The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607

https://doi.org/10.1007/s00170-022-08993-9

ORIGINAL ARTICLE

Monitoring of casting quality using principal component analysis

and self‑organizing map
Hocine Bendjama1 · Salah Bouhouche1 · Salim Aouabdi1 · Jürgen Bast2

Received: 2 November 2021 / Accepted: 27 February 2022 / Published online: 7 March 2022
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
The monitoring of casting quality is very important to ensure the safe operation of casting processes. In this paper, in order
to improve the accurate detection of casting defects, a combined method based on principal component analysis (PCA)
and self-organizing map (SOM) is presented. The proposed method reduces the dimensionality of the original data by the
projection of the data onto a smaller subspace through PCA. It uses Hotelling’s T ­ 2 and Q statistics as essential features for
characterizing the process functionality. The SOM is used to improve the separation between casting defects. It computes
the metric distances based similarity, using the ­T2 and Q ­(T2Q) statistics as input. A comparative study between conventional
SOM, SOM with reduced data, and SOM with selected features is examined. The proposed method is used to identify the
running conditions of the low pressure lost foam casting process. The monitoring results indicate that the SOM based on
­T2Q as feature vectors remains important comparatively to conventional SOM and SOM based on reduced data.

Keywords Condition monitoring · Principal component analysis · Self-organizing map · Casting quality · Hotelling’s ­T2 · Q
statistic

1 Introduction rejection from defect formation. It is therefore essential or

necessary to classify the casting defects in order to improve
Nowadays, the monitoring of casting quality plays a very the performance of the process. Traditionally, non-destructive
important role to ensure the steadiness of casting processes. testing techniques (eddy current, magnetic leakage, and so
Its main objective is to ensure the dependability of processes on) have been used and applied with human participation [3,
and increase their availability with lower cost. Modern 4]. However, these manual techniques are usually exhausting,
developments have led to a broad array of specialized cast- error-prone, and expensive. Today, many computer system-
ing processes to create complex castings at a high production based defect detection methods bring great convenience for
rate and good quality. Even in these controlled processes, quality control and provide operators with the process operat-
defects in the output products can occur. ing conditions.
Generally, casting defects fell into one or more of the Process condition monitoring is not an easy task; it is
established seven categories of casting defects such as metal- essentially a problem of process knowledge. In current meth-
lic projections, cavities, discontinuities, defective surface, ods, after data collection, features are first extracted and
incomplete casting, and incorrect dimensions or shape [1]. generated, which often require human expertise. Then, the
The formation of casting defects is linked to several factors defect identification takes place. The most effective feature
like pouring temperature, metal velocity, molding sand, and extraction and more accurate identification are needed to
refractory coating [1, 2]. Obviously, most castings suffer obtain the correct operating condition of the process.
In recent years, the use of hybrid form combining several
* Hocine Bendjama monitoring methods together is an effective tool for improv-
h.bendjama@crti.dz ing monitoring accuracy. It requires a critical processing
step designed to find the most informative feature. Various
1
Research Center in Industrial Technologies CRTI, combined methods have been widely developed and applied
P.O.Box 64, Cheraga, Algiers, Algeria
to identify casting defects, e.g., Dabade and Bhedasgaonkar
2
IMKF, TU Bergakademie Freiberg, Cotta Strasse 4, D‑9596, [5] presented a new combination method of casting defect
Freiberg, Germany

