Description

Description:

All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.

Share

Transcript

ORIGINAL ARTICLE
Optimizing sliver quality using Artiﬁcial NeuralNetworks in ring spinning
Samar Ahmed Mohsen Abd-Ellatif
*
Lecturer at the Textile Department, Faculty of Engineering, Alexandria University, ElShatby, Alexandria, Egypt
Received 22 April 2013; revised 20 September 2013; accepted 28 September 2013Available online 28 October 2013
KEYWORDS
Sliver;Neural Networks;Backpropagation;Quality;Optimization
Abstract
Sliver evenness is a very important parameter affecting the quality of the yarn produced.Therefore, controlling the sliver evenness is of major importance. Auto-levelers mounted on modernDrawing Frames should be accurately adjusted to help to achieve this task. The Leveling ActionPoint (LAP) is one of the important auto-leveling parameters which highly inﬂuence the evennessof the slivers produced. Its adjustment is therefore of a crucial importance. In this research work,Artiﬁcial Neural Networks are applied to predict the optimum value of the LAP under differentproductions and material conditions. Five models are developed and tested for their ability to pre-dict the optimum value of the LAP from the most inﬂuencing ﬁber and process parameters. As aﬁnal step, a statistical multiple regression model was developed to conduct a comparison betweenthe performance of the developed Artiﬁcial Neural Network model and the currently applied sta-tistical techniques.
ª
2013 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, AlexandriaUniversity.
1. Introduction
A major factor affecting yarn evenness is the evenness of theslivers from which they are produced. The sliver evenness isgenerally controlled by the auto-levelers mounted on themajority of the modern draw frames. The distance betweenthe scanning rollers pair and the point of draft is called theLeveling Action Point (LAP). The value of the LAP is oneof the most important parameters determining theperformance of auto-levelers. The accurate adjustment of thispoint should lead to an exact correction of the defectivematerial monitored by the auto-leveling system. This parame-ter is affected by a large number of material and productionparameters [1]. The selection of the suitable value of LAP istherefore very complicated and requires the consideration of all or at least the most relevant inﬂuencing parameters.Different techniques were adopted through the years todetermine the optimum value of LAP for different materialsand production conditions [2]. These techniques were generallytime-consuming and were accompanied by large material andproduction losses.In the common practice, the sliver samples had to beproduced with different settings, taken to laboratories andexamined on the evenness tester until the optimum LAP was
*Tel.: +20 1004472037.E-mail address: samar_eg@yahoo.com.Peer review under responsibility of Faculty of Engineering, AlexandriaUniversity.
Production and hosting by Elsevier
Alexandria Engineering Journal (2013)
52
, 637–642
Alexandria University
Alexandria Engineering Journal
www.elsevier.com/locate/aejwww.sciencedirect.com1110-0168
ª
2013 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria University.http://dx.doi.org/10.1016/j.aej.2013.09.007
found. Recent draw frames implement an automatic searchfunction for the optimum determination of the LAP. Throughthis function, the sliver is automatically scanned by adjustingthe different LAPs temporarily and the resulted values are re-corded. During this process, the quality parameters are con-stantly monitored and an algorithm automatically calculatedthe optimum LAP by selecting the point with the minimum sli-ver CV%. At present, a search range of 120 mm is scanned, i.e.21 points are examined using 100 m of sliver in each case; there-fore 2100 m of sliver is necessary to carry out the search func-tion. This is a time-consuming method accompanied bymaterial and production losses, and hence directly affectingthe cost parameters [1].Modeling was therefore a suggested solution for the prob-lem of the determination of the optimum value of LAP forspeciﬁc materials and production conditions. First-principlesmodeling, i.e. Mathematical tools have proved to be inade-quate. The mechanistic models overtly simplify the case tomake the equations manageable and pay the price with theirlimited accuracy. Empirical models that use statistical tech-niques have also shown up their limitations in use and arerarely used in any branch of the textile industry as a deci-sion-making tool. They require specialized knowledge of bothstatistical methods and techniques of design of experiments.Existing tools and techniques are thus inadequate for accu-rately modeling and optimizing complex nonlinear processes.An Artiﬁcial Neural Network is a promising step in this direc-tion. Neural networks are ideal for developing models in suchindustrial processes that are ‘‘data-rich and knowledge poor.’’Processes that have large quantities of historical data, but littleknowledge about the complicated interactions between the in-puts and outputs of the system, are ideal candidates for build-ing predictive models with ANNs [3].
2. Methodology
This research work intends to apply the technique of ArtiﬁcialNeural Networks as a tool for the prediction of the optimumLAP using a set of relevant material and production parame-ters. Artiﬁcial Neural Networks possess advantages that nei-ther conventional computing methods nor statistical analysistechniques possess.An Artiﬁcial Neural Network is an attempt to capture thefunctional principles of a biological neuron. There are twokinds of elements in the network – the neuron node and theconnection weight. A neuron node is the basic processing unitthat has an activation function. Neuron nodes are arranged ina layered structure, and the neuron nodes in consecutive layersare fully connected by connection weights [4]. The schematicrepresentation of the ANN model as envisaged by McCullochand Pitts is shown in Fig. 1.There are n inputs to the neuron from (
