A NEURAL NETWORK MODEL FOR SELECTION OF

MACHINING PARAMETERS IN CAPP

Burak Sari†, Saleh Amaitik† and S. Engin Kilic†

†
Middle East Technical University, Mechanical Engineering Department
Integrated Manufacturing Technologies Research Group-IMTRG, 06531, Ankara, Turkey
Email: burak@me.metu.edu.tr & e121259, engink@metu.edu.tr

Abstract
Machining parameters such as cutting speed, feed rate and depth of cut affect
the production rate, quality and cost of a component during a machining
operation. Selection of process parameters is traditionally carried out by
process planners or experienced machinists, either on the basis of their
experience on the shop floor or with the help of handbooks, for achieving the
desired surface quality on the component. In the past decade, process
planners have optimized the machining parameters using some optimization
algorithms. The optimization of machining parameters is a nonlinear
algorithm with constraints, so it is difficult for traditional optimization
algorithms to solve this problem because of the problems of convergence
speed or accuracy. Recently, process planners have started using artificial
intelligent techniques, such as neural networks, fuzzy logic and genetic
algorithms to select the machining parameters and have made some progress.
This paper presents a neural network model for CAPP to automatically select
machining parameters for rotational parts taking into consideration the
technological and economical constraints. A back-propagation algorithm is
used in the training of the network with more than 150 training patterns. The
applicability of the proposed model is illustrated by comparing the results
obtained for several samples with ones obtained from optimization model and
have shown good fit.

1 Introduction
In a Computer Aided Process Planning (CAPP) system, the machining parameters have to be
selected automatically. The approaches used in the existing computerized selection systems
for the machining parameters fall into one of the four categories, namely data storage and
retrieval, empirical method, expert system and mathematical modeling. The storage and
retrieval procedure requires a large amount of memory space for storing data, and the data it
provides are often too conservative and not optimal. The systems based on empirical
approaches reduce the data to an empirical form whereas the expert system approach uses the
stored data or the empirical equations along with a knowledge base [Shunmugam et al.,
2000]. The mathematical modeling is using optimization algorithms with constraints. It is
difficult for traditional optimization algorithms to solve this problem because of the
problems of the convergence speed and accuracy [Liu & Wong, 1999].

Recently, process planners have started using artificial intelligent techniques, such as neural
networks [Santoch & Dini, 1996], fuzzy logic [EL Baradie, 1997, Amaitik & Kilic, 2002]
and genetic algorithms [Shunmugam et al., 2000] to select the machining parameters and
have made some progress. In this paper, a neural network model for CAPP is proposed to
automatically select the machining parameters for rotational parts taking into consideration
the technological and economical aspects.

2 Neural Networks
Neural networks can be defined as a computing system made up of a number of simple,
highly interconnected processing elements, which process information by their dynamic state
response to external inputs. Neural networks are typically organized in layers. Layers are
made up of a number of interconnected nodes, which contain an activation function. Patterns
are presented to the network via the input layer, which communicates to one or more hidden
layers where the actual processing is done via a system of weighted connections. The hidden
layers the link to an output layer where the answer is output as shown in fig. 1.

Figure 1 Schematical structure of a neural network

A neural network is defined mainly by three features: topology, functionality and learning.
Topology refers to the number of nodes in each layer, and the way nodes are connected.
Functionality refers to the implemented activation function in each node. Learning refers to
the learning algorithm and the values of the learning parameters (e.g., learning rates, and
momentum rates) [Jain et al., 1999]. A neural network can be trained by presenting to the
network an appropriate number of examples, each of them formed by an input and output
vectors. By means of a training algorithm, the network self-learns the examples by
modifying step by step the connection weights in order to reduce as much as possible the
network error [Santochi & Dini, 1996]. Once a neural network is trained to a satisfactory
level it may be used as an analytical tool on other data. To do this, the user no longer
specifies any training runs and instead allows the network to work in forward calculation
mode only. New inputs are presented to the input pattern where they filter into and are
processed by the middle layers as though training were taking place, however, at this point
the output is retained and no learning occurs. The output of a forward calculation is the
predicted model for the data, which can then be used for further analysis and interpretation.

3 Neural network for selection of machining parameters

There are several types of neural network models widely used currently. These models can
be classified into supervised and unsupervised learning networks. The networks can also be
classified according to the input patterns, for example binary or continuous values. In any
case, a network has to go through two phases: training and application. The training of a
network is done by exposing the network to a number of examples, each of them formed by
an input vector and a target vector. By means of a training algorithm, the network self-learns
the examples by modifying step by step the connection weights in order to reduce as much as
possible the network error. The application is carried out by presenting to the network
several input vectors and its applicability can be verified [Santoch & Dini, 1996]. Some of
these network models are widely used to perform the process planning tasks. In this paper, a
multilayer perceptron neural network is adopted to select the machining parameters for
turning operations.

