See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324132141

Using regression models for predicting the product quality in a tubing

extrusion process

Article in Journal of Intelligent Manufacturing · August 2019

DOI: 10.1007/s10845-018-1418-7

CITATIONS READS

53 9,275

5 authors, including:

Vicente García Josep Salvador Sánchez

Universidad Autónoma de Ciudad Juárez Universitat Jaume I
117 PUBLICATIONS 3,052 CITATIONS 162 PUBLICATIONS 4,037 CITATIONS

SEE PROFILE SEE PROFILE

Luis Rodríguez-Picón Luis Carlos Méndez-González

Universidad Autónoma de Ciudad Juárez Universidad Autónoma de Ciudad Juárez
50 PUBLICATIONS 341 CITATIONS 41 PUBLICATIONS 248 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cost-Sensitive Neural Networks and Editing Techniques for Imbalance Problems View project

Classifier Ensembles View project

All content following this page was uploaded by Vicente García on 02 April 2018.

The user has requested enhancement of the downloaded file.

 J Intell Manuf manuscript No.
(will be inserted by the editor)

Using regression models for predicting the product quality in a tubing

extrusion process
Vicente Garcı́a · J. Salvador Sánchez · Luis Alberto Rodrı́guez-Picón · Luis Carlos
Méndez-González · Humberto de Jesús Ochoa-Domı́nguez

Received: date / Accepted: date

Abstract Quality in a manufacturing process implies that Keywords Regression Models · Extrusion Process ·
the performance characteristics of the product and the pro- Product Quality Prediction · support vector regression · k
cess itself are designed to meet specific objectives. Thus, ac- nearest-neighbor
curate quality prediction plays a principal role in delivering
high-quality products to further enhance competitiveness. In
tubing extrusion, measuring of the inner and outer diameters 1 Introduction
is typically performed either manually or with ultrasonic or
laser scanners. This paper shows how regression models can Tubings and pipes are manufactured using different meth-
result useful to estimate both those physical quality indices ods, but extrusion is probably the most efficient one. This
in a tube extrusion process. A real-life data set obtained complex thermoforming process involves heating a raw ma-
from a Mexican extrusion manufacturing company is used terial (usually plastic, metal, polymer, concrete or ceramic)
for the empirical analysis. Experimental results demonstrate and forming a final ring-shaped product. A primary advan-
that k nearest-neighbor and support vector regression meth- tage of extrusion over other manufacturing processes is its
ods (with a linear kernel and with a radial basis function) capability to create a very complex cross-sectional profile
are especially suitable for predicting the inner and outer di- object (Oberg et al, 2012). However, extrusion processing
ameters of an extruded tube based on the evaluation of 15 compresses many interdependent input parameters (both pro-
extrusion and pulling process parameters. cess and system variables) and output parameters. Process
variables refer to the operating conditions that can be con-
trolled and manipulated directly, whereas the system vari-
Vicente Garcı́a, Corresponding author
Departamento de Ingenierı́a Eléctrica y Computación, División Multi- ables are determined by the process parameters and have an
disciplinaria en Ciudad Universitaria, Universidad Autónoma de Ciu- impact on the output parameters (Chevanan et al, 2007).
dad Juárez, C. José de Jesús Macı́as Delgado, 18100 Ciudad Juárez, Unlike the cyclic techniques such as injection molding
Chihuahua,Mexico
or blow molding, extrusion is a steady-state or continuous
E-mail: [email protected]
process. This means that, for example, a change in the pa-
J. Salvador Sánchez
rameters of the extruder will disrupt the steady-state pro-
Institute of New Imaging Technologies, Universitat Jaume I,
Av. Vicent Sos Baynat s/n, 12071 Castelló de la Plana, Spain cess condition with a non-negligible effect on the quality of
the extruded product. Therefore, all input parameters must
Luis Alberto Rodrı́guez-Picón · Luis Carlos Méndez-González be identified, controlled and monitored to guarantee success
Departamento de Ingenierı́a Industrial y Manufactura, Universidad in the extrusion process since manufacturing quality predic-
Autónoma de Ciudad Juárez, Av. del Charro 450 Norte, 32310 Ciu- tion, control, and monitoring are critical (Khan et al, 2014).
dad Juárez, Chihuahua, Mexico
Common deficiencies of extruded products are related to vi-
sual or geometrical characteristics (e.g., diameter variations,
Humberto de Jesús Ochoa-Domı́nguez
color changes and rough surface) and physical or mechani-
Departamento de Ingenierı́a Eléctrica y Computación, Universidad
Autónoma de Ciudad Juárez, Av. del Charro 450 Norte, 32310 Ciu- cal properties (e.g., elasticity and rigidity).
dad Juárez, Chihuahua, Mexico Several intelligent and soft computing models (Witten
et al, 2011) have been applied to a large variety of manufac-
2 Vicente Garcı́a et al.

