Journal of Manufacturing Systems 62 (2022) 857–874

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

journal homepage: www.elsevier.com/locate/jmansys

Digital Twin-enabled Collaborative Data Management for Metal Additive

Manufacturing Systems
Chao Liu a, *, Léopold Le Roux a, Carolin Körner b, Olivier Tabaste c, Franck Lacan a,
Samuel Bigot a
a
School of Engineering, Cardiff University, Cardiff CF24 3AA, UK
b
Chair of Materials Science and Engineering for Metals, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91054, Germany
c
MSC Software part of Hexagon, Les Ulis 91978, France

A R T I C L E I N F O A B S T R A C T

Keywords: Metal Additive Manufacturing (AM) has been attracting a continuously increasing attention due to its great
Metal Additive Manufacturing advantages compared to traditional subtractive manufacturing in terms of higher design flexibility, shorter
Digital Twin development time, lower tooling cost, and fewer production wastes. However, the lack of process robustness,
data management
stability and repeatability caused by the unsolved complex relationships between material properties, product
data model
design, process parameters, process signatures, post AM processes and product quality has significantly impeded
machine learning
product lifecycle management its broad acceptance in the industry. To facilitate efficient implementation of advanced data analytics in metal
AM, which would support the development of intelligent process monitoring, control and optimisation, this
paper proposes a novel Digital Twin (DT)-enabled collaborative data management framework for metal AM
systems, where a Cloud DT communicates with distributed Edge DTs in different product lifecycle stages. A metal
AM product data model that contains a comprehensive list of specific product lifecycle data is developed to
support the collaborative data management. The feasibility and advantages of the proposed framework are
validated through the practical implementation in a distributed metal AM system developed in the project
MANUELA. A representative application scenario of cloud-based and deep learning-enabled metal AM layer
defect analysis is also presented. The proposed DT-enabled collaborative data management has shown great
potential in enhancing fundamental understanding of metal AM processes, developing simulation and prediction
models, reducing development times and costs, and improving product quality and production efficiency.

1. Introduction exist various types of metal AM technologies. Based on the distinctions

of AM processes, the international standard ISO/ASTM 52900:2015 [5]
Additive Manufacturing (AM) refers to the technologies used to identifies seven categories of AM technologies, including 1) powder bed
manufacture objects from 3D model data, in which materials are accu­ fusion, 2) directed energy deposition, 3) material jetting, 4) sheet
mulated layer by layer via specific techniques such as extrusion, sin­ lamination, 5) material extrusion, 6) binder jetting, and 7) vat photo­
tering, melting, photopolymerization, jetting, lamination, and polymerization. With regard to metal AM, the first four technologies can
deposition [1,2]. AM allows the manufacturing of products with com­ be applied to manufacture pure metal parts; the other three can be used
plex geometries, heterogeneous materials, and customisable material to produce metallic composite parts. While each technology has its
properties [3]. Compared to traditional subtractive manufacturing, AM distinct pros and cons, this paper focuses on the most mature and widely
enables higher design flexibility, shorter development time, lower used metal AM technology, i.e. metal Powder Bed Fusion (PBF). PBF
tooling cost, and fewer production wastes. The long-term impact of AM uses laser or electron beam to selectively fuse areas of a layer of pow­
is expected in highly customized manufacturing, where AM can be more ders, and then moves the powder bed downwards, adding another layer
cost-effective than traditional manufacturing [4]. of powders and repeating the process until the part has built up.
In recent years, metal AM has been attracting a continuously Commonly used metal PBF techniques include Direct Metal Laser Sin­
increasing attention in both academia and industry. Currently, there tering (DMLS), Selective Laser Melting (SLM) and Electron Beam

* Corresponding author.
E-mail address: [email protected] (C. Liu).

https://doi.org/10.1016/j.jmsy.2020.05.010
Received 12 February 2020; Received in revised form 8 April 2020; Accepted 18 May 2020
Available online 27 May 2020
0278-6125/© 2020 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Melting (EBM). Compared to traditional subtractive manufacturing AM and DT-supported data management in manufacturing. The research
which has been developed and significantly improved for over two gaps are then identified and discussed.
centuries [6], metal AM, which was first introduced in the 1990s [7], is
relatively new. Though metal AM has shown great potential and ad­ 2.1. Data management for metal AM
vantages, many problems still exist that seriously limit its broad
acceptance. The highly variable product quality (geometry, mechanical In the context of Industry 4.0 where information and communication
properties, physical properties, etc.) and the lack of process robustness, technology (ICT) plays a vital role, data management becomes a critical
stability and repeatability have been recognised as the main barriers for issue for any type of manufacturing system. In metal AM, it is estimated
the industrial breakthrough of metal AM systems [8]. Identifying the that PBF systems involve more than 130 variables [12], while over 50
complex relationships between material properties, product design, different process parameters in metal AM processes have influences on
process parameters, process signatures, post AM processes and product the final product quality [4,13]. The complex relationships between
quality remains a major challenge [4]. product design, process parameters, process signatures, post AM pro­
In order to address the aforementioned issues and challenges in cesses, and product quality present an urgent need for efficient data
metal AM, a prerequisite is to collect and analyse the metal AM data management in metal AM systems.
from all product lifecycle stages. A typical metal AM product lifecycle Since metal AM data comprises data from various sources in different
involves several different stages such as product design, process plan­ product lifecycle stages, collaborative data management systems are
ning, manufacturing, post processing and quality measurement. Each required for metal AM. Müller et al. [14] identified three main re­
stage generates huge amounts and various types of metal AM data that quirements for a metal AM product lifecycle data management system:
influences the final product quality. Since the development of standards 1) easy accessibility of the entire product knowledge, 2) traceability of
and certification for metal AM is still in the early stage, there is an urgent individual product information, and 3) automation of information flow.
need to develop efficient and collaborative data management systems Aiming to capture, store and manage data through the entire metal AM
for metal AM. lifecycle and value chain, the National Institute of Standards and
Recently, the advancements of Digital Twin (DT) have been exten­ Technology (NIST) [15,16] developed a collaborative AM data man­
sively studied in the domain of manufacturing. The implementation of agement system using the NoSQL (Not Only Structured Query Language)
DT technology in manufacturing systems has shown great potential in database technology. The cloud-based database is structured by an AM
enabling advanced manufacturing data management. In this context, data schema which enables meaningful data curation and data retrieval.
this paper proposes a novel DT-enabled collaborative data management An ontology-based web graphical user interface is developed for the data
framework for metal AM systems, which comprises a Cloud DT and management system, allowing users to perform data curating, exploring,
distributed Edge DTs in different product lifecycle stages. The proposed and downloading. The data management system also provides a
framework allows metal AM data from all product lifecycle stages to be Representational State Transfer (REST) interface for the integration with
collaboratively managed in the cloud, and hence enabling various types other applications. Uhlmann et al. [17] developed a data management
of advanced data analytics and product quality control to be realised. As system that can store data from different sources, such as the energy
a critical component for both data management system and DT, a metal system (regarding power consumption), internal machine tool sensors
AM product data model is developed by identifying and categorising the (platform temperature, process chamber pressure, etc.), external sensors
critical metal AM product lifecycle data that has influence on the final (microphone) and machine controller (process data). Their experi­
product quality. To demonstrate the feasibility and advantages of the mental results have proven that the developed data management system
proposed approach, an early implementation of the DT-enabled collab­ enables the analysis of the links between SLM machine status and AM
orative data management in a real metal AM system is introduced. A process conditions, though it works only in the local environment and
representative application scenario enabled by the proposed data man­ there is no data model to organise the data from different sources. Müller
agement system, i.e. cloud-based and deep learning-enabled metal AM et al. [14] developed a demonstrator tool as a metal AM product life­
layer defect analysis, is also developed and demonstrated. Results have cycle data management system, where the design, simulation,
shown that the proposed DT-enabled collaborative data management manufacturing, post processing and measurement data of each indi­
plays an important role in enhancing fundamental understanding of vidual product are stored. However, their demonstrator provides only
metal AM processes, developing simulation and prediction models, basic database and data visualisation functions without data analytics
reducing development times and costs, and improving product quality functions. Based on the material data repository named MAPTIS, NASA
and production efficiency. [18] developed a metal AM database that contains various types of metal
The remaining of this paper is organised as follows. Section 2 reviews AM data including build parameters, material conditioning, test envi­
the state-of-the-art research on data management for metal AM and DT- ronment, tensile results, fracture toughness and fatigue. They claimed
supported data management and identifies the research gaps. Section 3 that the developed database serves as the foundation of the metal AM
proposes a conceptual framework of the DT-enabled collaborative data data management, which strongly complements the development of
management for metal AM systems. Section 4 presents the developed standards and protocols for metal AM. Mies et al. [19] provided an
metal AM product data model. Section 5 demonstrates the early overview of the current status of AM informatics, in which the key
implementation of the DT-enabled collaborative data management in a technical requirements, the available AM informatics tools and the
real metal AM system as well as a representative application scenario. existing applicable solutions have been summarised. They concluded
Section 6 concludes the paper and discusses the future work. that while metal AM standards and certification are in the early devel­
opment stage, advanced data management for metal AM plays an
2. Literature Review important role in reducing development times and costs and improving
product quality and production efficiency.
Current research on metal AM focuses mainly on the material sci­ Data model is considered as the backbone of a data management
ence, processing science and process monitoring technologies [9–11]. system, especially in the case of metal AM where various types of data
Data management in the domain of metal AM has not been extensively from different lifecycle stages are involved. Utilising a Product Lifecycle
studied due to the relatively short development history of metal AM. On Management (PLM) data modelling method named PPR (product, pro­
the other hand, Digital Twin technology has just gained attention in cess and resource), Lu et al. [20] proposed a conceptual metal AM in­
manufacturing field in the last several years. Since this paper proposes a tegrated data model where the AM data are modelled as entities under
novel DT-enabled data management concept for metal AM, this section three categories, i.e. product, process and resource, and the fundamental
reviews the related work from two aspects: data management for metal relationships among the entities are defined. Based on the conceptual