13
Vol.:(0123456789)
3600 The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607

analysis based on Taguchi method and computer-aided cast-

ing simulation technique to analyze the sand-related and
methoding-related defects in green sand casting. Zhang and
Wang [6] optimized the low pressure die-cast process by a
combining neural network and genetic algorithm method.
Zhao et al. [7] extracted a robust randomly distributed tri-
angle feature for casting defects detection such as cracks,
blow holes, shrinkage porosities, and shrinkage cavities.
Sata and Ravi [8] developed a casting defect analysis sys-
tem to analyze and reduce defects in investment castings
based on Bayesian inference with pre-processing of the
input data. Galan et al. [9] proposed an inspection system
based on computations of independent quantities on sev-
eral areas within an image to identify defects on the surface
of a metal part produced by a casting process. Lee et al.
[10] developed a fault detection module based on an artifi- Fig. 1  Schematic representation of the low pressure lost foam casting
cial neural network with the pre-processing of time series process
temperature and pressure measurement data to reduce the
defect rate and production cost in die-casting industry. Niu not detect the faulty conditions, so additional algorithms
et al. [11] embedded a deep learning algorithm in a digital are necessary.
camera to quickly and automatically detect casting defects. The proposed hybrid form uses the advantages of the two
To effectively detect surface defects in continuous casting proposed methods: PCA and SOM. PCA is used for data
process, a methodology for automatic feature extraction and pre-processing, data reduction, and feature extraction. The
classification using vision-based sensing technology have features are obtained from casting data analysis using ­T2 and
been presented by Yang et al. [12]. Similarly, Lin et al. [13] Q ­(T2Q) statistics, which represent the input vectors to the
proposed a robust detection method of casting defects based SOM algorithm. The SOM computes the metric distances
on vision attention mechanism and deep learning of feature based similarity measure in order to identify the running
map. Ai and Xu [14] used contourlet transform and ker- conditions of the low pressure lost foam casting process.
nel locality preserving projections to extract sufficient and The main goal of this proposed method is to obtain more
effective features from metal surface data. Yu et al. [15] detailed information contained in the measured data than
developed a predictive control strategy using a heat transfer had been previously possible. The monitoring results using
model and quasi-Newton method to improve product quality real measurement data demonstrate that casting defects can
in continuous casting. In another application, Bouhouche be clearly identified by utilizing the proposed method.
et al. [16, 17] extracted a reduced data matrix by principal
component analysis (PCA) method, which is used as input to
self-organizing map (SOM) algorithm to evaluate the pick- 2 Experimental setup and data acquisition
ling process. Bendjama and Mahdi [18] suggested a com-
bined simulation approach between Monte Carlo method and The measurement data applied to condition monitoring
neural networks to bring essential information on the state of require different types and levels of equipment and tech-
materials. The work presented in this paper is motivated by niques. These depend on the investment and available
practical applications in the foundry industry, where a hybrid expertise.
form of PCA-SOM is used for defect assessment.
The multivariate statistical method PCA has become
widely used to improve fault detection accuracy in recent
years. PCA is a data compression method; it produces a Table 1  Test conditions
lower dimensional representation in a way that preserves
Test Pouring Measured Holding Mold filling Holding
the correlation structure among the process variables. The temperature temperature pressure time (s) time (s)
collected casting measures are used as inputs to the PCA (°C) (°C) (bar)
algorithm, which is in charge of the detection of abnormal
NR 735 711 0.24 6 90
situations and provide information about the state of the pro-
SC 750 700 0.24 6 90
cess using Hotelling’s ­T2 and Q statistics. These statistical
SH 730 730 0.24 6 90
parameters must not exceed threshold values in normal situ-
MP 725 725 0.24 6 90
ation; however, sometimes ­T2 and Q statistics alone could

