. 5. Acknowledgement Authors would like to acknowledge Prof. U. S. Dixit, Head of the Mechanical Engineering Department, IIT Guwahati for his help and encouragement. References 1. Azouzi R, and Guillot M, Int J Mach Tools Manuf 37(9) (1997) 1201-17. 2. Risbood K A, Dixit U S, and Sahasrabudhe A D, J Mater Process Technol 132 (2003) 203- 214. 3. Kohli A, and Dixit U S, Int J Adv Manuf technol 25 (2005) 118-129. 4. Pal S K, and Chakraborty D, Neural computing and applications 14 (2005) 319-324. 5. Özel T, and Karpat Y, Int J Mach Tools Manuf 45 (2005) 467-479.
  8. 8. 6. Ho S Y, Lee K C, Chen S S, and Ho S J, Int J Mach Tools Manuf 42 (2002) 1441-1446. 7. Jiao Y, Lei S, Pei Z J, and Lee E S, Int J Mach Tools Manuf 44 (2004) 1643-1651. 8. Abburi N R, and Dixit U S, Robotics and CIM 22 (2006) 363-372. 9. Sonar D K, Dixit U S, and Ojha D K, Int J Adv manuf Technol 27 (2006) 661-666. 10. Tsai Y H, Chen J C, and Lou S J, Int J Mach Tools Manuf 39 (1999) 583- 605. 11. Oktem H, Erzurumlu T, and Erzincanli F, Materials and Design 27 (2006) 735-744. 12. Bernardos P G, and Vosniakos G C, Robotics and CIM 18 (2002) 343-354. 13. Aguiar P R, Curz C E D, Paula W C F, Bianchi E C, Thomazella R, and Dotto F R L, Artificial intelligence and Applications (2007) 549. 14. Liu Q, and Altintas Y, Int J Mach Tools Manuf 39 (1999) 1945-1959. 15. Li X P, Iynkaran K, and Nee A Y C, J Mater Process Technol 89/90 (1999) 224-230. 16. Scheffer C, Kratz H, Heyns P, and Klocke F, Int J Mach Tools Manuf 43 (2003) 973-985. 17. Ghasempoor A, Jeswiet J, and Moore T N, Int J Mach Tools Manuf 39 (1999) 1883-1902. 18. Yao Y, Li X, and Yuan Z, Int J Mach Tools Manuf 39 (1999) 1525-1538. 19. Quan Y, Zhou M, and Luo Z, Engineering Applications of Artificial Intelligence 11 (1998) 717-722. 20. Choudhary S K, and Bartarya G, Int J Mach Tools Manuf 43 (2003) 747-753. 21. Lee J H, Kim D E, and Lee S J, Mechanical Systems and Signal Processing 10(3) (1996) 265-276. 22. Ko T J, and Cho D W, Int J Mach Tools Manuf, 12 (1996) 5-13. 23. Chen S L, and Jen Y W, Int J Adv Mach Tools Manuf 40 (2002) 381-400.
  9. 9. 24. Sanjay C, Neema M L, and Chin C W, J Mater Process Technol 170 (2005) 494-500. 25. Lin S C, and Ting C J, Int J Mach Tools Manuf 36 (1996) 465-475. 26. Mahfouz I A, Int J Mach Tools Manuf 43 (2003) 707-720. 27. Tansel I N, Arkan T T, Bao W Y, Mahendrakar N, Shisler B, Smith D, and McCool M, Int J Mach Tools Manuf 40 (2000) 599-608. 28. Tansel I N, Arkan T T, Bao W Y, Mahendrakar N, Shisler B, Smith D, and McCool M, Int J Mach Tools Manuf 40 (2000) 609-620. 29. Szecsi T, J Mater Process Technol 92-93 (1999) 344-349. 30. Zuperl U, and Cus F, J Mater Process Technol 153-154 (2004) 268-275. 31. Zuperl U, Cus F, and Mursec B, and Ploj T, J Mater Process Technol 175 (2006) 98-108. 32. Kadrigama K, and Abou-El-Hossein K A, Journal of Apllied Sciences 6(1) (2006) 31-34. 33. Hao W, Zhu X, Li X, and Turyagyenda G, J Mater Process Technol 180 (2006) 23-29. 34. Aykut S, Gölcü M, Semiz S, and Ergür H S, J Mater Process Technol 190 (2007) 199-203. 35. Aljami M S, and Alfares F, Artificial Intelligence and Applications (2007) 549. 36. Cus F, and Zuperl U, J Mater Process Technol 173 (2006) 281-290. 37. Tansel I N, Ozecelik B, Bao W Y, Chen P,Rincon D, Yang S Y, and Yenilmez A, Int J Mach Tools Manuf 46 (2006) 26-35. 38. Zuperl U, Cus F, Mursec B, and Ploj T, J Mater Process Technol 157-158 (2004) 82-90. 39. Chiang S T, Liu D I, Lee A C, and Chieng W H, Int J Mach Tools Manuf 34 (1995) 637-660. 40. Tandon V, El-Mounayri H, and Kishawy H, Int J Mach Tools Manuf 42 (2002) 595-605.
  10. 10. 41. EL-Mounayri H, Kishawy H, and Briceno J, J Mater Process Technol 166 (2005) 50-62. 42. Zuperl U, Kiker E, and Cus F, Industrial Technology, IEEE International Conference on 1 (2003) 393-398.

×