858
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

data model, Lu et al. [16] further developed an AM database schema managed, visualised and analysed. Cheng et al. [42] proposed a digital
using Unified Modelling Language (UML), which was used as the twin enhanced Industrial Internet (DT-II) reference framework towards
foundation of the metal AM data management system introduced pre­ smart manufacturing. Taking the production of a steam turbine as an
viously. To address the data interoperability issue in metal AM, Feng example, the authors claim that DT enables the interaction and
et al. [21,22] developed an activity model for structuring convergence between physical and virtual product, and hence shortens
process-related PBF data, by decomposing the entire PBF process production preparation time and production period, reduces production
(including design, process planning, fabrication, inspection and quality cost, and improves production efficiency and processing quality. Wang
control) into specific activities and identifying the inputs, outputs, et al. [43] developed a shop-floor DT that reflects real-time production
mechanisms and controls of each activity. Later, Kim et al. [23] devel­ status in the shop floor. Advanced data analytics functions were inte­
oped AM data models in the form of data packages (in XML format) upon grated in the DT to analyse the key production performance indicators
the activity model. Three case studies have been conducted to demon­ and predict the remaining processing time of work-in-processes.
strate how the data models improved data manageability, traceability In the context of Smart Manufacturing and Industry 4.0, data man­
and accountability in the metal AM processes. Based on the agement is closely correlated with PLM and PHM [44,45]. Applications
object-oriented modelling method, Bonnard et al. [24] proposed a new of DT-supported PLM and PHM have also been extensively developed.
AM digital chain model named Hierarchical Object-Oriented Model Kaewunruen et al. [46] developed a DT-enabled lifecycle management
(HOOM), which comprises seven levels corresponding to seven stages of system for railway turnout systems using the Building Information
the AM process from product design to post-production and validation. Modelling (BIM) technology. Data from different lifecycle stages of the
The model was implemented in a STEP-NC platform and tested through railway turnout system was integrated into a 6D model, such that
the manufacture of two parts. Their experimental results proved that the various lifecycle management functions (visualisation of product
proposed AM model improves transparency, interoperability and scal­ design/material/ component/project/schedule information, cost pre­
ability of the AM project data. Qin et al. [2] provided a comprehensive diction and carbon foot-print estimation) were realised in a big data
review of the existing representations of AM data. The data modelling sharing platform. Wang and Wang [31] developed a DT-based system for
strategies are categorised into six methods: STEP standards, coding the waste electrical and electronic equipment (WEEE) recovery to sup­
system method, digital thread method, integrated data schema method, port the manufacturing/ remanufacturing operations throughout the
unified storage file format and relational database method. Comparisons product’s lifecycle, from design to recovery. International
among the different strategies were conducted in terms of the coverage, standard-complaint data models were developed to support the
simplicity, interoperability, extensibility, inspectibility, accessibility DT-based PLM with high data interoperability. Aiming to improve the
and application. process planning for optimized machining solutions, Botkina et al. [38]
developed a digital twin of a cutting tool that could represent the
2.2. Digital Twin-supported data management in manufacturing properties of the cutting tool and perform precise process simulation,
control and analysis. The data format and structure of the digital twin
Recently, the Digital Twin concept has attracted great research in­ are based on the international standard ISO 13399 (Cutting tool data
terests in the domain of manufacturing. DT is considered as a key representation and exchange). Zheng and Sivabalan [47] proposed a
enabling technology for the envisioned Cyber-Physical Production Sys­ novel tri-model-based approach for product-level DT development. A DT
tems, Smart Factory and smart manufacturing systems in the context of of a 3D printer that works concurrently to simulate real-world physical
Industry 4.0 [25–29]. The potential of utilising DT to improve product behavior and characteristics of the digital model was developed.
design [30], manufacturing processes [31], process monitoring [32,33],
Prognostics and Health Management (PHM) [34,35], production man­ 2.3. Research gaps
agement [36] and PLM [37,38] has been extensively discussed and
studied. The literature review has revealed several research gaps. First, few
It has been identified that currently, some problems in studies have been conducted on data management for metal AM sys­
manufacturing systems have impeded the development of efficient data tems. There is a lack of collaborative data management systems for
management systems, including: 1) information islands between metal AM. Second, the literature review shows that data model is critical
different phases of product lifecycle caused by various data types and for both data management system and DT. However, there is a lack of a
tasks of different product lifecycle phases; 2) data waste and sharing metal AM product data model which contains a comprehensive list of
problems due to duplicate data in different product lifecycle phases; and specific data that has influence on the product quality. Third, the data
3) the absence of interaction and iteration between advanced data an­ communication between the field-level data sources and the data man­
alytics and various activities in the entire product lifecycle [26]. agement system has been rarely studied in the domain of metal AM.
To overcome these problems, recent advancements in DT technology Furthermore, though there has been a considerable amount of literature
have shown great potential and advantages in the form of DT-supported on DT in manufacturing field, in the domain metal AM, very few studies
product/manufacturing data management. In a comprehensive review (e.g. [48,49]) have applied the DT concept, and they all focus on the
on the development methods for DT, the enabling technologies and tools metal AM processes instead of data management.
for DT data management have been investigated in six categories, This paper attempts to bridge these research gaps by proposing a
including data collection, data transmission, data storage, data pro­ novel DT-enabled collaborative data management framework for Metal
cessing, data fusion, and data visualisation [39]. AM. A metal AM product data model is developed by identifying and
Applications of DT-supported data management for manufacturing categorising the critical metal AM data in different product lifecycle
systems have been widely reported. Lu and Xu [40] proposed a stages. Early implementation of the proposed framework in a real metal
DT-driven approach to enabling cloud-based manufacturing equipment AM system developed in the project MANUELA is conducted. Cloud-
and big data analytics. A DT for a roll forming machine, which mirrors based and deep learning-enabled metal AM layer defect analysis is
its near real-time status in the shop floor, was developed in the cloud to also demonstrated as a representative application scenario.
support various types of cloud-based data analytics such as machine
status monitoring and production reporting. Based on data modelling 3. Conceptual framework of the Digital Twin-enabled
methods provided by the open communication standards MTConnect collaborative data management for metal AM systems
and OPC UA, Liu et al. [41] developed DTs for CNC machine tools. The
DTs enable an interoperable data management environment where data Broadly, DT refers to an integrated multi-physics, multi-scale,
from different types of CNC controllers and sensors can be efficiently probabilistic simulation of a complex product and uses the best available

