Recent progress and future outlook of digital twins in structural health monitoring of civil infrastructure

Within a CPS framework, a DT serves as a virtual representation of a physical system, capturing its characteristics, behavior, and state in a digital format (Batty 2018). Until now, a unified consensus around the definition of a DT has been absent due to its multidisciplinary fields of implementation. However, some popular definitions exist, such as the one explained by Glaessgen and Stargel (2012): 'an integrated multiscale, multi-physics probabilistic simulation of a complex product that uses the best available physical model and sensing information to mirror the information of its physical twin'. Reifsnider and Majumdar (2013) explained DT as a 'high-fidelity simulation integrated with an onboard facility management system, maintenance log, and historical vehicle and aircraft fleet data'. A DT includes three main components: physical information, digital information, and data exchange between them. Unlike a digital model or a digital shadow, the data in DT is characterized by its two-way directional flow. A digital model is only attributed to the manual exchange of data and does not portray the real-time conditions of a physical structure.

In general, 'real-time' refers to the ability of the DT to provide instantaneous and up-to-date insights into the structural integrity of a physical system. This entails a synchronization of data acquisition, processing, and decision-making on a timescale that is sufficiently rapid to address potential structural issues promptly. In practical terms, 'real-time' implies continuous, high-frequency monitoring of the physical structure, with minimal latency in the transmission and processing of sensor data within the DT platform. Optimally, the DT must swiftly assimilate incoming information, update its model to accurately reflect the current state of the structure, and deliver actionable insights within a timeframe critical for making timely decisions. Therefore, a delicate balance between the timescales of data acquisition, the underlying processes and phenomena that the model in the twin describes, and the urgency of decision-making most influence the real-time characteristic. DT-based SHM implementation, thus, surpasses traditional approaches by offering real-time insights through accurate virtual replicas, facilitating predictive analytics and maintenance. Integrated with key elements of the Fourth Industrial Revolution, DTs provide comprehensive information on structural conditions, enabling remote monitoring and control. This leads to cost efficiencies by optimizing maintenance schedules and supporting data-driven decision-making. The detailed models and continuous monitoring contribute to accurate simulations, improving the overall understanding of structural behavior. In essence, DTs enhance the efficiency, accuracy, and proactive management of structural health compared to conventional monitoring methods, formulating the motive of this review.

From the perspective of civil engineering, this paper generates a description of a DT for an SHM application based on the accumulated literature implementations. A DT is a virtual real-time representation of the physical structure's past, current, and prospective behavior under exterior loading conditions while considering the structure's integrity and structural characteristics in the addressed period of study. For achieving a fully functional DT, a series of components are present in its framework of development, as represented in figure 2.

Zoom In Zoom Out Reset image size

IoT highly contributes to the development of a DT for its ability to network internet-connected devices such as sensors to measure continuous real-time data from physical elements. In IoT ecosystems, a plethora of communication protocols exist. To state a few, transmission control protocol/internet protocol (TCP/IP) (Shang 2016), for example, ensures reliable connectivity, while Message queuing telemetry transport (MQTT) (Singh et al 2015) offers a lightweight and efficient publish-subscribe model. Representational state transfer (REST) application programming interfaces (APIs) (Garg and Dave 2019) offer a web-based architecture, and hypertext transfer protocol/hypertext transfer protocol secure (HTTP/HTTPS) (Da Cruz et al 2019) provide standardized protocols for data transfer over the internet, enabling seamless interaction between interconnected devices. The data is stored in data hubs prior to sending the vast amounts of information to a cloud server allowing the preprocessing and querying of data securely in the early phases of data transmittal (Lea and Blackstock 2014). Data security plays an important role in the protection of sensitive data from being lost or vulnerable to cyber-attacks (Sharma et al 2022). A meticulous selection process of suitable communication protocols is therefore needed. In the cloud server, the data undertakes various forms of processing from data learning processes using ML techniques such as DL (Bruno et al 2018), support vector machines (SVMs) (Liu et al 2020), neural networks (Kwon et al 2021), etc or condition-based assessments (Scianna et al 2022). ML is a subset of AI technology that is fed with input data to accurately predict an output. It aids in forecasting the behavior of physical information based on its current or historic state. For building the DT's base repository of data, multiple models are present to aid in accumulating the information required for achieving the end platform. The models, explained further in the following section, portray a crucial role in representing the structure's behavior, static physical attributes (geometry, material, and structural properties), and real-time sensory data. Finally, the product is projected to a visualization interface such as web-browsers, augmented reality (AR) applications, and software applications supporting 3D model visualizations. The user interface generates warnings, preventive maintenance, and predictive maintenance strategies to support the decision-making process toward the physical structure's necessities.