13
The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607 3601

Table 2  Process variables National Instruments (NI) data acquisition board with a USB
Variable Description Unit
interface. The foam pattern was supported in the flask and
the thermocouples were wired to the data acquisition unit.
T Temperature °C When the vessel is pressurized, liquid metal rises through a
P Pressure bar steel pipe into the flask. All test part was cast using AlSi12
S Rise (height of filling) m alloy at temperatures between 725 and 750 °C (Table 1). NI
T1 Temperature 1 °C DASY Lab software was used to collect and analyze the sig-
T2 Temperature 2 °C nals from the temperature sensors. The measured variables
T3 Temperature 3 °C are listed in Table 2 and presented in Fig. 2.
T4 Temperature 4 °C Ten days of experimental measurements were collected
T5 Temperature 5 °C from 9 casting tests, including normal data and faulty data.
Each data set includes eight measurement variables and m
observations or samples at different times which make it
The experimental measurements presented in this paper possible to construct an m × 8 input matrix. The obtained
are entirely based on the casting data acquired from the low matrices, from all measured data, are then used to compute
pressure lost foam casting process. As shown in Fig. 1, the the PCA and SOM models to better control the quality of
casting process utilizes a resistance furnace capable of melt- castings.
ing standard aluminum base alloys. Air pressure is applied
to the chamber containing the crucible to push liquid metal
up into a flask containing the foam pattern and unbounded
sand. More details about the casting process were reported 3 Improvement of condition monitoring
in [19, 20]. using the hybrid form PCA‑SOM
The low pressure lost foam casting process is utilized
to create complex castings. However, filling the mold with 3.1 Related theories
molten metal can produce an undesired casting. Generally,
a casting defect is defined as all observable and unplanned 3.1.1 Principal component analysis
variation. If a defect occurs, measures must be adopted to
control and monitor the casting conditions. The data used The PCA [21] is a multivariate statistical analysis technique,
in this study were gained through the experimental meas- which reduces the original data space into a smaller dimen-
urements in normal (NR) or healthy operating conditions sion space in terms of protecting the main original data
and abnormal functionality of the casting process including information. Given data matrix X ∈ ℜ m×n composed of m
three defects: cracks (SC), holes (SH), and metal penetra- observations or samples and n variables, which have been
tion (MP). normalized to have 0 mean and unit variance, PCA is only
To acquire data, five temperature transducers were imple- interested in its variance and covariance. PCA actually relies
mented in the process by five thermocouples for temperature on eigenvalue/eigenvector decomposition of the covariance
input. The pressure and temperature inputs were wired to a or correlation matrix C given as follows:

Fig. 2  Measured data 752 0.25 0.25 600

750 500
0.2 0.2
Temperature1 (°C)
Temperature (°C)

Pressure (bar)

748 400
0.15 0.15
Rise (m)

746 300
0.1 0.1
744 200

742 0.05 0.05 100

740 0 0 0
0 50 100 0 50 100 0 50 100 0 50 100
Time (s) Time (s) Time (s) Time (s)

800 800 800 800

Temperature2 (°C)

Temperature3 (°C)

Temperature4 (°C)

Temperature5 (°C)

600 600 600 600

400 400 400 400

200 200 200 200

0 0 0 0
0 50 100 0 50 100 0 50 100 0 50 100
Time (s) Time (s) Time (s) Time (s)

