Abstract:
Yarn evenness is one of the important yarn quality
parameters, which affects the subsequent pro-
cessing and usage of the yarn. This research work
concentrated on the study of Kenyan ring spun yarn
evenness. Cotton and yarn samples were collected
from factories in Kenya. The data collected from the
samples was used to design yarn evenness predic-
tion models. The yarn evenness prediction models
were trained using Levenberg-Marquardt Backpropa-
gation algorithm (LMBP), one of the most commonly
used algorithms in the prediction of yarn properties.
This research work went a step further and attempt-
ed to improve the working of the evenness prediction
models using hybrid algorithms, namely Differential
Evolution and LMBP (LMBP-DE) and LMBP and
Particle Swam Optimization (PSO), christened LMBP
-PSO. A comparison of the performance of the yarn
unevenness prediction algorithms as the neurons in
the hidden layer were varied from 2 to 16 in steps of
twos was undertaken. In terms of speed (training
time) and performance (mse value) the hybrid algo-
rithms performed better than the LMBP algorithm.
LMBP also showed a higher CV of mse values which
could imply that it was more prone to getting stuck in
local minima than the hybrid algorithms.