A plethora of modeling approaches exist in the literature, driving the development of DTs. To state some, this section categorizes the diverse techniques implemented to serve DT-based SHM applications. For example, physics-driven modeling is another technique used to simulate and predict the behavior of physical systems based on fundamental physical principles and equations. It involves developing mathematical models that describe the relationships between different variables and parameters of the system. The physics-based models used in DT can be developed using various techniques, such as finite element modeling (FEM), or other numerical methods generated from advanced structural dynamics, such as state-space models (Kim et al 2023). FEM discretizes the system into a mesh composed of finite elements (FEs). It can predict how the system will respond to various inputs of structural characteristics. This prediction is crucial for the DT to simulate the behavior of the real-world system. Data-based models are other models that leverage available data from sensors, monitoring systems, historical records, or other sources to represent the behavior and characteristics of the physical system (Qi et al 2021). By analyzing and processing the data, data-based models can capture the patterns, correlations, and relationships to create a virtual representation of the system in the DT. Some common approaches to developing such models include ML algorithms (such as regression, classification, or clustering), statistical modeling, time series analysis, or artificial neural networks (ANNs). Wang et al (2023), for instance, utilized real-time numerical simulations, FEM updating, and DL for a DT-based SHM application.

On that note, recent research applications explore physics-guided ML techniques for SHM (Rizvi and Abbas 2023). The integration of both models brings DT's behavior closer to its real-life counterpart since reaching an accurate representation of real-life conditions while using single modeling techniques separately for complex infrastructure is unachievable. The combination of both data-driven information and physics-based models involves incorporating real-world data into the simulation to refine the model's predictions. It minimizes the inaccuracies from simplified and assumed mathematical models originating from traditional physics-based modeling. For example, Yin et al (2023) have recently explored physics-guided convolution neural networks (CNNs), while (De N Santos et al 2023) relied on ANNs. The process ameliorates damage identification accuracy with uncertainties in the developed numerical model. The accuracy was substantially improved compared to previous traditional ML approaches.

Surrogate models are additional simplified mathematical or statistical models that approximate the behavior of complex systems or processes using data-based approaches (Bárkányi et al 2021). They are introduced as a substitute when the physics of a structure is not fully understood. Physics-driven models are sometimes computationally expensive and time-consuming to obtain for complex systems. Therefore, surrogate models are constructed by training the model using the available input–output data obtained from simulations, experiments, or other sources. They aim to balance computational efficiency and reasonable accuracy.

Building information modeling (BIM) has emerged as a powerful tool for enhancing project quality, reducing costs, and increasing productivity (Azhar et al 2008). With the aid of BIM, data coordination is thus made easier through the collaboration of different disciplines under one single platform. In BIM-based SHM applications, Theiler and Smarsly (2018), among many others, explained the importance of a common semantic description for a structure's lifecycle operation and maintenance phases. The Industry Foundation Class (IFC) schema was developed to ensure interoperability between BIM (Laakso and Kiviniemi 2012). The IFC file format is a semantic approach that relies on an 'Entity-Relationship' concept where physical and virtual elements are linked to their corresponding attributes. Due to the presence of common definitions and attributes for both BIM and DTs, there exists some ambiguity around the distinction between the two concepts. As articulated by Deng et al (2021) and Opoku et al (2022), BIM deals with a structure's static information, while DTs are real-time representations of physical information. BIM is an effective tool for DT application for its rich data-fed models along with incorporating wireless sensor networks, data analytics, and ML to support the development of DTs in the construction industry. The relationship between both terms is shown in figure 3.

Zoom In Zoom Out Reset image size

On another side of the spectrum, various methods such as 3D scanning, photogrammetry, laser scanning, or computer-aided designs (CAD) modeling are used in the literature to generate a 3D model of the physical structure. These models are isolated from previously mentioned modeling techniques for their use of schemas incompatible with BIM software and data querying procedures presented by the IFC. Open-source toolkits and game engines are tools adopted for modeling such models. Other models can include geographic information systems (GISs), which represents a technology that involves capturing, managing, analyzing, and visualizing spatial or geographic data of the structure and its environment (Park and Yang 2020).

The DT's leverage over the functions of the standalone modeling approaches is key for SHM applications. The real-time data and model updating characteristics drive the decision-making process and boost the inspection and maintenance processes with reduced human intervention and error. The real-time experience that DTs provide is absent in present modeling approaches (Bárkányi et al 2021), however, the power of such data-rich base models is harnessed to achieve final DT platforms. This paper, therefore, articulates all DT applications that are built on BIM, FEM, GIS, surrogate models, physics-driven models, data-driven models, mathematical models, and 3D models.

In the literature, there has been extensive research in DTs across many fields of study, such as aerospace, healthcare, education, smart cities, construction, manufacturing, automotive, energy, agriculture, and more (de Freitas Bello et al 2021, Singh et al 2022). DTs, in the perspective of civil engineering, are still considered in their stages of infancy. Moreover, SHM information has faced impediments when linked with a DT model to finally attain an automated real-time monitoring framework. The literature, therefore, demands a state-of-the-art systematic review of technical case implementations serving the development of DTs for SHM-based applications.

