Professional Documents
Culture Documents
Xiaofen Ji2
Prediction of Garment Production Cycle
Time Based on a Neural Network
DOI: 10.5604/01.3001.0014.5036
1 Zhejiang Sci-Tech University, Abstract
College of Textile Science and Engineering, The process of garment production has always been a black box. The production time of
Hangzhou, 310000, P.R. China, different clothing is different and has great changes, thus managers cannot make a produc-
* e-mail: owencho@163.com tion plan accurately. With the world entering the era of industry 4.0 and the accumulation
2 Zhejiang Sci-Tech University, of big data, machine learning can provide services for the garment manufacturing industry.
School of International Education, The production cycle time is the key to control the production process. In order to predict
Hangzhou, 310000, P.R. China the production cycle time more accurately and master the production process in the garment
manufacturing process, a neural network model of production cycle time prediction is esta-
blished in this paper. Using a trained neural network to predict the production cycle time,
the overall error of 6 groups is within 5%, and that of 3 groups is between 5% and 10%.
Therefore, this neural network can be used to predict the future production cycle time and
predict the overall production time of clothing.
GENERAL PROBLEMS OF THE FIBRE AND TEXTILE INDUSTRIES
Key words: garment production, big data, cycle time, neural network, prediction.
Research methodology GST per capita: General sewing time the production cycle time is higher than
(GST) is a pre-set action time system spe- the GST per capita due to the low pro-
An artificial neural network (ANN) cially designed for the clothing industry, ficiency of workers. The initial cycle
is a mathematical model based on the which originated from the production time is the highest point on the produc-
working principle of a biological neural therblig analysis proposed by F.B. Gil- tion learning curve, which determines
network. A neural network is one of the breth. The GST programs the common the starting point of the learning curve.
most basic forms of machine learning. In actions in the sewing process and describe The initial cycle time needs to be de-
the textile and garment industry, a neu- the actions in code form. Each code has termined according to the bottleneck
ral network has been used to deal with a clear time value, and then a standard process on site, which cannot be calcu-
complex non-linear sewing process
problems, suchandas describe the actions in code form. Each code has a clear time value,
time is estimated according to the moving lated before production, and can reflect
prediction of fabric size
andand performance
then a standard time is estimated
distance and actionaccording
difficulty. to the moving distance
In a garment and action
the production difficulty of a garment to
[14], classification of fabric defects [15], factory, industrial engineers account for a certain extent.
difficulty. In a garment factory, industrial engineers account for the GST of each
human body recognition [16], evalua- the GST of each garment. The calculation
tion of the sewn thread consumption
garment. of formula
The calculation formula of GST
of GST perper capita
capita is: is: Building neural network
jean trousers and so on [17]. In this pa-
The structure of a BP neural network for
per, a BP (back propagation) neural net-
clothing production cycle time prediction
work is used to predict the learning curve
of the garment production cycle time. (1) is shown in Figure 3. There are three lay-
which reflects the production beat of the water line in an ideal ers: state.theHowever, in hidden layer and out-
input layer,
The curve is affected by many factors,
which reflects the production beat of the put layer. The input layer is responsible
but the logic relationactual production,
is difficult to ex-the GST per capita can only be used as a reference value. The final
water line in an ideal state. However, in for receiving data, with three variables
press with an accurate function relation.
stable cycle time may actual be larger than thethe
production, GSTGSTperpercapita
capitaduecan to factors
x1, x2 andsuch
x3 as the
representing the production
A BP neural network is a kind of forward
only be used as a reference value. The fi- quantity, GST per capita and initial cycle
tutor learning neural bottleneck
network withprocess
erroror complex style. It may also be smaller than the GST per capita due
nal stable cycle time may be larger than time.
reverse propagation. It does not need to
to factors such as workers the being
GST per more familiar
capita due to with the style
factors such and
as higher efficiency.
