Journal of Network Intelligence © 2022 ISSN 2414-8105 (Online)

Taiwan Ubiquitous Information Volume 7, Number 3, August 2022

Research on Quality Control Application of Whole

Process Intelligent Manufacturing in Steel Industry
4.0 Based on Big Data Analysis

Feng Zhao
School of Management Science and Engineering,
Anhui University of Technology
Xiushan District,Maanshan City, Anhui Province, China
[email protected]

Chen Yin∗
School of Management Science and Engineering,
Anhui University of Technology
Xiushan District,Maanshan City, Anhui Province, China
[email protected]

Xiaoxiao Huo
School of Aerospace Engineering,
Xiamen University
Xiang’an District, Xiamen City, Fujian Province, China
[email protected]

Yingying Xu
College of Science,
Hohai University
Nanjing China
[email protected]
∗
Corresponding author: Chen Yin
Received March 1, 2022, revised April 22, 2022, accepted June 10, 2022.

Abstract. In this paper, in the project of upgrading the intelligent manufacturing In-
dustry 4.0 of B steel plant, a large amount of measured historical sample data and real-
time data collected by sensors are processed to solve the need of the enterprise to change
from univariate to multivariate monitoring of the full process operation of the whole pro-
cess of the strip in the period of industrial development of big data cloud platform. In this
paper, the monitoring and diagnostic identification of the quality of the whole process are
achieved by applying a set of algorithm structures such as improved ZC4.5 decision tree
algorithm - MSPM algorithm (comparing performance with MICA algorithm) - improved
statistical result comparison analysis Granger analysis. This paper proposes to redevelop
the general idea and methodology of data processing based on the theoretical idea of the
MES system of steel plants to upgrade the big data technology is conducive to the trans-
formation of a single process to full process quality control mode, improve fault diagnosis
accuracy, reduce manual work, lay the foundation for product quality improvement of
enterprises, and can be extended to most steel enterprises.
Keywords: Big data analytics; intelligent Manufacturing; Whole process quality man-
agement; MSPM algorithm

685
686 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

1. Introduction. In recent years, because of the cyclical changes and adjustments in the
steel industry, a mega-combined enterprise has been gradually integrated and established,
and the production line has increased from the initial steel treatment to the subsequent
production processes. However, the corresponding companion is still very far apart. With
the advent of Industry 4.0 (including technologies related to Industrial Internet of Things
(IIoT), Artificial Intelligence (AI), process simulation and optimization, cognitive com-
puting, and cloud computing [1]) paradigm, the original MES (manufacturing execution
system) system can no longer meet the requirements of the existing information system
development[2]. In addition, this research experiment required is very high, because the
equipment for iron and steel enterprise investment is huge, so to realize the whole process
of monitoring all the way, must go to all involved in the process of plant collection in
each production line PLC (programmable logic controller) sensor data, the production
line PLC said there are hundreds of little quantity. The data of thousands of sensors need
to be synchronized with the clock (PLC data corresponds to the time of the same product
through different processes), and the transmission data is linked with the characteristic
equation processing, sorting out the dimensionless. Jaskó et al.[3] mentioned in their
literature review that next-generation MES solutions need to have such machine learning
(ML) data mining capabilities. The purpose of this paper is to show the development
and results of relevant data mining functions, as can be seen from the review of relevant
industrial informatics papers by Chen[4].
Due to the increase in production processes and the development of sensor technology,
the amount of industrial-level data has gradually entered into explosive growth, which
can no longer be solved by pure statistical models, and Kano and Nakagawa [5] pointed
out that statistical models are not suitable for processing large amounts of data when
looking at the future of the steel industry, so the research method of control charts based
on statistical theory is not a practical application development direction for future steel
enterprises.
The new quality control system must adapt to the existing MES system and ERP(Enterprise
Resource Planning) system, and realize the monitoring, control, and optimization of THE
manufacturing process by the MES system. The information provided by MES helps
decision-makers understand the interconnections among the various subsystems involved
in the production, and this knowledge can promote continuous improvement of the manu-
facturing industry. Therefore, it is necessary to refer to the previous development system
of quality management theory when developing a new multivariate monitoring model.
This paper mentions the use of Six Sigma theory about the whole product life cycle the-
ory, similar to what happens in the production chain mentioned by SzilardJasko, Adri-
ennSkrop. In product Lifecycle Management (PLM) systems, there is a common basic
idea about the industry 4.0 paradigm, everything must support computerization. Use
computer-based controls throughout the production chain.
Data-driven methods are complementary techniques that can be used in conjunction
with classical and advanced control algorithms. Their main focus in complex processes
is to ensure compliance of product quality [6], presents interactions between its different
variables, time-varying parameters of mass and heat transfer nonlinearities dynamics, and
large time delay due to the slow chemical reactions [7]. These intrinsic features make the
use of model-based techniques for fault detection and control purposes very difficult it
is not unrealistic [8]. Furthermore many cement plants worldwide still use centralized
manual control methods to ensure operation. Therefore, it would make sense to integrate
appropriate fault detection and diagnosis systems that can easily and accurately monitor
this type of industrial process.
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 687

