You are on page 1of 19

Journal of Intelligent Manufacturing (2023) 34:2889–2906

https://doi.org/10.1007/s10845-022-01973-6

Knowledge-embedded machine learning and its applications in smart


manufacturing
Farzam Farbiz1 · Mohd Salahuddin Habibullah1 · Brahim Hamadicharef1 · Tomasz Maszczyk1 ·
Saurabh Aggarwal1

Received: 19 October 2021 / Accepted: 21 May 2022 / Published online: 21 June 2022
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
Demands for more accurate machine learning models have given rise to rethinking current modeling approaches that were
deemed unsuitable, primarily due to their computational complexity and the lack of availability and accessibility to repre-
sentative data. In Industry 4.0, rapid advancements in Digital Twin (DT) technologies and the pervasiveness of cost-effective
sensor technologies have pushed the incorporation of artificial intelligence, particularly data-driven machine learning models,
for use in smart manufacturing. However, the persistent issue with such models is their high sensitivity to the training data
and the lack of interpretability in the outcomes, at times generating unrealistic results. The incorporation of knowledge into
the machine learning pipeline has been earmarked as the most promising approach to address such issues. This paper aims
to answer this call through a Knowledge-embedded Machine Learning (KML) framework for smart manufacturing, which
embeds knowledge from experience and, or physics information into the machine learning pipeline, thus making the outcomes
from these models more representative of real applications. The merits of KML were then presented through comparative
studies showing its capability to outperform knowledge-based and data-driven models. This promising outcome led to the
development of frameworks that can potentially incorporate KML for smart manufacturing applications such as Prognostics
and Health Management (PHM) and DT, further supporting the usefulness of the proposed KML framework.

Keywords Artificial intelligence · Smart manufacturing · Knowledge-embedded machine learning · Physics-based model ·
Data-driven model · Prognostics and health management · Digital twin · Inductive learning · Deductive learning · Abductive
learning

Abbreviations
CPS Cyber-Physical-Systems
AI Artificial Intelligence
CRM Customer Relation Management
BPFI Ball Pass Frequency at Inner
DDML Data-Driven Machine Learning
BPFO Ball Pass Frequency at Outer
DT Digital Twin
ERP Enterprise Resource Planning
B Farzam Farbiz IIoT Industrial Internet of Things
Farzam_Farbiz@ihpc.a-star.edu.sg
KML Knowledge-embedded Machine Learning
Mohd Salahuddin Habibullah MSE Mean Square Error
mohdsh@ihpc.a-star.edu.sg
PHM Prognostics and Health Management
Brahim Hamadicharef RUL Remaining Useful Life
brahim_hamad@ihpc.a-star.edu.sg
SM Smart Manufacturing
Tomasz Maszczyk SMAI Smart Manufacturing AND Artificial Intelligence
maszczyk_tomasz_karol@ihpc.a-star.edu.sg
SMAIK Smart Manufacturing AND Artificial Intelligence
Saurabh Aggarwal AND Knowledge
saurabh_aggarwal@ihpc.a-star.edu.sg
WIP Work-in-process
1 Department of Computing & Intelligence, WoS Web of science
Institute of High Performance Computing,
1 Fusionopolis Way, #16-16 Connexis,
Singapore 138632, Singapore

123
2890 Journal of Intelligent Manufacturing (2023) 34:2889–2906

Introduction The third approach is hybrid Liao and Kottig (2014);


Ahmad et al. (2018); Siddhpura and Paurobally (2013),
Over the last few decades, scientific and technological which combines knowledge about the physical process with
advancements have pushed the manufacturing industry information from the observed data. The aim is to improve
towards multiple stages: mechanization, electrification, and the model by incorporating the strengths and reducing the
now digitalization. Coined as “Industry 4.0”, there is a greater limitations associated with both approaches.
emphasis on moving the manufacturing industry towards These three approaches have found their applications in
smart factories. This is where accessibility to big data and the real-world problems, with the use of the data-driven approach
availability of digitalization technologies such as smart sen- being more prominent due to its independence from physi-
sors, Industrial Internet of Things (IIoT), and Cyber-Physical cal processes in the systems Bektas et al. (2020). However,
Systems (CPS) have transformed the industry into “smart the requirements for more robust and interpretable models in
manufacturing” Kusiak (2018, 2017). smart manufacturing to mimic real scenarios have prompted
Among these technologies, IIoT and big data are most a paradigm shift towards the use of knowledge-based and
commonly adopted as smart factories aim to use informa- hybrid approaches. This is primarily driven by the advance-
tion gathered from sensors attached to machines to manage ments made in DT technologies that combine physics and
the entire factory more efficiently Shrouf et al. (2014); Yan ML models to simulate these complex systems.
et al. (2017). In essence, smart manufacturing (or industrial The primary aim of this paper is also aligned with this
intelligentization Zhou (2013); Ding (2020) is the fusion direction, whereby the concept of Knowledge-embedded
of advanced manufacturing processes using Artificial Intel- Machine Learning (KML), is introduced as a bridge between
ligence (AI) that enables intelligent perception, analysis, traditional physics-based simulation models and current
reasoning, decision-making, and control. data-driven ML modeling approaches. The usefulness and
One of the impacts of adopting these technologies is in feasibility of the KML framework are illustrated through sim-
providing Prognostics and Health Management (PHM) solu- ulated and real-case studies. The results show more robust
tions Liu et al. (2018); Zhang et al. (2018). Such solutions and accurate models, in particular for smart manufacturing
offer better strategies for maintenance and repair, which will applications.
ultimately reduce the machines’ downtime and improve the The structure and technical contributions of this paper are
products’ quality. as follows: In ‘An overview of knowledge-based and data
Lee et al. (2018) proposed an industrial intelligence -driven models’, following a brief overview of Knowledge-
ecosystem comprising four key components: big data, data based (Kb) and Data-Driven Machine Learning (DDML)
analytics, cloud computing, and cyber technologies. These modeling approaches, the concept of Knowledge-embedded
key components play crucial roles in developing models for Machine Learning (KML) is introduced. Using simulation
prediction, optimization, and decision-making applications. studies, the improved performance of KML compared to Kb
Existing approaches for developing such models can be and DDML is shown. In ‘KML and smart manufacturing’,
classified into three main categories: knowledge-based, data- through a real-case example, the suitability of KML for
driven, and hybrid. Knowledge-based approaches Downey et smart manufacturing is presented, leading to a general
al. (2019); Daigle and Goebel (2013); Hosseinkhani and Ng KML framework for smart manufacturing applications. In
(2015) incorporate prior knowledge (e.g., physics or empiri- ‘Potential KML applications in smart manufacturing’, fol-
cal laws about the physical processes) to diagnose, optimize lowing a bibliometric study of “AI and knowledge”
and predict the system’s health. They can achieve accurate in smart manufacturing, potential ways of incorporating
solutions if physical simulation models of the processes and KML in Prognostics and Health Management (PHM)
interactions between the components are known. However, and Digital Twin (DT) are presented. Finally, in
the characteristics of real physical systems are generally ‘Conclusions and future work’, conclusions and future work
challenging to be mathematically formulated. Furthermore, are presented.
physics-based simulations are often too complex and com-
putationally expensive.
On the other hand, data-driven models Torabi et al. (2016); An overview of knowledge-based
Tsui et al. (2015) use Machine Learning (ML) techniques, and data-driven models
relying on collected data, to develop models for anomalies
and, or faults and, or predictions. Thus, these ML algorithms In this section, prior to introducing Knowledge-embedded
provide viable alternatives to those complex physics-based Machine Learning (KML), a definition of what is meant by
simulation models and can accurately model systems if the two key terms “knowledge” and “learning” is provided
substantial data under various health states is available for here. In general, it is difficult to have a precise definition of
training. the term knowledge, as there have been several definitions