859
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 1. Conceptual framework of DT-enabled collaborative data management for metal AM.

physical models, sensor updates, etc., to mirror the life of its corre­ cloud computing [51], edge computing [52] and fog computing [53] in
sponding twin [50]. Based on different types of physical objects and manufacturing systems, we propose the concepts of Cloud Digital Twin
their functional requirements, the DT of a specific object may be defined and Edge Digital Twin that communicate with each other to enable a
in different ways. For example, Mukherjee and DebRoy [48] defined the collaborative data management system.
DT of 3D printing hardware as an integration of a mechanistic model, a In general, the Edge DTs focus on specific computing tasks of
sensing and control model, a statistical model, big data and machine different product lifecycle stages locally and transfer the processed data
learning that can potentially improve part quality and shorten time to the Cloud DT; while the Cloud DT stores all the product lifecycle data
between the design and production; while Knapp et al. [49] claimed that in the cloud and provides data access as well as advanced data analytics
a DT of the AM process should provide accurate predictions of the spatial to different users and applications. The definitions and main functions of
and temporal variations of metallurgical parameters that affect the Edge DT and Cloud DT are explained as follows:
structure and properties of components.
In this work, focusing specifically on data management for metal AM • Edge DT: An Edge DT is a DT of a specific product lifecycle stage
systems, we propose a novel DT-enabled collaborative data management which resides in a distributed shop floor and focuses only on the
system that comprises a Cloud DT and distributed Edge DTs in different functions of that lifecycle stage. The Edge DTs have their own
product lifecycle stages. This section introduces the concepts of Cloud computing power (or local intelligence) to support various types of
DT and Edge DT and proposes a conceptual framework of DT-enabled local real-time data processing tasks such as machine control, process
data management for Metal AM systems. monitoring, in-process optimisation, simulation and prediction. The
Edge DTs store the raw data of a product lifecycle stage in local
databases and transfer the processed data which has influence on the
3.1. Cloud Digital Twin and Edge Digital Twin product quality to the Cloud DT through the Internet, thus enabling
efficient data communication by reducing the data traffic between
The collaborative data management system needs to manage data shop floors and the cloud.
generated in all product lifecycle stages in metal AM. Different stages • Cloud DT: The Cloud DT collects all the product lifecycle data from
contain different types of devices and software tools and perform the Edge DTs and records it in the cloud database which is structured
different tasks. Each stage may generate a huge amount of raw data (e.g. by a metal AM product data model. It resides in the cloud and
sensor signals) that needs to be processed locally. Hence, it is inefficient communicates with the Edge DTs as well as different users to allow
or even not possible to transfer all the field-level data to the cloud to collaborative data management for metal AM systems. Advanced
build a DT of the metal AM product. Inspired by the implementation of

860
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 2. Structure of the metal AM product data model.

data analytics are integrated in the Cloud DT to support analysing the the manufacturing stage. Various types of post AM processes are
relationships among the various historical product lifecycle data. monitored, controlled and optimised with local computing power. Pro­
Application interfaces are also embedded in the Cloud DT such that cess signatures and results are processes and transferred to the Cloud DT,
other applications can take advantage of the product lifecycle data. while the raw data is stored in local database.
In quality measurement stage, the Edge DT retrieves the measure­
ment tasks and product design files from the Cloud DT and supports the
3.2. Conceptual framework machine operators to measure the product qualities, analyse the
measured results and generate measurement reports. The processed
The main objective of the Cloud DT and Edge DT is to support product quality data is then transferred to the Cloud DT.
collaborative data management for metal AM. To provide guidance for In the collaborative data management platform, the Cloud DT col­
practical development and implementation, we propose a conceptual lects all the product lifecycle data from the Edge DTs and stores it in the
framework of DT-enabled collaborative data management for metal AM cloud database structured by a metal AM product data model. Different
systems, as shown in Fig. 1. The conceptual framework comprises six users (data analysts, project managers, customers, etc.) can access the
modules, i.e. a cloud-based collaborative data management platform Cloud DT through the Internet to collectively manage the metal AM
and five product lifecycle stages, including product design, process product data. Various types of functions such as project management,
planning, manufacturing, post processing and quality measurement. production planning, progress monitoring, advanced data analytics can
Each product lifecycle stage has an Edge DT that performs specific be developed to improve the product quality and the production
computing tasks to support its users and communicates with the Cloud efficiency.
DT to establish the collaborative data management for metal AM It is worth mentioning that this framework focuses on the metal AM
systems. product quality and its related production processes, though a full
In product design stage, the Edge DT supports the product designers product lifecycle may also include product usage stage and end-of-life
to design the metal AM product and perform CAE simulations with CAD treatment stage [54]. Furthermore, the proposed conceptual frame­
and CAE software. The design files and simulation results are transferred work can also be used as a generic reference framework for developing
to the Cloud DT to be recorded and shared. The Edge DT can also retrieve collaborative data management for other types of manufacturing
other data such as product quality data from the Cloud DT to aid the
product designers in design optimisation.
Table 1
In process planning stage, the Edge DT supports the process planners
Data items of design parameters in the data model
to generate support structures and conduct process planning (model
Sub-category Data item Description Data
slicing, selection of position, orientation, scan strategy, process param­
format
eters, etc.), process simulation and process optimisation. The process
plans, process parameters and simulation results are transferred to the CAD file of the designed
CAD file CAD file
product
Cloud DT. Part Material Material of the product string
In manufacturing stage, the Edge DT focuses on the AM building Part Volume (cm3) Volume of the designed part number
Design
processes of a specific metal AM product and performs various types of Minimum Feature Minimum feature size of the
Features number
data processing tasks such as process monitoring and control, in-process Size (mm) product in x, y, z axis
Surface Roughness Required surface roughness number
optimisation, machine learning model training and real-time decision-
Required dimensional and
making support. The monitored process signatures (machine logs, sensor Tolerances
geometric tolerances
CAD file
signals, images, etc.) are processed locally and the results are transferred Supports Designed support structure CAD file
to the Cloud DT. The raw data which is usually large in size is stored in Build
Build Position
Designed build position of the
number
the local database for further analysis. Preparation part on the powder bed
Build Orientation Designed build orientation CAD file
In post processing stage, the Edge DT performs similar functions as in

861
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Table 2
Data items of process parameters in the data model
Sub-category Data item Description Data Control-
Format ability