13
3602 The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607

C = (1∕(m − 1))X T X = PDPT (1) 3.1.2 Self‑organizing map

where D = diag(λ1….λn) is a diagonal matrix with diagonal Artificial neural networks (ANNs) are mathematical or com-
elements in decreasing magnitude order and P contains the putational models, inspired by biological nervous system.
eigenvectors. ANNs are comprised of an interconnected group of artifi-
PCA determines an optimal linear transformation of the cial neurons, and they have long been used for data-driven
data matrix X in terms of capturing the variation in the data decision-making. Based on their learning process (supervised
as follows: or unsupervised), ANNs are performed on the computer to
perform certain specific tasks like optimization and pattern
T = XP and X̂ = TPT (2) recognition. They can help draw useful conclusions and inter-
where T ∈ ℜ m×k is the principal component matrix and the pretations from observed data.
matrix, P ∈ ℜ m×k contains the principal vectors which are the ANNs using supervised learning like multi-layer per-
eigenvectors associated to the eigenvalues λi of the covari- ceptron (MLP), probabilistic neural networks (PNN), and
ance matrix and k denotes the principal components number radial basis functions (RBF) have proved to be advanta-
(PCs) of PCA model. A key issue to develop a PCA model geous in obtaining a good model with accurate predictions.
is to choose the adequate number of PCs. A number of well- However, they require that the output vector be known for
known techniques have been proposed for selecting the num- training phase. Unlike supervised learning, output vector is
ber of PCs [22]. In this study, the specific calculation method not required to be known with unsupervised learning, i.e.,
of selected k follows the cumulative percent variance (CPV) the network does not use training pairs consisting of input
principal [23]. It is a measure of the percent variance, such vector and desired output.
as 85%, captured by the first PCs. The SOM, invented by Kohonen [24], is a kind of ANNs
The difference between X and X̂ is the residual matrix E. that use unsupervised competitive learning to map a high
It can be calculated as follows: dimensional input space (the data space) onto a low dimen-
( ) sional output space, usually of two dimensions. It is a tech-
E = X − X̂ = X I − PPT (3) nique to group data with similar characteristics. Each neuron
or node comprises a vector of weights of the same dimension
where I is the unit matrix. as the input data vectors. The SOM is trained by presenting
To perform process fault detection, a PCA model of the the data repeatedly and upgrading the weights to learn the
normal operating conditions must be built. When a new structure of the data.
observation data is subject to faults, these new data can be The SOM consists of an input layer and a competitive
compared to the PCA model. The correlation of the new data or output layer, fully interconnected to each other. The
is detected by Hotelling’s T­ 2 and Q, called also as squared output layer consists of m neurons. Each neuron i (i = 1,
prediction error (SPE), statistics as follows: 2,..., m) is represented by an n dimensional weight vector
(4) wi = [wi1,....,win] where n is the dimension of the input vector.
T 2 = X T PD−1 PT X
In the output layer, the competitive process is done and the
( )T ( ) weight of connection is updated to choose a winner neuron.
Q = ET E = X − X̂ X − X̂ (5) The key steps in the SOM learning process (training) are
first, for each input vector, determining its best matching
The process is considered normal if Hotelling’s T ­ 2 and unit (BMU). The BMU is the node that is most similar to
Q statistics do not exceed threshold values. These statis- the input vector. If we denote b the BMU of input vector x
tics alone could not detect the faulty conditions. They are and wb the weight vector of this BMU, the identification is
used, in this study, as input to the SOM algorithm to further based on the minimum Euclidian distance, which is defined
improve the separation between casting conditions. as follows:
The computing steps using PCA method are summarized
{ }
as follows: ‖x − wb ‖ = mini=1,….,m ‖x − wi ‖
‖ ‖ ‖ ‖ (6)

1. Matrix X, Then, the process proceeds to update the weight vectors

2. Mean centring of X, of the BMU and those of its adjacent neurons (neighbor-
3. Calculation of covariance matrix, hood) to match the input data using the following equation:
4. Eigenvalues and eigenvectors, [ ]
5. Optimal number of PCs,
wi (t + 1) = wi (t) + 𝛼(t)hb (x)i (t) x(t) − wi (t) (7)
6. PCA Model, where t denotes time, 0 < α(t) < 1 is the learning rate, hb(x)i(t)
7. T2 and Q statistics. is the neighborhood function centered in the BMU b at time

13
The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607 3603

t. The neighborhood function is often taken as a Gaussian

Defective condition
function. Both α(t) and the width of hb(x)i(t) decrease gradu- Normal operating
with Defect 1
ally with the increasing step t. condition
Distance 1
Distance 0
The SOM algorithm is described as follows:
Current operating
data Distance 2
1. Set the initial random positions and the initial learning Defective condition
rate, Distance 3 with Defect 2
2. Provide an input data,
3. Computing and finding BMU, Defective condition
4. Adjust the weights, with Defect 3
5. Construct the output space,
Fig. 3  Condition monitoring using similarity measure