turing processes, such as production, fault detection, process optimize the technical parameters in the semi-solid extru-
planning and monitoring, machine maintenance, and quality sion of composite tubes and bars. Yu et al (2004) proposed
prediction and control (Charaniya et al, 2010; Choudhary a strategy based on a fuzzy neural-Taguchi network and a
et al, 2008; Harding et al, 2006; Köksal et al, 2011; Kusiak, genetic algorithm to determine the optimal die gap program-
2006; Pratihar, 2015; Yin et al, 2015). In particular, the use ming of extrusion blow molding processes. Oke et al (2006)
of these techniques for machinery fault detection and prod- optimized the flow rate of the plastic extrusion process in
uct quality prediction has received increasing attention over a plastic recycling plant with the application of a neuro-
the last years. fuzzy model. González Marcos et al (2007) introduced im-
Krömer et al (2010) showed the ability of genetic pro- provements in the rubber extrusion process by predicting the
gramming to evolve fuzzy classifiers on a real-world prob- characteristics of rubber with a multi-layer perceptron neu-
lem for detecting faulty products in an industrial production ral network. Sharma et al (2009) suggested a model of for-
process. Multi-layer perceptron neural networks were em- ward mapping for hot extrusion process using the ANFIS
ployed to predict errors in mold surface roughness (Erzu- neuro-fuzzy approach. Hsiang et al (2012) investigated the
rumlu and Oktem, 2007) and the product quality in a wave optimal process parameters that maximize the multiple per-
soldering process (Liukkonen et al, 2009). Support vector formance characteristics index for hot extrusion of magne-
machines (Jiang et al, 2013) and radial basis function neural sium alloy bicycle carriers through a fuzzy-based Taguchi
networks (Zhang et al, 2014) were used to predict the quality method. Ramana and Reddy (2013) proposed to make use
of propylene polymerization in industrial processes. Chien of clustering, naı̈ve Bayes, and decision trees to predict and
et al (2007) applied the K-means clustering algorithm and improve the final product quality in a plastic extrusion pro-
decision trees for the detection of defects in semiconduc- cess. Zhao et al (2013) employed a Pareto-based genetic al-
tor manufacturing. The rough set approach was applied to gorithm for optimization of porthole extrusion die. Support
find out solder defects in printed circuit boards (Kusiak and vector regression models and multi-layer perceptron neu-
Kurasek, 2001). Quality prediction in plastic injection mold- ral networks were compared for the prediction of specific
ing processes was tackled using back-propagation neural net- properties of rubber extruded mixtures (Urraca Valle et al,
works (Sadeghi, 2000), support vector machines (Ribeiro, 2013). Carrano et al (2015) employed an evolutionary com-
2005) and genetic algorithms (Meiabadi et al, 2013). A com- puting algorithm to optimize the operational and screw geo-
bined method based on artificial neural network and particle metrical parameters of a single screw polymer extrusion sys-
swarm optimization was proposed to improve the mechani- tem. One-class classification methods were used by Kohlert
cal performance of polymer products (Xu et al, 2015). Adly and König (2015) for yield optimization of an extrusion pro-
et al (2015) presented a simplified subspace regression algo- cess in a polymer film industry. Chondronasios et al (2016)
rithm for accurate identification of defect patterns in semi- introduced a feature extraction technique based on gradient-
conductor wafer maps. Two evolutionary fuzzy ARTMAP only co-occurrence matrices to detect blisters and scratches
neural networks were designed by Tan et al (2015) to deal on the surface of extruded aluminum profiles using a two-
with the class imbalance problem in semiconductor manu- layer feed-forward artificial neural network.
facturing operations. Ghorai et al (2013) developed a visual The main purpose of this paper, therefore, is to analyze
inspection system to localize defects on hot-rolled steel sur- the performance of some regression models in the predic-
faces employing some kernel classifiers, such as the sup- tion of product quality (regarding the inner and outer di-
port vector machine and the vector-valued regularized ker- ameters) in a tubing extrusion process. From an application
nel function approximation. Wu et al (2017) introduced a perspective, the novelty of this study is on the specific so-
method based on random forests for tool wear prediction lution proposed for product quality control in a plastic tube
and compared its performance with that of support vector manufacturing plant. To the best of our knowledge, there are
regression and feed-forward back-propagation neural net- no previous reports that analyze the use of parameters taken
works. Wang et al (2018) presented a comprehensive survey from the extrusion and pulling processes to predict the inner
of deep learning algorithms for smart manufacturing. and outer diameters of an extruded tube using the regression
With regards to the particular case of product quality methods considered here.
prediction in extrusion processes, we can pay attention to Henceforth the paper is organized as follows. Section 2
a set of works that have employed some soft computing describes the tubing extrusion process of a Mexican man-
techniques. For instance, Wu and Hsu (2002) combined the ufacturing company, which provided us with the database
finite-element approach, a polynomial network and a genetic used for the subsequent empirical analysis. Section 3 intro-
algorithm to develop a method for the design of the opti- duces the bases of the regression models that will be ex-
mal shape of an extrusion die. Li et al (2004) adopted the plored in this study. Next, Section 4 presents the experimen-
cooperation between a three-layer back-propagation neural tal set-up and the performance evaluation criteria used in
network and a genetic algorithm to set up the system and the experiments, while the results are given and discussed
Using regression models for predicting the product quality in a tubing extrusion process 3

in Section 5. Finally, Section 6 summarizes the main con- the thickness variation, the uneven wall thickness, the diam-
clusions and outlines some possible avenues for future re- eter variation, and the centering problem. In this work, the
search. extruded tube quality was defined regarding the inner diam-
eter (ID) and the outer diameter (OD), as shown in Fig. 3.
Although other characteristics could affect the quality of the
2 Description of the tubing extrusion process product (e.g., length of the tube, wall thickness, or color
uniformity), the only functional requirements for this appli-
This section provides a general description of the tube ex-
cation correspond to the inner and outer diameters because
trusion process used by a manufacturing company located
these are the critical characteristics that were stated by the
in Ciudad Juárez (Chihuahua, Mexico). Thus the extrusion
customer.
process consists of two stages. In the initial phase, the plas-
tic is fed into the heating chamber of the extruder to melt it
(see Fig. 1). Once molten, the plastic is pushed by a screw
device through the shaped die, which forms the plastic into
a tube-shaped form.