Regarding the methods applied in this paper for quality monitoring, Qin [9] mentioned
in a literature review that due to the data-based nature of the Statistical Process Moni-
toring SPM method, it is relatively easy to apply to real processes of considerable scale
compared to other methods based on system theory or rigorous process models. Tradi-
tional SPM methods include principal component analysis (PCA) and partial least squares
(PLS) [10], and it is generally assumed that the system variables are static and follow
a Gaussian distribution. Recent developments in SPM also include dynamic modelling,
non-Gaussian distributions, and non-linear methods [11, 12], they can be applied to a
wider range of industrial processes. Due to the data-based nature of the SPM methods,
it is relatively easy to apply to real processes of rather large scale comparing to other
methods based on systems theory or rigorous process models. The mature use of con-
trol charts tends to focus on univariate monitoring, and the commonly used univariate
control charts are mainly Shewhart control charts, cumulative sum control charts, and
exponentially weighted moving average control charts. The use of Multivariate Quality
Control techniques is usually avoided by practitioners because of the complexity involved
in the design, implementation, and maintenance of the control system. Sepú lveda and
Nachlas [13] proposed a simulation approach to multivariate quality control. From now
on, it does not apply to the application of large-scale industrial-level data, and the main
mechanism of control charts is based on statistical theory, so it must be discarded, also in
line with the academic trend that machine learning theory will surpass statistical theory
in future research. Bakdi et al. [14] used PCA analysis in chemical applications applied to
cement chemical production, some scholars believe that machine learning algorithms such
as PCA and some evolutionary principal meta-analysis methods cannot indicate which
process variables are responsible for the anomalies, the contribution of this paper lies
in such a large data magnitude of a day’s update in just over 100,000 industrial-grade
data, through decision tree algorithm-MSPM algorithm-Granger analysis, to achieve the
monitoring and diagnosis of the quality of the whole process Identifying, the experiments
locate the final process variables through a simplified Granger analysis to achieve the iden-
tification of abnormal sources. A set of identification variables for the process window is
provided for validation by researchers with similar product lines The analysis results are
shown in Table 1.

Table 1. A set of experimental process window monitoring variables

Number Case 1: Pressure oxygen quality monitoring point

1 Si, As content monitoring
Air-fuel ratio and heating time of the first heating section,
2
second heating section, third heating section and soaking section
3 Total heating time, air excess coefficient
4 Pressure, flow rate and effect of one-time phosphorus removal
5 Rough rolling phosphorus removal pass, pressure
Exit temperature of roughing mill and maximum
6
temperature of finishing mill entrance
7 F1-F4 slot spray open state
8 Rolling rhythm and predicted temperature of wheel face
9 Phosphorus removal by finishing rolling
10 F1-F4 working rod cooling water, etc.
688 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

Why combine decision tree algorithm with MSPM algorithm? Because multivariate
statistical methods use only few possible independent sources mostly causing the varia-
tion in the process and rendering these methods to be very efficient in monitoring large
scale processes [15]. The kernel entropy component analysis method can also be consid-
ered here in combination with the MSPM algorithm. However, for large scale systems
including many complex units, data-driven methods make a superior alternative solution
for fault detection and diagnosis systems [16, 17]. Generally, large size multivariate data
with highly correlated variables present low statistical rank. Since the number of process
variables on the production line is very large, even if the scope is reduced through metal-
lurgical experience, it is necessary to further determine the main variables. In this paper,
the decision tree algorithm is firstly used to screen out three main process variables: the
shape size of 2# rolling mill, steel passing rate, and the height of 1# ∼5# looper sleeve.
Because of the steel passing rate (PLC drawing speed), this process variable read all of
his PLC related data acquisition volume, including: Column1: Setting value of drawing
speed; Column2:Actual drawing speed; Column3: Set liquid level of molten steel; Col-
umn4: Actual liquid level of molten steel; Column5: Opening and closing value of plugrod.
Then the MSPM algorithm is analyzed for the data on the 1000 time points of these five
data collection points (note that this step of the collection should be clock synchronized,
because there is a time difference when the steel is run through different PLCs in the
production line). Finally, the source of the anomaly is introduced in reverse by Granger
analysis. This paper focuses on the structure of this algorithm, in which the formula
code of the decision tree algorithm - MSPM algorithm - Granger analysis is modified in
order to adapt to the line data of different process variables in different production lines,
whether the data is selected for dimensionless processing depends on the process variables
in the metallurgical process.
As general-purpose steel, the plate has a wide range of uses in various fields. The
plate can be processed into various components and products, so it has a large number of
customers. However, there are still many problems with sheet and strip products. There
are many individual customer needs, and the process of producing products through
the ”identify-verify-cure” process is extremely complex. The sheet market is still facing
high demand and low supply, despite the availability of raw materials, capital, and labor.
Technological progress is the main breakthrough direction, the whole process of integrated
design and personalization of the digital twin model needs to be established. The quality
control technology of big data analysis of the whole process studied in this paper can
provide an important means of accurate service and ultimately achieve the ultimate goal
of improving the efficiency of enterprises.

Figure 1. Actual production line

 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 689

1.1. Existing device information system. L1: Basic automation system refers to the
PLC control unit used for automatic equipment control in the strip production line. This
system is mainly used for recording process curve data, critical event status data, etc. L2:
Process automation systems cover many mechanized operations on the production line.
Including the mathematical model and material tracking system, storage material and
process parameter set value, measured value (feedback value), statistics, and other logical
correspondence. At the same time, it records the processing time information of products
in each piece of equipment. The steelmaking area includes (converter - desulfurization
- pouring - KR - refining - flame cleaning machine - 1, 2# continuous casting machine,
3# continuous casting machine) - hot rolling area - cold rolling area (acid rolling - 1#
galvanizing - 2# galvanizing - cover receding - leveling - shearing). Another external
property line is (acid rolling - 1# continuous rewind - 2# continuous rewind - recoil - 1#
annealing furnace - 2# annealing furnace). In the process of information transformation,
it is also a complex and arduous project to build the digital twin model of the site. This
paper focuses on data mining, in which the model is similar to human skeleton and data
mining is human blood. The actual production line is shown in Figure 1. The production
diagram of crude steel and its one - to - one ratio digital twin model is shown in Figure 2.