3.2 Similarity measure
current operating data and process conditions. The differences
After SOM training, the winner neuron is used to com- between two running conditions i and j (ΔDIN = ­DINi − ­DINj
pute metric distances. Distances or similarity measures are and ΔOGX = ­OGXi − ­OGXj) are used as indexes to distinguish
essential to solve many pattern recognition problems such the defective and healthy casting conditions.
as classification and clustering. Various distance measures
are applicable. In this work, we consider a similarity evalu- 3.3 Description of the proposed method
ation based on the Euclidian distances [16, 17]. There will
be used different formula to measure distances between two In order to construct a successful casting defects identi-
data points characterizing two monitoring conditions. The fication system, a combination using PCA and SOM is
followings distances can be applied: described in this work (Fig. 4). PCA is trained with input
matrix that contains the process variables, where the goal
• Distance between input and neuron data: is to establish the normal statistical correlation among the
√
( )
√ n
√∑ ( )2 measured data to characterize the operating conditions of
DIN x, wbj =√ xj − wbj (8) the casting process using Hotelling’s ­T2 and Q statistics.
j=1 Sometimes, these statistics cannot exactly detect the defec-
tive conditions. They are therefore proposed in this study
as characteristic features of the measured casting data. The
• Mean or gravity center of the input data:
n chosen features are used as input to SOM algorithm, which
∑
is used for casting quality evaluation. The main objective is
GCI = (1∕n) xj (9)
j=1 to compute the distances between the input vectors and win-
ning neuron to better identify casting defects. A compara-
tive study between conventional SOM, SOM with reduced
• Gravity center of the neuron data: data matrix (RD), and SOM with selected features ­(T2Q)
/ n
n
∑ ∑ is examined.
GCN = wbj xj wbj (10)
j=1 j=1

4 Results and discussion

• Distance between gravity centers (input and neuron data):
√
√ n The structure of the proposed fault identification method
√∑ involves two parts: one is the development and training
OGX = √ (GCI − GCN)2 (11)
j=1 the models, the other is testing the process fault based on
trained models. The measured data used in training represent
where xj is the input vector and wbj defines the winner a healthy casting condition.
neuron. In this study, the data matrices that contain the values
of measured variables are constructed from 930 sampling
The decision about running conditions can be performed instances of interval 0.1 s. Ten days of experimental meas-
using the computed metric distances: DIN and OGX. Figure 3 urements were collected from 9 casting tests as follows: two
shows the principle of condition monitoring using similarity; castings on the first day, two castings on the third day, two
the casting defect is determined by the distance separating the castings on the sixth day, and three castings on the last day.

13
3604 The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607

Fig. 4  Diagram of the proposed T2Q

method
Computing indexes
Matrix X PCA RD SOM
Decisions
method algorithm

Table 3  Eigenvalues and variances of PCs the performance of the SOM and PCA-SOM techniques;
PCs Eigenvalues Variances (%)
the corresponding indexes are thus calculated and discussed.
Through PCA, the eigenvalues of the covariance matrix,
1 6.4363 80.45 which are the variances of PCs, are listed in Table 3. The
2 1.0164 12.71 anterior 2 PCs explain over 85 of the total variance of the
3 0.2960 3.70 data. The PCA model is established making use of them, and
4 0.1143 1.43 then the monitoring performance is progressed. As shown
5 0.0700 0.87 in Fig. 5, all process variables are correctly estimated with
6 0.0395 0.50 this PCA model, except that some certain variables are less
7 0.0199 0.24 well estimated than others.
8 0.0076 0.10 According to SOM algorithm and PCA-SOM with RD as
input (PCA-SOM/RD), casting conditions are evaluated by
the metric distances DIN and OGX given by Eqs. (8) and
During this test period, three types of casting defects (SC, (11), respectively. Obtained results are represented in Figs. 6
SH, MP) were recorded and the remainder of the castings and 7. As we can see, it is not easy to distinguish between
were obtained without defects (NR). In total, 8 variables and normal and abnormal conditions and also between false and
930 data points at different times, for each condition, are missed alarms.
collected into a 930 × 8 matrix for computing the PCA and The proposed algorithm, PCA-SOM with ­T2Q as input
SOM models. The obtained matrices are used to evaluate (PCA-SOM/T 2Q), is tested using the same computed

4 2 1 1

0
0 0
Temperature1 (°C)
Temperature (°C)

2
Pressure (bar)

-1
Rise (m)

-2 -1
0 -2
-4 -2
-3
-2
-6 -4 -3

-4 -8 -5 -4
0 50 100 0 50 100 0 50 100 0 50 100
Time (s) Time (s) Time (s) Time (s)

