A knowledge graph based construction method for

Digital Twin Network

Yanhong Zhu, Danyang Chen, Cheng Zhou, Lu Lu*, Xiaodong Duan
Department of Basic Network Technology
China Mobile Research Institute
Beijing, China
{zhuyanhong, chendanyang, zhouchengyjy, lulu, duanxiaodong}@chinamobile.com

Abstract—With the advent of the digital age, digital twin as a model[2]. In recent years, digital twin technology has been used
2021 IEEE 1st International Conference on Digital Twins and Parallel Intelligence (DTPI) | 978-1-6654-3337-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/DTPI52967.2021.9540177

new potential technology has received much attention from both in satellite, health care, factory, energy and transportation,
academia and industry increasingly. Digital twin technology paves etc[3][4]. Similarly, the research of digital twin technology in
a way for achieving cost-efficient trial and performance-optimal the field of communication network, i.e. digital twin network
management, through creating a virtual representation of physical (DTN), is also gradually carried out[5]-[7]. A deep learning
network digitally to simulate and predict. In this paper, we architecture is proposed, where a digital twin of the real network
propose a novel network paradigm, called digital twin network environment is used to train the deep learning algorithm off-line
(DTN) architecture for the future network, which consists of at a central server. The paper has achieved lower normalized
Physical Network, Data Lake, Digital Twin Layer and Network
energy consumption with less computation complexity
Application Layer. A Sketch acquisition algorithm is introduced
compared with an existing method and approach to the
to meet the demand of large-scale real-time data collection in Data
Lake. A digital twin network construction method based on performance of the global optimal solution[8]. In order to
knowledge graph is developed, which can represent entity address the challenge that network management is more
attribute and various topologies according to physical network. complex due to stricter requirements on virtualized
Finally, the proposed DTN is introduced to realize fault self- infrastructure and quality of service (QoS), a scalable DT of
healing in the future network, and the feasibility and novelty is network slicing is developed, aiming to capture the intertwined
proved. relationships among slices and monitor the end-to-end (E2E)
metrics of slices under diverse network environments[9]. A
Keywords—digital twin network, knowledge graph, artificial double closed-loop architecture with three layers and three
intelligence domains is proposed by China Mobile. At the same time, this
paper introduces four core elements and five key technologies
I. INTRODUCTION in detail, which lays a solid foundation for subsequent research
With the deployment of Internet of Things, cloud computing work on DTN[10]. In addition, there is a network architecture
and data center, etc., the scale of the current network is expanded called cybertwin which is similar to DTN in concept, consisting
gradually. At the same time, users put forward higher network of a cloud-centric network architecture and a radio access
resource requirements and the network needs to support more network architecture. It serves as communications assistant,
emerging applications. However, the increase of network scale network behavior logger, and digital asset owner[11].
leads to the increasing complexity of the current network, and
As a conclusion, there have been preliminary studies on
that induces plenty of problems.
DTN in industry and academia at present, but there is a lack of
In order to improve the autonomy ability of network and research on DTN architecture and construction method. At the
reduce the negative effect on physical network, many experts same time, the construction of DTN also faces many difficulties.
and scholars have proposed various solutions, such as building So the challenges we need to address in this article are listed as
network slices, constructing automatic driven network or intent follows:
based network and so on. We consider that an endogenous
x Digital twin technology has been put forward for a long
intelligent and autonomous network architecture which achieves
time and used in industry widely. The successful
self-optimization and decision is indispensable, especially for
application cases fully prove the feasibility of this
6G networks. Digital twin, as an innovative technology, has the
technology, such as aircraft manufacturing. However,
potential to realize this architecture because it can optimize and
there are rarely studies on application of digital twin in
validate policies through real-time and interactive mapping with
network. In order to solve the problems faced by
physical entities.
complex networks more effectively in the whole network
The concept of digital twin was first proposed by Dr. M. life cycle of planning, construction, maintenance and
Grieves[1], who defined it as a three-dimensional model operation, an innovative DTN architecture needs to be
including physical products, virtual products and the connection proposed.
between them. And in 2012, the United States Air Force
Research Laboratory and the National Aeronautics and Space x Data is the cornerstone of DTN construction. In the face
Administration proposed digital twin as a highly integrated of large network scale, data collection, storage and
multi-physical field, multi-scale, multi-probability simulation
Supported by National Key Research and Development Program of China
(2020YFB1806801, 2020YFB1806800).
* The corresponding author.