Fig. 3 Extruded tube. Quality indices: inner and outer diameters

To guarantee the quality of the manufactured tube (i.e.,

Fig. 1 Extrusion process. Process parameters that determine the qual-
ity of the tube: [1-4] zone temperatures, [5] melting temperature, [6-9] the inner and outer diameters have to be within design speci-
die temperature, [10] revolutions per minute of the screw, and [11] base fications set by the customer), every process parameter must
hopper temperature be identified, controlled, and monitored throughout the ex-
trusion process. For example, in the extruder zone, there are
several input parameters that might yield significant devia-
In the second phase depicted in Fig. 2, the extruded tube
tions in the characteristics of the product: the base hopper
is pulled by a mechanism through a water tank or a blowing
temperature (BHT), the zone temperature (ZT), the die tem-
system to cool it down and get the final form.
perature (DT), the melting temperature (MT), and the revo-
lutions per minute of the screw (SRPM). In the case of the
pulling stage, the set of parameters are the tank tempera-
ture (TT), the vacuum pressure (VP), and the tension of the
pulling mechanism (TPM).
In total, there are 15 process parameters that may pro-
duce deviations in the functional requirements of the ex-
Fig. 2 Pulling process. Process parameters that determines the qual- truded tube: four zone temperatures, four die temperatures,
ity of the tube: [1-2] tank temperature, [3] vacuum pressure, and [4] the melting temperature, the revolutions per minute of the
tension of the pulling mechanism
screw, the base hopper temperature, two tank temperatures,
the vacuum pressure, and the tension of the pulling mecha-
A defect can be defined as a deviation of the product nism. As a result, each sample will be described by these 15
characteristics from the specifications set up by the manu- input parameters and the two output variables mentioned in
facturing process (Khan et al, 2014), or the difference be- the previous paragraph (ID and OD).
tween the desired product and the resulting product (Dhafr Table 1 reports the main characteristics of the database
et al, 2006). It can be caused by a single source or the cumu- used in the empirical analysis: the attribute number, the at-
lative effect of several factors, which may arise at any stage tribute description and some statistics, such as the minimum
of the extrusion process. Some defects can be found in ex- and maximum values of the attribute, the mean and the stan-
truded parts such as the rough surface, the extruder surging, dard deviation.
4 Vicente Garcı́a et al.

Table 1 Characteristics of the tube-extrusion data set used in the experiments

Extrusion process
No. Description Minimum Maximum Mean Std. Dev.
1 Zone 1 Temperature (◦ F) 435 448 440.68 2.39
2 Zone 2 Temperature (◦ F) 432 445 440.53 2.45
3 Zone 3 Temperature (◦ F) 429 463 433.69 4.09
4 Zone 4 Temperature (◦ F) 425 440 433.48 3.05
5 Melting Temperature (◦ F) 436 450 441.78 2.96
6 Die 1 Temperature (◦ F) 442 454 447.99 2.88
7 Die 2 Temperature (◦ F) 441 459 448.39 3.28
8 Die 3 Temperature (◦ F) 445 456 448.97 2.31
9 Die 4 Temperature (◦ F) 444 458 449.10 2.46
10 Revolutions per Minute of the Screw (rpm) 52 75 62.00 5.94
11 Base Hopper Temperature (◦ F) 46 80 58.50 7.84
Pulling process
No. Description Minimum Maximum Mean Std. Dev.
1 Tank 1 Temperature (◦ F) 49 71 64.40 5.70
2 Tank 2 Temperature (◦ F) 0.2 10.1 1.01 6.23
3 Vacuum Pressure (inHg) 59 65.6 61.42 1.42
4 Tension of the Pulling Mechanism (ft/min) 40 56 50.34 2.13
Output Variables
No. Description Minimum Maximum Mean Std. Dev.
1 Inner Diameter (in) 0.227 0.233 0.230 0.001
2 Outer Diameter (in) 0.228 0.331 0.328 0.014

