Digital Signal Processing Reference
In-Depth Information
Programmable function generator
RS-232C
ISA
Computer IBM PC
Signal Processor board
with A/D converter
Fig. 12.23
Layout of the laboratory stand with TMS signal processor
such a low calculation speed was the exponent function, realization of which is
quite time-consuming in real time. Execution of the algorithm with activation
function satlins (ANN4) took ca. 960 ls which is a bit less than the time between
two consecutive signal samples. Possibilities of time savings were searched by
modernizing the algorithm from the numerical point of view. Time-consuming
loop instructions were replaced with multiple repeated simple operations. This
enabled shortening the execution times ca. by half, i.e. to 650 ls for the network
ANN2 and to 420 ls for the network ANN4. It is to be stressed that the real-time
tests were done with relatively slow signal processor and even by such a hardware
arrangement the required time for execution of the algorithms of the neural CT
compensators within one time-step were close to 0.4 ms. Deeper entry in
assembler could bring about considerable shortening of the algorithm execution
time. With faster signal processors (available at present above 200 MHz) and the
entire program code written in the machine language even 10-time higher sam-
pling rates could be used. That means that practical implementation of the ANN-
based CT saturation compensators is utterly attainable.
References
1. Bothe HH (1998) Neuro-Fuzzy-Methoden. Einfuehrung in Theorie und Anwendungen.
Springer-Verlag GmbH, Berlin Heidelberg, ISBN: 978-3-540-57966-4
2. Bunyagul T, Crossley P, Galac P (2001) Overcurrent protection using signals derived from
saturated measurement CTs. In: Proceedings of PES summer meeting, vol 1. Vancouver,
pp 103-108
3. Chen KW, Glad ST (1991) Estimation of the primary current in a saturated transformer. In:
Proceedings of the 30th conference on decision and control, vol 3. Brighton, England,
pp 2363-2365
Search WWH ::
Custom Search