362
978-1-6654-3337-2/21/$31.00 ©2021 IEEE
Authorized licensed use limited to: VTU Consortium. Downloaded on March 24,2025 at 05:20:32 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. The architecture of Digital Twin Network

management are faced with great challenges. If the full- cleaning and processing of network data, so as to facilitate the
data collection method is adopted, huge storage space efficient and distributed storage of large-scale data. The Digital
and bandwidth resource is needed, especially for Twin Layer is the digital representation of the Physical Network,
complex scenarios that require real-time data and traffic which can be interactive mapped with the physical network in
from multi-source heterogeneous devices. Therefore, it real time to realize the double closed-loop control. It is
is extremely import to promote a lightweight and high composed of Physical entity modeling, Requirement modeling
precision data collection method. and Twin management&control center. The main purpose of
Physical entity modeling is to complete the digital modeling of
x A physical network usually consists of lots of physical Physical layer devices, topologies and virtual network elements.
entities, and the topology among those entities is very Requirement modeling is used to build a variety of scenario
complex. How to build the physical entities of the models such as network resource prediction, anomaly detection,
network in a digital way and extract the complex automatic operation through AI algorithms. Twin
relations of the network topology is a burning question management&control center has two functions, one is to drive
in DTN construction. the requirement model to simulate and verify on the model of
In order to address those three challenges, this paper designs physical entity, the other is to issue the feasible strategy from the
a novel architecture of DTN and introduces a Sketch algorithm Digital Twin Layer to the Physical Network.
to achieve the data collection. What’s more, a method of DTN
construction based on knowledge graph is demonstrated. III. DATA COLLECTION
Data is the basis of constructing digital twin network. The
The rest of the paper is organized as follows. Section Ċ fidelity of physical network restoration depends on whether the
introduces the novel architecture of DTN. In section ċ, we data is comprehensive and accurate. However, with the advent
present a Sketch algorithm to collect data. Section Č shows the of the Internet of Everything, the amount of data is growing
details of the DTN construction method based on knowledge explosively. If the full-data collection method is adopted, huge
graph. Section č concludes this paper. storage space and bandwidth resources may be insatiable,
especially for scenarios that require mapping real-time data such
II. DIGITAL TWIN NETWORK ARCHITECTURE as digital twin network construction. To solve the above-
As shown Fig. 1, we propose a novel network paradigm for mentioned problems, designing a high precision and lightweight
future network, called digital twin network (DTN), which network data acquisition method is extremely necessary. As is
includes Physical Network, Data Lake, Digital Twin Layer and shown in Fig. 2, sketch algorithm is proposed to solve this
Network Application Layer. Data Lake have three main puzzle. In addition, in the future network architecture such as
functions, namely collecting, storing and data pre-processing. It digital twin network, employing the distributed data center in the
collects and stores various configuration data, network operation physical network may be more in line with the development
status data and user business data of network entities through the trend. Therefore, how distributed collaborative algorithms can
southward interface, to provide network data for the be applied to network data collection is also worthy of exploring.
construction of DTN model. The main function of data pre- The current data collection algorithm mainly stores the
processing is to complete the extraction, conversion, loading, collected data through a hash array, and the main resource