It is important to point out that the different input pa- set of stored instances and therefore, it predicts the output of
rameters were measured and recorded using specific sensors a new sample based on its closest neighbor.
during the extrusion and pulling processes, and an opera- The concept of the NN rule can be generalized for re-
tor collected the data at a fixed time. Analogously, the input gression because the nearest neighbor method assigns a new
and output diameters of tubes were measured manually with sample y the same target value as the closest instance in
a vernier caliper by the operator. At each shift, these tasks T , according to a particular dissimilarity measure (gener-
were carried out three times, thus obtaining a data set as the ally, the Euclidean distance). An extension of this procedure
one shown in the example of Table 2. is the k-NN decision rule, in which the algorithm retrieves
the k closest instances in T . When k = 1, the target value
assigned to the input sample is the target value indicated
3 Regression models by its nearest neighbor. For k > 1, the k-NN regression
model (k-NNR) estimates the target value f (y) of a new in-
In this section, we briefly introduce the regression methods put sample y by averaging the target values of its k nearest
that will be further applied to product quality prediction for neighbors (Biau et al, 2012; Guyader and Hengartner, 2013;
the tubing extrusion process just described. Kramer, 2011; Lee et al, 2014):
Let T = {(x1 , a1 ), . . . , (xn , an )} ∈ (x × a)n be a data
set of n independent and identically distributed (i.i.d.) ran- k
dom pairs (xi , ai ), where xi = [xi1 , xi2 , . . . , xiD ] repre- 1X
f (y) = ai (1)
sents an instance in a D-dimensional feature space and ai k i=1
denotes the continuous target value associated to it. The aim
of regression is to learn a function f : y → a to predict the where ai denotes the target value of the i-th nearest neigh-
value a for a new sample y = [y1 , y2 , . . . , yD ]. bor.

3.1.1 Distance-weighted k-NN regression

3.1 Nearest neighbor regression
When the basic k-NN algorithm estimates the target value
One of the most popular and successful supervised learn- for the new sample, it ignores some relevant information
ing methods corresponds to the nearest neighbor (NN) rule that each of the k nearest neighbors might provide regard-
due to its algorithmic simplicity and high prediction perfor- ing their distance (Batista and Silva, 2009). To overcome
mance. This non-parametric technique works under the as- this shortcoming, Dudani (1976) proposed a weighting func-
sumption that new samples share similar properties with the tion, which weights more heavily closer neighbors than dis-
Using regression models for predicting the product quality in a tubing extrusion process 5

Table 2 An example of collected data for the tubing extrusion process

Time ZT1 ZT2 ZT3 ZT4 MT DT1 DT2 DT3 DT4 SRPM BHT TT1 TT2 VP TPM ID OD
1 438 437 431 434 442 448 451 452 455 64 71 62 1.58 63.4 47 0.228 0.327
2 441 439 452 437 450 452 457 453 455 68 78 56 3.05 61.7 51 0.230 0.311
..
.
t 440 443 461 429 444 445 452 448 448 70 74 60 1.83 64.0 52 0.229 0.330

tant neighbors, depending on their corresponding distances the independent variables yi , and  is the residual or fitted
to the new sample. error.
In general, a weighting function has to work based upon The regression coefficients βi are estimated by curve fit-
the premise that the weights should decrease with increas- ting based on the least square method with the aim of mini-
ing sample-to-neighbor distance (Dudani, 1976). Let xi (i = mizing the fitted error (the difference between the observed
1, . . . , k) be the closest instances to an input sample y, and and estimated values). Eq. 5 indicates how the average re-
let di = d(xi , y) be the distance between xi and y. A com- sponse of the output variable changes with the independent
mon weighting technique computes wi for the i-th nearest variables. Thus the LR model can be used to predict the tar-
neighbor as the inverse of its distance (Dudani, 1976): get value a from new observed values of y.

1 3.3 Support vector regression

wi = di 6= 0 (2)
di
Another possible weighting function (Batista and Silva, The foundations of support vector machines are well-known
2009) can be defined as for both classification and regression problems. Smola and
Schlkopf (2004) published an excellent tutorial on support
vector machines for regression (SVR). The objective of the
wi = 1 − di (3) SVR model is to define a linear regression function to map
the input data to a high-dimensional feature space, in which
Once the weights wi have been computed, the distance- input data can be separated easier than in the original input
weighted k-NN approach for regression (k-NNRw) estimates space (Chou et al, 2017; Ma et al, 2003),
the target value as follows (Hall et al, 2009):

f (x) = WT Φ(x) + b (6)

k
P
wi ai where W is a weight vector, Φ(x) maps the input sample x
i=1
f (y) = k
(4) to the high-dimensional feature space, and b is a bias term.
P
wi The W and b can be obtained by solving an optimization
i=1
problem (Ma et al, 2003):

3.2 Linear regression 1 T Xn

minimize W W+C (ξi + ξi∗ )
W,b 2
Multiple linear regression (LR) attempts to model the rela- i=1
tionship between two or more independent variables (in this subject to ai − (WT Φ(x) + b) ≤ (ξi + ξi∗ ) (7)
case, the input attributes reported in Table 1) and an output T
(W Φ(x) + b) − ai ≤ (ξi + ξi∗ )
or response variable by fitting a linear equation to the ob-
served data (Draper and Smith, 1998). Every value of the in- ξi , ξi∗ ≥ 0, i = 1, . . . , n
dependent variable is associated with a value of the response where C is a regularization parameter, ξi and ξi∗ are non-
variable. The general form of the multiple linear regression negative slack variables to penalize for errors that are greater
equation can be written as follows: than  in magnitude
By introducing the Lagrange multipliers α, α∗ , and a
D
X kernel function K , the model form in the dual space can be
f (y) = α + βi yi +  (5) written as:
j=1
n
where α is a constant (the point where the regression line f (x) =
X
(αi − αi∗ )K(xi , x) + b (8)
intercepts the Y -axis), βi are the regression coefficients on i=1
6 Vicente Garcı́a et al.

