Neural networks - Signals, Processes, and Systems - page 448

signals time domain si g nal ro uter time del ay mo m. s i gnal par amet er net progn osi s net outp ut formula act i on

_pa rame ter01 counter01 cyc les01

P ro gnos i s

Triangle

*.nn-file

*.nnd-file

Illustration 289: Neuronal network to recognize 4 different periodic signals
The Illustration shows the crucial processes of collecting training data, identifying the network structure
and network training. With the help of a block diagram, you can see: at first, the four signals are generated
with the generator module, then one after another is selected by a switch, then fed to the parameter
module. The 6 parameters (max. pos. gradient, max. neg. gradient, integral, average, final value and
standard irregularity) are filed in the *.nnd file. Every single signal receives a “target value”, e.g. triangle
= 1, sinus = 2 etc. . At the moment, the triangle signal with the target value 1 (at the exit of the neuronal
network) is processed.
Looking at the *.nnd file on the left shows how reasonable the parameters were chosen. They make the four
signals perfectly distinguishable. Each of the four signals has been read 10 times. The target values each
stand on the right side and represent (coded) the signal type.
The very clearly arranged *.nn- (network) file provides any data about network structure and network
size. This example shows very clearly the direct dependence of the network´s size from the number of
target values. Not yet recognizable is the strong similarity between sinus and triangle. This is shown in
Illustration 289 .