1 1 1 2

0 0 0
Temperature4 (°C)
Temperature2 (°C)

Temperature3 (°C)

Temperature5 (°C)

0
-1 -1 -1

-2 -2 -2 -2

-3 -3 -3
-4
-4 -4 -4

-5 -5 -5 -6
0 50 100 0 50 100 0 50 100 0 50 100
Time (s) Time (s) Time (s) Time (s)

Fig. 5  Estimation of the measured data

13
The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607 3605

indexes. The isolation level of the normal and abnormal To evaluate the performance of the SOM, PCA-SOM/
casting conditions is clearer than the SOM and PCA-SOM/ RD and PCA-SOM/T2Q techniques, a comparative study
RD. The tested distances indicate that the casting defects using ΔDIN and ΔOGX distances is carried out. Differ-
are successfully identified (Fig. 8). The output results can ences between computed indexes of defective and healthy
be used to describe the capability of fault isolation. casting conditions are listed in Table 4. From obtained

Fig. 6  Evolution of DIN and 2000 1600

OGX indexes using SOM 1800

NR
1400
algorithm SC
1600
SH
1200
1400 MP
1000
1200

OGX index
DIN index

1000 800

800
600
NR
600
SC 400
400 SH
MP 200
200

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Sample number Sample number

Fig. 7  Evolution of DIN and 22 70

OGX indexes using PCA-SOM/ 20 NR NR
RD technique SC 60
SC
18
SH SH
16 MP 50
MP

14
OGX index
DIN index

40
12
30
10

8 20

6
10
4

2 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Sample number Sample number

Fig. 8  Evolution of DIN and 250 150

OGX indexes using PCA-SOM/ NR NR

T2Q technique SC SC
200 SH SH
MP MP
100
150
OGX index
DIN index

100
50

50

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000

Sample number Sample number

13
3606 The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607

Table 4  Performance evaluation ΔDIN ΔOGX

Defects SC SH MP SC SH MP Computing
Methods time (s)