The use of a kernel function allows to deal with feature

spaces of arbitrary dimensionality without having to com- v
u n
pute the mapping function Φ(x) explicitly (Yang and Shieh, u1 X
RM SE = t (pi − ai )2 (9)
2010). The kernels most commonly used are linear, polyno- n i=1
mial, sigmoid, and radial basis functions.
and the mean absolute error (MAE),

4 Experimental set-up
n
1X
As already stated, this study aims to evaluate the perfor- M AE = | pi − ai | (10)
n i=1
mance of some regression models for product quality pre-
diction in the tubing extrusion process of a manufacturing Both these metrics show how far away the predicted val-
plant. Thus we conducted a pool of experiments on a data ues pi are from the target values ai by averaging the magni-
set with 260 samples that were collected using the proce- tude of individual errors without taking care of their sign.
dure described in Section 2. It has to be remarked that all
input attribute values (process parameters) were normalized
to the range [0, 1]. 5 Results and discussion
We focused our study on the simple k-NNR (no weight-
ing), two weighted versions of k-NNR using Eq. 2 and 3, Since the quality product was predicted here using the in-
in the sequel called k-NNRw1 and k-NNRw2 respectively, ner and outer diameters of the extruded tubes, the experi-
the LR model, the SVR technique with three different ker- ments and the subsequent analysis of results were performed
nels and the multi-layer perceptron (MLP) neural network. according to these two physical quality indices. Hence, for
The kernels used in the SVR model were a linear function each database (i.e., the outer and inner diameter databases),
(SVR-1), a polynomial function of degree 2 (SVR-2) and a we compared the average of the two performance measures
radial basis function (SVR-RBF). For the regression algo- (RMSE and MAE) achieved by each regression method.
rithms based on the k-NN rule, twenty-five odd values of k
(1, 3, . . . , 29) were tested. All regression models were taken
from the WEKA toolkit (Hall et al, 2009). 5.1 Outer diameter database
Following the standard strategy used to evaluate the per-
formance of regression models when databases are small- Table 3 reports the average results in terms of RMSE and
or medium-sized, the 10-fold cross-validation method was MAE across the 100 runs for each regression technique.
adopted (Buza et al, 2015; Hall et al, 2009; Hu et al, 2014). In the case of the k-NN methods, the values correspond to
The original data set was randomly divided into ten parts those of the best k. Based on the root mean square error, one
of size n/10 (where n denotes the total number of samples can observe that the three k-NN algorithms, the linear sup-
in the data set); for each fold, nine blocks were gathered as port vector (SVR-1) and the SVR-RBF obtained the lowest
the training set for learning the model, and the remaining error rates (very close to 0) when predicting the outer diam-
fold was used as an independent test set. Additionally, with eter. In addition, the same behavior can be viewed in terms
the aim of increasing the statistical significance of the ex- of MAE.
perimental scores, ten repetitions were run for each trial and If we consider that the output values of the outer diame-
the results from predicting the output of test samples were ter are in a range from 0.228 to 0.331, then all these regres-
averaged across the 100 runs. sion models appear to be suitable for predicting the quality
of extruded tubes. However, the small differences in both
RMSE and MAE results do not allow us to draw signifi-
4.1 Evaluation criteria cant conclusions about which method is the best numerical
prediction technique. In fact, even the LR and MLP models
In the framework of regression, the purpose of most perfor- could be applied to tackle this problem because their errors
mance evaluation metrics is to estimate how much the pre- were also close enough to 0.
dictions (p1 , p2 , . . . , pn ) deviate from the actual target val- Figure 4 shows the performance measures for the three
ues (a1 , a2 , . . . , an ). These metrics are minimized when the versions of k-NN regression when varying the value of k
predicted value for each test sample agrees with their true from 1 to 29. The graphical results suggest that when k in-
value (Caruana and Niculescu-Mizil, 2004). Two of the most creases, the k-NNRw1 shows a steady behavior along the
popular performance measures that have frequently been em- X-axis. In the case of k-NNR and k-NNRw2 models, their
ployed to assess the model performance in regression prob- error rates increase along with the value of the parameter k.
lems are the root mean square error (RMSE), In summary, it appears that k-NNRw1 performed the best
Using regression models for predicting the product quality in a tubing extrusion process 7

Table 3 Average results in terms of RMSE and MAE (with standard deviations) on the outer diameter database
RMSE MAE
k-NNR 0.0096912371 ± 0.0075398296 0.0030848789 ± 0.0021710948
k-NNRw1 0.0096563665 ± 0.0072890820 0.0030848789 ± 0.0021710948
k-NNRw2 0.0096912371 ± 0.0075398296 0.0030848789 ± 0.0021710948
LR 0.0138590570 ± 0.0194352190 0.0062838398 ± 0.0045738531
SVR-1 0.0099222666 ± 0.0101735078 0.0029027012 ± 0.0024793246
SVR-2 0.2603216450 ± 0.9783462937 0.0523638108 ± 0.1920644156
SVR-RBF 0.0095693860 ± 0.0103128854 0.0027710505 ± 0.0024827469
MLP 0.0130784745 ± 0.0076458498 0.0066302173 ± 0.0053236816

averaged across the 100 runs for each technique. Results are
10 -2 conceptually similar to those of the outer diameter database:
(i) the methods based on k-NN, the SVR-1 and the SVR-
10-2 RBF yielded very low error values (≈ 0); (ii) here MLP also
10-2 appears to be among the best performing algorithms; and
(iii) except the SVR-2 method, differences in the results of
RMSE