determine the mapping relationship be-
the bottleneck process or complex style. The hidden layer is responsible for data
tween the input and output in advance,
It may also be smaller than the GST per decomposition and processing. There are
but only learns someInitial
rules cycle
through its when the first garment of a production order is produced, the
time: capita due to factors such as workers be- 10 neurons, each of which is a processing
own training. Its basic idea is the gradi-
production cycle time is inghigher
more than
familiar
the with
GSTthe perstyle anddue
capita high-
to the unit
low receiving input
proficiency of from the upper layer.
ent descent method, which uses gradient
er efficiency. The activation function, also called the
search technology to workers.
minimiseThe the initial
error cycle time is the highest point on the production learning curve,
mapping function tansig, can transform
mean square deviation of the actual out-
which determines Initial cycle time: when the first gar-
the starting point of the learning curve. The initial the cycleresult of calculation
time needs to in a non-line-
put value and the expected output value
ment of a production order is produced, ar way, so as to improve the expression
of the network. be determined according to the bottleneck process on site, which cannot be calculated
Training parametersbefore production, and can reflect the production difficulty of a garment to a certain
Input layer Hidden layer Output layer
Production quantity: extent.
the larger the pro-
duction quantity is, the more skilled
the workers can be. The stable cycle
time will be greatly Building
reduced compared
neural network
with the initial cycle time, but there is
The structure of a BP neural network for clothing production cycle time prediction is
also a limit. The smaller the production
shown
quantity is, the shorter in Figuretime
the learning 3. There are three layers: the input layer, hidden layer and output layer.
of the workers is. It is possible that the
The input layer is responsible for receiving data, with three variables 、 ,
production has not reached the ultimate 10 160
representing
efficiency and the production hasthe production quantity, GST per capita and initial cycle time.
been
completed. Figure 3. Structure of cycle time prediction BP neural network.
ç l å ŷi 2 – � åççŷŷlii2å
ç çl å ÷� ÷çŷçiå
2
åŷçi ÷å
l� –çi ÷lçå
yi 2÷ŷ� å÷÷ ççyylii2å
÷� ÷çyiå�yçi ÷å÷yi ÷ ÷
2
which10
cycle time from is divided
pieces to into1600
the cycle
piecestime
in from final prediction
the production processdataof awere obtained ac- ç i=1 ç è i=1 ø ÷øè çèi=1i=1è øi=1÷ø çè øè i=1 ø ÷øè i=1 è øi=1÷ø ø ÷ø
è è i=1è i=1 ø
i=1 è i=1
10 pieces to 1600 pieces in the produc- cording to the fitting curve, and finally
t. Accordingtion to process
experience (4)
of a [18],
certainthe outputAc-
garment. layer compared
activationwith function
the real data.Where,
Where, represents
Where,represents the
represents
truethe value theoftrue
true thevalue
value ofsample,
the of the sample, -sample,
the predicted - the- predic
theval p
cording to experience [18], the output Where, yi represents the true value of the
function, layer
whichactivation
has a faster convergence speed than
function selects the Relu 1) Select 89 groupssample,
andof production
andsample,
sample, isdata,
theand
and number
isi the is number
sample, the
of samples.
number
y^i – of of
The
thesamples. samples.
predicted range The ofThe
value range coefficient
ofrange the of coeffici
of i coefficient is be
the weight of function, which layer,
the output has a faster
and convergence including
is the output layer five categories of dress, skirt,
threshold. sample, and l is the number of samples.
The closer
TheitThe is tocloser
closer 1,
it the
is it
tobetter
is1,tothe 1,thethefitting
better better
thedegree the fitting
fitting is.
degree degree is. is.
speed than sigmoid and tansig. wmn is the shirt, knitwear and coat; the minimum The range of coefficient R2 is between
ion is expressed
weight as of follows:
the output layer, and bmn is the production quantity Theisresults
181 show
The The
resultsresults
that
pieces, showtheshowfitting
that
0-1. that
thedegree
The closer the fitting
fitting of
it the
degreeis degree
totrained
of1, the ofneural
the the
trained
better trained network
neural
the neural is good
networknetw
output layer threshold. The Relu function the maximum production quantity – fitting degree is.