Laser/Beam Type The type of the laser/beam string Predefined

Laser/Beam Mode Continuous wave or pulsed string Predefined
Spot Diameter (μm) Diameter of the laser/beam spot number Predefined
Wavelength (μm) Laser/beam wavelength number Predefined
Beam Quality Factor The degree of variation of a beam from an ideal Gaussian beam number Predefined
Pulse Frequency (Hz) Pulses per unit time (in pulsed mode) number Predefined
Laser/Beam and Scanning Parameters Pulse Width (μm) Length of a laser pulse (in pulsed mode) number Predefined
Peak Power (W) Maximum power in a laser pulse (in pulsed mode) number Predefined
Laser/Beam Power (W) Power of the laser/beam number Controllable
Scan Speed (m/s) Velocity at which the laser/beam moves across the build surface number Controllable
Hatch Spacing (μm) Distance between the centres of two adjacent scan paths number Controllable
Pattern in which the laser/beam is scanned across the build
Scan Pattern CAM file Controllable
surface
Powder Material Material of the powder string Predefined
Material Thermal Conductivity (W⋅ m-
1 Measure of material’s ability to conduct heat number Predefined
⋅ K-1)
Material Specific Heat Capacity (J⋅ K-1⋅ Measure of energy required to raise the temperature of the
number Predefined
kg-1) material
Material Melting Temperature (◦ C) Temperature at which the material melts number Predefined
Chemical Composition Measured powder chemical composition file (txt) Predefined
Impurity Element Characterization of material impurity string Predefined
Max Allowed Impurity Concentration
The controlled impurity concentration number Predefined
(wt%)
Phase Composition X-ray diffractogram image Predefined
Powder Material Properties Particle Morphology SEM images of powder image Predefined
Measure of how closely the shape of an object resembles that of a
Particle Sphericity number Predefined
perfect sphere
Particle Aspect Ratio Measure of roundness of the particle number Predefined
Particle Surface Roughness Measure of surface roughness of the particle number Predefined
Particle Presence of Defect The presence of defects in particles (Y/N) string Predefined
Particle Inhomogeneity The homogeneity of particles (Y/N) string Predefined
Particle Size Distribution Particle size distribution plot image Predefined
Powder Apparent Density (g/cm3) Apparent density of the powder number Predefined
Powder Tap Density (g/cm3) Tapped density of the powder number Predefined
Powder Cohesive Force Index Plot of flowability as a function of rotating drum speed image Predefined
Powder Hausner Ratio Hausner ratio from packing dynamic measurements number Predefined
Density (g/cm3) Measure of packing density of powder particles number Predefined
Thermal Conductivity (W⋅ m-1⋅ K-1) Measure of powder bed’s ability to conduct heat number Predefined
Measure of energy required to raise the temperature of the
Heat Capacity (J/K) number Predefined
powder bed
-1 -1
Absorptivity (L⋅ mol ⋅ cm ) Measure of absorbed laser energy number Predefined
Powder Bed Properties and Recoating Emissivity Ratio of energy radiated to that of black body number Predefined
Parameters Recoater Type Type and mechanism of the recoating system string Predefined
Recoater Speed (mm/s) Velocity at which the recoater moves during recoating number Controllable
Dosing per Layer (%) Dosing of powders during recoating number Controllable
Layer Thickness (μm) Height of a single powder layer number Controllable
Powder Bed Preheating Temperature
Preheating (bulk) temperature of the powder bed number Controllable
(◦ C)
Type of Shield Gas Argon, Nitrogen, Helium, etc. string Predefined
Shield Gas Molecular Weight (g/mol) Molecular weight of shield gas number Predefined
Shield Gas Viscosity (Pa⋅s) Viscosity of shield gas number Predefined
Shield Gas Thermal Conductivity (W⋅
Thermal conductivity of shield gas number Predefined
m-1⋅ K-1)
Shield Gas Heat Capacity (J/K) Heat capacity of shield gas number Predefined
Convective cooling of just melted part by gas flowing over the
Convective Heat Transfer Coefficient number Predefined
Build Environment parameters surface
Surface Free Energy (mJ/m2) Surface free energy between melted powders and shield gas number Predefined
Build Plate Material Material of the build plate string Predefined
Build Plate Thickness Thickness of the build plate number Predefined
Oxygen Level (%) Percentage of Oxygen in the build chamber number Controllable
Pressure (kPa) Pressure in the build chamber number Controllable
Gas Flow Velocity (m3/s) Gas flow velocity in the build chamber number Controllable
Ambient Temperature (◦ C) Ambient temperature in the build chamber number Controllable

862
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Table 3
Data items of process signatures in the data model
Sub- Data item Description Sample rate Data Format
category

signal file/
Melt Pool Temperature Temperature signatures of melt pool (maximum temperature, gradient temperature, etc.) > 10 kHz
image
Melt Pool
signal file/
Melt Pool Geometry Geometry signatures of melt pool (width, length, depth, shape, area, intensity, etc.) > 10 kHz
image
Plume Signature Signatures of plume (plume consists of metal vapor and plasma) 50 Hz to 5 kHz image
Signatures of spatter (spatter comes from vapor jets and bubbles, generated by the ablation
By-products Spatter Signature 50 Hz to 5 kHz image
pressure to the vapor on the melt pool surface)
Acoustic Emission Acoustic emission in the build chamber > 100 kHz signal file
Track Width Width of a single track ex-situ image
Track Continuity Continuity of a single track ex-situ image
Track Depth Depth of a single track ex-situ image
Single
Track Overlap Overlap between two adjacent tracks ex-situ image
Track
Track Microcracks Microcracks in a single track > 50 kHz image
Track Consolidation Consolidation characteristics of a single track, such as continuous, discontinuous, weak, ball-
> 10 kHz image
Characteristics shaped, very little consolidations.
Melt Pool Intensity Variation Melt pool intensity variation across a single layer > 10 kHz image
per layer to 70 signal file/
Layer Surface Topography Topography of a single layer (layer geometry, height, waviness, roughness etc.)
Single kHz image
Layer Layer/Powder Bed Surface per layer to > 10 signal file/
Temperatures of layer surface
Temperature kHz image
Layer Defects Layerwise defects (porosity, uneven, under-melt, over-melt, etc.) per layer image
Powder Bed Surface
Surface topography of the powder bed per layer image
Topography
Powder Bed
Various types of powder bed anomalies (recoater hopping, recoater streaking, debris, super-
Powder Bed Anomalies per layer image
elevation, part failure, incomplete spreading, etc.)

systems. The data items of design parameters, along with their description and
data format, are listed in Table 1.
4. Metal AM product data model
4.2. Process parameters
Data model is a critical element for both DTs and data management
systems. The proposed DT-enabled collaborative data management Process parameters refer to the ‘input’ parameters for the
system requires a data model as the underlying database structure in the manufacturing processes in metal AM. They determine the rate of laser/
Cloud DT which can organise the data acquired from all the Edge DTs of electron beam energy delivered to the powder bed surface and how that
different product lifecycle stages in a clear and logical manner, such that energy interacts with the powders. Based on different components
advanced data analytics can be efficiently applied to analyse the com­ involved in the manufacturing process, the critical data of the process
plex relationships among the product lifecycle data. parameters is categorised into four sub-categories: 1) laser/beam and
In this work, we propose a metal AM product data model by iden­ scanning parameters, 2) powder material properties, 3) powder bed
tifying and categorising the critical metal AM product lifecycle data that properties and recoating parameters, and 4) build environment
has influence on the final product quality, based on a review of existing parameters.
literature as well as experts’ knowledge. To ensure the data model can be The controllability of process parameters is critical for metal AM
easily implemented in the data management system, a product-centric process monitoring, control and optimization. Process parameters can
and object-oriented modelling strategy is utilised. be characterised as either predefined or controllable. Predefined pa­
The structure of the proposed data model is shown in Fig. 2. The data rameters are set before each build and cannot be modified during
model is developed as a tree structure. The highest-level root element printing; while controllable parameters are possible to be continuously
represents a metal AM product which contains five categories of data tuned during printing. It is noted that different machine manufacturers
(design parameters, process parameters, process signatures, post pro­ may provide users with different levels of access rights to control the
cessing parameters and product quality), divided based on the five process parameters during the build. Based on some previous summaries
product lifecycle stages mentioned in the conceptual framework (Fig. 1). [55,56], Table 2 lists the data items of process parameters in the data
Each category contains several sub-categories representing different model, including their description, controllability and data format.
aspects of that category. Finally, each sub-category contains a compre­
hensive set of specific critical metal AM data. The details of the critical 4.3. Process signatures
data in each category are listed and explained in the following
subsections. Process signatures refer to the dynamic characteristics of the powder
heating, melting and solidification processes occurred during the build.
4.1. Design parameters Unlike process parameters that can be directly set or controlled, process
signatures reflect the dynamic processing results that can only be
Design parameters refer to the product design parameters that are controlled indirectly by modifying the process parameters. Based on
defined in the product design stage. Design parameters can only be different scopes of the signatures, the critical data of the process sig­
modified at the design stage before manufacturing processes. The crit­ natures is categorised into five sub-categories: 1) melt pool, 2) by-
ical data of the design parameters include two sub-categories: 1) design products, 3) single track, 4) single layer, and 5) powder bed.
features, and 2) build preparation. Measurement of process signatures is critical for process and product
Design parameters have a significant influence on the product quality improvement. The sample rate required for measuring a process
quality, especially in metal AM processes where inappropriate designs signature directly determines the feasibility, cost and efficiency of the
can directly lead to build failures and damage to machine components. metal AM process monitoring and optimisation. Based on a literature

