Reviewer 4

The paper presents an excellent overview of challenges and potential solutions to the issues related to digital twin data.

Question 1: The authors should consider including references to the ISO23247 standard which would complement the presented work very nicely. In particular, Part 3 of the document provides a standardised framework for digital representation of the elements in a digital twin and Part 4 provides the framework to underpin information exchange in Digital Twins.

Response: Thanks for the comments. The reference to the ISO 23247 has been added in the revised version. The information classification for the manufacturing elements in Part 3 is referenced in the section “Composition of digital twin data” as follows:

Section “Composition of digital twin data”, Para. 2 

According to ISO 23247-3 series (Digital Twin manufacturing framework-Part 3: Digital representation of manufacturing elements), physical entity-related data can be grouped in two categories: the first describing the static information concerning the physical entity (e.g. identification, characteristics, schedule) and the second reflecting the dynamic states (e.g. status, location, relationship). These data can help to represent the physical entity in a digital way.

The suggested data formats and protocols for data exchange in Part 4 are cited in the section “Digital twin data interaction”, and the section “Key enabling technologies for digital twin data” as follows:

Section “Digital twin data interaction” Then the data can be transferred between the DTD parts through some types of data transmission protocols and formats suggested by the ISO 23247-4 (Digital Twin manufacturing framework- Part 4: Information exchange).  

Section “Key enabling technologies for digital twin data”, Para. 3 ISO 23247-4 provides some options for data transmission protocols and formats. For example, the virtual entity can access data of the physical entity by using protocols such as MTConnect, OPC-UA, HTTP and CoAP; and the virtual entity can be connected with the service by using an application interface (i.e. open API) or by employing shared memory. The suggested data formats include XML, JavaScript object notation (JSON), quality information framework (QIF) and so forth.

Question 2: There are a few other well-cited references which contain digital twin data issues that should be used in the paper to position the work better. These include: Wagner et al. (2017) - Provides an overview of the Asset Administration Shell as an interoperability wrapper that enables data transfer between the physical and digital counterparts in a manner that virtualises the asset and allows various twins to communicate in a similar manner. Jones et al. (2020) - Provides a contextual framework for Virtual-Physical and Physical-Virtual twinning to underpin the various purposes of data transfer and storage in Digital Twins (an article of mine). Boschert et al. (2016) - Provides a view of the Digital Twin and its data across the lifecycle of a product and its production system that provides scope for the topic of discussion in the presented paper.

Response: Thanks for the comments. These references have been added in the sections of “Principles for digital twin data”, “Digital twin data gathering”, and “Key enabling technologies for digital twin data” as follows:

Section “Principles for digital twin data”, Para. 3 It aims at ensuring the rapid twinning process, namely the process of synchronizing the virtual and physical states, thus to enable parts of the DT to act synchronously ³¹. Section “Digital twin data gathering” In addition, the DTD would also extract essential information from the existing IT systems, such as product lifecycle management (PLM) system, product data management (PDM) system, manufacturing execution system (MES) and supervisory control and data acquisition (SCADA) system, which store large amounts of data across the lifecycle of the physical entity ³².

Section “Key enabling technologies for digital twin data”, Para. 3 In addition, the asset administration shell, which virtualizes the asset and enables the access to asset-related data via application interfaces, also provides a way for data exchange ⁵².

[31] Jones D, Snider C, Nassehi A, et al.: Characterising the digital twin: a systematic literature review. CIRP Journal of Manufacturing Science and Technology. 2020; 29, Part A: 36-52.

[32] Boschert S, Rosen R: Digital twin-the simulation aspect. In Mechatronic Futures. Springer, 2016.

[52] Wagner C, Grothoff J, Epple U, et al.: The role of the Industry 4.0 asset administration shell and the digital twin during the life cycle of a plant. 2017 22nd IEEE International Conference on Emerging Technologies and Factory Automation, September 12-15, 2017, Limassol, Cyprus, 2017, DOI: 10.1109/ETFA.2017.8247583.

Reviewer 3

This paper tries to propose a new concept, DTD, to describe the related requirements, principles and enabling techs. It is interesting in the attempt to draw a new DT data processing framework or guideline. However, from reviewer's perspective on operation in practice, there are still some places that are unclear and hard to follow.