SOM 27.45 - - - - - 54.33

PCA-SOM/RD 7.06 1.16 3.81 0.57 1.89 24.58 55.62
PCA-SOM/T2Q 148.74 50.38 97.16 84.91 25.32 42.69 50.80

results, important variations are detected using PCA-SOM/ References

T2Q. These variations show the detection capacity of cast-
ing conditions. Moreover, SOM and PCA-SOM/RD tech- 1. Beeley P (2001) Foundry technology, 2nd edn. Butterworth-Henemann,
niques take more computing time than PCA-SOM/T2Q. Oxford
2. Rao P (2008) Manufacturing technology: foundry, forming and
The proposed method therefore shows a significant high welding. Tata McGraw Hill
performance compared to the considered methods (SOM 3. Hughes RR, Dixon S (2014) Eddy-current crack detection at
and PCA-SOM/RD). frequencies approaching electrical resonance. AIP Conf Proc
On the other hand, the proposed method also proves its 1581(1):1366–1373
4. Kim HM, Rho YW, Yoo HR, Cho SH, Kim DK, Koo SJ, Park GS
ability to ensure high defect detection performance com- (2012) A study on the measurement of axial cracks in the mag-
pared with other applications that use PCA-SOM with RD netic flux leakage NDT system. In: IEEE International Conference
as feature [16, 17]. on Automation Science and Engineering (CASE): 624–629
5. Dabade UA, Bhedasgaonkar RC (2013) Casting defect analysis
using design of experiments (DoE) and computer aided casting
simulation technique. Procedia CIRP 7:616–621
6. Zhang L, Wang R (2013) An intelligent system for low-pressure
5 Conclusion die-cast process parameters optimization. Int J Adv Manuf Tech-
nol 65:517–524
7. Zhao X, He Z, Zhang S, Liang D (2015) A sparse-representation-
In this paper, a new statistical method to improve the based robust inspection system for hidden defects classification
monitoring of casting conditions based on a combination in casting components. Neurocomputing 153:1–10
of PCA and SOM is proposed. The PCA is exploited to 8. Sata A, Ravi B (2017) Bayesian inference-based investment-casting
extract statistical features from experimental data. Using defect analysis system for industrial application. Int J Adv Manuf
Technol 90:3301–3315
Hotelling’s ­T2 and Q statistics as input to SOM algorithm, 9. Galan U, Orta P, Kurfess T, Ahuett-Garza H (2018) Surface defect
the casting defects are separated using Euclidian distances identification and measurement for metal castings by vision sys-
based similarity evaluation. This combined method is tem. Manuf Lett 15:5–8
accurate in casting quality monitoring; it can be seen from 10. Lee J, Lee YC, Kim JT (2021) Migration from the traditional to
the smart factory in the die-casting industry: novel process data
the analysing results that the normal and the abnormal acquisition and fault detection based on artificial neural network.
casting conditions can be clearly identified. It is clear that J Mater Process Technol 290:116972
the proposed method outperforms the other approaches 11. Niu B, Qu X, Guan X, Zhang F (2021) Rapid detection of highly
and algorithms considered in this work. reflective surface defects based on digital micromirror device. Opt
Commun 501:127385
12. Yang Q, Jin J, Chang TS (2010) Automatic feature extraction and clas-
Acknowledgements This research acknowledges the German Aca- sification of surface defects in continuous casting. Transactions North
demic Exchange Service (in German, Deutscher Akademischer American Manufacturing Research Institution of SME 38:563–570
Austausch Dienst (DAAD)). The authors would like to extend their 13. Lin J, Yao Y, Ma L, Wang Y (2018) Detection of a casting defect
greatest gratitude to the Institute of Mechanical Engineering, Freiberg tracked by deep convolution neural network. Int J Adv Manuf
University of Mining and Technology, Freiberg, Germany. Comments Technol 97:573–581
and suggestions from anonymous reviewers and the Editor are greatly 14. Ai Y, Xu K (2013) Surface detection of continuous casting slabs
appreciated to improve the manuscript. based on curvelet transform and kernel locality preserving projec-
tions. J Iron Steel Res Int 20(5):80–86
Declarations 15. Yu Y, Luo X, Liu Q (2017) Model predictive control of a dynamic
nonlinear PDE system with application to continuous casting. J
Ethics approval Not applicable. Process Control 65:41–55
16. Bouhouche S, Yahi M, Bast J (2011) Combined use of principal
Consent to participate Not applicable. component analysis and self-organization map for condition moni-
toring in pickling process. Appl Soft Comput 11:3075–3082
Consent for publication Not applicable. 17. Bouhouche S, Laïb DLY, Bast J (2014) Evaluation-based closed-
loop errors using principal component analysis and self-organisation
map with an application to a pickling process. Int J Adv Manuf
Conflict of interest The authors declare no competing interests. Technol 70:1033–1041

13
The International Journal of Advanced Manufacturing Technology (2022) 120:3599–3607 3607

18. Bendjama H, Mahdi D (2016) Computer code for materials diag- 22. Tamura M, Tsujita S (2007) A study on the number of principal
nosis using Monte Carlo method and neural networks. J Fail Anal components and sensitivity of fault detection using PCA. Comput
Prev 16(4):931–937 Chem Eng 31:1035–1046
19. Lang L (1999) Development and testing of a low pressure lost 23. Jackson JE (2003) A user’s guide to principal components. John
foam casting machine and an examination of the process. PhD Wiley & Sons, New York
thesis, Freiberg University, Germany 24. Kohonen T (2001) Self-organizing Maps. Springer-Verlag, Berlin
20. Bast J, Aitsuradse M, Hahn T (2004) Advantages of the low pres-
sure lost foam casting process. AFS Transactions 112:1131–1144 Publisher's Note Springer Nature remains neutral with regard to
21. Anderson TW (2003) An introduction to multivariate statistical jurisdictional claims in published maps and institutional affiliations.
analysis. John Wiley & Sons, California

13