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2891

and debates among scientists in the field of epistemology sions based on pre-defined sets of rules to be applied to
(the branch of philosophy dealing with the theory of knowl- new incoming data. For example, in simulating physical and
edge Steup and Neta (2020); Zagzebski (2017); Machamer chemical properties, it has been shown Durkin (1990) that ES
and Silberstein (2002)). Although the precise definition of performed very well when those properties can be accurately
knowledge is beyond the scope of this paper, the defini- defined by pre-defined sets of rules/theories. However, ES
tion provided by Poole and Mackworth (2017) is adopted, does not perform well when these properties are too complex.
whereby “knowledge” is defined as “the information about a In contrast to deductive-based approaches, inductive-
domain that can be used to solve problems in that domain”. based ML Aggarwal (2021) provided AI researchers with
As for “learning”, the definition provided by Gross (2005) as novel ways to deal with modeling complex systems for appli-
“learning is the process of acquiring knowledge” is adopted. cations such as Computer Vision (CV) and Natural Language
Hence, learning can be defined as the process of acquiring Processing (NLP). In this approach, at first, a ML model is
skills that are used to solve problems in a particular domain. trained using a large dataset. Once the training process is
completed, new data is presented to the model to estimate the
model’s performance and compare it with the expected out-
Knowledge-based vs data-driven models
comes (i.e., ground truth). Furthermore, due to the model-free
nature of these so-called data-driven models, these methods
As mentioned by Read (2016), humans learn in 2 ways.
have the ability for online learning. This means that these
Firstly, deductive approach, whereby concepts are first intro-
models can be changed over time as they learn from new
duced and the instructed theories/hypotheses are then applied
incoming data.
to the specific cases. Secondly, inductive approach, whereby
The comparison between knowledge-based and data-
learners discover rules/theories by observing examples and
driven models is shown in Table 1. It can be observed that
sample data before reaching broad generalizations. In the
purely knowledge-based models generally provide reason-
field of computer science, both deductive and inductive learn-
able interpretation and can often be generalized for adoption
ing approaches have been applied to ML Aggarwal (2021).
across different systems. However, these models tend to suf-
For deductive-based ML models, often called Expert Systems
fer from low accuracy, especially when modeling complex
(ES) Stefik (1995), knowledge sources or rules are pre-coded
systems.
into a computer program, enabling the program to make deci-

Table 1 Comparisons between knowledge-based and data-driven models (summarized from Karpatne (2017), Yang (2016), Yang et al. (2021), Liu
et al. (2021)
Knowledge-based Model

Learning method Deductive learning Read (2016)


Data requirements Minimal or not required
Advantages Independent on data quality and quantity
Provide good interpretability for domain experts to comprehend
Provide good generalizability across different systems
Successfully applied to real applications such as fluid simulation, material discovery, and product design
Disadvantages Requires deep domain knowledge
Lack of flexibility as it tends to be domain-specific
Computationally expensive
Low accuracy, especially for complex systems
Data-Driven Model
Learning Method Inductive learning Aggarwal (2021)
Data Requirements Quantity and quality are paramount
Advantages Flexible for applications in different problems without (or minimal) domain knowledge
Accuracy can be improved with the availability of more data
Successfully applied to real applications such as smart manufacturing, aerospace, and maritime
Disadvantages Requires large amount of data that might not be available in real applications
Possible lack of interpretability as output might be scientifically inconsistent, unrealistic, or unpredictable
Lack of generalizability, especially in scenarios not represented in the training phase
Models tend to be sensitive to training data

123
2892 Journal of Intelligent Manufacturing (2023) 34:2889–2906

On the other hand, data-driven models are more flexible. (2017)) a complex physical object is often simulated by
They can be used in various domains. In addition, the model’s a simplified physical approximated model, whose hyper-
accuracy can be improved simply by adding more data into parameters values are estimated from observed data Chan
the training phase. However, despite the rapid advancements and Elsheikh (2017). ML models have been applied in
in data-driven models Pecht and Jaai (2010), they gener- this domain to achieve more efficient optimization of those
ally suffer from inherited problems: The models’ outputs hyperparameters (so called “parametrization”). For instance,
are generally unpredictable, and they are very sensitive to Hansen (2013) showed that ML-based parametrization for
the training data, whereby outliers can potentially lead to atomic energy use-case in computing quantum-chemical
unrealistic results. properties can be achieved in a few seconds compared to
a few days using traditional methods.
Combining knowledge-based and data-driven Thus, it can be observed that such a way of combin-
approaches ing knowledge-based and data-driven is useful in optimizing
parameters of the physics-based model and achieving higher
Based on the individual merits of knowledge-based accuracy. However, this approach is still limited by the overall
(deductive-learning) and data-driven (inductive-learning) structure of the physics-based model, as they cannot accu-
approaches, researchers have attempted to combine them rately simulate complex systems Willard et al. (2020).
to extract utmost benefits from both approaches Liao and
Kottig (2014); Atamuradov et al. (2017); Lee (2014). These Fusing knowledge-based and data-driven models
combinations can be broadly undertaken in the following
approaches. The second approach is the so-called hybrid model, whereby
the outputs from both knowledge-based (extracted from
Optimizing models’ hyperparameters using data experience-based/rules-based or physics-based models) and
data-driven models are fused together, as shown in Fig. 2.
As illustrated in Fig. 1, the first approach involves optimiz- For example, Erge and van Oort (2020) used a physics-based
ing the hyperparameters in a knowledge-based model, often model together with a data-driven deep learning model. The
represented by physics-based or experience-based models, authors then included these two models into a hidden Markov
using data collected from the system. For example, Chinnam model, to predict the pressure in a well construction. The
and Baruah (2004) developed an experience-based model for results showed higher accuracy when these two models were
monitoring and estimating the Remaining Useful Life (RUL) combined as compared to cases where each model was used
of a cutting tool. This was done using an Adaptive Neuro- independently.
Fuzzy Inference System (ANFIS) method based on domain In another case study Bikmukhametov and Jäschke (2020),
expert knowledge. The authors demonstrated the feasibility at first a physics-based model was developed to estimate mul-
of the approach for use in online reliability estimation of tiphase flowrate in a petroleum production system. Then, a
high-speed-steel drill-bits. In another example, Chen et al. data-driven model was trained to learn the mismatch between
(2012), Chen et al. (2011) used an ANFIS, trained with his- the output of the physics-based model and the target, as
torical data to simulate a degradation process. Applied to shown in Fig. 3a. Finally, the outputs of these two models
both machine condition predictions and RUL case studies, were combined to improve the prediction accuracy, as shown
the authors concluded that their approach outperformed the in Fig. 3b.
classical predictors. Due to the decoupled nature of physics-based and data-
For modeling physical properties (e.g., in geology driven models, many fusing methods have been proposed in
Zhang (2019) or atmospheric sciences Lopatin and Zhirov

Fig. 1 Optimizing hyperparameters in a knowledge-based model using Fig. 2 General architecture of a hybrid model in which knowledge is
offline data and ML model represented through experience—based or physics—based models

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2893

physics information) into the ML pipeline. The aim is to


bring the results of the ML models closer to theory, thus
making them more realistic and interpretable. In this paper,
we call this approach Knowledge-embedded Machine Learn-
ing (KML). Based on our adopted definition of knowledge
as “information about a domain that can be used to solve
(a)
problems in that domain”, the main challenge in developing
a KML solution is how the domain information can be effec-
tively integrated/embedded into the ML pipeline. As shown
in Fig. 4, von Rueden et al. (2021) showed that knowledge
can generally be embedded into the following four stages of
a ML pipeline:

(b)
1. Training Data It is the most common way for knowl-
Fig. 3 Hybrid model architecture proposed by Bikmukhametov and edge integration, where knowledge is used to generate
Jäschke (2020) with a training pipeline for learning mismatch between simulated/augmented data. This augmented data is then
the physics-based model and the expected target, and b deployment
pipeline for using the estimated mismatch for improving the prediction added to the available data for training the ML model. For
example, Stoecklein et al. (2017) used simulated data on
physics-based models in training deep neural networks.
2. Hypothesis Set It involves designing the ML architecture
the lietrature. For instance, Liao and Kottig (2014) provided and its associated hyperparameters, whereby an expert
a comprehensive review of using hybrid models for RUL pre- chooses a proper architecture for the model based on
diction in engineering systems. Similarly, in the field of PHM, his/her experience, using the nature of the problem and
Atamuradov et al. (2017) presented a detailed overview of details about the types of data (image, voice, time series,
implementation steps in PHM systems and highlighted areas ...) as the criteria. A known example in this field is the
where a hybrid approach can be adopted in different stages architecture of convolution neural networks Krizhevsky et
of the systems. al. (2017), where knowledge is used in making the model
However, despite the potential advantages of a hybrid translation-invariant to the objects’ location in the image.
approach, it still suffers from three potential challenges: (i) 3. Learning Algorithm It includes the selection, design of
the selection of knowledge/information and source of data the loss function, and optimization algorithms to be used.
to be considered in the hybrid model, (ii) the dependence For instance, Jia et al. (2019) used a physics-based model
on proper fusing mechanism on the available knowledge and of lake temperature as a loss function in the training phase
data, and (iii) the need to manage uncertainties in the final of a neural network, aiming for the model’s prediction for
predictions as each model may produce different prediction unlabeled data to be more consistent with physical sim-
results Liao and Kottig (2014). ulation model of the lake. The main advantage of such
knowledge incorporation is to have more physically con-
Knowledge-embedded machine learning (KML) model sistent solutions even when the training data is limited.
4. Final Hypothesis In this stage, the output of the ML model
Recently, a third approach has emerged in the research com- is tested against existing or known knowledge, and will be
munity Karpatne et al. (2017); Camps-Valls (2018); Raissi et discarded if the result for a particular input is inconsistent
al. (2019), which incorporates knowledge (experience and/or with prior knowledge. For example, Fang et al. (2017)

Fig. 4 Information flow in a


ML pipeline, where knowledge
can be integrated von Rueden et
al. (2021)

123
2894 Journal of Intelligent Manufacturing (2023) 34:2889–2906

prior knowledge about the system that has been embedded


into its modeling pipeline. For example, simulated data gen-
erated from the Kb model can be fed into the KML model
during the training stage to improve its accuracy. However,
this add-on knowledge also limits the accuracy of KML as the
number of available data increases. Knowledge-based mod-
els (e.g., physical equations to model the system behavior)
are only approximations of the real systems. When integrated
into KML, they cause the KML model output to remain
close to those of approximated knowledge-based models.
Although this reduces the accuracy of KML as compared
to DDML when dealing with a high amount of data, it helps
Fig. 5 General expectation of accuracy improvement in DDML, Kb, the model not to generate unrealistic outputs.
and KML models over increasing data To verify the accuracy of different modeling approaches
shown in Fig. 5, a simulated test case example is studied.
used a graph model to adjust the probability outputs from Assume a system defined by Eq. 1, and the Kb model defined
a deep learning model for object classification in a image. by Eq. 2 as an approximation of this system. In conducting
In their method, based on objects detected with high con- the comparative studies, as shown in Fig. 6, three sets of
fidence levels, the knowledge graph model will assist to training data with 10, 100, and 1000 data points were used.
reclassify remaining objects in the image. Results from the models trained using DDML, and the
two models developed using KML are shown in Fig. 7. For
Comparison between DDML, Kb, and KML models the KML–Data–Augmented model, extra 100 data samples
were generated using Equation 2 and added to the three
A schematic representation comparing the three models, datasets shown in Fig. 6 for training the model. For the KML–
namely, Data-Driven Machine Learning (DDML) model, Knowledge–Loss model, the Kb model was used as extra
Knowledge-based (Kb) model, and Knowledge-embedded loss function for the model training, as described by Eq. 3.
Machine Learning (KML) model is shown in Fig. 5. It can be As expected, and shown in Fig. 7, all models’ performance
observed that the accuracy of DDML is improved by adding improve as the number of training data is increased. The
more training data to the model. In contrast to DDML, the results from this study, as summarized in Table 2 and Fig. 8,
Kb model does not rely on the amount of data, and hence its support what presented earlier in Fig. 5.
accuracy remains flat (i.e., plateau).
Similar to DDML, the accuracy of KML also improves
with additional data, but with higher accuracy for the low y = sin (2π x) + 0.2sin (10π x) + 0.1sin (100π x) (1)
amount of data. This is because the KML model also uses y K b = sin (2π x) (2)

(a) (b) (c)

Fig. 6 Training datasets used in the simulated test-case study with a 10 data, b 100 data, and c 1000 data, randomly selected from Eq. 1

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2895

(a) (b)

(c) (d)

Fig. 7 Results from the proposed two KML models and Data-Driven Machine Learning (DDML) model, compared to the actual system model
(using Eq. 1)

T rainingLoss = M S E_Loss ( pr ediction, target) shows a comparison among abductive reasoning vs. deduc-
+M S E_Loss ( pr ediction, y K b ) (3) tive vs. inductive.
A typical application of abductive learning is in fault
detection, whereby using knowledge about the faults, abduc-
tive reasoning can derive sets of faults that are likely to
KML and abductive learning happen Dai et al. (2019). Indeed, this can be interpreted as
the case of generating results from a ML-based model based
The authors also see a synergy between the concept of on an incomplete set of data, such that the results are consis-
KML and abductive learning, which refers to the process tent with the background knowledge. This can be achieved
of inferring hypotheses that give the best explanation to cur- by embedding the knowledge as part of the constraints or
rent observations based on background knowledge Magnani loss functions in the KML pipeline. However, this synergy
(2009). Figure 9 adapted from Ketokivi and Choi (2014)

123
2896 Journal of Intelligent Manufacturing (2023) 34:2889–2906

Table 2 Results of the losses (in


Training Kb DDML KML-data augmented KML-knowledge loss
MSE) for different models used
data model model model model
to simulate the system described
by Eq. 1 10 0.025 0.2956 0.0314 0.1598
100 0.025 0.0215 0.0137 0.0143
1000 0.025 0.0092 0.0116 0.0104

sensors’ information and some physical models are also


available for simulating the machines and processes.
It is also important to note it is not economically feasible to
generate large amount of data needed for training the DDML
models for such applications. This is because in many man-
ufacturing environments data collection is not trivial, as it
requires proper Design of Experiment (DoE) planning, which
is generally costly and time-consuming.
On that basis, it can be argued that the information flow
von Rueden et al. (2021) described in Fig. 4 lacked crucial
first steps where knowledge can also be embedded in the
ML pipeline. Essentially, this is in relation to data collection
where expert knowledge of the manufacturing processes is
necessary. Thus, the authors propose a more comprehensive
Fig. 8 Comparative results from the different approaches in modeling
ML pipeline for manufacturing environment, as illustrated in
the system of Eq. 1 against increasing the number of training data
Fig. 10, by adding the following three stages, where knowl-
edge can also be embedded into them.
study requires further investigations, which is not the focus
of this paper. (i) Sensorization: It is used to extract representative key
process variables from the machines. This includes i)
the type and suitability of sensors to be deployed, ii)
mounting of identified sensors at suitable positions on
KML and smart manufacturing the machines, and iii) the design and running of cus-
tomized DoE for data collection.
Studies conducted in Comparison between DDML, Kb, and
(ii) Data Preparation: This step is crucial as it ensures
KML models shows that the use of KML is more suitable
the integrity of the collected data. It verifies that the
in cases where i) there are some data available for model
captured raw data can represent the entire machining
training, and ii) some physics-based models (or expert knowl-
process. It also includes suitable techniques to han-
edge) are available to be used for training the model. Thus,
dle incomplete or missing data and prepare a ‘cleaned’
KML can be a promising candidate for smart manufacturing
dataset for feature engineering Habibullah et al. (2021).
applications, where some data is usually available through
(iii) Feature Engineering: As the collected raw data is usu-
ally not useful for model development, some features
from the raw data can be extracted to reduce the data
size while preserving the important characteristics of
the original data. Typical feature extraction techniques
include time-domain Wu et al. (2012), frequency-
domain and time-frequency domain Me (2015). Once
the key features are extracted, feature selection can then
be used to identify the most important features.