Question 1: There is no clear definition about DTD, but more about data classification in the context of DT. It is possible to give a new definition to show the difference of data in the context of DT? The definition difference would focus on more about contents or format or application?

Response: Thanks for the comments. A definition about DTD is added in the first paragraph of the section “Composition of digital twin data” as follows:

As cyber-physical integration is a distinguishing feature of DT, cyber-physical data fusion is a major characteristic of the DTD. It could be reflected on the data contents, as the DTD involves data from both the physical and virtual spaces, and tends to merge these data to generate fusion data for more accurate and comprehensive information. DTD refers to a wide spectrum of data closely related to the DT. It focuses equally on data collected from the physical space (e.g. sensor data from equipment, materials, and workers) and data produced by the virtual space (e.g. data mimicked by simulation models, data generated by algorithms embedded in services, and data deduced based on knowledge), and tends to make the data from the two spaces corrected and supplemented by each other through data fusion, to achieve more accurate and comprehensive information for the DT-related applications.

Question 2: Is there an exact logic for the listing of DTD requirements? Mining is before or after fusion, or both directions are possible?

Response: The listing of DTD requirements is corresponding to the logic of data gathering, interaction, storage, association, fusion, evolution and usage. There is a one to one correspondence between the DTD requirements and the principles (Fig.2), as well as the principles and methods (Fig.3). As data mining is the basis for data fusion, mining is before fusion. Data fusion is developed based on data with different relations, such as fusing similar data for data correction and fusing complementary data for getting supplementary information. While, data association relations (e.g. causality, similarity, and complementation) are extracted based data mining algorithms, such as Apriori algorithm, K-means, and Pearson correlation analysis, etc.

Question 3: It is hard to understand 'data interaction'. It is more similar to service as we often encountered in IOT or CPS systems

Response: Thanks for the comments. In the paper, 'data interaction' refers to the transmission and comparison of connection data from any two parts of the DT. It aims at making the data from different sources consistent with each other to enable the coordinated operations. In the section “Digital twin data interaction”, an example about the data interaction is added as follows: 

To better illustrate the interaction, take an autoclave (a key equipment for composites curing) as an example. During curing, temperature distribution of the mold within the autoclave is an important quality indicator, since it largely affects the deformation degree of the composite components ¹⁹. To monitor the component quality, the real mold temperatures are collected by thermocouples. While, a virtual model can also be built to mimic the expected mold temperature distribution by using the CFD simulation. The data interaction between the physical equipment and the virtual model is used to ensure that the actual curing is executed as simulated. In this situation, a major feature for reflecting the mold temperature distribution, i.e. the difference between the maximum and minimum temperatures on the mold upper surface, can be selected as the connection data, leaving out other redundant or unimportant features. Then the collected and simulated mold temperature differences can be exchanged and compared. If their deviation is always within a predefined threshold, the expected composite quality would be achieved. Otherwise, perturbing factors causing the inconsistency need to be found, such as uneven airflow field in the autoclave, deteriorating insulation of the autoclave wall, and abrupt changes of curing conditions. These factors should be eliminated by adjusting parameters of the physical autoclave (e.g. heating speed, air speed, pressure) or by maintaining the equipment, which aims at reducing negative effects caused by the disturbances, to keep the collected data close with the simulated ones.

In the section “Key enabling technologies for digital twin data”, more technical details for data interaction are also added in the part (2). Besides, as the service is a relatively broad concept, I think the “data interaction” could also be treated as a kind of service, which exchanges and compares data from different sources to ensure the data consistency. Perhaps it can be encapsulated as a service and then invocated to realize the data exchange and comparison when necessary.

Question 4: What exactly is 'iterative optimization'? As you mentioned in the paper: iterative optimization refers to a cyclic process of “data increase - data fusion - information increase”, through which, new data is fused with historical data to generate new information continuously. This means to generate new data sets. It is more about data evolution? However, as you also mentioned in the later section of the paper, 'optimization of DTD'. So, how to optimize DTD? What are the objective, process and constraints?