Figure 2. Production diagram of crude steel and its one - to - one ratio
digital twin model

1.2. Problems existing in the strip process quality system. (1) The existing con-
trol systems lack attention to the quality determination process data, and each data is
relatively isolated in the whole process flow operation process of the strip. Since there are
many factors affecting product quality, there is a high-dimensional and multivariate cou-
pling problem in the production process. It is difficult to discover the root causes affecting
quality through simple threshold measurements. In particular, the relationship between
the problems occurring in hot rolling and steelmaking, and the relationship between the
problems occurring in cold rolling and hot rolling require further cluster analysis and
cross-process correlation analysis of the process data. There is a lack of effective corre-
lation analysis tools for exploring quality problems in the analysis of multidimensional
factors.
(2) The production depletion of sheet and strip is high. It is difficult to match the in-
dividual needs of customers with the whole process of product quality production. The
only remaining transactions are still in the form of paper agreements. The entire strip
production process cannot be quality controlled for different needs. Because of the ex-
isting problems, there is an urgent need to improve the way of exchanging opinions to
increase efficiency in the sheet trade.
(3) External disturbances in process input and quality variation in the process window
are still difficult to control. At the production site, the formulation of process windows
such as speed, temperature, heating temperature, arrival time, arrival composition, etc.
is still determined empirically. The above-mentioned effects on the surface quality of the
strip have not been included in the analysis. The strip production quality lacks long-term
690 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

stability, data traceability is inefficient, and the boundaries between processes and depart-
ments are unclear. Due to departmental divisions, quality problems cannot be effectively
tracked and analyzed between different processes

2. General idea. (1) The researchers improved the PCA algorithm and designed the
algorithm multivariate statistical process monitoring (MSPM), which is suitable for de-
termining the surface quality of plates.
(2) To solve the problems arising in the process, the researchers used decision trees to
filter and analyze the relational data, screened the three most representative operational
process variables to provide the basis for the new algorithm, and selected the training
data set.
(3) To meet real-time monitoring requirements, the screened data are entered into a
database and modelled to visualize the data (process curve control charts). The plate
production data is monitored for abnormalities during the production process.
(4) The MSPM algorithm and supporting software were developed by solving the actual
problem of plate production in company B, thus verifying the purpose of the MSPM al-
gorithm. As a result, it both meets the actual production needs and accelerates the pace
of moving towards Industry 4.0.

2.1. Concept Statement. (1) Decision tree model: the introduction of penalty param-
eter Z improves the ID3 algorithm to form a new C4.5 algorithm more suitable for process
variable screening.
The decision tree algorithm used in this paper for screening key factors is a very classi-
cal machine learning algorithm, which is suitable for integrated learning such as random
forest algorithm and can be used as a regression algorithm, and also can be used as a
classification. The decision tree model is a typical tree structure, and its learning process
consists of feature selection, decision tree generation and pruning. Since this paper applies
the decision tree algorithm to screen the process variables that play a major role in the
quality problem, the pruning process is not considered in the paper.
The decision tree algorithm uses a tree model, where the linear model is a linear model
where all features are given weights to sum up to get a new value, while the tree model is
a partition for each feature. Decision trees can find non-linear divisions. The tree model
is closer to the human way of thinking and can produce visual classification rules that
produce models with interpretability. The function fitted by the tree model is a step
function of the partition. The decision tree starts at the root node, which iterates from
the top down to produce multiple internal nodes and leaf nodes. Each node of the tree
represents a test of a feature, and the branches of the tree represent the results of each
test of that feature. Each leaf node of the tree represents a category, and the final child
node that cannot be split is called a leaf node and can represent the final category.
As can be seen above, the core of the decision tree algorithm lies in how to make the
optimal attribute selection, and there are three main criteria for the optimal selection of
decision trees, which are maximum information gain, maximum information gain ratio
and Gini coefficient. The algorithms corresponding to these three criteria are ID3 algo-
rithm, C4.5 algorithm and CART algorithm.
(2) Information gain: the difference between the empirical entropy of set D and the em-
pirical conditional entropy H(D|A) of set D under the given conditions of feature A.H(D)
denotes the empirical entropy of data set D, H(D|A) denotes the empirical conditional
entropy of set D under the given conditions of feature A, g(D, A) denotes the information
gain, and the information gain is calculated as follows:
Let A be - a discrete random variable with finite values and its probability distribution
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 691

is :
P (A = xi ) = pi , i = 1, 2, · · · , n (1)
Then the first degree of the random variable A is defined as:
X n
H(A) = − pi log pi (2)
i=1

The conditional entropy function H(D|A) is defined as the uncertainty of a sample set D
of random variables conditional on a random variable.
H(D|A) is the mathematical expectation of the entropy of the conditional probability
distribution of D of A under the given conditions on A.
X
H(D | A) = p(a)H(D | A) (3)
a∈A

Where p(a) denotes the probability of occurrence of A = a.

