Designed by Manisha Mogare
Modeling yarn properties from fiber parameters has been a theme
of research for many years. Mechanistic and statistical approaches have been
dominating the area. The limitations and strengths of both approaches have
been appreciated and presently neural networks, fuzzy logic, and computer
simulations are being explored. The use of artificial neural networks for
predicting yarn properties from fiber parameters is discussed in this chapter
Key words: yarn engineering, yarn property modeling, neural networks in
textiles.
The industrial relevance of the topic of yarn property modeling is obvious.
Predicting product performance and its properties from raw material
characteristics has been a theme of research in many areas including textiles.
The outcome of a process in the textile industry could be a fiber, yarn, fabric
or garment.
Manufacturing each product is an industry by itself. The yarn
manufacturing industry involves a large number of processes such as opening,
cleaning, carding, drawing, combing, roving preparation and spinning that
lead to the production of yarns of various counts and blends. Knowing what is
expected from a raw material is important to both the supplier of raw material
and the purchaser.
A cotton grower would like to know what
sort of yarn quality can be produced from his crop so that he can claim the
right price for his produce. The buyer, a spinning mill, would be interested
in knowing whether it is possible to attain the desired yarn properties from
a particular variety of cotton it intends to buy. The user of the yarn, either
a knitter or a weaver, will be interested in knowing the performance of the
yarn from its physical and mechanical properties.
Hence, there is a need for
a reliable method for predicting yarn properties from fiber characteristics
and relevant yarn parameters.
Properties such as fiber length, fineness, strength, elongation, yarn uniformity, thin places, and twist can be estimated by modern instrumentation. It is very difficult to establish a definite relationship between fiber, process and yarn parameters as their exact relationship is yet to be established, since they are highly nonlinear, complex and interactive. Hence there is a need to follow a non-traditional approach to model them.
1. Characterized by a large number of simple neuron-like processing elements and a large number of
weighted connections between them which can accurately capture the nonlinear relationship between
different process and material parameters.
2. Have good predictive power.
3. Require fewer data sets than conventional regression analysis.
4. The neural network can be easily updated with both old and new data.
5. Cannot be reliably used to predict the parameter outside
the range of data
6. Do not provide any insight about the mechanics of the
relationship between the parameters.
The problem of trying to predict yarn properties from fiber properties & process parameters can be
viewed
as one of function.
yarn property = f (fiber properties, fiber configuration and their arrangement, yarn count, yarn
twist).
Fiber configuration and their arrangement in a way is a reflection of the dynamics of the process.
If
the process remains constant then it may be considered to be a black box and the yarn property then
becomes a function of fiber properties and yarn count.
So that, ANN can predict yarn properties approximately.
1. Select feed-forward network if possible.
2. Select input and output nodes equal to the number of input and output
signals.
3. Select appropriate input and output scale factors for normalization and
denormalization of input and output signals.
4. Create input–output training data based on experimental results.
5. Set up network topology assuming it to be a three-layer network. Select
hidden layer nodes equal to average of input and output layer nodes.
Select transfer function.
6. Select an acceptable training error. Initialize the network with random
positive and negative weights.
7. Select an input–output data pattern from the training data set and change
the weights of the network following the back-propagation training
principle.
8. After the acceptable error is reached, select another pattern and repeat
the procedure until all the data pattern is completed.
9. If a network fails to converge to an acceptable error, increase the hidden
layers neurons or increase the number of hidden layers as one may feel
necessary. Usually problems are solved by having three hidden layers
at most.
10. After successful training, test the network’s performance with some
intermediate data input.
| Sr.no | Input fibre Properties | Output fibre Properties |
|---|---|---|
| 1 | Upper half mean length, uniformity index, short fiber %, strength, fineness, maturity, grayness, yellowness. | Count Strenth Product |
| 2 | Percentage Polyester in blend , count , first and second nozzle pressure | Breaking load and breaking elongation. |
| 3 | Upper quartile length , mean fibre length , % short fibres, diameter, neps, total trash. | Yarn Irregularity. |
| 4 | Blend ratio, count, first and second nozzle pressure. | Yarn Tenacity. |
| 5 | 2.5% span length, uniformity ratio, fiber fineness, bundle strength, trash content, nominal yarn count. | Total imperfection. |
A feed-forward neural network can be constructed with six input units, five corresponding to the
fiber
properties and one to the yarn count.
Two types of architecture that we can choose
A network with one output unit corresponding to one of the other six yarn properties at a time.
A network with six output units corresponding to six yarn properties
| Sr.no | It should contains fibre properties like | and yarn properties should be like |
|---|---|---|
| a | 2.5% span length(in mm) | count(tex) |
| b | Uniformity Ratio(%) | lea strength(kg) |
| c | fineness (µg/inch) | CSP |
| d | Bundle strength (cN/tex) | CV % of count |
| e | trash content(%). | CV% of strength, |
2.1 We need to use different python libraries like numpy, pandas, keras,
tensorflow, matplot,
streamlit, etc
to design a neural network and prediction results format.
2.2 To create a proper responsive and functional web application for this
prediction,
we will work with different scripting languages like HTML, CSS, JavaScript and
tools like git and github to deploy this web app on web server.