10-3
the regression models seem not to be significant.
10-3 Figure 5 depicts the performance results for all versions
of the k-NN regression models as a function of k. One can
10-3
observe that k-NNRw1 achieved very similar performance
k-NNR
10-3 k-NNRw1 results regardless of the k value. In the case of plain k-NNR
k-NNRw2 and k-NNRw2, when k increases, the error rates decreases
100
1 3 5 7 9 6 13 15 17 19 21 23 25 27 29 slightly. These results suggest that the three k-NN regres-
k sion models are suitable for predicting the inner diameter
(a) Root Mean Square Error
of extruded tubes, although the k-NNR and k-NNRw2 ap-
proaches seem to be the best techniques.
10-3
10-3
10-3 6 Conclusions and future work
10-3
10
-3 The present paper has focused on predicting two quality in-
MAE

-3 dices in a tubing extrusion process. A thoroughly experi-

10
mental study has been carried out on a real-life data set
-3
10 provided by an extrusion tube manufacturing plant located
10-3 in Ciudad Juárez (Chihuahua, Mexico). More specifically,
k-NNR
10-4 k-NNRw1 three k-NN regression methods (the straightforward algo-
k-NNRw2 rithm and two distance-weighted approaches), the linear re-
100
1 3 5 7 9 6 13 15 17 19 21 23 25 27 29 gression model, three SVR configurations (SVR-1, SVR-
k 2, and SVR-RBF), and a multi-layer perceptron have been
(b) Mean Absolute Error
used to predict the inner and outer diameters of an extruded
Fig. 4 Outer diameter database: Performance regression measures tube based on the evaluation of 15 process parameters.
with k-NN regression models when varying k Experimental results suggest that distance-weighted k-
NN regression models along with the linear and the RBF-
based support vector regression methods were the most ef-
with all values of k > 1, demonstrating the benefits of ap-
fective techniques for the prediction of extruded tube qual-
plying this technique to predict the quality of extruded tubes
ity, achieving RMSE and MAE rates close to 0. From our
concerning their outer diameter.
analysis when varying the k values, we found out that when
k increases, the performance regression rates are (almost)
5.2 Inner diameter database stable.
Future research will be mainly addressed to incorporate
As in the previous section, we analyzed the behavior of the a feature selection phase to remove any attribute that might
regression models to predict the inner diameter of an ex- be considered noisy or irrelevant. Another avenue for further
truded tube. Table 4 shows the results of RMSE and MAE investigation concentrates on developing some regression
8 Vicente Garcı́a et al.

Table 4 Average results in terms of RMSE and MAE (with standard deviations) on the inner diameter database
RMSE MAE
k-NNR 0.0010767157 ± 0.0001233918 0.0008922620 ± 0.0001050170
k-NNRw1 0.0012095069 ± 0.0001328387 0.0010122028 ± 0.0001120801
k-NNRw2 0.0010783177 ± 0.0001229690 0.0008940031 ± 0.0001048994
LR 0.0011882847 ± 0.0009202389 0.0009338954 ± 0.0002309976
SVR-1 0.0013502332 ± 0.0011949720 0.0009922225 ± 0.0002923969
SVR-2 0.1471915327 ± 0.5658384933 0.0296721074 ± 0.1110184843
SVR-RBF 0.0010781128 ± 0.0001270223 0.0008608249 ± 0.0001151820
MLP 0.0013352608 ± 0.0002263802 0.0010851685 ± 0.0001695045

References
-3
10
Adly F, Alhussein O, Yoo PD, Al-Hammadi Y, Taha K,
10-3 Muhaidat S, Jeong YS, Lee U, Ismail M (2015) Simpli-
10-3
fied subspaced regression network for identification of de-
fect patterns in semiconductor wafer maps. IEEE T Ind
RMSE

10-4 Inform 11(6):1267–1276

10-4 Batista GEAPA, Silva DF (2009) How k-nearest neighbor
parameters affect its performance. In: Argentine Sympo-
10-4
sium on Artificial Intelligence, Mar de Plata, Argentina,
k-NNR
10-4 k-NNRw1 pp 1–12
k-NNRw2 Biau G, Devroye L, Dujmović V, Krzyzak A (2012) An
100
1 3 5 7 9 6 13 15 17 19 21 23 25 27 29 affine invariant k-nearest neighbor regression estimate. J
k Multivariate Anal 112:24–34
(a) Root Mean Square Error Buza K, Nanopoulos A, Nagy G (2015) Nearest neighbor
regression in the presence of bad hubs. Knowl-Based Syst
86:250–260
10-3
Carrano EG, Coelho DG, Gaspar-Cunha A, Wanner EF,
10-3 Takahashi RH (2015) Feedback-control operators for im-
10-3 proved Pareto-set description: Application to a polymer
extrusion process. Eng Appl Artif Intell 38:147–167
10-4
MAE