Selection of machining parameters such as cutting speed, feed rate and depth of cut affect the
production time, cost and quality of a product during a machining operation. Rational
selection of machining parameters cannot be achieved unless technical specifications,
limitations and economical aspects are taken fully into consideration. Moreover, the number
of input parameters makes this selection more difficult. Neural network can be used
effectively to select the required machining parameters. To verify the applicability of neural
network in this task, a fully connected multi-layer neural network has been designed as
shown in fig. 2. The network has six input neurons receiving the coded values of the material
hardness, surface finish, tool material (high speed steel and carbide), tool life, operation type
(rough and finish cut) and economic criteria (minimum production cost and time); the output
vector is formed by three elements corresponding to the cutting speed, feed rate and depth of
cut. In the selection of machining parameters, only tool carbon steel is considered as a
workpiece material.

4 Training the neural network

Back-propagation (BP) is a widely used algorithm for training of multilayer perceptron for
the following reasons: (1) it is most representative and commonly used algorithm and is
relatively easy to apply; and (2) it has been proven to be successful in practical applications
[Tsai, 1999]. The implementation of this algorithm for training the designed neural network
is summarized by the following steps.

Step 1 Set all the necessary network parameters such as the number of input neurons, the
number of hidden layers and the number of neurons included in each hidden
layer, the number of output neurons, etc.
Step 2 Set all network weights to small random values, positive and negative.
Step 3 Initialize the iteration (epoch) number and presentation (example) number.
Step 4 Apply one training sample to the input layer and note the corresponding desired
output.
Step 5 Calculate the output of the neurons layer by layer through the network, from input
to the output layer.
Step 6 Calculate the error gradient for every neuron in every layer in backward order
from output to the first hidden layer.
Step 7 Adjust the network weights starting from the output layer neurons working back
to the first hidden layer.
Step 8 Repeat actions in steps 4 to 7 for every training sample.
Step 9 Calculate the average sum-squared error resulted at the end of every training
cycle.
Step 10 Compare the average sum-squared error with the tolerance value of the error, if it
is less then stop. Otherwise, increase number of iterations and randomize the
order in the training set and return to step 4.

Figure 2. Fully connected multi-layered neural network

4.1 Selection of training patterns

A successful neural network requires that the training data set and training procedure be
appropriate to the problem. The training data set must span the total range of input patterns
sufficiently well so that the trained network can generalize about the data. In order to have
extrapolation and interpolation capabilities, neural networks must be trained on a wide
enough set of input data to generalize from their training sets. To achieve this goal and
demonstrate the applicability of the designed neural networks, more than 150 training
patterns (each pattern is formed by input and output vectors) are generated to select
machining parameters for turning operation. The input vector values of these patterns are
selected from within specified range for each input parameter. Table 1 presents selected
samples of these patterns.
 Input vector Output vector
Material Surface Tool Tool Operation Econ. Cutting Feed Depth
Hardness Finish Mat. Life Type Criteria Speed Rate of Cut
BHN µm - min. - - m/min mm/rev mm
70 3 HSS 15 Rough Cost 125.95 0.1 1.1
90 3 Carbide 15 Rough Cost 80.02 0.32 1.1
110 3 HSS 25 Rough Cost 94.5 0.25 1.1
130 3 Carbide 25 Rough Cost 74.63 0.41 1.1
150 3 HSS 15 Rough Time 139.05 0.38 1.2
170 3 Carbide 15 Rough Time 112.54 0.58 1.2
190 3 HSS 25 Rough Time 119.68 0.6 1.2
210 3 Carbide 25 Rough Time 98.81 0.87 1.2
230 0.6 HSS 15 Finish Cost 52.08 1.49 0.9
250 0.6 Carbide 15 Finish Cost 56.26 0.98 0.9
270 0.6 HSS 25 Finish Cost 50.19 1.7 0.9
290 0.6 Carbide 25 Finish Cost 56.26 0.98 0.9
310 0.6 HSS 15 Finish Time 84.54 1.7 0.7
330 0.6 Carbide 15 Finish Time 94.68 0.98 0.7
350 0.6 HSS 25 Finish Time 84.54 1.7 0.7

Table 1 Training patterns for turning machining parameters.

Several training experiments have been performed to select the optimal structure and training
parameters of the neural networks. In this work, the optimal structure of the designed
network, which gave minimum error, is summarized as follows: the input layer consists of 6
neurons, two hidden layers with 8 neurons each and the output layer consists of three
neurons. The learning parameters are 0.9 for momentum and 0.95 for learning rate. An
average error of 3x10-5 has been reached after the 160 training samples (selected for training
the designed network) have been presented to the network 15000 times. Figure 3 shows the
training progress of the network. The graph illustrates downward movement of the error rate
as learning progressed, indicating that the average error decreased between actual and
predicted results.

Figure 3 Training progress of the network

5 Implementation of the proposed model in CAPP
The trained neural network can be implemented in CAPP system to carry out the task of
machining parameters selection. Compared with traditional approaches, neural network
model has the advantage of fast inference and high working efficiency. This implementation
can be done by training one neural network for each workpiece material and the trained
weights are stored in the CAPP system. Based on the input parameters, the appropriate
network will be used. But, before implementing the trained network in the CAPP system, the
performance of the network output and the generalization ability have to be verified.
Usually, this can be done by presenting to the network several input vectors. The network
output is recorded and compared with the actual one. If the error is lower then the
permissible one, the network is then validated and can be used. Table 2 shows the results
obtained by presenting selected number of input parameters.