863
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Table 4 Table 5
Data items of post processing parameters in the data model Data items of product quality in the data model
Sub-category Data item Description Data Sub-category Data item Common measurement Data Format
Format devices/methods

Powder Vibration Machine vibration settings string CT dimensional

measurement
Removal Setting Dimensional metrology; coordinate
reports (PDF,
Rotational Machine rotational settings file (txt) Accuracy measuring machine
Geometry and CSV, XLS)
Setting (CMM); 3D laser scanner
Dimension
Time (min) Time of the removal process number CT dimensional measurement
Geometric
Heat Treatment Sequence of heat treatment steps metrology; CMM; 3D reports (PDF,
file (txt) Accuracy
Sequence for a single component laser scanner CSV, XLS)
The heat treatment technique Surface Contact/optical surface number; profile
Technique used (e.g. annealing, quenching, string Roughness (μm) profilometers graphs
case hardening) Surface Waviness Contact/optical surface number; profile
Media (nitrogen, argon etc.) and (μm) profilometers graphs
Atmosphere string
pressure Surface Surface
Contact/optical surface number; profile
Heating Rate (K/ Quality Deformation
Heat Heating rate number profilometers graphs
min) (μm)
Treatment
Final X-ray photoelectron
Surface Chemical report (image,
Temperature Final temperature number spectroscopy; scanning
composition table)
(◦ C) electron microscopy
Holding Time Holding time at final temperature Archimedes drainage
number
(h) in hours method; Gas
Cooling Rate (K/ Part Density (g/ pycnometry;
Cooling rate number number
min) cm3) Microscopy of cross-
Cooling Medium Air, Nitrogen, Argon, Water, etc. string sections; X-ray scanning
Milling, drilling, fly cutting, of cross-sections
Machining
grinding, turning, shaping, string Archimedes method; gas
Operation
slotting, planning etc. pycnometry; CT scan images;
Part Porosity
Process plan of the machining microscopic analysis of 3D model (STL)
Physical
operation, including the cross-sections; X-ray CT
Properties
machining strategy, tool path Diffraction based
Post Process Plan plan, G code, and all process CAM file Residual Stress methods; Mechanical
number/report
Machining parameters such as cutting speed (MPa) strain relaxation-based
(m/min), feed rate (mm/min), methods
depth of cut (mm), etc. Cracks and Camera; Visual
image
Specification of cutting tool Delamination observation
Tool Geometry CAD file
geometry. SEM, transmission
Part
Specification of cutting tool electron microscopy image
Tool Material string Microstructures
material. (TEM)
For example, painting, etching, Yield Strength Tensile tests based on
Process Name string number
plasma treatment etc. (N/m2) standards
Non-standard
A description of the non-standard Tensile Strength Tensile tests based on
Processes Process number
processes and process parameters file (txt) (N/m2) standards
Specification
used. Tensile tests based on
Elongation (%) number
standards
Mechanical Fatigue tests based on
Fatigue report
Properties standards
review on process monitoring of metal AM, the required sample rates of Rockwell/Brinell/
the process signatures are summarised. Table 3 lists the data items of Hardness (kgf/ Vickers/Knoop hardness
number
mm2) testing; Instrumented
process signatures in the data model, including their description, Indentation Testing
required sample rate and data format. Toughness (J/ Fracture Toughness
number
m3 ) Testing

4.4. Post processing parameters

Post processing parameters refer to the designed post processes and for each data item are summarised from literature. Table 5 lists the data
their related process parameters. Post processing is an optional process items of product quality in the data model, including their common
depending on specific requirements. There exist various types of post measurement devices/methods and data format.
processing technologies that can be generally divided into four sub-
categories: 1) powder removal, 2) heat treatment, 3) post machining, 4.6. Summary
and 4) non-standard processes. Table 4 lists the data items of post pro­
cessing parameters in the data model, including their description and The features and advantages of the proposed metal AM product data
data format. model are summarised as follows:

4.5. Product quality • The data model contains a comprehensive set of specific metal AM
product data and divides them into five categories corresponding to
Product quality refers to the final quality of the metal AM product. It five product lifecycle stages, thus allowing collaborative data man­
is the joint results of all the other critical data introduced previously. agement and advanced data analytics to be realised efficiently.
Product quality can be categorised into four sub-categories: 1) geometry • The data model is designed to be product-centric and object-
and dimension, 2) surface quality, 3) physical properties, and 4) me­ oriented. Each specific metal AM product can be generated as a
chanical properties. Accurate and efficient measurement of the product new instance based on the data model, which contains only the
quality is critical for product quality control, design optimisation and product-specific data while omitting the data items that are not of
process optimization. Commonly used measurement devices/methods interest or not available.

864
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 3. Overall system architecture of DT-enabled collaborative data management in MANUELA.

Fig. 4. Main functions of Dashboard and Dashboard Users.

• The data model allows data of different formats to be stored, stacks of images, a link to another database could be stored under the
including number, string, CAD file, CAM file, image, signal file, re­ data item.
ports, and so on. Database developed based on this data model needs • The data model can be easily extended by adding additional data
to be able to store data in different formats. For large data sets such as items or sub-categories into their related categories.

865
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 5. Manual data input using the Dashboard.

• Implementation of the data model in databases or data management 5.1. System architecture and early implementations
systems is flexible due to its simple hierarchical structure.
The conceptual framework proposed in Section 3.2 describes the
5. Early implementation of DT-enabled collaborative data general functions and high-level data communication structures. To
management in MANUELA metal AM system implement the DT-enabled collaborative data management in a real
metal AM system, specific data communication and storage strategies in
To demonstrate the feasibility and advantages of the proposed the field level must be developed correspondingly to establish the con­
approach, this section introduces an early implementation of DT- nections among the field-level data, the Edge DTs, the Cloud DT as well
enabled collaborative data management in a metal AM system estab­ as different users. Based on the shop floors and hardware equipment
lished under the European Union’s H2020 project MANUELA, which involved in MANUELA, the overall system architecture of the DT-
involves 20 partner organisations from both research institutions and enabled collaborative data management in MANUELA metal AM sys­
industrial companies. The goal of MANUELA is the development of a tem is developed as shown in Fig. 3.
metal AM pilot line service covering the full AM development cycle The system comprises five main modules: 1) Dashboard, 2) Dash­
including simulation, robust AM manufacturing and on-line process board Users, 3) Pilot Line, 4) Post AM, and 5) Quality Measurement. In
control, characterization, real-time feedback, post-treatment, AM qual­ correspondence with the conceptual framework (Fig. 1), the Dashboard
ification protocols and associated business model. MANUELA contains module represents the cloud-based collaborative data management
all the five product lifecycle stages described in the conceptual frame­ platform. The Dashboard Users module represents different users who
work. However, the shop floors and hardware equipment are utilise software tools and local intelligence of their Edge DTs to collab­
geographically distributed in several countries in Europe. Thus, to oratively manage the metal AM product lifecycle data in the cloud. The
realise collaborative data management among all partners throughout product design and process planning stages described in the conceptual
the product lifecycle, the DT-enabled collaborative data management is framework are integrated in the Dashboard Users module since they do
implemented in MANUELA metal AM system in an early stage. not directly work on the field-level manufacturing equipment. The other
First, the system architecture of the DT-enabled collaborative data three modules, i.e. Pilot Line, Post AM and Quality Measurement, are
management in MANUELA is introduced. Detailed data communications respectively corresponding to the manufacturing stage, post processing
in each product lifecycle stage are explained. Examples of early imple­ stage and quality measurement stage in the metal AM product lifecycle.
mentations are also demonstrated. Second, a representative application The following sub-sections explain the detailed functions and data
scenario of the collaborative data management system, i.e. cloud-based communications in each module.
and deep learning-enabled layer defect analysis, is presented.
5.1.1. Dashboard and Dashboard Users
The Dashboard resides in the cloud and communicates with

866
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 6. Bulk data upload using the Dashboard.