Caruana R, Niculescu-Mizil A (2004) Data mining in met-

10-4 ric space: An empirical analysis of supervised learning
performance criteria. In: Proceedings of the 10th ACM
10-4
SIGKDD International Conference on Knowledge Dis-
k-NNR
10-4 k-NNRw1 covery and Data Mining, New York, NY, pp 69–78
k-NNRw2 Charaniya S, Le H, Rangwala H, Mills K, Johnson K,
100
1 3 5 7 9 6 13 15 17 19 21 23 25 27 29 Karypis G, Hu WS (2010) Mining manufacturing data for
k discovery of high productivity process characteristics. J
(b) Mean Absolute Error Biotechnol 147(3–4):186–197
Fig. 5 Inner diameter database: Performance regression measures with Chevanan N, Muthukumarappan K, Rosentrater KA (2007)
k-NN regression models when varying k Neural network and regression modeling of extrusion pro-
cessing parameters and properties of extrudates contain-
ing DDGS. Trans Am Soc Agr Biol Eng 50(5):1765–1778
Chien CF, Wang WC, Cheng JC (2007) Data mining for
algorithms based on the surrounding neighborhood concept. yield enhancement in semiconductor manufacturing and
Finally, we are also interested in analyzing the behavior of an empirical study. Expert Syst Appl 33(1):192–198
ensembles of regression models. Chondronasios A, Popov I, Jordanov I (2016) Feature se-
lection for surface defect classification of extruded alu-
minum profiles. Int J Adv Manuf Tech 83(1):33–41
Acknowledgements The authors would like to acknowledge the fi- Chou JS, Ngo NT, Chong WK (2017) The use of artificial
nancial support from the Spanish Ministry of Economy, Industry and
intelligence combiners for modeling steel pitting risk and
Competitiveness [TIN2013-46522-P], and the Generalitat Valenciana
[PROMETEOII/2014/062]. corrosion rate. Eng Appl Artif Intell 65:471–483
Using regression models for predicting the product quality in a tubing extrusion process 9

Choudhary AK, Harding JA, Tiwari MK (2008) Data mining Kramer O (2011) Unsupervised K-Nearest Neighbor Re-
in manufacturing: A review based on the kind of knowl- gression. ArXiv e-prints 1107.3600
edge. J Intell Manuf 20(5):501–521 Krömer P, Snášel V, Platoš J, Abraham A (2010) Evolving
Dhafr N, Ahmad M, Burgess B, Canagassababady S (2006) fuzzy classifier for data mining - an information retrieval
Improvement of quality performance in manufacturing approach. In: Proceedings of the 3rd International Con-
organizations by minimization of production defects. ference on Computational Intelligence in Security for In-
Robot Cim-Int Manuf 22(5–6):536–542 formation Systems, León, Spain, pp 25–32
Draper NR, Smith H (1998) Applied Regression Analysis. Kusiak A (2006) Data mining: manufacturing and service
John Wiley & Sons, Hoboken, NJ applications. Int J Prod Res 44(18–19):4175–4191
Dudani SA (1976) The distance-weighted k-nearest- Kusiak A, Kurasek C (2001) Data mining of printed-circuit
neighbor rule. IEEE T Syst Man Cy SMC-6(4):325–327 board defects. IEEE T Robotic Autom 17(2):191–196
Erzurumlu T, Oktem H (2007) Comparison of response sur- Lee SK, Kang P, Cho S (2014) Probabilistic local recon-
face model with neural network in determining the sur- struction for k-nn regression and its application to virtual
face quality of moulded parts. Mater Design 28(2):459– metrology in semiconductor manufacturing. Neurocom-
465 puting 131:427–439
Ghorai S, Mukherjee A, Gangadaran M, Dutta PK (2013) Li HJ, Qi LH, Han HM, Guo LJ (2004) Neural net-
Automatic defect detection on hot-rolled flat steel prod- work modeling and optimization of semi-solid extrusion
ucts. IEEE T Instrum Meas 62(3):612–621 for aluminum matrix composites. J Mater Process Tech
González Marcos A, Pernı́a Espinoza AV, Alba Elı́as F, 151(1–3):126–132
Garcı́a Forcada A (2007) A neural network-based ap- Liukkonen M, Hiltunen T, Havia E, Leinonen H, Hiltunen
proach for optimising rubber extrusion lines. Int J Comput Y (2009) Modeling of soldering quality by using artificial
Integ Manuf 20(8):828–837 neural networks. IEEE T Electron Pack M 32(2):89–96
Guyader A, Hengartner N (2013) On the mutual near- Ma J, Theiler J, Perkins S (2003) Accurate on-line support
est neighbors estimate in regression. J Mach Learn Res vector regression. Neural Comput 15(11):2683–2703
14:2361–2376 Meiabadi MS, Vafaeesefat A, Sharifi F (2013) Optimization
Hall M, Frank E, Holmes G (2009) The WEKA data mining of plastic injection molding process by combination of
software: An update. SIGKDD Explor 11(1):10–18 artificial neural network and genetic algorithm. J Optim
Harding JA, Shahbaz M, Srinivas S, Kusiak A (2006) Data Ind Engi 6(13):49–54
mining in manufacturing: A review. J Manuf Sci Eng Oberg E, Jones F, Horton H, Ryffel H, McCauley C (2012)
128(4):969–976 Machinery’s Handbook. Industrial Press, New York, NY
Hsiang SH, Lin YW, Lai JW (2012) Application of fuzzy- Oke SA, Johnson AO, Charles-Owaba OE, Oyawale FA,
based Taguchi method to the optimization of extru- Popoola IO (2006) A neuro-fuzzy linguistic approach in
sion of magnesium alloy bicycle carriers. J Intell Manuf optimizing the flow rate of a plastic extruder process. Int
23(3):629–638 J Sci Technol 1(2):115–123
Hu C, Jain G, Zhang P, Schmidt C, Gomadam P, Gorka T Pratihar DK (2015) Expert systems in manufacturing pro-
(2014) Data-driven method based on particle swarm opti- cesses using soft computing. Int J Adv Manuf Tech
mization and k-nearest neighbor regression for estimating 81(5):887–896
capacity of lithium-ion battery. Appl Energ 129:49–55 Ramana EV, Reddy PR (2013) Data mining based knowl-
Jiang H, Yan Z, Liu X (2013) Melt index prediction using edge discovery for quality prediction and control of ex-
optimized least squares support vector machines based trusion blow molding process. Int J Adv Manuf Tech
on hybrid particle swarm optimization algorithm. Neuro- 6(2):703–713
computing 119:469–477 Ribeiro B (2005) Support vector machines for quality moni-
Khan JG, Dalu RS, Gadekar SS (2014) Defects in extrusion toring in a plastic injection molding process. IEEE T Syst
process and their impact on product quality. Int J Mech Man Cy C 35(3):401–410
Eng Robot Res 3(3):10–18 Sadeghi BHM (2000) A BP-neural network predictor model
Kohlert M, König A (2015) Large, high-dimensional, het- for plastic injection molding process. J Mater Process
erogeneous multi-sensor data analysis approach for pro- Tech 103(3):411–416
cess yield optimization in polymer film industry. Neural Sharma RS, Upadhyay V, Raj KH (2009) Neuro-fuzzy mod-
Comput Appl 26(3):581–588 eling of hot extrusion process. Indian J Eng Mater Sci
Köksal G, Batmaz I, Testik MC (2011) A review of data 16:86–92
mining applications for quality improvement in manufac- Smola A, Schlkopf B (2004) A tutorial on support vector
turing industry. Expert Syst Appl 38(10):13,448–13,467 regression. Stat Comput 14(3):199–222
10 Vicente Garcı́a et al.