g(D, A) = H(D) − H(D | A) (4)
k
X |Ck | |Ck |
H(D) = − log2 (5)
k=1
|D| |D|
n n k
X |Di | X |Di | X |Dik | |Dik |
H(D | A) = H (Di ) = − log2 (6)
i=1
|D| i=1
|D| k=1
|Di | |Di |
For the sample set D, the random variable A is the category of the sample, i.e., the sample
is assumed to have k categories and the probability of each category is |Ck|/|D|, where
|Ck| denotes the number of samples of category k and |D| denotes the total number of
samples.
Information gain - G(D, A) indicates the degree of uncertainty reduction in the classifi-
cation of the dataset due to the selection of feature A. The more the reduction, the lower
the uncertainty in the classification of the dataset. The more the reduction, the lower the
uncertainty of the dataset classification.
H(D) indicates the entropy of the label category of the dataset, i.e., the uncertainty when
each label takes the value of a different category.
H(D|A) denotes the entropy of the category label of the dataset under the condition
that features A is selected. In this case, it can also represent the mutual information
of category labels and features. (3) ID3 algorithm calculates the information gain of all
node’s technical features and selects the feature with the largest information gain for
splitting.ID3 algorithm tends to select features with more values, and sometimes this ten-
dency brings some error in the construction of decision trees. Wang et al. [18] proposed
an RLBOR algorithm, which considers The number of nodes in the decision tree model to
optimize the decision tree optimization ratio, Decision Tree Optimization Ratio but still
can not avoid the error, such as extreme conditions, based on a certain attribute after
splitting, a subset corresponds to a data, when the information gain is maximum and the
information entropy is 0, but this division has no value. Because metallurgical process
variables have a relatively small range of data variation due to the variables themselves,
there is continuous data, and ID3 will favor that feature A. To correct this error in the
ID3 algorithm, the C4.5 algorithm that uses the information gain ratio as the optimal
attribute selection index is proposed. Mu et al. [19] mentioned the application of the C4.5
algorithm and pointed out that in supervised classification, large training data are very
common and decision trees are widely used. However, many supervised classifiers (includ-
ing the classical C4.5 tree) cannot directly handle large data due to some bottlenecks such
692 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

as memory limitations, time complexity, or data complexity. In supervised classification,

large training data are very common, and decision, However, as some bottlenecks such
as memory restrictions, time complexity, or data complexity, many supervised classifiers
One solution for this problem is to design a highly parallelized learning algorithm. One
solution for this problem is to design a highly parallelized learning algorithm. And in this
paper, we need a C4.5 algorithm for feature selection of metallurgical process variables,
and then choose other machine learning algorithms for big data processing by combining
the characteristics of metallurgical data. The proposed ZC4.5 algorithm is that when a
feature corresponds to too many values, and improved penalty parameter Z is taken to
multiply the information gain so that the information gain ratio is small. The information
gain ratio is defined as the ratio of the information gain brought by feature A to set D to
the entropy of feature A itself.
Information gain ratio C4.5: The ratio of the information gain brought by feature A to set
D to the entropy of feature A itself. The information gain ratio is calculated as follows:
gr (D, A)
gr (D, A) = (7)
HA (D)
n
X |Di | |Di |
HA (D) = − log2 (8)
i=1
|D| |D|
Information gain ratio = penalty parameter * information gain.
Penalty parameter info=1/HA (D) ∗ Z

 0 R < 1%
Z= 1 1% < R < 30%
null 30% < R

When the value of Z changes in the range of less than 1%, the code set Z = 0, to facilitate
the writing of the model code calculation automatically screens invalid process variables
to reduce the pressure of the calculation.
when the range of value change of Z is greater than 30% (this PLC has an error), the
code sets Z = N ull at this time to indicate that the data model this PLC has an error.
When the value of Z changes in the range of 1 to 30%, indicating that the data can be
operated normally, Z = 1.
In the formula, n indicates the number of subsets after splitting based on feature A, and
Di indicates the number of samples in each subset.
(4) PCA(Principal Component Analysis), PCA, is a technique for data analysis.PCA can
extract the main factors of multivariate things, remove the noise and redundancy in the
whole data, and reveal its essential characteristics.

3. Experimental analysis. In this paper, a new multivariate statistical process moni-

toring (MSPM) algorithm is formed by adding Hotelling- T2 and squared prediction error
(SPE) as indicators for fault determination based on PCA dimensionality reduction, which
transforms the univariate monitoring algorithm of metallurgical production lines into a
multivariate monitoring algorithm.
3.1. Initial screening model of indicators based on decision tree algorithm. In
this paper, the historical data of 13 production process parameters were obtained from
steel plant B. The decision tree algorithm was used to feature select these parameter
variables, and the three most representative and independent operation variables were se-
lected from these variables , to retain the main features for the next step of model training
and ignore some minor factors. The actual data of each operational variable are limited
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 693

to a certain range according to the needs of the actual production environment. For the
acquired 60,000 sample data, the quality parameters of qualified samples are marked as 1,
and the quality parameters of unqualified products are marked as 0. The truncated part
of the data is shown in Table 2 below. We used the C4.5 classification tree for feature
Table 2. Historical data of process parameters of strip production line

Sample
1 2 3 4 5 6 7 8
number
Steel speed 128.53 159.92 156.64 141.45 125.03 117.87 168.81 151.6
Descaling
16.18 15.3 16.94 14.88 14.74 17.78 15.09 15.72
pressure
1#∼5#Looper
110.66 126.38 109.25 116.37 129.91 133.62 128.17 113.56
height
Tension control 1.98 1.42 0.76 0.88 0.63 0.72 0.86 1.37
Heating
section 1140.52 1153.61 1173.83 1152.34 1121.4 1137.25 1125.36 1120
temperature
Average heat
segment 1165.74 1195.66 1166.72 1194.48 1181.12 1185 1197.23 1162.36
temperature
Advance
quantity of 9.45% 5.07% 5.93% 2.31% 3.98% 9.11% 4.16% 7.18%
pinch roll
Finishing mill
81.66 19.36 37.73 16.88 83.21 91.84 82.88 39.5
speed
2#Rolling mill
17.16 17 16.9 16.84 17.31 17.23 17.05 16.75
material size
Roller speed 78.98 71.83 61.35 79.2 82.87 60.56 83.54 60.55
Fan delivery 11.97 23.02 0.37 39.53 64.08 28.61 85.09 94.92
Guideway
alignment 0.2 -0.32 0.33 0.16 -0.08 -0.38 -0.06 -0.48
accuracy
Water tank 2.06 1.45 2.01 0.3 2.14 0.84 1.39 2.47
Quality 1 0 0 1 1 1 0 1