x
1
to
x
n
). The bias isa ﬁxed input. The inputs are multiplied by weights – from
w
k
1
to
w
kn
, respectively. The weighted sum of the inputs, which isdenoted by
u
k
, is given by the following:
u
k
¼
X
n j
¼
0
w
kj
x
j
The parameter
u
k
now becomes the input to the activationfunction and is modiﬁed according to the nature of the activa-tion function. The output
y
k
is given by the following:
y
k
¼
w
ð
u
k
Þ
where
w
is a nonlinear function of
u
k
and is termed an activa-tion function.The simplest ANNs consist of two layers of neurons Fig. 2.The introduction of another layer of neurons (known as a hid-den layer) between the input and output layers results in a mul-tilayer network Fig. 3.With a set of input-training data that correspond to adeﬁnite signal pattern, the network can be trained to give acorrespondingly desired pattern at output.The actual output pattern can be compared with the desiredoutput pattern, and an algorithm can adjust the weights untilpattern-matching occurs – i.e. the error becomes acceptablysmall. At the completion of training, the network should becapable not only of recalling all the input–output patterns usedfor training, but also of interpolating and extrapolating them.A variety of training algorithms have been developed eachone with its own strength and weaknesses The most widelyused is the ‘‘back propagation algorithm’’ which is used todetermine the weights.In order to train a feed-forward Neural Network properly,various network parameters have to be optimized. Theseparameters include the number of hidden layers, the numberof units in the hidden layer, the learning rate and the numberof training cycles (also known as epochs). The ﬁrst decisionconcerning network parameters is deciding the number of hid-den layers. It is known that a feed-forward Neural Networkwith only one hidden layer can approximate any function toan arbitrary degree of accuracy. However, many authors havereported using more than one hidden layer and obtaining goodresults. The number of hidden layers is not likely to produce asigniﬁcant improvement in network performance providedthat all the network parameters are optimized [3]. The numberof units in the hidden layer is decided upon by conducting trialand error techniques.
3. Experimental work
A large number of production and material parameters wouldaffect the selection of the value of LAP. A development of anANN model holding all these parameters is unpractical andwould scarcely prove any success. Too many input variableswould affect the effectiveness of the ANN negatively. There-fore, a systematic method for the selection of the input vari-ables was necessary. A new technique was adopted to selectthe most relevant parameters to be taken as input parametersfor the developed Artiﬁcial Neural Network model [5].
X
0
X
1
X
2
X
n
Σ Ψ
(.)
nputs Bias Input Summing Junction Activation Function OutputY
k
U
k
W
k0
W
k1
W
k2
W
kn
Weights
Figure 1
Artiﬁcial neuron [3].
638 S.A.M. Abd-Ellatif
Conducting this technique to the totality of the material andprocess parameters, the most relevant parameters to be takeninto consideration while developing the ANNs could besummed up into the following: ﬁber length, length uniformityratio, ﬁber ﬁneness, sliver linear density, the required outputsliver evenness and the main draft gauge. Other parameterssuch as machine speeds, main draft setting, feeding tension,break draft and settings of sliver deﬂection bar were foundto be of relatively lower effect on the CV of the slivers pro-duced and were therefore not taken into consideration. Thecoefﬁcient of friction of the different types of cotton used inthe investigation is an important parameter. It was, neverthe-less, not included in the model developed as the coefﬁcient of friction is known to be highly correlated with the ﬁber ﬁneness,which is part of the model developed. Two varieties of extralong and long staple Egyptian cotton (G.87 and G.90) andtwo varieties of medium staple cotton (Russian and GreekAkala cottons) were investigated.The production line adopted consisted of the secondpassage Rieter Drawing Frame RSB-D 35 equipped with anauto-leveler:Fig. 4 shows an overview of the drafting system of theDrawing Frame used in the experiments and shows the Level-ing Action Point (LAP) to be adjusted according to the prop-erties of the ﬁbers used.In Fig. 4, GB represents the range of settings for LevelingAction Point, total 939–1059 mm,
E
represents the LevelingAction Point,
VV
represents the break draft,
HV
representsthe main draft,
HVD
represents the main drafting distance,
V
represents the total draft,
VE
represents the entry tensionand B91 is the impulse generator (=88 impulse/revolution).BL is a constant length (=crow ﬂight distance) betweencenter of scanning rollers (point
O
