Transcript Power point

An Artificial Neural Network
Approach to Surface Waviness
Prediction in Surface Finishing
Process
by Chi Ngo
ECE/ME 539 Class Project
Problem description: In order to control the
surface finishing results in automated surface
finishing process, it is important to specify the
relationship between surface waviness and
cutting parameters such as feed rate, cross
feed, tool displacement, and spindle speed. In
this project, an artificial neural network is
used to predict surface waviness (output) from
feed rate, cross feed, tool displacement, and
spindle speed (inputs).
Outline: The data file consists of six columns.
The first four columns represent feed rate,
cross feed, tool displacement, and spindle
speed, and are used as feature vectors. The
last two columns represent depth of cut and
surface waviness, and are used as target
vectors. I have collected twenty-eight real
sets of data performed on actual CNC
machine and workpieces. These sets of data
will be used to train and test the neural
network.
Method: Back propagation perceptron
learning would work best for this application.
A multi-layer perceptron neural network back
propagation algorithm with one hidden layer
is implemented in this project. Since this is
application problem, bpappro.m is used with
some modifications.
output
.3
.4
.1
.9
.6 Target vector
.1
.1
.1
.2
.3
Multi-layer perceptron
Input vector
Results: parameters used that produce best
results:
For depth of cut prediction:
-Learning rate (between 0 and 1) = 0.3
-Momentum constant (between 0 and 1) = 0.8
-Hidden neurons: 15
-Maximum number of epochs to run = 800
-Epoch size (# of samples) = 25
Training data set results:
-Mean square error = 0.07362
-Maximum absolute error = 0.000963
For waviness prediction:
-Learning rate (between 0 and 1) = 0.2
-Momentum constant (between 0 and 1) = 0.8
-Hidden neurons: 24
-Maximum number of epochs to run = 800
-Epoch size (# of samples) = 26
Training data set results:
-Mean square error = 7.8705
-Maximum absolute error = 3.839
Conclusion/Discussion:
Multi-layer perceptron neural network back
propagation algorithm seems to work better
than time series prediction and other methods
that could be used for this project. But still,
the mean square error is big. It is very
difficult for the program to converge. Other
existing mathematical models to predict depth
of cut and waviness seem to work much better
than my neural network in this project.