To the best of our knowledge, this is the first time


such a framework been proposed for smart manufacturing
Fig. 9 Graphical comparison of deductive vs. inductive vs. abductive applications. In showing the benefits of adopting the KML
in using knowledge and data for their inferences (adapted from Ketokivi framework presented in Fig. 10, its application on roller
and Choi (2014)) element bearing, a component often used in smart manufac-

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2897

Fig. 10 General framework of


KML for smart manufacturing
applications

turing, is investigated. The objective is to build a ML model outer race fault, and 7 recordings with inner race fault. For
to detect potential faults in roller element bearing using each recording, only a few seconds of signal from accelerom-
an accelerometer sensor. In general, a typical rolling ele- eters were made available. Thus, the recordings were split
ment bearing can experience four common faults (or cracks), into one—second segments and time—domain features, as
namely, outer race, inner race, cage, and rolling element. defined in Table 3, were applied to them. In simulating a
Each of these faults affects the performance of the bearings real environment, where the data is always contaminated
differently, and evaluating the lifetimes of these bearings will with some noise, 15% Gaussian noise was also added to the
require comprehensive studies, which is not the focus of this dataset. The detail of the experimental settings is shown in
investigation. Table 4.
Figure 11 shows the diagram of a typical roller element Figure 12a shows the DDML pipeline for the analysis.
bearing, with induced inner race and outer race faults where Using the equations in Table 3, time-domain features were
Bd is the ball diameter, Pd is the pitch diameter, Nb is the extracted from the raw data. The data was then split into train-
number of balls, S is the rotational speed and  is the contact ing (70%) and testing (30%) and fed to an eXtreme Gradient
angle. In our analysis, the data from the Machinery Failure Boosting (XGBoost) Chen and Guestrin (2016) classifier.
Prevention Technology Challenge Data Bechhoefer (2013), The prediction accuracy against the size of the training data
containing 23 recordings collected from a machine with dif- was then evaluated. As shown in Fig. 13, the model accuracy
ferent roller bearing conditions was used. increases as we increase the number of training data.
Most recordings were sampled at 48828Hz, (based on a As mentioned in Randall and Antoni (2011), when rolling
12.5MHz system clock followed by a clock divider with a elements strike a local fault on the outer or inner race, a
ratio of 128 or 256. Recordings sampled at 97656Hz were shock is introduced that excites high-frequency resonances
downsampled to 48828Hz using MATLAB signal process- of the whole structure between the bearing and the response
ing toolbox functions. The dataset contained 3 recordings transducer. These shocks modulate a signal at the associ-
with healthy bearing condition (baseline), 10 recordings with ated bearing pass frequencies, such as Ball Pass Frequency

Fig. 11 Roller bearing with


outer race and inner race faults

123
2898 Journal of Intelligent Manufacturing (2023) 34:2889–2906

Table 3 List of time-domain features Table 4 Settings for the roller element bearing experiment
Feature Formula ∗
Item Setting value
k
Average amplitude p1 = 1
k i=1 s(i) Number of samples 23
 1 Sampling rate 48,828 Hz
k 2
i=1 (s(i)− p1 )
2
Standard deviation p2 = k−1 Segmentation length 1s
  1 Shaft speed 1500 rpm
k 2
Root mean square amplitude p3 = 1
k i=1 s(i)
2
Load weight 0–300 lbs
  √ 2
k
Square mean rooted absolute amplitude p4 = 1
k i=1 |s(i)| Ball diameter (Bd ) 0.235 in
Peak value p5 = max(| s(i) |) Pitch diameter (Pd ) 1.245 in
k Number of balls (Nb ) 8
i=1 (s(i)− p1 )
3
Skewness coefficient p6 = (k−1)( p2 )3
k
Contact angle () 0
i=1 (s(i)− p1 )
4
Kurtosis coefficient p7 = (k−1)( p2 )4
p5
Peak factor (Crest factor) p8 = p3
Margin factor p9 = p5 be predicted.
p4
Waveform factor p10 = k
p3  
1
s(i) Nb Bd
BPFI = ∗S∗ 1+ ∗ cos 
k i=1

Impulse factor p11 = 


p5 (4)
1 k 2 Pd
k i=1 s(i)
 
Min amplitude p12 = min(s(i)) Nb Bd
BPFO = ∗S∗ 1− ∗ cos  (5)
Max amplitude p13 = max(s(i)) 2 Pd
Max-min p14 = p13 − p12
Peak-mean p15 = p5 − p1 This knowledge about roller bearings can also be used to
create a Kb model for the problem. In doing this, the Power
∗ s(i) is the i th sample of the signal, k is length of the signal Spectral Density (PSD) for B P F I and B P F O frequencies
were calculated, and knowledge-based rules from Eq. 6 were
then applied for fault detection. Figure 12b shows the Kb
pipeline based on these rules. The result of the fault prediction
for the same test data is shown in Fig. 13. As expected, the
Outer Race (BPFO), and Ball Pass Frequency Inner Race accuracy of the Kb model is not affected by increasing the
(BPFI) Bechhoefer (2018). Thus by calculating these two training data.
frequencies (see Eqs. 4 and 5), the rolling bearing fault can

Fig. 12 Different pipelines for


roller element bearing fault
detection with a DDML, b Kb,
and c KML

(a)

(b)

(c)

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2899

Fig. 13 Comparison of the accuracy for the different models


⎨ IF (P S D(B P F I )  P S D(B P F O)) THEN inner race f ault Bibliometric study
IF (P S D(B P F I )  P S D(B P F O)) THEN outer race f ault

ELSE H ealthy bearing
From the previous sections, it was clearly shown that out of
(6) the different ways of combining knowledge-based and data-
driven approaches, the usability and applicability of KML
Finally, in developing the KML pipeline, as shown are evident. To understand its adoption in smart manufactur-
in Fig. 12c, two additional attributes, P S D(B P F I ) and ing, a multi-level bibliometric state-of-the-art study on AI
P S D(B P F O), were added to the time–domain features, in smart manufacturing was carried out. Each search was
which is categorized under the “Feature Engineering” part performed using Web of Science (WoS) Clarivate Analytics
of the KML’s general framework shown in Fig. 10. It can (2021), with the combination (additive with AND operator)
be observed in Fig. 13 that KML was able to achieve higher of keywords in the topic. Records for each year were saved
accuracy by using less training data. The encouraging results as text files, with full records and cited references, and pro-
from this real case example demonstrated the usefulness cessed by analysis scripts to create the figures in this study.
and feasibility of the KML framework, thereby opening Two searches were performed. One with “Smart Manufac-
the doors for its potential use in other smart manufacturing turing AND Artificial Intelligence (SMAI)” and another with
applications. ‘Smart Manufacturing AND Artificial Intelligence AND
Knowledge (SMAIK)”. The growth of SMAI literature fol-
lows a power law while the growth of SMAIK literature
Potential KML applications in smart follows a linear progression, as shown in Fig. 14a and b,
manufacturing respectively. It is worth noticing that the data available for
SMAIK is only limited to 4 years. This trend is not surpris-
In this section, potential applications of KML in smart man- ing due to the performance improvements of AI compared
ufacturing are proposed. Starting with a bibliometric study to known conventional methods, driven mainly by initiatives
on the current landscape in this field to obtain a better such as Industry 4.0 Bassi (2017) and IIoT Whitmore et al.
understanding of the current trends, the use of KML in two (2015).
important areas of smart manufacturing: Prognostics and In the SMAI search, it was observed that 30 countries con-
Health Management (PHM) and Digital Twin (DT) are then tributed to this research field (WoS has 110 entries with the
reviewed. field country of origin), with USA (36), China (13), Taiwan

123
2900 Journal of Intelligent Manufacturing (2023) 34:2889–2906

(a) (b)

Fig. 14 Growth of the literature in a smart manufacturing with AI and b with both AI and Knowledge