) and center delivery roll(1005 mm).At the most modern draw frames, the thickness varia-tions in the fed sliver are continually monitored by amechanical device (a tongue–groove roll) and subsequentlyconverted into electrical signals. The measured values aretransmitted to an electronic memory with a variable the timedelayed response. The time delay allows the draft betweenthe mid-roll and the delivery roll of the draw frame toadjust exactly at the moment when the defective sliver piece,which had been measured by a pair of scanning rollers, ﬁndsitself at a point of correcting the draft. At this point, a servomotor operates and adjusts its speed depending upon theamount of sliver mass variation detected in the fed slivercross section. The distance that separates the scanning roll-ers and the point of draft is called the zero point of regula-tion or the Leveling Action Point (LAP). This leads to thecalculated correction on the corresponding defective material[1].Moreover, in the case of a change in ﬁber material, themachine settings and process controlling parameters such asproduction speed, material, break draft setting, main draftsetting, feeding tension, break draft, and setting of the sliverguiding rollers, LAP is needed to be altered.During the course of the experiments, the productionparameters were kept at the usual values used for the rawmaterials adopted and were kept constant on all the machinesthroughout the production line except for the Leveling ActionPoint (LAP) on the Drawing Frame which was varied through-out the experiments.The material properties and production parametersadopted throughout the investigations are summarized inTable 1.Five Artiﬁcial Neural Network (ANN) models were devel-oped and tested for their convenience for the case studied inthe research work.
u
i1
u
o1
u
i2
u
i3
u
in
u
o2
u
o3
u
on
Input Layer of Neurons Output Layer of Neurons
Figure 2
A Neural Network with no hidden layers [3].
u
i1
u
h1
u
i2
u
i3
u
in
u
h2
u
h3
u
on
Input Layer of Neurons Hidden Layer of Neurons
u
o1
u
o2
u
o3
u
on
Output Layer of Neurons
Figure 3
A Neural Network with one hidden layer [3].
Optimizing sliver quality using Artiﬁcial Neural Networks in ring spinning 639
For the training of the developed Artiﬁcial Neural Networkmodel, the most commonly used learning technique (backpropagation algorithm) is applied.Back propagation was created by generalizing theWidrow–Hoff learning rule to multilayer networks and nonlin-ear differentiable transfer functions. Input vectors and thecorresponding target vectors are used to train a network untilit can approximate a function, associate input vectors with spe-ciﬁc output vectors, or classify input vectors in an appropriateway as deﬁned by you. Networks with biases, a sigmoid layer,and a linear output layer are capable of approximating anyfunction with a ﬁnite number of discontinuities. Properlytrained backpropagation networks tend to give reasonable an-swers when presented with inputs that they have never seen.Typically, a new input leads to an output similar to the correctoutput for input vectors used in training that are similar to thenew input being presented. This generalization property makesit possible to train a network on a representative set of input/target pairs and get good results without training the networkon all possible input/output pairs [6].As a multilayer network, the tan-Sigmoid transfer functionwas used for the ﬁrst layer while the linear transfer functionwas used for the output layer. Both functions are consideredas the most commonly used transfer functions for backpropa-gation networks. Nevertheless, combinations with thelog-sigmoid transfer function were tried but showed nopromising results. The network was initialized as a multilayernetwork with three layers: the ﬁrst being the input layer, thesecond being the hidden layer while the third being the outputlayer. Only one hidden layer was selected as in general, accord-ing to Kolmogorov’s theorem, networks with a single hiddenlayer should be capable of approximating any function toany degree of accuracy [4]. Neural Network models with twohidden layers were also tested but showed limited success.The batch training mode was adopted. In this mode, theweights and biases of the network are updated only after theentire training set has been applied to the network. The gradi-ents calculated at each training example are added together todetermine the change in the weights and biases.Afasttrainingalgorithmusingstandardnumericaloptimiza-tion techniques was adopted: the Levenberg–Marquardt algo-rithm. The srcinal description of the Levenberg–Marquardtalgorithm and the application of Levenberg–Marquardt toNeural Network training are described in [7,8]. This algorithmappears to be the fastest method for training moderate-sizedfeed-forwardNeuralNetworks(uptoseveralhundredweights).It also has a very efﬁcient MATLAB implementation, since thesolution of the matrix equation is a built-in function, so itsattributes become even more pronounced in a MATLABsetting.The MATLAB Neural Networks Toolbox function ‘‘train-lm’’ was applied. An abstract depiction of the experimentalmodel is shown in Fig. 5.Five different Neural Networks were trained to predict theLAP with network structures as shown in Table 2. The datasets were divided into three sets: training, validation and test
Figure 4
Overview of the Leveling Action Point (LAP) [1].
Table 1
Summary of the ﬁber and production parameters used in the investigation.
Cotton variety Average ﬁber properties Production parametersFiber length UHML mm Uniformity ratio Micronaire value Sliver count Ne Main draft gauge mmGiza 87 37 87.2 3 0.15 47Giza 90 29.8 85 3.9 0.14 38.7Russian 28 77 4.5 0.13 36.5Greek Akala 30 80 4.5 0.14 39
LAPFiber FinenessFiber LengthUniformity RatioSliver linear density Main Draft GaugeSliver Evenness CV%
Figure 5
Abstract Neural Network model.
640 S.A.M. Abd-Ellatif

Search

Similar documents

Tags

Related Search

We Need Your Support

Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...Sign Now!

We are very appreciated for your Prompt Action!

x