Dashboard Users through the Internet to achieve collaborative data the Dashboard through the Internet. The main tasks of each type of
management in MANUELA. Fig. 4 illustrates the main functions and Dashboard Users are summarised in Fig. 4.
tasks of the Dashboard and different Dashboard Users. The Dashboard The Dashboard application for MANUELA project is developed using
provides five main functions as follows: MSC’s product lifecycle management software named MaterialCenter.
MaterialCenter perfectly meets the requirements of the proposed DT-
1) The Dashboard provides a Graphical User Interface (GUI) that allows enabled collaborative data management since it provides flexible and
efficient data management and data visualisation. Customised GUIs efficient functions for 1) developing customised Graphical User In­
are developed for different users based on specific requirements. terfaces (GUIs) and databases, 2) integrating customised data analytics
2) An interactive metal AM workflow is designed in the Dashboard for functions, and 3) interfacing with external software applications. In
each product/project by the project managers. The workflow, based addition, MaterialCenter allows secure cloud-based collaborative data
on the product lifecycle stages, defines the specific tasks and access management by granting different administration rights to different
rights for each type of Dashboard Users to establish a safe and users.
collaborative project management platform. The database in the Dashboard is developed based on the metal AM
3) The Dashboard contains a cloud-based relational database which is product data model introduced in Section 4. XML schema files are
structured by the metal AM product data model introduced in Sec­ developed for MaterialCenter to define the structure as well as the name,
tion 4. This database stores all the product lifecycle data retrieved data type, description and properties of each individual data item of the
from the Edge DTs. database. The Dashboard allows users to upload data to the database in
4) Advanced data analytics tools can be embedded in the Dashboard to two manners, i.e. manual input and bulk upload. Fig. 5 shows the
provide cloud-based decision-making support, such as analysis of developed Dashboard GUI for manual data input where users can input
material-geometry-process-property relationships and non-real-time values or files to each individual data item in the database.
product quality control. The GUIs and procedures for bulk data upload using the Dashboard
5) The Dashboard also provides application interfaces for a wide range are described in Fig. 6. A template Excel file that contains all the data
of software (CAD/CAM/CAE software, data visualisation and ana­ items in the database is developed as shown in the bottom of Fig. 6. An
lytics software, etc.) that allow different Dashboard Users to perform XML mapping schema is developed for mapping all the data in the
various types of specialised data analysis. template Excel file to the database through the MaterialCenter plugin in
Excel software. In the Dashboard GUI, the user chooses the Excel file to
There are mainly six types of Dashboard Users, including 1) product be imported and the mapping schema to perform the bulk data upload.
designers, 2) process planners, 3) production planners, 4) project man­ In this way, various types of field-level manufacturing data can be
agers, 5) data analysts, and 6) customers. These users utilise the local uploaded to the cloud database, and hence enabling efficient and
intelligence of their Edge DTs to perform their tasks and interact with collaborative metal AM product data management.

867
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 7. Detailed data communications and tasks in Pilot Line.

5.1.2. Pilot Line communications among the devices. All the devices send handshakes to
Fig. 7 describes the detailed data communications and tasks in the the PLC and the PLC responds with control commands of the next
Pilot Line. This current early stage involves three metal AM machines in actions.
three distributed shop floors, including an in-house developed EBM The Edge DT downloads all the post processing tasks from the Cloud
machine [57], an EOS M290 DMLS machine and an EOS M270 DMLS DT and sends the process sequences and process parameters to the main
machine. These machines are equipped with different sensors or moni­ PLC and the AGV to coordinate the post AM processes. The AGV is
toring systems. For example, a backscatter electrons (BSE) detector is controlled via the AGV interface (HTTP REST API) embedded in the
installed in the EBM machine to acquire electron optical (ELO) images of Edge DT. Ethernet TCP is used to send and receive process data between
the printed layers; while the EOS M290 machine has a built-in process the Edge DT and the devices. During post processes, computational tasks
monitoring system called EOSTATE ExposureOT and MeltPool. such as process simulation and real-time data processing are conducted
An Edge DT is developed for each machine in the local shop floor. with the local intelligence of the Edge DT. The process events and pro­
During printing, the machine logs and sensor signals generated by the cess data are stored in the local database and transferred to the Cloud
machines and sensors are transferred to the Edge DT through field bus, DT. The main tasks of the Edge DT and the Cloud DT are listed in Fig. 8.
Ethernet, WiFi, etc. The operators in the shop floors utilise the local A Virtual Reality (VR)-based DT application is developed for the Post
intelligence provided by the Edge DT to conduct various computation­ AM shop floor as shown on the right of Fig. 9. The application allows
ally intensive tasks as listed in Fig. 7. The main tasks of the Cloud DT for real-time field-level data to be communicated between the
the Pilot Line are also listed in Fig. 7. The interactions between the Edge manufacturing facilities and their virtual models, and hence providing
DTs in Pilot Line and the Dashboard will be demonstrated in a repre­ real-time post AM process monitoring functions. An Application Pro­
sentative application scenario in Section 5.2. gramming Interface (API) is developed in the MaterialCenter to allow
data communication between the Dashboard and the DT application.
5.1.3. Post AM Fig. 9 shows the developed GUI in the Dashboard for post AM machine
Fig. 8 describes the detailed data communications and tasks in Post status monitoring. More advanced data analytics for the Post AM pro­
AM. The Post AM module comprises a post AM cell, an automated supply cesses will be integrated in the Dashboard in the future.
chain, a main Programmable Logic Controller (PLC) and an Edge DT.
The post AM cell consists of a machine tool (for post machining), a 5.1.4. Quality Measurement
furnace (for heat treatment), a robot arm and a laser scanner (for quality Fig. 10 describes the detailed data communications and tasks in
assurance). The automated supply chain consists of an Automated Quality Measurement. The Quality Measurement module comprises
Guided Vehicle (AGV) and a robot. The main PLC is used to control all product quality measurement and analysis devices, including an X-ray
the devices via field bus such as EthernetIP or Profinet. ProfiSafe (a Computed Tomography (CT) scanner, a tactile Coordinate Measurement
safety layer on top of the field bus) is implemented to ensure safe data Machine (CMM), their control PCs and an Edge DT. The CT scanner take

868
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 8. Detailed data communications and tasks in Post AM.

Fig. 9. Interaction between Dashboard and Edge DT in Post AM.

cross-sectional images of the product and send the radiographs to the The Edge DT retrieves the measurement tasks and the related prod­
control PC for image analysis and 3D reconstruction; while the tactile uct design files from the Cloud DT via the Internet. All the measurement
CMM measures the dimensional accuracy of the product and send the results are sent to the Edge DT and processed into product quality data
tactile probe data to the control PC for data processing. and measurement reports corresponding to the data items in the metal

869
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 10. Detailed data communications and tasks in Quality Measurement.

Fig. 11. Overview of the cloud-based and deep learning-enabled metal AM layer defect analysis.

870
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 12. Snapshot of the uploaded product lifecycle data in the Dashboard.

871
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

Fig. 13. Procedures of the on-line layer defect detection in the Dashboard.