(7), Germany (6), Spain (5), U.K. (5), Italy (4), South Korea the machines’ maintenance. The challenge here is to obtain
(4), France (3), Ireland (3), Canada (2), Finland (2), Greece accurate health and lifetime prediction of the machines’ com-
(2), etc. For the SMAIK search, it was found that only 9 coun- ponents through direct or indirect sensing and then analyze
tries contributed to this research field (WoS has 22 entries the current and historical sensors’ data. This has become fea-
with the field country of origin), with USA (10), China (2), sible due to the rapid advancements made in big data analytics
France (2), Germany (2), Spain (2), Finland (1), Japan (1), as well as the availability of cost-effective sensing technolo-
Singapore (1) and Taiwan (1). These findings showed that gies. Atamuradov et al. (2017) and Lee (2014) provide a
the incorporation of knowledge in smart manufacturing is comprehensive description of PHM systems, covering the
still in its infancy, and would indicate a potential for research different types of tools, critical components, and algorithms
in these domains. used.
Results from the state-of-the-art bibliometric studies One of the machine components that requires close mon-
showed that KML (and/or even AI) adoption in smart man- itoring and directly affects the quality of the products being
ufacturing is still early, with its value and benefit not fully machined is the degradation of the tool. Typically, this
established. Still, there have been already a few attempts in involves the need to continuously detect anomalies, which
incorporating knowledge and/or physics into ML for smart could potentially lead to faults, followed by the prognosis
manufacturing applications. As shown in Fig. 14b, the main of the lifetimes of the tool. In order to estimate the lifetime
research activities can be grouped into the following three of these tools using data-driven approaches, it is essential to
groups: Prognostics and Health Management (PHM), Digital collect the complete degradation data during the machining
Twin (DT), and others (including IIoT, networking, security, process. However, in reality, it is often timely and expensive
etc.). Since PHM and DT are the two largest, a deeper dive to generate such complete degradation data, either due to the
into these two topics were undertaken. The aim is to better cost of the workpieces involved or the time required for the
understand how researchers have incorporated knowledge tool to be completely worn out.
and/or physics into ML as well as proposing KML frame- Literature Camci et al. (2013); Diamanti and Soutis (2010)
works for these two research topics. have proposed the use of specific accelerated DoEs to gen-
erate representative data for analysis. This implies the need
Prognostics and health management (PHM) to incorporate knowledge in these experiments that aims to
“best-mimic” the degradation of the tool, aligning to the
One of the primary aims of PHM in the manufacturing KML’s general framework under the “sensorization” stage in
industries is to achieve better planning for the scheduling of Fig. 10. For instance, in Das et al. (2019), the incorporation of

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2901

Fig. 15 Proposed KLM framework for PHM applications

knowledge led to experiments on the new-tool, medium-worn was first built, with the input features extracted from multiple
tool, and full-worn tool to be conducted on a limited number raw sensory signals, and subsequently, the estimated HI was
of workpieces. These experiments were then repeated on five mapped to the RULs. Multiple motors’ data were utilized in
different tools and the same tool was “worn out” for all of the learning process. Each motor’s data is used for training
the three experiments. Using limited degradation data from a single model and then predicting the RUL. The predicted
these knowledge-embedded experiments, a ML model was RULs from the multiple single models were then combined in
trained. The results showed the ability of the model to esti- an ensemble process to produce the final RUL prediction. The
mate the current wear value of a tool and to predict the time authors showed that the smoothed HI has lower variations and
to full-worn. higher closeness to the true HI of the motor over its lifetime.
In predicting a machine’s RUL, knowledge can also be This generic prognostics framework for RUL prediction has
incorporated as a Health-Index (HI), with the assumption since been extended to different types of HI degradations
that the HI is continuously and gradually decreasing over Yang et al. (2021).
time. For instance, Wang et al. (2008) used a linear regression Based on the above study, the proposed KML general
model for HI calculation, and then a similarity-based prog- framework in Fig. 10 can also be revised to include a HI
nostics approach was used to estimate the RUL. As another module for PHM applications, as illustrated in Fig. 15. This
example, Riad et al. (2010) used a similar linear regression framework also includes the prediction pipeline, which is
model smoothed with a third-order polynomial curve fitting the user-facing portion of the KML framework. This pipeline
in calculating the HI. The HI was fed as an input to a neu- captures the new data, performs feature engineering, and uses
ral network for RUL prediction. Liu et al. (2018) developed the developed KML model in the training pipeline to predict
a composite HI through the weighted summation of mul- the current state of the equipment as well as the RUL. Fur-
tiple degradation-based sensor data. In this approach, the thermore, while the results are being shown to the users,
weights were determined by optimizing a quadratic program- performance evaluation of the system can be concurrently
ming problem. A probability model with the composite HI done regularly if the system parameters are non-stationary.
for RUL prediction Xi et al. (2013) was then employed. Bech-
hoefer et al. (2011) proposed the construction of a HI using Digital twin
the Cholesky decomposition-based whitening transform for
a condition monitoring application. Digital Twin (DT) is defined as a physical and/or virtual
In another example where knowledge was incorporated, machine or computer-based model that simulates, emulates,
Yang (2016) used the HI to represent the health status of an mirrors, or twins the life of a physical entity, which can be a
induction motor, by assuming that the HI degrades linearly process, an object, a human, or a human-related feature Kir-
from 100% to 0%. In their work, a model for HI estimation itsis (2011); Grieves (2015). In essence, each DT is linked

123
2902 Journal of Intelligent Manufacturing (2023) 34:2889–2906

While authors in Alam and El Saddik (2017), Negri et


al. (2017), Tao and Qi (2019), Lu et al. (2020), Gabor et
al. (2016), Zheng (2018), Rios et al. (2016) have proposed
different DT frameworks for specific applications, a generic
DT framework, as shown in Fig. 16), can be defined with the
specific spaces defined as follows:

1. Physical Space: A container that brings manufacturing


resources together. Here, the manufacturing resources
Fig. 16 Generic framework for a Digital Twin include Work-In-Process (WIP), and smart manufactur-
ing devices (such as wired or wireless sensors, and smart
gateways), where the status of WIP and smart manufac-
to its physical twin through a unique key that first identifies
turing devices are perceived by sensors or other input
the physical twin and then establishes a bijective relation-
mechanisms. Thus, this space captures individual aspects
ship (i.e., one-to-one correspondence) between the DT and
and parameters for which DT is being implemented, often
the physical asset.
referred to as the physical world Barricelli et al. (2019)
It has been argued that a DT is more than a simple
or physical things Alam and El Saddik (2017). Theoret-
model or simulation Grieves (2015); Kritzinger et al. (2018);
ically, DT aims to include a comprehensive structuring
Boschert and Rosen (2016). Rather, it is an intelligent, liv-
of heterogeneous kinds of data available on the manu-
ing, and evolving model, being the virtual counterpart of a
facturing shop floor. This data will then be connected to
physical entity or process. Typically, it follows the lifecy-
their respective semantics in order to facilitate retrieval,
cle of its physical twin asset, which monitors, controls, and
interpretation, and exploitation.
optimizes its processes and functions. It continuously pre-
2. Social Space: This space integrates a variety of service
dicts future statuses such as defects, damages, and failures
systems, such as Customer Relation Management (CRM)
and allows for simulating and testing novel configurations.
and Enterprise Resource Planning (ERP), which bridges
Specifically, the twinning process is made possible due to the
the gap between the supply of DT and the demand of cus-
continuous interactions, communications, and synchroniza-
tomers in service-oriented manufacturing. In essence, this
tions (closed–loop optimization) between the DT, its physical
space will cognize and analyze customer demands based
twin, and the external environment Barricelli et al. (2019).
on social data from customs and historical knowledge in
Authors in Alam and El Saddik (2017), Negri et al. (2017),
dynamic knowledge bases. It will then provide manufac-
Tao and Qi (2019) defined DT as an enabling technology
turing services by producing production orders through
of CPS. It contains high-fidelity digital models of physical
service systems, guiding the production process of the
equipment based on physical laws, which acquires real-time
DT.
sensing data during the operation of the equipment and stores
3. Digital Space: A container made up of virtual DT models
the historical running data for further utilization.
such as virtual WIP, virtual manufacturing devices, and
Uhlemann et al. (2017) pointed out that DT as a core
virtual machining processes. This is done using real-time
component of CPS makes it able to sense and accurately
manufacturing data published by data space and historical
reflect the behavior and real-time state of production systems
knowledge in knowledge space, whereby it will simulate,
so that the processes can be analyzed, simulated, predicted,
understand, predict and finally optimize the performance
and optimized Lu et al. (2020). These technologies enable a
of the DT. This space consists of the data space and AI
large volume of data from various manufacturing activities
space for data and information management.
to be collected and managed. In its implementation, DT con-
sists of various spaces that need to communicate with each (a) Data space: It is the core of the DT, often referred
other seamlessly, especially the perpetual communications to as the world model Barricelli et al. (2019), cyber
between the physical asset (or physical space) and the dig- model Gabor et al. (2016), and digital mock-up Zheng
ital asset (or digital space). This continuous data exchange (2018). It is a container for real-time manufactur-
can be direct or through cloud-based connections Zhang et ing data related to the status of WIP, manufacturing
al. (2019). Its importance is in the knowledge space where devices, and the manufacturing environment. It is
data needs to be stored, processed, and analyzed for decision responsible for pre-processing and publishing real–
making. Furthermore, in systems that need to be monitored time data from the physical space and subscribing
regularly, the social space is a necessity as it integrates the control orders from the digital space or social space
variety of service systems connected to the physical world. through smart gateways.