selection on the above sample data. Since the samples were classified as qualified (noted
as 1) and unqualified (noted as 0), a binomial classification tree was used as the training
model. We calculated the information gain ratios corresponding to the 13 feature vari-
ables and sorted them in order from largest to smallest, and selected three main factors
based on the information gain ratio of each feature variable.
The experimental environment was set up as Intel(R) Core(TM) i5-5200U CPU @3.60GHz,
and the operating system was Windows 10, which was implemented in the Jupyter note-
book platform using the Python language. The analysis results are shown in Table 3.
According to the quality inspection department’s records of historical samples, the main
factors affecting the surface quality of the strip are the size of the mill shape (2#), the
oversteel rate, and the height of the live sleeve (1#∼5#) produced by the equipment. The
variation of the above three variables is mainly monitored at the time of going online.
We define thresholds in our experiments, above which an alarm is generated. Through
694 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

Table 3. Information gain analysis results

Characteristic variables Information gain

of process parameters ratio
2# rolling mill material shape size 0.1645
Steel speed 0.1389
1#∼5# looper height 0.1229

process control, surface quality problems such as periodic wax hanging, iron oxide, and
cracks can be reduced for this process parameter. To meet the needs of the enterprise,
the univariate monitoring problem is transformed into a multivariate monitoring problem
after extracting the important influencing variables, and a multivariate monitoring model
is developed.
3.2. Dimensionality reduction model for quality inspection data based on prin-
cipal component analysis. Principal component analysis (PCA) is one of the basic
projection models in multivariate statistical analysis. PCA transforms the original data
into a set of linearly independent representations of each dimension through a linear trans-
formation, which can be used to extract the main characteristic components of the data
and is often used for dimensionality reduction of high-dimensional data. If there are n
variables in the original data table, PCA will consider the information in this data table
to readjust the combination and extract P main feature variables from it (p < n).
Let x ∈ Rm be the measurement samples of m sensors, and each sensor has n indepen-
dent samples. Construct matrix X = [x1 , x2 , . . . . . . , xn ]T ∈ R(n×m) where each column
represents a measurement variable and each row represents a sample. Treat the columns
of X as variables with unit variance and zero mean. Define S as the normalized sample
X covariance matrix. Decompose the eigenvalues and arrange them in descending order.
The improved PCA model is used to decompose X as follows:
X = X̂ + E = T P T + E (9)
T = XP (10)
m×A
Where P ∈ R is the load matrix composed of the first A feature vectors of S. T ∈ Rn×A
is the scoring matrix. The columns of T are called principal variables, and A represents
the number of principal elements.
PCA model divides the variable space into two orthogonal and complementary subspaces.
The subspace composed of all columns of P is called the principal component space (PCS),
and the orthogonal complement of PCS is called the residual subspace (RS). Any sample
vector can be decomposed into a projection on a host subspace and a residual subspace:
x = x̂ + x
e (11)
x∧ = PPT x ∈ Rp ≡ P (12)
T ⊥


e = I − PP x ∈ R ≡ P
x (13)
∧
x denotes the projection of the sample in the principal element space that is modeled,
and xe denotes the projection of the sample in the residual space that is not modeled.
The input: N dimensional sample set D = (x(1) , x(2) , . . . , x(m) ), the dimension reduced
to p;
The output: Sample set D′ after dimensionality
Pm reduction;
(i) (i) 1 (j)
1) Centralize all samples: x = x − m j=1 X ;
2) Calculate the covariance matrix of the sample xxT ;
3) Do the eigenvalue decomposition of xxT ;
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 695

4) Take the eigenvectors {ω1 , ω2 , . . . . . . , ωn′ }, after all the feature vectors are normalized,
the feature vector matrix W is formed.
5) To a new sample Z (i) = W T x(i) ;
6) Get D′ = (z (1) , z (2) , . . . , z (m) ).
3.3. Calculation of statistical indicators and principles of multivariate process
monitoring.
(1) T2 statistics
Hotelling-T 2 statistics measure the variation of variables in the principal metric space.
T 2 = xT P/pT x ≤ Tα2 (14)
∧ = diag {λ1 , . . . . . . , λA } , Tα2 2
stands for T limit of control with a confidence of α. As-
suming that the samples under normal operation of the process obey multivariate normal
distribution, the control limits can be calculated by following the equation.
A (n2 − 1)
Ta2 = FA,n−A;a (15)
n(n − A)
FA,n−A;a is the critical value of F distribution with A and N − A degrees of freedom and
the confidence interval of α.
(2) Square prediction error index
The SPE metric measures the change in the projection of the sample vector in the residual
space.
SP E =∥ I − P P T x ∥2 ≤ δa2


(16)
δa2 represents the control limit with the confidence interval of α. The process is considered
normal when the SPE is within the limits of control. The calculation formula of the
control limit is
p ! h1
2 2
c a 2θ h
2 0 θ h
2 0 (h 0 − 1) 0

δa2 = +1+ (17)

