Prediction of Surface Roughness for CNC Turning of EN8 Steel Bar Using Artificial Neural Network Model

In this world of cut-throat competition, significance of quality products has grown manifold. Producing quality products at reduced cost has become a challenge to manufacturing organization. Machining parameter not only affect the manufacturing cost and productivity but also quality of the final product. There are several parameters, which affect the surface roughness of a component and can be classified into four main groups, machining parameters, like spindle speed, feed rate, depth of cut etc., machining and machine tool conditions, like dry or wet turning, type of cutting fluid, cutting tool parameters, such as tool geometry, tool material, tool wear, tool nose radius and material properties of work piece [1]. Surface roughness can be optimized for given set of parameters using techniques such as Taguchi method, response surface method, genetic algorithm, ANN etc. [2-6]. Mathematical modeling of a process reduces the effort, save money and time for optimal and efficient implementation of any process. Many investigations have been made in this direction using ANN modeling; however, modeling is machine tool and process parameters specific and needs to be done for specific process.

Tsai et al. (1999) developed an ANN model to predict the surface roughness of End-milled part. An accelerometer and a proximity sensor were employed during cutting to collect the vibration and rotation data. Spindle speed, feed rate, depth of cut and the average vibration per revolution as input layer and surface roughness (R_a) value as output layer were used for the ANN [7]. Karayel (2009) developed ANN model to Predict and control of surface roughness for steel on CNC lathe. He obtained average absolute error of 2.29 % for ANN model and 10.75 % absolute error for regression model [8]. Patel and Pal employed the Taguchi design method and ANOVA to optimize the surface roughness quality in milling of aluminium. They found that spindle speed is most and depth of cut is least significant parameter [9].

From the literature review, it was found that there are various techniques used for model preparation and surface roughness prediction like response surface methodology (RSM), Genetic algorithm (GA), Statistical regression technique, ANN etc. However, in all the cases, the efficiency and robustness of the model play an important role to predict the accurate results.

In this paper, authors have discussed the effect of machining parameters on surface roughness. The study has been carried out on computer numeric control (CNC) lathe “MCL10”. The significant machining parameters namely spindle speed, feed rate and depth of cut were chosen to be optimized to get better surface finish of EN8 steel work-piece in turning operation. The objective of the present study is to develop an artificial neural network (ANN) model for CNC turning operation to predict surface finish of EN8 bar.

Design of Experiments (DOE) refers to the process of planning, designing and analyzing the experiment so that valid and objective conclusions can be drawn effectively and efficiently.

The methodology of DOE is fundamentally divided into four phases; 1) Planning phase: Problem recognition, formulation and selection of process and response parameters, determination of levels etc. 2) Designing phase: selecting the most appropriate design for the experiment, 3) Conducting phase: Conduction of experiments as per design and 4) Analysing phase: Analysis of results from different perspectives.

Factorial design enables an experimenter to study the joint effect of the factors on a response. A factorial design can be either full or fractional factorial. A full factorial design consists of all possible combinations of levels for all factors. Total number of experiments for studying k factors at N-levels is N^k. In present work, three most influential turning parameters namely spindle speed, feed rate and depth of cut, each with three levels are selected as shown in Table 2. Full factorial design is chosen, so that a good number of results can be obtained to train and test the ANN model. Total 27 (3³) experiments were designed and conducted as shown in Table 3.

Table 2. Machining parameters with levels

figure_1.jpg

Figure 1. MCL10 Lathe, CIM Lab, JUET

Artificial neural network (ANN) is an artificial intelligence tool for machine learning. ANN consists of input, hidden and output layers as shown in Figure 4.

figure_4.jpg

Figure 4. Artificial neural network model

Hidden layer consists of units which convert the input values into some intermediate values which are further used by output layer. Initially, a network is created in the training of a feed forward network [10-11]. The ANN knowledge in interconnection weights is adjusted during training process. Independent variables in the model is equal to input neurons, whereas the output neurons are in same number as the functions to be approximated. Here, inputs to the ANN model are the cutting parameters, namely depth of cut, feed rate and cutting speed. The ANN model output parameter is surface roughness. The input layer has three neurons, hidden layer has ten neurons and output layer has one neuron. ANN training and architectural parameters have to be appropriately selected so that an optimal ANN model for the prediction of surface roughness is developed [12]. The most elementary method for selection of parameter settings is the use of a trial and error method [13]. The most important training and architectural parameters are: learning rate, momentum, number of hidden layers, and number of neurons in hidden layer. The number of neurons in the hidden layer is a primary parameter of the ANN architecture. Due to the full interconnection used between neurons, any increase in the neuron number leads to increasing net complexity as well as training time. On the other hand networks with more neurons in the hidden layer may solve more complex problems. However, an ANN with too many hidden neurons will generalize very badly, which is a reflection of well-known over-fitting problem.