363

Authorized licensed use limited to: VTU Consortium. Downloaded on March 24,2025 at 05:20:32 UTC from IEEE Xplore. Restrictions apply.
 Ontology Language)[16]. RDF can only define simple triples
(subject, predicate, and object), and RDF(S) provides the most
basic description metalanguage for classes and attributes. Both
RDF and RDF(S) can express some simple semantics, but their
expressive ability seems to be incompetent for complex
scenarios. Due to the complexity of the network, OWL is
employed for knowledge representation.
According to the characteristics of physical network, a six
elements structure of ontology is introduced to represent
physical network devices, which can be denoted as <Instance,
class, property, relationship, rule, axiom>. As shown in the
Tabel 1, the OWL representation details of Switch 2(S2) are
initially proposed.
Fig. 2. Sketch based collective data collection TABLE I. SWITCH2 KNOWLEDGE REPRESENTATION INFORMATION
BASED ON OWL
overhead of the algorithm is the storage space used by hash array.
In fact, collecting full global spatiotemporal data will face the <OWL: >
problem of information redundancy, and we can store the sketch <OWL: Instance >
count on demand instead of storing all the elements from hash <OWL: Name = “S2”>
</OWL: Instance >
array. Sketch algorithm is suitable for asymptotic statistical <OWL: Class>
distribution under sparse and dense conditions. Quantile sketch <OWL: subClassOf =“10gigabit Ethernet Switch”>
addresses this problem of the expensive computation and space <OWL: subClassOf = “Modular Switch”>
cost for a large volume of data by using a small data structure to <OWL: subClassOf = “Three Layer Switch”>
approximate the exact distribution of item value in a single pass </OWL: Class >
<OWL: Property>
over the items[12]. Frequency Sketch, such as Count-min sketch, <OWL: Sise = “442*476*442mm”>
is proposed to estimate the frequency of different values of <OWL: Exchange Mode= “store-and-forward”>
items[13]. Since distributed Machine Learning or Federated <OWL: Protection Level= “IP30-IP67”>
Learning algorithms trained by stochastic gradient descent </OWL: Property>
(SGD) need to communicate gradients, SketchML is proposed <OWL: Relationship >
<OWL: connectTo = “R&S1&S3”>
to compress the communicated gradients and accelerate training </OWL: Relationship >
speed[14]. In summary, it is very effective for traffic statistics, <OWL: rule>
state distribution, entropy estimation, etc. The sketch data <OWL: bandwidth Range = “19.84-86.4Tbps”>
structure can improve the execution efficiency and estimation <OWL: Packet Rate Range = “3240-26400Mbps”>
accuracy of the data flow algorithm, and reduce the calculation </OWL: rule>
and storage overhead. At the same time, Sketch can detect <OWL: axiom>
<OWL:security = “Command line hierarchical protection,
abnormal data and discover network faults timely. After using unauthorized users can not intrude”>
the sketch algorithm to collect data from different local data </OWL: axiom>
centers, it needs to be aggregated and stored in the data lake with </OWL: >
the help of the distributed collaborative algorithm.
B. Network Relation Extraction
IV. DTN CONSTRUCTION BASED ON KNOWLEDGE GRAPH In the face of the complex connection of network, it is
Knowledge graph is a technical method that uses graph always necessary to extract the relation with the help of expert
models to describe knowledge and model the relationship experience or artificial intelligence. Relation extraction methods
between everything in the world[15]. It is composed of nodes can be roughly divided into template-based extraction and
and edges. Nodes denote entities, and edges represent the supervised learning extraction. Since the template-based
relationship between entities. The complex relationship of the extraction method is not sufficient to deal with complex network
network is more suitable to be represented by Graph, while the relationships, a network relation extraction (NRE) algorithm
knowledge graph can be understood as Multi-relational Graph. composed of Long-Short Term Memory (LSTM) and
Describing complex digital twin network by knowledge graph Convolutional Neural Networks (CNN) is proposed to mine the
can not only represent the physical network attributes, but also complex relationships between topologies or network elements.
mine the network relations from different dimensions. Therefore, Fig. 3 describes the architecture of the neural network that we
a method of knowledge graph is introduced to construct digital use for relation extraction. The network takes an input sentence
twin network. and accomplish feature extraction. It primarily includes the
following three components: Word Representation, Feature
A. Network Knowledge Representation Extraction and Output. The algorithm does not need any
Knowledge representation can represent a traditional data complicated syntactic or semantic preprocessing, and the input
format into a data structure that can be calculated efficiently. of the model is a sentence with the marked nouns. The word
World Wide Web Consortium proposed three standard tokens are transformed into vectors by One-Hot Encoding. If the
languages of knowledge representation, namely RDF (Resource vector dimension is too large, Principal Component Analysis
Description Framework), RDFS (RDF Schema) and OWL (Web (PCA) will be used for dimensionality reduction. One-hot