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2903

Fig. 17 An example of DT
architecture that includes
simulation, visualization, and
support for decision-making

(b) AI space: It acts as the brain of DT by integrat- a DT for corrective maintenance developed using distance
ing multiple AI algorithms for self-decision making sensor values and corresponding error classification in the
to handle various manufacturing problems in physi- simulation. These developed models act as a sub-module or
cal, digital, and social spaces. One of the requests is entire module in the DT block of Fig. 16.
to have AI algorithms that are dynamic and adapt- Thus, one way of using KML in DT applications is through
able so it can learn and be improved by itself and the development of a model that can simulate the manufac-
adapt to new conditions and environments. This is turing machines. Here, knowledge from essential aspects of
where KML can play an important role, as domain the physical machines should be included in the KML mod-
knowledge can be embedded into the ML models for eling. As suggested by Aivaliotis et al. (2019), these aspects
those decision-making tasks. Such a KML model can include dynamic behavior of the machine, modeling of vir-
reduce unstable behaviors of commonly used DDML tual sensors needed for simulation, and settings of necessary
models. A close example of this concept is the DT hyperparameters related to physics equations to be used in
presented in Qiao et al. (2019) that combined knowl- KML model.
edge management and data analytics systems. An Such a simulation model can be considered as a subset
ontology-based knowledge model was used as a com- of AI space or as a separate module in a DT architecture.
mon repository, which is based on different types For example, Ladj (2021) used a similar simulation model
of insights such as industrial data and knowledge, that relied on core components of AI space. It enables the
business rules provided by experts, and knowledge management and analysis of real-time data coming from the
learned from data analytics. The employed knowl- actual physical counterpart Schuh et al. (2019), to commu-
edge management method enabled the integration of nicate with the simulation model as shown in Fig. 17. As
inferred knowledge. Furthermore, the data analytics another example, using a similar architecture, Zhang et al.
method relied on unsupervised machine learning for (2019) proposed and implemented a DT for an industrial
the classification of machining contexts and detection robot as a case study that is able to perceive and simulate the
of detrimental incidents. As a case study, the feasibil- manufacturing process using real-time data published from
ity of the DT framework for tool failure detection the data space while understanding, predicting and optimiz-
was presented, where DT enhanced production per- ing the manufacturing performance using AI space.
formance by supporting decision-making.

Indeed, knowledge is often integrated and embedded in the Conclusions and future work
development of DT. For example, researchers in Aivaliotis
et al. (2019) and Aivaliotis et al. (2017) presented a physics This paper introduced Knowledge-embedded Machine
modeling approach used for predictive maintenance appli- Learning (KML), being one of the three possible methods
cations and RUL prediction. Aivaliotis et al. (2019) also of combining knowledge-base and data driven models into
developed a Kb model in predictive maintenance applica- the ML pipeline. Using data from simulations and real test
tions. Along the same lines, Aivaliotis et al. (2017) then cases, the comparative performance improvements in KML
presented an approach for predicting the condition and status models against Kb and DDML models were evident. The
of machines using simulations without ceasing the operation choice of modeling approach based on the data availability
of these machines. Vathoopan et al. (2018) also presented is summarized in Table 5.

123
2904 Journal of Intelligent Manufacturing (2023) 34:2889–2906

Table 5 Modeling approach recommendation from data availability perspective


Information available Recommended modeling approach

None or little data is available Knowledge-based (Kb) model


Large amount of data is available Data-Driven Machine Learning (DDML) model
Knowledge from experience-based rules and/or physics equations about the Knowledge-embedded Machine Learning (KML) model
system coupled with representative data is available

Applied to smart manufacturing case studies, our findings based modelling in predictive maintenance. Procedia of the 52nd
showed the suitability of adopting KML frameworks when- CIRP Conference on Manufacturing Systems, 81, 417–422.
Alam, K. M., & El Saddik, A. (2017). C2PS: A digital twin architecture
ever physical (or experience-based) models together with
reference model for the cloud-based cyber-physical systems. IEEE
DoE data are available. The outcome of these findings also Access, 5, 2050–2062.
led to a general KML framework. Atamuradov, V., Medjaher, K., Dersin, P., Lamoureux, B., & Zerhouni,
Bibliometric studies, conducted on smart manufacturing N. (2017). Prognostics and health management for maintenance
practitioners - review, implementation and tools evaluation. Inter-
articles, which contained knowledge and AI, showed that the
national Journal of Prognostics and Health Management, 8(3),
majority of them are related to PHM or DT. Hence, how 1–31.
these applications can potentially benefit from KML were Barricelli, B. R., Casiraghi, E., & Fogli, D. (2019). A survey on digital
then discussed. Although it can be argued that some of the twin: Definitions, characteristics, applications, and design impli-
cations. IEEE Access, 7, 167653–167671.
referred publications on DT are closer to hybrid rather than
Bassi, L. (2017). Industry 4.0: Hope, hype or revolution? Proceed-
KML models, the expectation is that KML will add value ings of the 2017 IEEE 3rd international forum on research and
when applied in this field. However, this realization requires technologies for society and industry (RTSI’2017), Modena, Italy,
further investigation and is left as future work. September 11–13.
Bechhoefer, E. (2013). Condition based maintenance fault database for
The main limitation of the study is the need for representa- testing of diagnostic and prognostics algorithms. https://mfpt.org/
tive data about the system that might not always be available, fault-data-sets/.
or even if it is available, the quantity of the representative Bechhoefer, E. (2018). A quick introduction to bearing envelope
data might be insufficient. One potential solution using com- analysis. (Society for Machinery Failure Prevention Technology).
https://mfpt.org/wp-content/uploads/2018/03/MFPT-Bearing-
puter simulation models to generate representative simulated Envelope-Analysis.pdf.
datasets that can best represent the system that is being con- Bechhoefer, E., He, D., & Dempsey, P. (September, 2011). Gear health
sidered. These simulated datasets can then be used together threshold setting based on a probability of false alarm. Proceedings
with knowledge from experience-based rules and/or physics of the annual conference of the Prognostics and Health Manage-
ment Society, Montreal, Canada, pp. 275–281.
equations about the system, as demonstrated by KML–Data– Bektas, O., Marshall, J., & Jones, J. A. (2020). Comparison of compu-
Augmented model in Comparison between DDML, Kb, and tational prognostic methods for complex systems under dynamic
KML models. The authors aim to conduct this study in the regimes: A review of perspectives. Archives of Computational
near future. Methods in Engineering, 27, 999–1011.
Bikmukhametov, T., & Jäschke, J. (2020). Combining machine learn-
Acknowledgements The authors would like to thank Mr. Eric Bech- ing and process engineering physics towards enhanced accuracy
hoefer (Green Power Monitoring Systems (GPMS) International, Inc., and explainability of data-driven models. Computers & Chemical
USA) for the detailed information on the bearing dataset, data acquisi- Engineering, 138, 106834.
tion system, and the related sampling rate. Boschert, S. & Rosen, R. (2016). Digital Twin – The Simulation Aspect.
Mechatronic Futures, pp. 59–74.
Camci, F., Medjaher, K., Zerhouni, N., & Nectoux, P. (2013). Feature
evaluation for effective bearing prognostics. Quality and Reliabil-
References ity Engineering International, 29(4), 477–486.
Camps-Valls, G., et al. (2018). Physics-aware Gaussian processes in
Aggarwal, C. C. (2021). Machine learning: The inductive view. Artifi- remote sensing. Applied Soft Computing, 68, 69–82.
cial Intelligence, 1, 167–210. Chan, S. & Elsheikh, A. (2017). Parametrization and generation
Ahmad, W., Khan, S. A., & Kim, J.-M. (2018). A hybrid prognostics of geological models with generative adversarial networks.
technique for rolling element bearings using adaptive predictive arXiv:1708.01810
models. IEEE Transactions on Industrial Electronics, 65(2), 1577– Chen, T. & Guestrin, C. (August 2016). XGBoost: A scalable tree boost-
1584. ing system. Proceedings of the 22nd ACM SIGKDD international
Aivaliotis, P., Georgoulias, K., & Chryssolouris, G. (2017). A RUL cal- conference on knowledge discovery and data mining (KDD’16),
culation approach based on physical–based simulation models for San Francisco, USA, pp. 785–794.
predictive maintenance. Proceedings of the 23rd international con- Chen, C., Vachtsevanos, G., & Orchard, M. E. (2012). Machine remain-
ference on engineering, technology and innovation (ICE/ITMC), ing useful life prediction: An integrated adaptive neuro-fuzzy and
Madeira, Portugal, June 27–29, pp. 1284–1287. high-order particle filtering approach. Mechanical Systems and
Aivaliotis, P., Georgoulias, K., Arkouli, Z., & Makris, S. (2019). Signal Processing, 28, 597–607.
Methodology for enabling digital twin using advanced physics-

