Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
1|Page
ABSTRACT
2|Page
CONTENTS :
TITLE Pg.no
1. INTRODUCTION………….……………………………………………………………………………………………….4
1.1 INTRODUCTION TO EDM………………………………………………………………………………………………………4
3. CONCLUSION…………………………………………………………………………………………………………….17
4. REFERENCES……………………………………………………………………………………………………………..19
3|Page
INTRODUCTION
EDM founds its wide applicability in manufacturing of moulds, forging dies, press
tools, die castings, automotive, aerospace and surgical components. As EDM does not make
direct contact (an inter electrode gap is maintained throughout the process) between the
electrode and the workpiece it’s eradicates mechanical stresses, chatter and vibration
problems during machining. Various types of EDM process are available, but here the
concern is about die sinking type EDM machines which require the electrode to be
machined in the exact opposite shape as the one in the workpiece.
4|Page
As the workpiece remain fixed by the fixture arrangement, tool helps in focusing the
discharge or intensity of generated heat at the place of shape impartment. Application of
focused heat raises the temperature of workpiece in the region of tool position, which
subsequently melts and evaporates the metal. In this way the machining process removes
small material of workpiece material by the mechanism of melting and vaporization during
a discharge. The material removed by a single spark is in the range of 10−6 ‑10−4 𝑚𝑚3, but
this basic process is repeated typically 10,000 times per second. Figure 1.1 shows a layout
of Electric Discharge Machining.
Figure 1.1
5|Page
The erosion process due to a single electric spark in EDM generally passes through
the following phases in order:
(1) Pre-breakdown:
In this phase the electrodes move close to the workpiece and the voltage is applied
between the electrodes. Open circuit voltage Vo .
(2) Breakdown:
When the applied voltage crosses the boundary limit of dielectric strength of used
dielectric fluid, the breakdown of the dielectric is initiated. Generally, the dielectric starts to
break near the closest point between tool and workpiece, but it will also depend on
conductive particles between the gap if present any [3]. When the breakdown occurs the
voltage falls and a current rises abruptly. In this phase the dielectric gets ionized and a
plasma channel is introduced between the electrodes and also there is possibility of
presence of current.
(c) Discharge:
In this phase the discharge current is maintained at constant level for a continuous
bombardment of ions and electrons on the electrodes. This will cause strong heating of the
work-piece material (and also of the electrode material), rapidly creating a small molten
metal pool at the surface. Also, a small amount of metal can have directly vaporized due to
the tremendous amount of heat.
During the discharge phase, the plasma channel grows; therefore, the radius of the
molten metal pool also increases with time. In this phase some portion of the work-piece
will be evaporated and some will be remaining in molten state. The Inter Electrode Gap
(IEG) is an important parameter throughout the discharge phase. It is estimated to be
around 10 to 100 micrometres (IEG increases with the increase in discharge current).
At the end of the discharge, current and voltage supply is shut down. The plasma
collapses (since current is supply is stopped, there will be no more spark) under the
pressure enforced by the surrounding dielectric.
6|Page
(e) Post-discharge:
There will be no plasma in this stage. Here a small portion of metal will be machined
and a small thin layer will be deposited because of plasma is rapidly cooling. This layer is
(20 to 100 microns) is known as white layer. Consequently, the molten metal pool is
strongly sucked up into the dielectric, leaving a small crater on the work-piece surface
(typically 1-500 micrometers in diameter, depending on the current).
7|Page
1.3 INTRODUCTION TO ANN :
ANN refers to the computing systems whose fundamental concept is taken from
analogy of biological neural networks. Many days to day tasks involving intelligence or
pattern recognition are extremely difficult to automate, but appear to be performed very
easily by animals. The neural network of an animal is part of its nervous system, containing
a network of specialized cells called neurons (nerve cells).
Each artificial neuron receives signals from the environment, or other artificial
neurons, gather these signals, and when fired transmits a signal to all connected artificial
neurons. Input signals are inhibited or excited through negative and positive numerical
weights associated with each connection to the artificial neuron. The firing of an artificial
neuron and the strength of the exciting signal are controlled via. a function referred to as
the activation function.
The summation function of artificial neuron collects all incoming signals, and
computes a net input signal as the function of the respective weights and biases. The net
input signal serves as input to the transfer function which calculates the output signal of
artificial neuron.
8|Page
ANN PERFORMANCE ANALYSIS
The most familiar process parameters that are varied to obtain an efficient ANN
model are ANN architectures; learning/training algorithms and nos. of hidden neuron.
These parameters have been chosen here as process parameters to a full factorial design.
The process parameter ANN architecture at two levels, learning/training algorithm at three
levels and nos. of hidden neuron at four levels have been selected as shown in Table 3.1.
The performance parameters for evaluating the ANN model are taken as training Mean
squared error (MSE), testing MSE, training Correlation coefficient (R) and testing R, which
are the default performances evaluating parameters assumed by MATLAB.
