Introduction

Although widely used statistical and simulation models are excellent methods for data analysis, they have limitations in analyzing the complex real world. Statistical analysis can build high-fidelity models through calibration using big data, but it is difficult to implement a wide range of models (Richardson, 2015). This is because, as the model becomes more complex, programming becomes more difficult, and excessive time is consumed for analytical calculations. Simulation can model and analyze a wide world based on real data and theories. Although these models have advantages in verification, validation and optimization aspects, they have limitations in distribution assumptions or are not easy to apply during simulation run-time (Liu et al., 2021). Consider analyzing the logistics supply chain. The models implement a large-scale logistics system that not only reflects the various processes of SCM, but also considers precise demand patterns and raw material supply conditions. In simulation-based analysis, it is possible to build, verify, and optimize the model of the entire SCM, but it is difficult to include detailed parts such as failure of logistics equipment such as conveyor, palletizer, and folk lift. Their failures can become a bottleneck and affect the entire SCM, but in general, in the simulation model, these accidents are reflected as random event occurrences based on statistical theory. e.g., it is assumed that the failure of conveyor occurs randomly in a Normal (5, 10). On the other hand, using a statistical model, it is possible to analyze the failures of logistics equipment in a highly realistic way using big data. However, if the SCM is extended to a wide range, the complexity becomes too large, making it very difficult to build a model. Digital twin (DT) can simulate a wide world with high fidelity. DT is not a term referring to a specific technology. It is an idea that simulates the fusion of many existing cutting-edge technologies. It becomes a replica of reality by implementing the real world as closely as possible and calibrating it with data. It was originally proposed as a concept to support decision-making in the design phase of a product (Grieves and Vickers, 2017), but it is being used as an analysis tool of the total life cycle (Tao et al., 2019). The use of DT has the advantage of being able to understand anomalous events or unknown phenomena (Tao et al., 2018), but two major issues have not been established in academia. First, it is difficult to integrate multi-scale, multi-physics, and interfaces. DT studies should integrate models across lifecycle stages, taking into account various levels of detail and all relevant disciplines (Boschert and Rosen, 2016). There are 4 phases (design, manufacturing, service, retire) in the total life of a product (Liu et al., 2021). The units of various influence factors such as people, equipment, and systems are all different. It is difficult to build an integrated model considering their interfaces and protocols during the total lifecycle. Second, due to integration difficulties, a standardized process widely used for DT modeling has not yet been established. Existing DT studies utilize tools that are easy to implement for each phase of the total lifecycle. Many types of tools such as Predix, ANSYS, Bluemix, and MindSphere are being used. Most of the past DT studies have only integrated these multiple models into simple input/output (Liu et al., 2021). Therefore, there is a need to establish a standardized process that can fundamentally solve the mutual influence of different disciplines, time and space, and different formats and protocols. As a method to solve the two problems, this study proposes the construction of a digital twin using system dynamics (SD). In SD, it is possible to implement a complex world with a feedback loop composed of root causes (Sterman, 2010). Multiple data scales of heterogeneous physics existing on various platforms can be integrated, and multiple time horizons can be controlled, making it easy to build an integrated model. The latest SD tools have dramatically increased the reality by overcoming the limitations of traditional simulation models such as distributional assumptions, and a method to analyze the system model analytically is also provided. In particular, it is possible to support WinBurg-based MCMC in the system dynamics model, and to support interworking with programming languages ​​R and Python. Application during run-time became possible, and data calibration became easy, resulting in very high fidelity (Richardson, 2015). In other words, if a DT model is built using SD, data of different formats and protocols and multi-disciplinary interface support are possible. Since SD models for each phase of the total life have already been studied a lot, it is possible to build a DT model by integrating them. Therefore, in this study, we propose to construct a DT model using SD. The proposed method is applied to the operation and maintenance system of ROK Naval ships. This system is the service phase of its total life. It is often known that the implementation of the service phase is the most difficult. This is because the service target is decentralized and it is difficult to consider all utilization in various environments (reliability, convenience, real-time operation status, maintenance strategy, etc.). Through the model building process proposed in this study, the process of integrating multi-scales and multi-disciplinary of multi-physics is confirmed, and anomalous events or unknown latent effects are identified. The remainder of this paper consists of six sections. The title of each section (Fig. 1) is the DT construction process using SD we propose. After selecting the target system, we explore and analyze the root causes. Various methods such as statistical modeling (B-spline, Bayesian estimation, phase type distribution fitting, etc.) and simulation are applied for analysis (Chapter 2). The dynamic variables related to the analyzed root causes are implemented as a system dynamics model (Chapter 3). The models built in Chapter 3 are applied as one module constituting the integrated model. In Chapter 4, the model is integrated and validated according to the causal relationship between modules. Simulate and analyze the integrated DT model to identify potential problems and latent effects (Chapter 5). Chapter 6 summarizes the results and limitations of the study and suggests future research directions.