123
Journal of Intelligent Manufacturing (2023) 34:2889–2906 2905

Chen, C., Zhang, B., Vachtsevanos, G., & Orchard, M. E. (2011). Mining (Society for Industrial and Applied Mathematics), May
Machine condition prediction based on adaptive neuro-fuzzy and 2019, pp. 558–566.
high-order particle filtering. IEEE Transactions on Industrial Elec- Karpatne, A., Watkins, W., Read, J., & Kumar, V. (2017). Physics-
tronics, 58(9), 4353–4364. guided Neural Networks (PGNN): An application in lake temper-
Chinnam, R. B., & Baruah, P. (2004). A neuro-fuzzy approach for esti- ature modeling. arXiv:1710.11431
mating mean residual life in condition-based maintenance systems. Karpatne, A., et al. (2017). Theory-guided data science: A new
International Journal of Materials and Product Technology, 20(1), paradigm for scientific discovery from data. IEEE Transactions
166–179. on Knowledge and Data Engineering, 29(10), 2318–2331.
Clarivate Analytics. Web of Science (2021). http://apps. Ketokivi, M., & Choi, T. (2014). Renaissance of case research as a
webofknowledge.com. scientific method. Journal of Operations Management, 32(5), 232–
Daigle, M. J., & Goebel, K. (2013). Model-based prognostics with 240.
concurrent damage progression processes. IEEE Transactions on Kiritsis, D. (2011). Closed-loop PLM for intelligent products in the era
Systems, Man, and Cybernetics: Systems, 43(3), 535–546. of the Internet of things. Computer Aided Design, 43(5), 479–501.
Dai, W.-Z., Xu, Q., Yu, Y., & Zhou, Z.-H. (2019). Bridging machine Kritzinger, W., Karner, M., Traar, G., Henjes, J., & Sihn, W. (2018).
learning and logical reasoning by abductive learning. Advances in Digital Twin in manufacturing: A categorical literature review and
Neural Information Processing Systems, 32, 2019. classification. IFAC-PapersOnLine, 51(11), 1016–1022.
Das, A. K., Feng, Y., Habibullah, M. S., Yu, Z. & Farbiz, F. (October Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet classifi-
2019). Tool Wear Health Monitoring with Limited Degradation cation with deep convolutional neural networks. Communications
Data. Proceedings of 2019 IEEE region 10 conference (TENCON– of the ACM, 60(6), 84–90.
2019), Kochi, India, pp. 17–20. Kusiak, A. (2017). Smart manufacturing must embrace big data. Nature,
Diamanti, K., & Soutis, C. (2010). Structural health monitoring tech- 544, 23–25.
niques for aircraft composite structures. Progress in Aerospace Kusiak, A. (2018). Smart manufacturing. International Journal of Pro-
Sciences, 46(8), 342–352. duction Research, 56(1–2), 508–517.
Ding, H., et al. (2020). State of AI-based monitoring in smart manufac- Ladj, A., et al. (2021). A knowledge-based Digital Shadow for machin-
turing and introduction to focused section. IEEE/ASME Transac- ing industry in a Digital Twin perspective. Journal of Manufactur-
tions on Mechatronics, 25(5), 2143–2154. ing Systems, 58(Part B), 168–179.
Downey, A., Lui, Y.-H., Hu, C., Laflamme, S., & Hu, S. (2019). Physics- Lee, J., et al. (2014). Prognostics and health management design
based prognostics of lithium-ion battery using non-linear least for rotary machinery systems-Reviews, methodology and appli-
squares with dynamic bounds. Reliability Engineering and Sys- cations. Mechanical Systems and Signal Processing, 42(1–2),
tem Safety, Elsevier, 182(C), 1–12. 314–334.
Durkin, J. (1990). Research review: Application of expert systems in Lee, J., Davari, H., Singh, J., & Pandhare, V. (2018). Industrial arti-
the sciences. The Ohio Journal of Science, 90(5), 171–179. ficial intelligence for industry 4.0-based manufacturing systems.
Erge, O. & van Oort, E. (January 2020). Combining Physics–Based Manufacturing Letters, 18, 20–23.
and Data–Driven Modeling for Pressure Prediction in Well Con- Liao, L., & Kottig, F. (2014). Review of hybrid prognostics approaches
struction. Proceedings of the 19th Python in Science Conference for remaining useful life prediction of engineered systems, and
(SCIPY–2020), pp. 125–131. an application to battery life prediction. IEEE Transactions on
Fang, Y., Kuan, K., Lin, J., Tanand, C., & Chandrasekhar, V. (2017). Reliability, 63(1), 191–207.
Object detection meets knowledge graphs. Proceedings of the 26th Liu, C., Wang, K., Wang, Y., & Yuan, X. (2021). Learning deep multi-
international joint conference on artificial intelligence (IJCAI-17), manifold structure feature representation for quality prediction
Melbourne, Australia, August 19–25, pp. 1661–1667. with an industrial application. IEEE Transactions on Industrial
Gabor, T., Belzner, L., Kiermeier, M., Beck, M. T., & Neitz, A. (2016). Informatics. https://doi.org/10.1109/TII.2021.3130411.
A Simulation–Based Architecture for Smart Cyber–Physical Sys- Liu, T., Zhu, K., & Zeng, L. (2018). Diagnosis and prognosis of
tems. Proceedings of the 2016 IEEE international conference on degradation process via hidden Semi-Markov Model. IEEE/ASME
autonomic computing (ICAC2016), Wuerzburg, Germany, July 17– Transactions on Mechatronic, 23(3), 1456–1466.
22 . Lopatin, D. V., & Zhirov, A. I. (2017). Geomorphology in the system of
Grieves, M. (2015). Digital twin: Manufacturing excellence through Earth sciences. Geography and Natural Resources, 38, 313–318.
virtual factory replication. Digital Twin White Paper. Lu, Y., Liu, C., Wang, K.I.-K., Huang, H., & Xu, X. (2020). Dig-
Gross, R. (2005). Psychology: The science of mind & behaviour (5th ital Twin-driven smart manufacturing: Connotation, reference
ed.). London: Hodder Education Publishers. model, applications and research issues. Robotics and Computer-
Habibullah, M. S., et al. (2021). A perspective into analysing tool wear Integrated Manufacturing, 1, 101837.
condition in hard-turning process - The key lessons learnt. In Machamer, P., & Silberstein, M. (2002). The Blackwell guide to the
C. Toro, W. Wang, & H. Akhtar (Eds.), Implementing Industry philosophy of science (Vol. 19). Oxford: Blackwell.
4.0. Intelligent Systems Reference Library (Vol. 202, pp. 79–111). Magnani, L. (2009). Abductive cognition: The epistemological and eco-
Cham: Springer. cognitive dimensions of hypothetical reasoning (1st ed.). Berlin:
Hansen, K., et al. (2013). Assessment and validation of machine Springer.
learning methods for predicting molecular atomization energies. Me, M. (2015). A review on applications of the wavelet transform tech-
Journal of Chemical Theory and Computation, 9(8), 3404–3419. nique in spectral analysis. Journal of Applied & Computational
Hosseinkhani, K., & Ng, E. (2015). A combined empirical and numeri- Mathematics, 4(4), 1–6.
cal approach for tool wear prediction in machining. Procedia CIRP, Negri, E., Fumagalli, L., & Macchi, M. (2017). A review of the roles of
31, 304–309. digital twin in CPS-based production systems. Proceedings of the
Jia, X., et al. (2019). Physics guided RNNs for modeling dynamical 27th international conference on flexible automation and intelli-
systems: A case study in simulating lake temperature profiles. gent manufacturing (FAIM–2017), Modena, Italy, June 27–30, pp.
Proceedings of the 2019 SIAM International Conference on Data 939–948.