Response: Thanks for the comments. Before the “iterative optimization”, “optimization of DTD” is first explained. “Optimization of DTD” mainly aims at making the data more informative through data fusion. In different situations, the objective, process and constraints would be different. For example, assume that there are data of several variables related to the remaining useful life (RUL) of an equipment component, such as vibrations, speeds and forces collected by sensors, as well as deformation and stress distribution simulated by models. We can fuse the data by using the neural network as the fusion algorithm to obtain more informative data. Inputs of the network are all or part of the RUL-related variables and the output is the merging result, i.e. the fusion data, which would be more relevant with the RUL than any other single variable. The relevance between the fusion data and the RUL can be measured by the mutual information (a measure to quantify the shared information between two variables). Higher relevance means that the data carry more RUL-related information. In this situation, the optimization objective would be maximizing the mutual information between the fusion data and the RUL. Since selecting different subsets of the RUL-related variables as inputs of the neural network would lead to different fusion data, the optimization process mainly needs to search out the best subset from the candidate ones to make the objective optimal. Some constraints would be imposed on the variables to be selected. For example, the relevance between each variable and the RUL would be larger than a lower limit, to eliminate the less correlated ones; and the relevance between any two selected variables would be lower than an upper limit, to reduce unnecessary redundancy. The iterative optimization refers to the repeated process of the above optimization. When new data from different parts such as the physical entity, virtual model and service are injected into the DTD, a new optimization would begin to decide what new data to be added to the fusion to make the objective more optimal than before. It means that the newly generated fusion data would carry more information. Besides, the “iterative optimization” is closely related with the data evolution. The logic in the paper is as follows. “Iterative optimization” is a requirement on DTD. To meet the requirement, the “information growth principle” proposed in the later section should be observed, to ensure that the “data evolution” step in the section “Methodology for digital twin data” would keep the fusion data generated and information carried by the fusion data increased. To better illustrate the iterative optimization, some explanations are added in the “Requirements on iterative optimization” in Section 1 as follows: 

In the cyclic process, each optimization, triggered by the injection of new data, aims at adding appropriate data into the fusion to increase the amount of valuable information carried by the fusion data, while being subject to possible constraints on the data to be merged, such as keeping the data redundancy lower than an upper bound.

Question 5: The requirement of data universality seems an ideal solution in every domain. However, DT is usually customized system. Hence, is there any conflict between this requirement with the application? What does it mean by universality, contents or format?

Response: It could not be conflicted. As the DT system is customized, it would develop different sensors, models and services for different applications, which bring a wide variety of data with different types and contents. Although it is hard to make the data unified in every aspect, we can still make them transformed into some unified formats (e.g. XML) for data storing and sharing, and transmitted by some universal interfaces and communication protocols for easier communication. In the paper, universality mainly refers to unified data format, interface and communication protocol.

Question 6: What is fusion data, data of data (well structured and associated) or 'green data' (intermediate data)?

Response: Fusion data is a merging result of data from different sources by using data fusion algorithms. Take the healthy monitoring for an equipment as an example. From the physical space, data such as the vibrations, speeds, forces and pressures can be collected by sensors. While, from the virtual space, data such as deformation, stress distribution, and airflow distribution can be simulated by virtual models. Since all the data would be relevant with the equipment conditions, a merging result of these data (i.e. the fusion data) can be used to support a more accurate monitoring, as it involves more comprehensive information, compared with data from a single source. The fusion data is obtained through processing the original data from the sensors, simulations or other sources via data fusion algorithms, such as the weighting method, neural network, Bayesian method, etc. To clarify it, some explanations are added in the second to last paragraph in the section “Composition of digital twin data” as follows: 

For example, the healthy condition of an equipment can be reflected by various sensor data such as the vibrations, speeds, forces and pressures, as well as the simulated data such as deformation, stress distribution, and airflow distribution. The fusion data is a merging result of the above data, which is obtained by processing the data from different sources via data fusion algorithms, such as the weighting method ²⁸, neural network ²⁹, and Bayesian method ³⁰. As a result, the fusion data would be more informative for the condition monitoring.

Question 7: Are the components of DTD organized in a hierarchical way or not?