A single neuron is not enough to solve real life problems (any linear or nonlinear
computation) efficiently, and networks with more number of neurons arranged in particular
sequences are frequently required. This particular ways /sequences of arrangement of
neurons are coined as neural architecture. The sequences of arrangements of neurons
determine how computations will proceed and also responsible for the effectiveness of the
model.
9|Page
Here a single hidden layer has been used throughout, as single hidden layer is
sufficient for back-propagation neural network to define the input-output mapping. The
nos. of neuron in the input layer and the output layer are fixed as the input parameters and
response parameters of EDM process are fixed for this particular investigation and also the
mapping is to be established between input and output parameters .So to investigate the
effect of nos. neuron on model performance nos. of hidden neuron has been varied at four
levels.
The error function that has been used here for supervised training is average mean
squared error function (Eavg ). Mathematically it can be expressed as:
Since we are operating with multiple outputs in neural networks the MSE is
expressed as the following
Where dnk the desired output for exemplar n at neuron k of output layer and ank is
the network output for exemplar n at neuron k of output layer (i.e. the predicted output of
neural network); K is the number of neurons in the output layer and N is the number of
exemplars in the data. Mean squared error (MSE) is two times of the average mean squared
error function. The factor 0.5 is multiplied here with the mean squared error function to
make the differentiation of this function easier. Lower value of MSE is preferable for a
superior ANN model
CORELATION FACTOR:
Correlation coefficient can be used to determine how well the network output fits
the desired output. The correlation coefficient between a network output (a) and a desired
output (d) can be mathematically defined as:
10 | P a g e
Similarly for the neural networks the correlation coefficient is defined as follows.
where n = exemplar or run number, an and dn are the network output and desire
output respectively at a particular exemplar, and are the data mean of network output and
desire output respectively. Higher value of R is desirable for an effective ANN model.
The following data set is obtained for EDM process of AISI D2 Steel. The following
data has 120 data points and the data is separated as 70 : 15 : 15 for training, validation and
testing data for the network.
RESULTS :
The input data is loaded in MATLAB. The neural network tool is initiated and no of
neurons in hidden layers is varied from 6 to 16. The MSE and R of each network is obtained
and tabulated as follows.
No of neurons r mse
6 0.990628 4.0914
8 0.997973 0.978745
10 0.996883 1.3272
16 0.99724 2.01249
MSE
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
6 8 10 16
MSE
From this we can observe that the MSE is minimum for 8 neurons in hidden layer.
14 | P a g e
REGRESSION PLOT
A higher R value indicates there is a good correlation between the input and the
output from the given regression analysis. The R value is 0.99797 for the given neural
network system.
15 | P a g e
The following MATLAB code is used to create the NN.
%%createnetwork.m
data = xlsread('edm_nn_data.xlsx');
x = data(:,1:4);
x = x';
ymrr = data(:,5);
ymrr = ymrr';
ytwr = data(:,6);
ytwr = ytwr';
nnstart;
16 | P a g e
Once the neurons if finalised, the training data is trained and saved for further
reference. The test sample which was not part of training was used to check the accuracy of
the system. The results are as follows. The MATLAB code for analysing the data,
%%test.m
load nnmrr;
load nntwr;
testdata = xlsread('edm_test.xlsx');
testx = testdata(:,1:4);
testx = testx';
wear = nntwr(testx);
mrrate = nnmrr(testx);
S.NO Ip Ton Tau V MMR TWR ANN pred. ANN %error %error
(A) (microsec) (%) (volt) (mm3/min) (mm3/min) MRR pred. in MRR in TWR
TWR
The average error for the ANN predicted MRR and the actual MRR is 3.96% and the
error in tool wear rate is 36.1%. The errors in the MRR is well below the acceptable range.
The tool wear error can be reduced by adding more data points to network and retrain it as
only 120 data points are used for training the NN. The actual vs predicted MRR and TWR
are plotted and shown.
17 | P a g e
ANN PREDICTED vs ACTUAL MRR
70
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
1.8
1.6
1.4
1.2
0.8
0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
actual wear precdicted
18 | P a g e
FURTHER SCOPE:
(1) Since this project only considered MLP algorithm with varying neurons, we can use
some other algorithms like genetic algorithm and perform a similar analysis.
(2) This can be used for optimisation of the EDM process reducing cost and material
waste.
(3) Surface roughness can also be added as output to this network and a relation can be
developed. Since we had only limited data source it has been restricted to MRR and
TWR.
(4) The most important proceedings is we can do the inverse problem of finding the
optimal parameters for given MRR and Surface roughness values. If sufficient data
points are available/collected the network can be used for finding the cause for
effect without any prior knowledge about the inverse calculations.
REFERENCES
19 | P a g e