Mat. Sur. Tool Tool Ope. Econ. Optimization Neural network

Hard. Fin. Mat. Life Type Cri. v F d v f d
- µm - min - - m/min mm/rev mm m/min mm/rev mm
70 3 HSS 15 Rough Cost 125,95 0,1 1,1 124,59 0,12 1,11
90 3 HSS 15 Rough Cost 106,76 0,17 1,1 106,85 0,18 1,12
110 3 HSS 15 Rough Cost 94,5 0,25 1,1 92,688 0,25 1,12
130 3 HSS 15 Rough Cost 85,38 0,34 1,1 81,735 0,35 1,12
150 3 HSS 15 Rough Cost 77,71 0,45 1,1 73,385 0,47 1,12
170 3 HSS 15 Rough Cost 71,87 0,57 1,1 67,022 0,61 1,12
190 3 HSS 15 Rough Cost 66,75 0,71 1,1 62,136 0,77 1,12
210 3 HSS 15 Rough Cost 62,91 0,85 1,1 58,337 0,94 1,11
230 3 HSS 15 Rough Cost 59,46 1,02 1,1 55,347 1,12 1,11
250 3 HSS 15 Rough Cost 56,33 1,19 1,1 52,967 1,31 1,11
270 3 HSS 15 Rough Cost 53,47 1,38 1,1 51,057 1,51 1,11
290 3 HSS 15 Rough Cost 51,46 1,59 1,1 49,516 1,7 1,11
310 3 HSS 15 Rough Cost 48,99 1,8 1,1 48,271 1,89 1,11
330 3 HSS 15 Rough Cost 47,29 2,03 1,1 47,268 2,08 1,12
350 3 HSS 15 Rough Cost 45,71 2,27 1,1 46,466 2,27 1,12

Table 2 Results generated by neural networks and optimization techniques

In fig. 4, the cutting speed values generated from neural network are plotted with the values
calculated using optimization model [Sari et al., 2002]. The results presented in fig. 4 show a
very good correlation between the cutting speed values calculated form the optimization and
those generated using neural network. This shows the applicability of neural networks in
machining parameters selection of automated process planning.

6 Conclusion
The neural network model proposed in this paper demonstrates the possibility of application
neural networks in CAPP systems to select machining parameters. The main advantages of
this application, in comparison with traditional approaches are outlined as follows: (1) fast
inference and high working efficiency, (2) Its ability to perform the new inference procedure
by training neural networks with new examples, and (3) Its adaptability to the dynamic
manufacturing environment, owing to efficient knowledge acquisition capability. Further
work in this research will include more workpiece and tool material combinations and
implementation of the model in a hybrid neural-fuzzy automated process planning system.

Figure 4 Comparison of cutting speed values

7 Acknowledgement
We would like to express our gratefulness and appreciation to the Turkish State Planning
Agency, Turkish Scientific and Technique Research Association and Graduate School of
Natural and Applied Sciences of METU for their support throughout this research.

8 References
Amaitik, S. M. & Kilic, S. E. (2002), Fuzzy logic model for selection of machining
parameters in CAPP", Annals of DAAAM 2002, 11-12.
Chao, P. & Hwang, Y.D. (1997), An improved neural network model for the prediction of
cutting tool life, Journal of Intelligent Manufacturing, 8: 107-115.
EL Baradie, M. A. (1995). A fuzzy logic model for machining data selection, International
Journal of Machine Tool & Manufacture, 37, 9: 1353-1372.
Jain, R.K., Jain, V.K & Kalra, P.K. (1999), Modelling of abrasive flow machining process: a
neural network approach, Wear, 231:242-248.
Liu, Y. & Wang, C. (1999). A modified genetic algorithm based optimization of milling
parameters, Advanced Manufacturing Technology, 15: 296-799.
Santochi, M. & Dini, G. (1996), Use of neural networks in automated selection of
technological parameters of cutting tools, Computer Integrated Manufacturing Systems,
9, 3:137-148.
Sari, B., Kilic, S.E, Anlagan, O, and Sen, D.T. (2002), Web-based optimization system for
machining operations, UMTIK 2002 The 10th International Conference on Machine
Design and Production, 259-278.
Shunmugam, M.S., Reddy, S.V. & Narendran, T.T. (2000), Selection of optimal conditions
in multi-pass face-milling using a genetic algorithm, International Journal of Machine
Tool & Manufacture, 40: 401-414.
Tsai, Y., Chen, J.C. & Lou, S-J. (1999), An in-process surface recognition system based on
neural networks in end milling cutting operations, International Journal of Machine
Tools & Manufacture, 39: 583-605.
Zhang, H.C. & Huang, S.H. (1995), Applications of neural networks in manufacturing: a
state-of-the-art survey, International Journal of Prod. Res., 33, 3: 705-728.