AM product data model. Then the product quality data and measure­ Network (CNN) model. The training data includes over 16,000 ELO
ment reports are uploaded to the Cloud DT to be stored and shared with images that are labelled into three categories: 1) good, 2) porous and 3)
the Dashboard Users. bulging, corresponding to three types of layer defects appeared during
the metal AM process. Since the development of CNN model is not the
focus of this application, we applied the transfer learning technique to
5.2. Representative application scenario: cloud-based and deep learning-
train the AlexNet [58] with the ELO images and achieved 95.0% test
enabled layer defect analysis
accuracy. The trained CNN model is finally integrated into the Dash­
board as an embedded data analytics function that is accessible for
The implementation of DT-enabled collaborative data management
different users through the Internet.
in MANUELA enables various possibilities for advanced cloud-based
The developed application enables both off-line product quality
metal AM data analytics, thus improving the metal AM process perfor­
optimisation and on-line layer defect detection. On the one hand,
mance and product quality. This sub-section briefly introduces a
product designers and process planners can effectively correlate the
representative application scenario, i.e. cloud-based and deep learning-
layer defect analysis results with specific product design features and
enabled metal AM layer defect analysis, to demonstrate the great ad­
process parameters (part geometry, laser power, layer thickness, scan
vantages and potential of the proposed approach.
pattern, etc.), and hence perform design and process optimisations to
The overview of the application scenario is described in Fig. 11. First,
improve the product quality. On the other hand, during metal AM
the product designers and process planners upload the design and pro­
building process, the ELO image can be uploaded to the Dashboard once
cess parameters of the product to the cloud database through the
it has been generated and processed locally, then the embedded CNN
Dashboard GUI. Then the operator in the shop floor retrieves the process
model can detect the layer defects on-line.
parameters from the Dashboard to configure the machine and conduct
The on-line layer defect detection function has been developed by
the experiment. The experiment is conducted on the in-house developed
integrating the CNN model into the Dashboard’s backend. Fig. 13 il­
EBM machine mentioned in Section 5.1.2. Ti-6Al-4 V ELI (grade 23)
lustrates the designed GUIs and procedures of the on-line layer defect
powders are utilised to build eight cuboid samples with different scan
detection in the Dashboard. Briefly, it includes four steps as follows:
speeds. During the metal AM process, the BSE detector installed in the
building chamber of the EBM machine generates the ELO images of each
• Step 1: Import ELO images to the Dashboard.
printed layer. The raw ELO images are processed with image processing
• Step 2: Launch the embedded deep learning-enabled layer defect
techniques in the Edge DT and then uploaded to the cloud database in
detection function.
the Dashboard. Fig. 12 shows a snapshot of the cloud database in the
• Step 3: Display runtime status of the data analytics function and
Dashboard that contains the product lifecycle data uploaded by different
generate the prediction results as a. csv file.
users.
• Step 4: Import the predicted defect result as a child object of the ELO
Data analysts download the processed ELO images from the Dash­
image in the database.
board and use them as the training data to train a Convolutional Neural

872
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

This application allows distributed users to monitor the layer quality Acknowledgements
during the metal AM building process. The early identification of layer
defects not only protects the recoating system from being damaged, but This research was performed within the project Additive
also reduces production costs in terms of material waste and production Manufacturing using Metal Pilot Line (MANUELA), which received
time. funding from the European Union’s Horizon2020 research and innova­
tion programme under grant agreement No 820774. The authors would
6. Conclusions and future work like to thank Mr. Emil Johansson, Dr. Benjamin Bircher, Dr. Vaclav
Pejchal and Dr. Zhuoer Chen for their support in developing the data
Metal AM has been attracting a continuously increasing attention model and the MANUELA system architecture, and Mr. Daniel Gage for
due to its great advantages compared to traditional subtractive coordinating the experiments.
manufacturing in terms of higher design flexibility, shorter development
time, lower tooling cost, and fewer production wastes. However, the References
lack of process robustness, stability and repeatability caused by the
unsolved complex relationships between material properties, product [1] Gibson I, Rosen D, Stucker B. Additive manufacturing technologies: 3D printing,
rapid prototyping, and direct digital manufacturing. second edition 2015.
design, process parameters, process signatures, post AM processes and [2] Qin Y, Qi Q, Scott PJ, Jiang X. Status, comparison, and future of the representations
product quality has significantly impeded its broad acceptance in the of additive manufacturing data. CAD Computer Aided Design 2019;111:44–64.
industry. To enable efficient implementation of advanced data analytics, [3] Atzeni E, Salmi A. Economics of additive manufacturing for end-usable metal parts.
International Journal of Advanced Manufacturing Technology 2012;62:1147–55.
this paper proposes a novel DT-enabled collaborative data management [4] Mani M, Lane B, Donmez MA, Feng SC, Moylan SP, Fesperman R. Measurement
approach for metal AM systems. The main contributions of this work are Science Needs for Real-time Control of Additive Manufacturing Powder Bed Fusion
summarised as follows: Processes. NIST Interagency/Internal Report (NISTIR) - 8036. 2015.
[5] ISO/ASTM. INTERNATIONAL STANDARD ISO / ASTM 52900 Additive
manufacturing — General principles — Terminology. International Organization
• Proposed a novel DT-enabled collaborative data management for Standardization; 2015.
framework for metal AM systems, where a Cloud DT communicates [6] Xu X. Machine Tool 4.0 for the new era of manufacturing. International Journal of
Advanced Manufacturing Technology 2017;92:1893–900.
with distributed Edge DTs in different product lifecycle stages.
[7] Kruth JP, Leu MC, Nakagawa T. Progress in additive manufacturing and rapid
• Proposed a metal AM product data model that contains a compre­ prototyping. CIRP Annals - Manufacturing Technology 1998;47:525–40.
hensive list of specific product lifecycle data that has influence on the [8] Grasso M, Colosimo BM. Process defects and in situ monitoring methods in metal
powder bed fusion: A review. Measurement Science and Technology 2017;28:
product quality.
044005.
• Practically implemented the DT-enabled collaborative data man­ [9] Frazier WE. Metal additive manufacturing: A review. Journal of Materials
agement in a distributed metal AM system. Early development of the Engineering and Performance 2014;23:1917–28.
data management system has demonstrated efficient data commu­ [10] Sames WJ, List FA, Pannala S, Dehoff RR, Babu SS. The metallurgy and processing
science of metal additive manufacturing. International Materials Reviews 2016;61:
nications between the distributed shop floors and the cloud-based 315–60.
data management system. [11] Everton SK, Hirsch M, Stavroulakis PI, Leach RK, Clare AT. Review of in-situ
• Developed a representative application scenario of cloud-based and process monitoring and in-situ metrology for metal additive manufacturing.
Materials and Design 2016;95:431–45.
deep learning-enabled metal AM layer defect analysis, which enables [12] Yadroitsau I. Selective laser melting: Direct manufacturing of 3D-objects by
both off-line product design and process optimisation and on-line selective laser melting of metal powders. Lambert Academic Publishing 2009.
layer defect detection. [13] Craeghs T, Clijsters S, Yasa E, Kruth JP. Online quality control of selective laser
melting. 22nd Annual International Solid Freeform Fabrication Symposium - An
Additive Manufacturing Conference, SFF 2011. 2011. p. 212–26.
The proposed DT-enabled collaborative data management has shown [14] Müller JR, Panarotto M, Malmqvist J, Isaksson O. Lifecycle design and
great potential in enhancing fundamental understanding of metal AM management of additive manufacturing technologies. Procedia Manufacturing
2018:135–42.
processes, developing simulation and prediction models, reducing
[15] Lu Y, Witherell P, Lopez F, Assouroko I. Digital solutions for integrated and
development times and costs, and improving product quality and pro­ collaborative additive manufacturing. Proceedings of the ASME Design
duction efficiency. Engineering Technical Conference. 2016.
[16] Lu Y, Witherell P, Donmez A. A collaborative data management system for additive
Moreover, the proposed DT-enabled conceptual framework is not
manufacturing. Proceedings of the ASME Design Engineering Technical
limited to the application of metal AM systems. It can be treated as a Conference. 2017.
generic reference framework for developing collaborative data man­ [17] Uhlmann E, Pontes RP, Laghmouchi A, Bergmann A. Concept of Sustainable Data
agement for other types of manufacturing systems. In those cases, for a Selective Laser Melting Machine. Procedia Manufacturing 2018:655–62.
[18] Prater T. Database development for additive manufacturing. Progress in Additive
domain-specific product data models need to be developed by identi­ Manufacturing 2017;2:11–8.
fying the specific product quality-related manufacturing data in each [19] Mies D, Marsden W, Warde S. Overview of Additive Manufacturing Informatics: “A
lifecycle stage. Efficient data communications between field-level Digital Thread.”. Integrating Materials and Manufacturing Innovation 2016;5:
114–42.
manufacturing devices, the Edge DTs and the Cloud DT need to be [20] Lu Y, Choi S, Witherell P. Towards an integrated data schema design for additive
established with consideration of the specific manufacturing facilities manufacturing: Conceptual modeling. Proceedings of the ASME Design
involved. Other product lifecycle stages that have not been considered in Engineering Technical Conference 2015.
[21] Feng SC, Witherell P, Ameta G, Kim DB. Activity model for homogenization of data
this work such as product usage stage and end-of-life treatment stage sets in laser-based powder bed fusion. Rapid Prototyping Journal 2017;23:137–48.
could also be included to enable various types of PLM services. [22] Feng SC, Witherell PW, Ameta G, Kim DB. Fundamental requirements for data
Our future work will focus on the development and implementation representations in laser-based powder bed fusion. ASME 2015 International
Manufacturing Science and Engineering Conference, MSEC 2015 2015.
of machine learning-enabled advanced data analytics in the DT-enabled
[23] Kim DB, Witherell P, Lu Y, Feng S. Toward a Digital Thread and Data Package for
collaborative data management system in MANUELA. Recent advance­ Metals-Additive Manufacturing. Smart and Sustainable Manufacturing Systems
ments of machine learning will be studied and applied for metal AM 2017;1:75.
[24] Bonnard R, Hascoët JY, Mognol P, Zancul E, Alvares AJ. Hierarchical object-
process monitoring, control and optimisation. Various types of machine
oriented model (HOOM) for additive manufacturing digital thread. Journal of
learning-enabled closed-loop metal AM applications based on the DT- Manufacturing Systems 2019;50:36–52.
enabled collaborative data management system will be developed to [25] Tao F, Qi Q. Make more digital twins. Nature 2019:490–1.
improve the metal AM product quality and production efficiency. [26] Tao F, Cheng J, Qi Q, Zhang M, Zhang H, Sui F. Digital twin-driven product design,
manufacturing and service with big data. International Journal of Advanced
Manufacturing Technology 2018;94:3563–76.
Declaration of Competing Interest [27] Uhlemann THJ, Lehmann C, Steinhilper R. The Digital Twin: Realizing the Cyber-
Physical Production System for Industry 4.0. Procedia CIRP 2017:335–40.

