تطوير وتصميم شبكة عصبية اصطناعية للتحكم الصوتي فينظومة التحكم
Files
Date
2008
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
جامعة النيلين
Abstract
Abstract
This thesis has two main goals, the first goal is development and design
artificial neural network architecture for specific application which is voice
control, in addition to increase network generalization when it deals with
new cases, this because human sound is not constant and vary each time he
utterances the same word according to his health and emotional state. To
achieve this goal, number of feed forward back propagation neural networks
was built, and three types of transfer function was used which are Log
sigmoid, Hyperbolic tangent sigmoid and Pure Liner to test performance for
each of them, the experiences explained that Log sigmoid is much suitable
for this application. Also this work follow effective manner in choosing
numbers of layer and hidden nodes depending on trail and error but with
specific strategy.
The second goal is using the designed neural network to work as commands
recognizer in voice control system, this system is a pressure control system
which controls on petroleum pipelines and consist of pump, tank and
electronic valve. Three experiments was applied to design this system, the
designing on first experience depend on using isolated commands (Open,
Close) and training neural network to recognize it, and then measured
performance of the recognizer. The neural network here used log sigmoid
for activate layer output and contain two hidden layer, number of nodes in
them are 120 and 50 nodes. This design achieve performance ratio
98.5714%. The second designing using commands from two part
(Open pump ,Close pump, Open Valve , Close valve ), the neural network
used here has 70 nodes and one hidden layer between input and output layer.
This design achieve perfonnance ratio 96.67%.The recognizer in last
designing used phonemes of command to recognize it, this for make the
system more general. The neural network used here has one hidden layer
contain 100 nodes. This design achieve performance ratio only 23.3%.
The three experiences were designed using MATLAB program, it used also
for write preprocessing program which was use to extract word features.
Description
Keywords
الشبكات العصبية, تصميم الشبكات