θ1 θ12
Where θi = m i 2θ1 θ3
P
j=A+1 λj (i = 1, 2, 3), h0 = 1 − 3θ12 , λj is the eigenvalue of the covariance
matrix of X, Ca is the threshold of the standard normal distribution at the confidence
level, A is the number of principal elements of the PCA model, and m is the dimension
of sample X. If SP E ≤ δa2 occurs, then this part of the metallurgical process is normal in
terms of the pull speed value process. When a fault occurs, the projections of the faulty
problem part and the normal part are combined into a process process problem sample
vector X. This makes SP E > δa2 , δa2 denote the control limit of SP E, and this expression
is similar to the third order moment distribution of SP E.
(3) Model Building
We divide the multivariate process monitoring modeling process into four steps: model
building, data selection, data processing, and model updating. The variable names in
the data table are first defined and displayed on the development interface, and then
the production operator selects the desired variables and the corresponding data ranges
to build the model. For this experiment, we select the five process variables in the table
shown in Figure 3 for analysis. After determining the data range of the model, the operator
saves the information of the created model in the form of an XML file to a specified address
for the next call to the model information. Using c# programming software to import the
above multivariate process monitoring model based on the improved PCA algorithm, and
develop the normality curve module, and then use Minitab and other software to test the
normality of the data of the above five variables again (Stat-Basic Statislics-Normality
Test), resulting in data P = 0.46, which is greater than 0.05. This means that the data
are normally distributed. If the operator chooses not to introduce new variables in the
696 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

Figure 3. Range of data for analysis of five process variables for the se-
lected example

data processing interface, the data processing interface will directly call the model created
by himself, and the following results will be obtained for the training sample projection
data processing in Figure 4, as well as the results of the indicator analysis in Figure 5
and Figure 6. We monitored T 2 to determine whether the data were abnormal or not,

Figure 4. Two-dimensional scatter of PCA projection data for multivari-

ate monitoring data processing interface

and judged whether a fault occurred according to the SP E index, and combined with the
PCA principal component analysis method to construct the MSPM model. After testing,
as shown in Figure 5, the five variables of pulling speed test value, actual pulling speed
value, level setting value, actual level value, and plug bar opening and closing value were
stable in the first 1000 sample points of the principal element projection T 2 , and rarely
exceeded the upper limit value. As shown in Figure 6, the SP E has a larger variation
and easily exceeds the upper limit. After theoretical analysis, this is because most of
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 697

the data of these five variables tend to be normally distributed in the steel metallurgical
process, and due to a large amount of sample size data, the T 2 control limit Tα2 not only
tends to be F-distributed but also will tend to be more chi-squared. This leads to the
possibility that the control limits of the SP E indicators for theoretical process monitoring
are not realistic, but we have included the examination of the local sample range in writing
the entire module code, which can be used to observe whether the specific fault variable
samples are consistent with the SP E and T 2 alarms through statistical test plots.

Figure 5. T 2 control chart of test sample data for multivariate monitoring

data processing interface

Figure 6. SP E control chart of test sample data for multivariate moni-

toring data processing interface

As shown in Figure 7, the sample data with problematic SPE statistics are 11, 20, and
31, respectively, while the variables with T 2 overruns on the left side after local Tα2 control
limit adjustment are also: 11, 20, and 31, in line with the practical test accuracy. Figure
698 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

8 shows the respective contributions of the two indicators in this projection, while some
papers propose the use of a composite indicator to combine the two indicators, which is
not in line with the actual operation of the production staff and is not conducive to their
judgment, so it is not used.

Figure 7. Comparison of SP E and T 2 statistic consistency for training samples

Figure 8. Contribution values of T 2 and SP E statistics of training samples

From the data obtained from the production operation and management department,
the problem of quality monitoring variation of each batch of steel strip may be related
to the influence of special differential working conditions such as duty personnel and
process procedures, while the previous univariate monitoring obviously had many other
indicators that chose to be ignored when a quality problem occurred, and therefore could
not detect the problem of its correlation with important process variables. The improved
PCA algorithm model provides a multivariate monitoring approach, which well improves
the detection accuracy of the model and effectively solves the problem.
(4) Model Validation
To study and verify the effect of the algorithm, we collected records of batches with
warpage in steel strip production in a quarter of B steel enterprise and analyzed the
data. In the process of analysis and training, 200 faulty batches, as well as 100 normal
batches, were collected as the test data set, and 14 process values such as mill speed
difference, live sleeve set amount, and temperature difference for each batch in this data
set were collected as training data set, as shown in Table 4. The previous univariate
fault detection rate FDR as well as the fault false alarm rate FAR are compared, and
the kernel parameters of the conventional PCA algorithm for unilateral quantities as well
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 699

as the improved MSPM algorithm are unit normalized, and the kernel density is used
to estimate the control limits, i.e., the empirical values are used to estimate the specific
values. The corresponding values of the Fault Fault Reporting Rate are given in Table 5.

Table 4. The 14 process value indicators collected

Process value Variable name Variable number

1zone∼3zone Temperature value of heating zone 3
procedure
1#∼4# Looper height 4
parameter
1#∼8#frame speed frame speed difference 7

Table 5. Univariate detection statistics of warped skin and detection rate

of MSPM software
Algorithm Variable indexes Texting indexes The rate of checkout(%)
FAR 0.0
SPE
FDR 18.5
MPSM Algorithm
FAR 0.0
T2
FDR 12.0
FAR 6.0
I2
FDR 11.5
MICA Algorithm
FAR 7.0
T2
FDR 3.0

Figure 9. T 2 versus SPE data in MPSM algorithm (red line is alarm threshold)

(5) Comparison of the improved MPSM algorithm with the traditional mul-
tiplexed independent component analysis MICA algorithm
Figures 9 and 10 above demonstrate that the fault detection rate FDR of MSPM reaches
18.5%, while the detection result of traditional MICA is only a modest 3%. MPSM is
obviously better than the traditional monitoring means, and the false alarm rate has also
improved significantly, with 0% false alarms. Because there is a multivariate projection
in the monitoring process of the analysis of the operational process, so after getting the
fault alarm data, can not immediately determine which variable is the problem, the 14
700 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

Figure 10. T 2 versus SPE data in MICA algorithm (red line is alarm threshold)

variables there are multivariate coupling role through Granger causality analysis, the vari-
able at the beginning of the arrow triggered the change of the variable at the end of the
termination, the number of times each variable triggered other variables in descending
order ranking, locating the root cause as shown in Figure 11 of.