No potential conflict of interest was reported by the authors.

873
C. Liu et al. Journal of Manufacturing Systems 62 (2022) 857–874

[28] Lu Y, Liu C, KI-K Wang, Huang H, Xu X. Digital Twin-driven smart manufacturing: [44] Xia T, Dong Y, Xiao L, Du S, Pan E, Xi L. Recent advances in prognostics and health
Connotation, reference model, applications and research issues. Robotics and management for advanced manufacturing paradigms. Reliability Engineering and
Computer-Integrated Manufacturing 2020;61:101837. System Safety 2018;178:255–68.
[29] Lim KYH Zheng P, Chen C-H. A state-of-the-art survey of Digital Twin: techniques, [45] Xia T, Xi L. Manufacturing paradigm-oriented PHM methodologies for cyber-
engineering product lifecycle management and business innovation perspectives. physical systems. Journal of Intelligent Manufacturing 2019;30:1659–72.
Journal of Intelligent Manufacturing 2019:1–25. [46] Kaewunruen S, Lian Q. Digital twin aided sustainability-based lifecycle
[30] Tao F, Sui F, Liu A, Qi Q, Zhang M, Song B, et al. Digital twin-driven product design management for railway turnout systems. Journal of Cleaner Production 2019;228:
framework. International Journal of Production Research 2019;57:3935–53. 1537–51.
[31] Wang XV, Wang L. Digital twin-based WEEE recycling, recovery and [47] Zheng P, Sivabalan AS. A generic tri-model-based approach for product-level
remanufacturing in the background of Industry 4.0. International Journal of digital twin development in a smart manufacturing environment. Robotics and
Production Research 2019;57:3892–902. Computer-Integrated Manufacturing 2020;64:101958.
[32] Liu C, Vengayil H, Zhong RY, Xu X. A systematic development method for cyber- [48] Mukherjee T, DebRoy T. A digital twin for rapid qualification of 3D printed
physical machine tools. Journal of Manufacturing Systems 2018;48:13–24. metallic components. Applied Materials Today 2019;14:59–65.
[33] Liu C, Hong X, Zhu Z, Xu X. Machine Tool Digital Twin: Modelling methodology [49] Knapp GL, Mukherjee T, Zuback JS, Wei HL, Palmer TA, De A, et al. Building blocks
and applications. Proceedings of International Conference on Computers and for a digital twin of additive manufacturing. Acta Materialia 2017;135:390–9.
Industrial Engineering, CIE 2018. [50] Glaessgen EH, Stargel DS. The digital twin paradigm for future NASA and U.S. Air
[34] Tao F, Zhang M, Liu Y, Nee AYC. Digital twin driven prognostics and health force vehicles. Collection of Technical Papers - AIAA/ASME/ASCE/AHS/ASC
management for complex equipment. CIRP Annals 2018;67:169–72. Structures. Structural Dynamics and Materials Conference 2012:1818.
[35] Wang J, Ye L, Gao RX, Li C, Zhang L. Digital Twin for rotating machinery fault [51] Xu X. From cloud computing to cloud manufacturing. Robotics and Computer-
diagnosis in smart manufacturing. International Journal of Production Research Integrated Manufacturing 2012;28:75–86.
2019;57:3920–34. [52] Georgakopoulos D, Jayaraman PP, Fazia M, Villari M, Ranjan R. Internet of Things
[36] Zhuang C, Liu J, Xiong H. Digital twin-based smart production management and and Edge Cloud Computing Roadmap for Manufacturing. IEEE Cloud Computing
control framework for the complex product assembly shop-floor. International 2016;4:66–73.
Journal of Advanced Manufacturing Technology 2018;96:1149–63. [53] Wu D, Liu S, Zhang L, Terpenny J, Gao RX, Kurfess T, et al. A fog computing-based
[37] Gerhard D. Product lifecycle management challenges of CPPS. Multi-Disciplinary framework for process monitoring and prognosis in cyber-manufacturing. Journal
Engineering for Cyber-Physical Production Systems: Data Models and Software of Manufacturing Systems 2017;43:25–34.
Solutions for Handling Complex Engineering Projects. 2017. p. 89–110. [54] Peng T, Kellens K, Tang R, Chen C, Chen G. Sustainability of additive
[38] Botkina D, Hedlind M, Olsson B, Henser J, Lundholm T. Digital Twin of a Cutting manufacturing: An overview on its energy demand and environmental impact.
Tool. Procedia CIRP 2018:215–8. Additive Manufacturing 2018;21:694–704.
[39] Qi Q, Tao F, Hu T, Anwer N, Liu A, Wei Y, et al. Enabling technologies and tools for [55] O’Regan P, Prickett P, Setchi R, Hankins G, Jones N. Metal Based Additive Layer
digital twin. Journal of Manufacturing Systems 2019. Manufacturing: Variations, Correlations and Process Control. Procedia Computer
[40] Lu Y, Xu X. Cloud-based manufacturing equipment and big data analytics to enable Science 2016:216–24.
on-demand manufacturing services. Robotics and Computer-Integrated [56] Spears TG, Gold SA. In-process sensing in selective laser melting (SLM) additive
Manufacturing 2019;57:92–102. manufacturing. Integrating Materials and Manufacturing Innovation 2016;5:
[41] Liu C, Vengayil H, Lu Y, Xu X. A Cyber-Physical Machine Tools Platform using OPC 16–40.
UA and MTConnect. Journal of Manufacturing Systems 2019;51:61–74. [57] Arnold C, Pobel C, Osmanlic F, Körner C. Layerwise monitoring of electron beam
[42] Cheng J, Zhang H, Tao F, Juang CF. DT-II:Digital twin enhanced Industrial Internet melting via backscatter electron detection. Rapid Prototyping Journal 2018;24:
reference framework towards smart manufacturing. Robotics and Computer- 1401–6.
Integrated Manufacturing 2020;62:101881. [58] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep
[43] Wang W, Zhang Y, Zhong RY. A proactive material handling method for CPS convolutional neural networks. Advances in Neural Information Processing
enabled shop-floor. Robotics and Computer-Integrated Manufacturing 2020;61: Systems 2012:1097–105.
101849.

874