Response: The components of DTD can be organized in a hierarchical way. In the previous work of our team [1], the hierarchical way for data organization has been preliminarily discussed, combining with the edge computing, fog computing and cloud computing. According to it, the original data from the physical entity, virtual entity and service belongs to the edge level. In this level, the incoming data can be preprocessed (e.g. filtering, redundancy removing, missing data filling) and then used to identify some potential threat patterns (e.g. equipment fault). The edge level can provide timely responses to the threats, as it generally has low latency. Then the preprocessed data would be transferred to the fog level, where more complex computing tasks are enabled to support the data feature extraction, data relation mining and fusion data generation. In this level, the fusion data can be employed to provide more comprehensive information for monitoring, prediction and optimization. The cloud level stores amounts of historical data from the edge and fog levels and supports the large-scale computation. It helps to mine more domain knowledge from the mass data to support long-term prediction and planning. In the paper, the hierarchical organization way for the DTD is not particularly discussed. More technical details about the DTD organization would be explored in our future work.

[1] Qi Q., Zhao D., Liao T. W., Tao F., Modeling of cyber-physical systems and digital twin based on edge computing, fog computing and cloud computing towards smart manufacturing, ASME 2018 13 ^th International Manufacturing Science and Engineering Conference, Texas, US, 2018, V001T05A018.

Reviewer 2

In this paper, the authors provide a future prospect and couple of requirements of data applications in the digital twin domain from structure, principle, and methodology views. With the illustration, the potential coming research directions are summarized and pointed out. This topic and corresponding learning areas are promising and worth investigating deeper. Here are some minor issues that may be considered for improving this paper:

Question 1: In the third requirement of digital twin data, could you explain more on the seamless mean? Such requirement is too blurry and what kind of performance do you want to achieve?

Response: Thanks for the comments. The “seamless” is used to express that data carrying different information and coming from different sources are fused to maximize improving the data quality, such as improving data accuracy, reducing data uncertainty, and increasing the amount of information carried by data. For example, to improve accuracy of the simulated data, the fusion would aim at reducing the root mean squared error (RSME) between the simulated and measured data. For reducing data uncertainty, it would minimize the information entropy of the data. To improve the amount of information carried by data, taking product quality-related information as an example, the fusion will aim at improving the relevancy between the data and the quality indicator.

To clarify it, some words are added to this requirement (the fifth requirement in the revised version) as follows:

It would bring lots of benefits, such as reducing information entropy of the sensor data (i.e. reducing data uncertainty), reducing root mean squared error (RMSE) between the simulated data and measured data (i.e. improving data accuracy), and improving relevancy between the data and a certain target indicator, such as product quality, remaining useful life of a key component, and equipment healthy condition (i.e. increasing the amount of information related to the indicator).

Besides, as the “seamless” may seem blurry to express the above content, this word is deleted in the revised version.

Question 2: In the composition of digital twin data section, the explanation of the connection data is provided theoretically. However, it lacks the intuitive example demonstration, therefore, could more cases or concrete technologies be provided to help better understanding?

Response: Thanks for the comments. In the section “Composition of digital twin data”, a simple example is added in the last paragraph as follows:

For example, the real temperature collected by sensors and the expected temperature simulated by virtual models can be selected as the connection data between the physical entity-related data and virtual entity-related data. If their deviation exceeds a predefined threshold, perturbing factors (e.g. manual interferences, environment changes) causing the inconsistency would be found and eliminated, which helps to restore the data consistency.

Then a more detailed example about the connection data of an autoclave is provided in the section “Digital twin data interaction” as follows:

To better illustrate the interaction, take an autoclave (a key equipment for composites curing) as an example. During curing, temperature distribution of the mold within the autoclave is an important quality indicator, since it largely affects the deformation degree of the composite components ¹⁹. To monitor the component quality, the real mold temperatures are collected by thermocouples. While, a virtual model can also be built to mimic the expected mold temperature distribution by using the CFD simulation. The data interaction between the physical equipment and the virtual model is used to ensure that the actual curing is executed as simulated. In this situation, a major feature for reflecting the mold temperature distribution, i.e. the difference between the maximum and minimum temperatures on the mold upper surface, can be selected as the connection data, leaving out other redundant or unimportant features. Then the collected and simulated mold temperature differences can be exchanged and compared. If their deviation is always within a predefined threshold, the expected composite quality would be achieved. Otherwise, perturbing factors causing the inconsistency need to be found, such as uneven airflow field in the autoclave, deteriorating insulation of the autoclave wall, and abrupt changes of curing conditions. These factors should be eliminated by adjusting parameters of the physical autoclave (e.g. heating speed, air speed, pressure) or by maintaining the equipment, which aims at reducing negative effects caused by the disturbances, to keep the collected data close with the simulated ones.