Tan SC, Watada J, Ibrahim Z, Khalid M (2015) Evolution-

ary fuzzy ARTMAP neural networks for classification of
semiconductor defects. IEEE T Neural Networ Learn Syst
26(5):933–950
Urraca Valle R, Sodupe Ortega E, Antoñanzas Torres J,
Alonso Garcı́a E, Sanz Garcı́a A, Martı́nez de Pisón As-
cacı́bar FJ (2013) Comparative methodology of non-
linear models for predicting rheological properties of rub-
ber mixtures in industrial lines. In: Proceedings of the
17th International Congress on Project Management and
Engineering, Logroño, Spain, pp 1346–1357
Wang J, Ma Y, Zhang L, Gao RX, Wu D (2018) Deep learn-
ing for smart manufacturing: Methods and applications. J
Manuf Syst DOI 10.1016/j.jmsy.2018.01.003
Witten IH, Frank E, Hall MA (2011) Data Mining: Practical
Machine Learning Tools and Techniques. Morgan Kauf-
mann Publishers, Burlington, MA
Wu CY, Hsu YC (2002) Optimal shape design of an ex-
trusion die using polynomial networks and genetic algo-
rithms. Int J Adv Manuf Tech 19(2):79–87
Wu D, Jennings C, Terpenny J, Gao RX, Kumara S (2017)
A comparative study on machine learning algorithms for
smart manufacturing: Tool wear prediction using random
forests. J Manuf Sci Eng 139(7):071,018–071,026
Xu Y, Zhang Q, Zhang W, Zhang P (2015) Optimization of
injection molding process parameters to improve the me-
chanical performance of polymer product against impact.
Int J Adv Manuf Tech 76(9):2199–2208
Yang CC, Shieh MD (2010) A support vector regression
based prediction model of affective responses for product
form design. Comput Ind Eng 59(4):682–689
Yin S, Li X, Gao H, Kaynak O (2015) Data-based tech-
niques focused on modern industry: An overview. IEEE
T Ind Electron 62(1):657–667
Yu JC, Chen XX, Hung TR, Thibault F (2004) Optimization
of extrusion blow molding processes using soft comput-
ing and Taguchi’s method. J Intell Manuf 15(5):625–634
Zhang Z, Wang T, Liu X (2014) Melt index prediction by
aggregated RBF neural networks trained with chaotic the-
ory. Neurocomputing 131:368–376
Zhao G, Chen H, Zhang C, Guan Y (2013) Multiobjec-
tive optimization design of porthole extrusion die using
pareto-based genetic algorithm. Int J Adv Manuf Tech
69(5):1547–1556

View publication stats