364

Authorized licensed use limited to: VTU Consortium. Downloaded on March 24,2025 at 05:20:32 UTC from IEEE Xplore. Restrictions apply.
 Application Layer. The proposed sketch algorithm solves the
puzzle of data acquisition in Data Lake. The proposed DTN
construction method based on knowledge graph can represent
entity attribute and various topologies. The feasibility and
novelty of the DTN is proved in fault self-healing scenarios of
the future network.
REFERENCES
[1] Grieves M. Digital twin: Manufacturing excellence through virtual
factory replication[Online], available: https://www.3ds.com/fileadmin/-
PRODUCTS/SERVICES/DELMIA/PDF/Whitepaper/DELMIAAPRIS-
Digital-Twin-Whitepaper.pdf
[2] Glaessgen E H, Stargel D S. "The digital twin paradigm for future NASA
and U.S. air force vehicles", in Proceedings of the 53rd Structures,
Structural Dynamics, and Materials Conference: Special Session on the
Digital Twin, Hawaii pp. 1-14, Apr., 2012.
[3] F. Tao, H. Zhang, A. Liu and A. Y. C. Nee, "Digital Twin in Industry:
State-of-the-Art," in IEEE Transactions on Industrial Informatics, vol. 15,
no. 4, pp. 2405-2415, 2019.
[4] F. Tao, J. Cheng, Q. Qi, M. Zhang, H. Zhang, F. Sui. "Digital twin-driven
product design, manufacturing and service with big data." in International
Journal of Advanced Manufacturing Technology, vol. 94, no. 9, pp. 3563-
3576, 2018.
[5] K. Zhang, J. Cao, S. Maharjan and Y. Zhang, "Digital Twin Empowered
Content Caching in Social-Aware Vehicular Edge Networks," in IEEE
Transactions on Computational Social Systems, pp 1-13, 2021.
Fig. 3. NRE algorithm based relation extraction [6] W. Sun, H. Zhang, R. Wang and Y. Zhang, "Reducing Offloading Latency
for Digital Twin Edge Networks in 6G," in IEEE Transactions on
Vehicular Technology, vol. 69, no. 10, pp. 12240-12251, 2020.
vectors are feet into the NRE model, the temporal features are
[7] Q. Yu, J. Ren, Y. Fu, Y. Li and W. Zhang, "Cybertwin: An Origin of Next
extracted using LSTM model, and the spatial features are Generation Network Architecture," in IEEE Wireless Communications,
extracted using CNN model. Finally, the extracted feature vol. 26, no. 6, pp. 111-117, 2019.
vectors is fed into a softmax classifier to compute the confidence [8] R. Dong, C. She, W. Hardjawana, Y. Li and B. Vucetic, "Deep Learning
of each relation. The output of the classifier is a vector which for Hybrid 5G Services in Mobile Edge Computing Systems: Learn From
can represent the relationship between entities accurately. a Digital Twin," in IEEE Transactions on Wireless Communications, vol.
18, no. 10, pp. 4692-4707, 2019.
C. DTN based Fault Self-healing [9] H. Wang, Y. Wu, G. Min and W. Miao, "A Graph Neural Network-based
In this section, the application of using DTN to realize fault Digital Twin for Network Slicing Management," in IEEE Transactions on
Industrial Informatics, pp. 1-1, 2020.
self-healing in the future network is demonstrated. When the
[10] T. Sun, C. Zhou, X. Duan, Lu Lu, and etc., "Digital twin network (DTN):
network encounters faults, it is extremely risky to distribute the concepts, architecture, and key technologies", in Acta Automatica Sinica,
maintenance orders to the physical network directly. Therefore, vol. 47, no. 3, pp. 569-582, 2021.
digital twin network is employed to simulate and verify the [11] Q. Yu, J. Ren, H. Zhou and W. Zhang, "A Cybertwin based Network
feasibility of repair policy. Once the physical network has fault, Architecture for 6G", in Proceddings of 2020 2nd 6G Wireless Summit
the data lake collects the fault data from the physical network (6G SUMMIT), pp. 1-5, Finland, Mar., 2020.
and sends it to DTN. The Requirement Modeling Module of [12] M. Greenwald and S. Khanna, "Space-efficient online computation of
DTN trains the fault detection model using AI algorithm based quantile summaries", in ACM SIGMOD Record, Vol. 30, pp. 58-66, 2001.
on the fault data. According to fault type, the corresponding [13] G. Cormode and S. Muthukrishnan, "An improved data stream summary:
repair policy is generated. Then, the policy is simulated and the count-min sketch and its applications", in Journal of Algorithms vol.
55, no. 1, pp. 58-75, 2005.
verified based on the constructed knowledge graph. The
[14] J. Jiang, F. Fu, Y. Tong, and etc., "SketchML: Accelerating Distributed
feasibility of this policy can be examined through Deductive Machine Learning with Data Sketches" in Proceddings of the 2018
Reasoning or Inductive Reasoning. If feasible, the policy is International Conference, May, 2018.
deployed to the physical network for fault repair. If not, it [15] Qingjie Lv, Lingyu Xu, Jie Yu, Suixiang Shi, Yang Liu, "Research on
continues to be iteratively optimized until achieving network domain knowledge graph based on the large scale online knowledge
autonomy. fragment," 2014 IEEE Workshop on Advanced Research and Technology
in Industry Applications (WARTIA), pp. 312-315, Sep., 2014
V. CONCLUSION [16] D. Allemang, J. Hendler, “Semantic web for the working ontologist:
effective modeling in RDFS and OWL”, in Semantic Web for the
This paper demonstrates a novel network architecture for the Working Ontologist, Elsevier, vol. 14, no. 3, pp. 343-346, 2011.
future network called digital twin network, which consists of
Physical Network, Data Lake, Digital Twin Layer and Network

365

Authorized licensed use limited to: VTU Consortium. Downloaded on March 24,2025 at 05:20:32 UTC from IEEE Xplore. Restrictions apply.