123
2906 Journal of Intelligent Manufacturing (2023) 34:2889–2906

Pecht, M., & Jaai, R. (2010). A prognostics and health management Industry 4.0. Procedia CIRP (24th CIRP Conference on Life Cycle
roadmap for information and electronics-rich systems. Microelec- Engineering), 61, 335–340.
tronics Reliability, 50(3), 317–323. Vathoopan, M., Johny, M., Zoitl, A., & Knoll, A. (2018). Modular fault
Poole, D. L., & Mackworth, A. (2017). Artificial intelligence: Founda- ascription and corrective maintenance using a digital twin. IFAC-
tions of computational agents (2nd ed.). Cambridge: Cambridge PapersOnLine, 51(11), 1041–1046.
University Press. von Rueden, L., et al. (2021). Informed machine learning— A taxonomy
Qiao, Q., Wang, J., Ye, L., & Gao, R. X. (2019). Digital Twin for machin- and survey of integrating prior knowledge into learning systems.
ing tool condition prediction. Procedia CIRP, 81, 1388–1393. IEEE Transactions on Knowledge and Data Engineering.
Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics- Wang, T., Yu, J., Siegel, D., & Lee, J. (October, 2008). A similarity-
informed neural networks: A deep learning framework for solving based prognostics approach for remaining useful life estima-
forward and inverse problems involving nonlinear partial differen- tion of engineered systems. In Proceedings of the international
tial equations. Journal of Computational Physics, 378, 686–707. conference on prognostics health management (PHM–2008)
Randall, R. B., & Antoni, J. (2011). Rolling element bearing diagnostics pp. 1–6.
- A tutorial. Mechanical Systems and Signal Processing, 25(2), Whitmore, A., Agarwal, A., & Xu, L. D. (2015). The internet of things
485–520. - A survey of topics and trends. Information Systems Frontiers, 17,
Read, C. (2016). Logic deductive and inductive (2nd ed.). Scotts Valley: 261–274.
CreateSpace Independent Publishing Platform. Willard, J., Jia, X., Steinbach, M., Kumar, V., & Xu, S. (2020). Inte-
Riad, A. M., Elminir, H. K., & Elattar, H. M. (2010). Evaluation of neu- grating physics-based modeling with machine learning: A survey.
ral networks in the subject of prognostics as compared to linear arXiv:2003.04919v4
regression model. International Journal of Engineering & Tech- Wu, Q., Yang, X., & Zhou, Q. (2012). Pattern recognition and its appli-
nology, 10(6), 52–56. cation in fault diagnosis of electromechanical system. Journal of
Rios, J., Morate, F. M., Oliva, M., & Hernandez, J. C. (2016). Frame- Information and Computational Science, 9(8), 2221–2228.
work to support the aircraft digital counterpart concept with an Xi, Z., Jing, R., Wang, P., & Hu, C. (2013). A Copula-based sampling
industrial design view. International Journal of Agile Systems and method for data–driven prognostics and health management. Pro-
Management, 9(3), 212–231. ceedings of the 2013 IEEE conference on prognostics and health
Schuh, G., Kelzenberg, C., Wiese, J., & Ochel, T. (2019). Data struc- management (PHM–2013), Gaithersburg, USA, June 24–27, pp.
ture of the digital shadow for systematic knowledge management 2–82.
systems in single and small batch production. Procedia CIRP, 84, Yan, J., Meng, Y., Lu, L., & Guo, C. (2017). Big–Data–Driven Based
1094–1100. Intelligent Prognostics Scheme in Industry 4.0 Environment. Pro-
Shrouf, F., Ordieres, J., & Miragliotta, G. (2014). Smart Factories ceedings of the 2017 Prognostics and System Health Management
in Industry 4.0: A Review of the Concept and of Energy Man- Conference (PHM–2017), Harbin, China, July 9–12 1–5 .
agement Approached in Production Based on the Internet of Yang, F., et al. (2016). Health index-based prognostics for remaining
Things Paradigm. Proceedings of the IEEE International Con- useful life predictions in electrical machines. IEEE Transactions
ference on Industrial Engineering and Engineering Management on Industrial Electronics, 63(4), 2633–2644.
(IEEM’2014), Selangor, Malaysia, December 9–12 697–701 . Yang, F., Habibullah, M. S., & Shen, Y. (2021). Remaining useful
Siddhpura, A., & Paurobally, R. (2013). A review of flank wear predic- life prediction of induction motors using nonlinear degradation
tion methods for tool condition monitoring in a turning process. of health index. Mechanical Systems and Signal Processing, 148,
International Journal of Advanced Manufacturing Technology, 65, 1–17.
371–393. Zagzebski, L. (2017). What is Knowledge? The Blackwell Guide to
Stefik, M. (1995). Introduction to knowledge systems (1st ed.). San Epistemology (Book Editors: J. Greco, E. Sosa)(Chapter 3) (pp.
Francisco: Morgan Kaufmann Publishers Inc. 92–116). (ISBN: 9780631202912) .
Steup, M. & Neta, R. (2020). The Stanford Encyclopedia of Phi- Zhang, C. et al. (2018). A multi-state diagnosis and prognosis
losophy: Epistemology (Stanford Encyclopedia of Philosophy framework with feature learning for tool condition monitoring.
(Online), Stanford, CA, USA, 2020). https://plato.stanford.edu/ arXiv:1805.00367 (pp. 1–12).
entries/epistemology. Zhang, G., et al. (2019). Effects of uniaxial and triaxial compression
Stoecklein, D., Lore, K. G., Davies, M., Sarkar, S., & Ganapathysub- tests on the frozen sandstone combining with CT scanning. Inter-
ramanian, B. (2017). Deep learning for flow sculpting: Insights national Journal of Physical Modelling in Geotechnics, 19(5),
into efficient learning using scientific simulation data. Scientific 261–274.
Reports, 7, 1. Zhang, C., Zhou, G., He, J., Li, Z., & Cheng, W. (2019). A data- and
Tao, F., & Qi, Q. (2019). Make more digital twins. Nature, 573, 490– knowledge-driven framework for digital twin manufacturing cell.
491. Procedia CIRP, 83, 345–350.
Torabi, A. J., Er, M. J., Li, X., Lim, B. S., & Peen, G. O. (2016). Appli- Zheng, P., et al. (2018). Smart manufacturing systems for Industry 4.0:
cation of clustering methods for online tool condition monitoring Conceptual framework, scenarios, and future perspectives. Fron-
and fault diagnosis in high-speed milling processes. IEEE Systems tiers of Mechanical Engineering, 13, 137–150.
Journal, 10(2), 721–732. Zhou, J. (2013). Digitalization and intelligentization of manufacturing
Tsui, K. L., Chen, N., Zhou, Q., Hai, Y., & Wang, W. (2015). Prognos- industry. Advances in Manufacturing, 1, 1–7.
tics and health management: A review on data driven approaches.
Mathematical Problems in Engineering, 1(17), 1–17.
Uhlemann, T.H.-J., Lehmann, C., & Steinhilper, R. (2017). The Dig- Publisher’s Note Springer Nature remains neutral with regard to juris-
ital Twin: Realizing the Cyber-Physical Production System for dictional claims in published maps and institutional affiliations.

123
Journal of Intelligent Manufacturing is a copyright of Springer, 2023. All Rights Reserved.

You might also like