Figure 11. Interaction of 14 variables

For this situation, we can analyze by Granger analysis, Granger mentioned that his the-
ory has a very limited scope of application and cannot be used as a real method to obtain
the final results, Tiwari, and Aviral. [20] used a modified asymmetric Granger analysis
to solve the relationship of the actual energy consumption, so this paper considers the
complexity of the actual production process, the product anomalies are often influenced
by multiple variables or coupled between multiple variables. Obviously, if only one failure
variable is identified as the root cause of the failure, it cannot meet the actual produc-
tion needs. Therefore, Granger analysis should be improved here to adapt metallurgical
process variables for analysis, and the results of statistical analysis should be compared
with Granger causality analysis to analyze the root cause of failure in detecting faulty
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 701

batches for further analysis. Figure 12 shows the statistical appearance of the 14 process
variables that produce the impact of the abnormal warpage situation against Figure 11.

Figure 12. Statistics of 14 process variables producing effects on abnormal warpage

The statistics show that the most influential variables are 2, 7, and 8. The sample size
of this experiment was then expanded to 100 data, as shown in Figure 13.

Figure 13. 100 statistics of 14 process variables producing effects on ab-

normal warping cases

The results show that 2, 7, and 8 are still the most influential variables, and 2, 7, and
8 correspond to the temperature difference of the average heat section, 2# live sleeve
height, and 7# rack speed difference, respectively. This shows that the MSPM algo-
rithm can monitor the data abnormality more accurately, so as to alarm the fault, but
also through Granger analysis and statistical analysis combined to further determine the
specific variables that cause the abnormality, so as to reduce the false alarm rate.
4. Conclusion. The target set by the China Industry and Information Technology Bu-
reau by 2020 is to compress crude steel production capacity by 100-150 million tons, while
according to the research in this paper, it is expected that China’s crude steel consumption
will also decline by about 3% per year after 2020, and the current oversupply situation in
the steel industry will not change shortly. In this context, the concept of whole-process
quality management is proposed and practiced, forcing enterprises to continuously up-
grade their quality management level and improve the core competitiveness of products is
702 F. Zhao, C. Yin, X.X. Huo and Y. Y. Xu

the most urgent need at present. At present, the Internet technology continues to break
through, injecting new momentum for industrial development, Chinese steel enterprises
to achieve the curve to overtake, need to be in the process of information technology will
be customized data mining big data analysis technology and production process for deep
integration, thus helping China in the process of production line automation information
technology to develop and master more core independent technology.
The main problems we want to solve now are the accurate implementation of customer
requirements, multi-module collaboration to achieve in-event control of product quality
and consistent quality, one-click traceability of defects, rapid adjustment and optimiza-
tion of process parameters with the help of AI technology and visualization technology
such as digital twin, and the provision of customized business to achieve collaborative
manufacturing of quality, process, and equipment operation. Nowadays, the generalized
Chinese steel mill informatization architecture is mainly a five-layer hierarchical quality
information system architecture. The current idea is to make it a flat quality information
system architecture, which can meet both the production requirements in management
and the quality needs in production, in line with the currently recognized more main-
stream trend of steel mill informatization. The contribution of this paper is to narrow
down the selection range of influencing process variables through the improved Z4.5 al-
gorithm (the limitation of the algorithm does not allow for large data processing), to
apply the MSPM algorithm in chemical production to metallurgical process data analysis
and compare the algorithm performance with the MICA algorithm, to monitor multiple
variables of production data simultaneously, and finally to accurately diagnose anomalies
through the improved Granger analysis of comparative statistical results. The anomalies
are diagnosed accurately and a set of monitoring points is provided for the engineers of
the relevant product lines to verify. This paper also combines experimental kernel entropy
component analysis (KECA) with DISSIM, a statistical process monitoring method for
process data variability analysis (DISSIM) proposed by KANO et al. [21] to replace the
combination of the Z4.5-MSPM algorithm to test the actual effect.
Because the experimental conditions in this paper are very demanding, so the research
workers conducting related aspects of the published literature is very small, and many of
them declare it as a patent, or a means of industry profitability, for smart manufacturing
Industry 4.0 many scholars who are not practicing on the front line consider the further
upgrade of the MES system [22]. This paper tends to integrate many of the original
upstream and downstream factories in the steel manufacturing industry, and inject the
original design MES and ERP concepts into the full computer control system by inte-
grating the concept of the whole product life cycle, and in the future, by establishing a
series of algorithmic models, data analysis and data The future will be through the es-
tablishment of this series of algorithmic models, and data analysis and data visualization,
to achieve the digital twin, which can be a large amount of today’s data for rapid cloud
computing processing, so as to guide production. Consider network security issues for
enterprise cloud data, enterprises can refer to key Agreement Protocol in Cloud-based
Smart Healthcare Environments proposed by Wu et al. [23].
The general idea and methodology of this paper for plant data processing can be of
guidance and reference value in continuous production process continuous optimization,
yield improvement, quality improvement, equipment warning, energy optimization man-
agement, etc.