In addition, more related technical details are added in the part (2) of the section “Key enabling technologies for digital twin data” as follows:

(2) With respect to data interaction, data transmission technology is required to exchange connection data among different parts of DTD. ISO 23247-4 provides some options for data transmission protocols and formats. For example, the virtual entity can access data of the physical entity through protocols such as MTConnect, OPC-UA, HTTP and CoAP; and the virtual entity can be connected with the service by using application interfaces (i.e. open API) or by employing shared memory. The suggested data formats include XML, JavaScript object notation (JSON), quality information framework (QIF) and so forth. The asset administration shell, which virtualizes the asset and enables the access to asset-related data via application interfaces, also provides a way for data exchange ⁵². Then to reduce the redundancy in the exchanged data, measures such as Pearson correlation coefficient ³⁸ and mutual information ⁵³ can be employed to judge the overlapping of the data. Besides, Euclidean distance, Manhattan distance and dynamic time warping ⁵⁴, etc. can be used to evaluate the discrepancy between data from different parts of the DT to evaluate their consistency.

Question 3: In the principle part, could more explanation of figure 2 be provided? What’s the serial numbers and arrows in the figure mean? What do the different levels stand for in the figure?

Response: Thanks for the comments. More explanation for Fig. 2 is provided in first paragraph of the section “Principles for digital twin data” as follows, which explains the different levels in the figure. As in Figure 2, the outer layer shows the data requirements in seven aspects as raised in Section 1. While, the corresponding principles, which should be observed for the sake of better fulfilling the requirements, are given one by one in the inner layer. As arrows in the figure have no practical significance, they are deleted. In addition, serial numbers in the figure are used to distinguish the principles, since in the previous version, these principles are mentioned as “Principle (1), Principle (2)…” in Fig.3. But as in this version, such statements are deleted, the serial numbers in this figure are deleted accordingly.

Question 4: In figure 3, the quality of figures and expression way may be refined.

Response: Thanks for the comments. The Fig.3 has been refined.

Question 5: In key enabling technologies, some technologies concepts are adopted to enhance the development of DTD. However, could the authors provide more technical/algorithm details to achieve the corresponding goal/proposed performances?

Response: Thanks for the comments. More technical and algorithm details have been added in the section “Key enabling technologies for digital twin data” in the revised version.

Reviewer 1

Data plays critical roles in constructing virtual models, building cyber-physical connections, and executing intelligent operations in digital twins. This paper explores some basic principles and methods for digital twin data gathering, storage, interaction, association, fusion, evolution and servitization, as well as the key enabling technologies. The theoretical underpinnings for DTD proposed in this paper, are imperative for the further promotion and application of DT.

Question 1: The barriers or challenges to digital twin data could be addressed.

Response: Thanks for the comments. The challenges to digital twin data are discussed in the last section as follows:

For the DTD, challenges ahead are mainly concerning simulation data generation and cyber-physical data fusion. Firstly, since DTD emphasizes on the supplementary role of simulation data, how to build virtual models to generate the simulation data accurately remains a fundamental challenge. The commonly used models include data model and physical model. However, the former relies heavily on the modeling data and it usually has poor interpretability. Whereas, although the latter is established based on physical mechanisms and constraints, it tends to be over-simplistic sometimes and thus leads to non-ignorable deviations with the practice. The hybrid modeling method that merges the data model and the physical model would be a promising way for addressing this challenge. Secondly, even though cyber-physical integration is a distinguishing feature of DT, cyber-physical data fusion remains to be a challenge. As data from the physical and virtual spaces may have their own limitations and flaws, how to ensure that the fusion result would always be superior to data from either space is a concerned problem.

Question 2: If possible, an example or a scenario could be used to enhance the paper’s contributions.

Response: Thanks for the comments. This paper mainly provides a theoretical method for DTD gathering and processing, aiming at helping more DT researchers to incorporate DTD into their works. The related research is still at the primary stage. Based on this, more applications examples and scenarios would be provided in our future works, such as examples of DTD applications in product design optimization, product quality control and equipment prognosis. They will further validate and refine the proposed methods.