f (x) = max(0, x) of production of production
of productiondata ofdata prediction
data of samples
of prediction prediction aresamples
samples inputtedare are to inputted
inputted the to neural to th
the ne
is expressed as follows: 1823 pieces, the minimum GST per cap-
f (x) = max(0, x). (3) ita – 45 seconds, thesimulation maximum prediction,
simulation
simulation
GST per The results
and
prediction,
prediction, 10 groups
and 10 show
and of
groups that
10prediction
groups ofthe offitting
cycle
predictionprediction degree
time cycle can
cycle be time
time obtaine
can be ca
capita – 277 seconds, the minimum ini- of the trained neural network is good.
al 2)Due 2) to the2)
Due disturbance
Due
to thetodisturbance
the disturbance
of neural ofnetwork
tial cycle time – 60 seconds, and the max- 10 groups of production data of predic-
of neural
neural prediction
network network data,
prediction prediction the
data, latter
data,
thepredi the
latt
Experimental imum initial cycle may time is 450 seconds. tion samples are inputted towhichtheviolatesneural
, the production data of a women's clothing enterprise in Hangzhou, bemay larger
maybe thanbe larger
larger thethan former
than
the the one,
former former
which one,one, violates
which the violates
actual the produc
the
actual act
In this research, the production data of Because the input data size needs to be network for simulation prediction, and
Therefore, Therefore,
curve fitting
Therefore, curve curve
of thefitting
fitting neural
of the ofnetwork
the neural
neural output
network network data
output output
is needed
data data befor
is needeis n
d to carry outa women’s clothing 120
the experiment. enterprise
groupsin of
Hang- unified
production in the data
process neuralofnetwork, the cycle 10 groups of prediction cycle time can be
zhou, China, was used to carry out the time of different production prediction quantities
results
prediction
prediction results obtained.
are results
obtained.
are are According
obtained. obtained. According According
to the to production
theto the productio
productionquantity,qu
om May to July 2019 were obtained. 89 groups were selected for neural
corresponding
corresponding
prediction
corresponding cycle
prediction
predictiontime,
cyclecycle
such
time,as
time,
the such
such first as
group
as the the of
first first
data.
groupgroup
ofThe
of
dat
g and 10 for a simulation test. The data experiment was carried out in
180 quantity0.15
is 407,
quantity
quantity therefore
is 407,
is 407, select
therefore
thereforea select
prediction
select acycle
prediction
time
a prediction from
cycle
cycle 10time
time to 410,
from from
10 and
to10
41ft
7b. Through the training of the neural network, 10 groups of prediction
predicted cycle time
actual cycle time
in MATLAB inthrough
MATLAB
in MATLAB the through
4th degree
through thepolynomial,
the 4th 4th degree
degree polynomial,
and get and
polynomial, the following
andthe
get getfollowi
thefittin
foll
170 0.1
ned, and curve fitting was carried out for the prediction data. Then the 4
f (x) = 3.495e =f (x)
f (x) � 09 �= x3.495e
3.495e ��09 4
��x09
4.196e x 4 ��x4.196e
��06
4.196e 3 3
+ 06 ��x06
� +� x 3 +
160 0.05
data were obtained according to the fitting curve, and finally compared
Cycle time, s
� 0.4517
0.0019860.001986 � 0.4517
ta. 150 0
Figure 4Figure
shows
Figure
4 the
4 fitted
showsshows
theproduction
the fitted
fitted production
learninglearning
production curve.
learning
Takecurve.
41Take
curve. groups
Take of
41
41 gro
oups of production
140 data, including five categories of dress, skirt, shirt,
quantity intoquantity
theinto
quantity
-0.05 formula
into the
to get
formula
the formula the
to final
tothe
get get
41final
the
groups
final
41 of41predicted
groups
groups ofcycle
predicted
of predictedtime.
cycl
oat; the minimum
130 production quantity is 181 pieces, the maximum -0.1