REFERENCES

[1] M. Hermann, T. Pentek, B. Otto, “Design principles for industrie 4.0 scenarios,” in 49th Hawaii
International Conference on System Sciences (HICSS), 2016,pp. 3928-3937
 Quality Control Application of Whole Process Intelligent Manufacturing in Steel Industry 4.0 703

[2] K. Wang; P. Liu; A. Zhao; Q. Zhang; L. Wang; Y. Xue; X. Gao; D. Gao, “Development and
Application of MES Based on Cloud Platform for Steel Structure Enterprises,” in IEEE International
Conference on Industrial Engineering and Engineering Management .IEEM , 2019,pp. 521-525.
[3] S.Jaskó, A.Skrop, T. Holczinger, T. Chován, J. Abonyi, “Development of manufacturing execution
systems in accordance with Industry 4.0 requirements: A review of standard- and ontology-based
methodologies and tools,” Computers in Industry, vol. 123, 103300, 2020.
[4] Y. Chen, “Industrial information integration—A literature review 2006–2015,” Journal of Industrial
Information Integration, vol. 2, pp.30-64, 2016.
[5] M.Kano, Y. Nakagawa, “Data-based process monitoring, process control, and quality improvement:
Recent developments and applications in steel industry,” Computers and Chemical Engineering, vol.
32, pp.12–24, 2008.
[6] A.-K.Pani, H.-K. Mohanta, “Online monitoring of cement clinker quality using multivariate statistics
and Takagi-Sugeno fuzzy-inference technique,” Control Engineering Practice, vol. 57, pp. 1-17, 2016.
[7] K.-S.Stadler, J.Poland, E.Gallestey, “Model predictive control of a rotary cement kiln,” Control
Engineering Practice, vol. 19,pp. 1-9, 2011
[8] J.Zhang, G.Rizzoni, A.-C.Arenas, A. Amodio, B.-A.Guvenc, “Model-based diagnosis and fault tol-
erant control for ensuring torque functional safety of pedal-by-wire systems,” Control Engineering
Practice, vol. 61 , pp. 255-269, 2017.
[9] S. -J. Qin, “Survey on data-driven industrial process monitoring and diagnosis,” . Annual Reviews
in Control . vol. 36, pp. 220–234, 2012.
[10] S. -J. Qin. “Statistical process monitoring: Basics and beyond,” Journal of Chemometrics, vol. 17
, no. 8-9, pp. 480-502, 2003.
[11] Z. Li, U. Kruger, L. Xie, A. Almansoori, H. Su. “Adaptive KPCA modeling of nonlinear systems,”
IEEE Transactions on Signal Processing, vol. 63, no. 9, pp. 2364-2376, 2015.
[12] L. Xie, J. Zeng, U. Kruger, X. Wang, J. Geluk. “Fault detection in dynamic systems using the
Kullback–Leibler divergence,” Control Engineering Practice, vol. 43, pp. 39-48, 2015.
[13] A .Sepúlveda, J .-A. Nachlas, “A simulation approach to multivariate quality control,” Computers
& industrial engineering, vol. 33 no.1, pp. 113-116, 1997.
[14] A.Bakdi,A.Kouadri ,A.Bensmail, “Fault detection and diagnosis in a cement rotary kiln using PCA
with EWMA-based adaptive threshold monitoring scheme,” Control Engineering Practice, vol. 66,
pp. 64 -75, 2017.
[15] L.-H.Chiang, E.-L.Russell, R.-D.Braatz ,” Fault diagnosis in chemical processes using Fisher dis-
criminant analysis, discriminant partial least squares, and principal component analysis,” Control
Engineering Practice, vol. 50, no.2, pp.243-252, 2000.
[16] S. Joe Qin. “Statistical process monitoring: Basics and beyond,” Journal of Chemometrics,
vol.17,no. 8–9, pp. 480-502, 2003.
[17] V. Venkatasubramanian, R. Rengaswamy, S.N. Kavuri, K. Yin. “A review of process fault detection
and diagnosis: Part III: Process history based methods,” Computers & Chemical Engineering, vol.27
vol. 3, pp. 327-346, 2003.
[18] H. Wang, T. Wang, Y. Zhou, L. Zhou, H. Li. “Information classification algorithm based on decision
tree optimization,” Computers & Chemical Engineering vol 22, pp .7559–7568, 2019.
[19] Y.-S.Mu, X.-D. Liu, Z.-H. Yang, X.-L. Liu, “A parallel C4.5 decision tree algorithm based on MapRe-
duce,” Concurrency and Computation-practice & Experience. 2017.
[20] Tiwari, Aviral ,” The asymmetric Granger-causality analysis between energy consumption and income
in the United States.” Renewable and Sustainable Energy Reviews . vol. 36.pp.362-369.
[21] M. Kano,H. Ohno,S. Hasebe,I. Hashimoto , “ Process-Monitoring Based on Dissimilarity of Time
ScricsData.” Kagaku Kogaku Ronbunshu. vol. 25 no.6,pp.1004-1009,1999.
[22] A. Moeuf , R. Pellerin, “The industrial management of SMEs in the era of Industry 4.0. International,
” Journal of Production Research vol .56 no.3, pp. 79-87,2018.
[23] T.-Y. Wu, L. Yang, J.-N. Luo, J. M.-T. Wu. “A provably secure authentication and key agreement
protocol in cloud-based smart healthcare environments,” Security and Communication Networks,
vol. 2021, 2299632, 2021.