This section articulates existing review papers on DTs to depict the novelty of the current review paper. For example, several review papers exist regarding DT applications; however, they do not tackle the essence of SHM-oriented DT applications alone. Sharma et al (2020) shed light on the essential and elementary components of DTs and their leverage over existing technologies. The power of ML and Big Data in DT applications was highlighted, and its efficient use in multiple industries was demonstrated. Cao et al (2020) highlighted the efficiency of integrating IoT technology in improving the structural monitoring process by reviewing a series of applications. The study explored BIM applications and importance over the years and classified its maturity levels into four phases. Wang et al (2021) accumulated the DT studies revolving around off-shore wind turbines. Hämäläinen (2021) proposed a qualitative research methodology to explore the modern concept of DTs in smart city implementations. The importance and efficiency of CityGML used in Helsinki, which stored and exchanged multiple virtual 3D city models, was highlighted while stating its concluding benefits. Shahzad et al (2022) presented fields of implementing DTs in a real-life environment. They highlighted the use of DTs in smart cities, decision-making applications, manufacturing, facility management, and monitoring construction progress. Salem and Dragomir (2022) proposed a framework for developing a state-of-the-art literature review for DTs used in construction management. DT applications were classified based on their use of BIM, their use of existing monitoring technologies, and their use of AI. Pregnolato et al (2022) represented case studies of DT for multiple orientations over the whole architecture, engineering and construction (AEC) sector. Opoku et al (2022) concentrated on highlighting the motifs for adopting DT concepts, while Jones et al (2020) focused on representing the general characteristics of a DT.

Other recent existing review papers were more inclined toward SHM-oriented DT applications. The review developed by Hosamo et al (2022) specifically focused on studies that implemented 3D laser scanning for DT applications. The studies were classified into four categories: ML, bridge management system (BMS), bridge information modeling (BrIM), and 3D modeling. Saback et al (2022) reviewed the studies that integrated BIM for DT applications regarding the asset management of concrete bridges. The article divided studies into four themes: bridge inspection, BIM, DTs, and BMS. The key properties of creating a DT for BMS were highlighted, and the role of the IFC to manage and query data in a shareable file format was stressed. Hosamo et al (2022) proposed a literature review for DT applications in all of the AEC-facility management (FM) industry. Li et al (2022) discussed the implementation of advanced information technologies in phases of construction and maintenance of infrastructure. Zabin et al (2021) focused on finding papers that implemented ML techniques to data extracted from a BIM-based project. The fields of applications varied from all general stages of design, construction, operation and maintenance, and other areas related to safety, evacuation, and disaster management. Bado et al (2022) published an exploratory review regarding distributed sensing updating for creating a DT in civil engineering. The motifs of their use, the methodology for creating a DT, and the importance of distributed optical fiber sensing on DTs were highlighted throughout the review.

Previous related literature reviews of DTs have, therefore, dealt with general DT applications in multiple industries, tackled general stages of the structure's lifecycle, focused on specific types of infrastructures, and considered only specific technologies such as BIM, laser scanning, fiber optics, etc. This unprecedented review paper specifically reviews the state-of-the-art studies relating to the development of DTs for SHM for a wide range of civil engineering infrastructure. In addition, this paper extended to encapsulate all technologies used in DT-based SHM applications. Applications that incorporated the data repositories used (BIM, FEM, surrogate models, GIS, 3D representations), IoT, AI, and diverse types of sensory devices were all considered as part of the systematic review instead of focusing on one specific technology. This review, therefore, assembles the required studies for achieving prosperity in the field of SHM in civil engineering. As a result, the current systematic review provides the users with the outcome of experiencing the literature's use of technologies, their efficiency, and the attributes obstructing the creation of a fully developed DT for SHM.

Given the plethora of models used to accumulate DT-compatible information in the literature, the articles collected are classified thematically based on the types of data repositories used for building DT-based SHM applications. In the case of achieving DTs, the methods adopted to process acquired data are crucial in the framework. In the end, DT aims to twin the physical structure with the least uncertainty, computing power, and processing duration. Therefore, the entire methodology is highly sensitive to the type of data processing chosen. On that account, a corresponding table is generated to encapsulate the key data collection types and techniques along with the type of analysis used in the framework at the end of each section. The proposed categorization will help the readers to easily pinpoint the case study of interest by navigating to the corresponding section dealing with that specific type of infrastructure from all relevant papers' findings.

3.1.1. Bridges and tunnels

3.1.2. Roads and pavements

3.2.1. BIM approaches.

3.2.2. Physics-based modeling approaches.

3.2.3. Hybrid modeling approaches.

3.2.4. Surrogate-based modeling approach.

3.2.5. 3D representation modeling approach.

3.3.1. BIM approaches.

3.3.2. Physics-based modeling approaches.

Extending outside the scope of specific infrastructure's uses of DTs, there exist case implementations on other structural or non-structural civil engineering components in the literature. This section covers case studies on elements such as structural beams, cable domes, steel structures, bar structures, plate specimens, non-structural civil engineering utilities, and single-degree-of-freedom systems. For example, D'Amico et al (2019) focused on introducing a framework of a DT for a general complex engineering system. A central unit bonded all the modules via a common language through an attached interface, and no particular modeling approach was represented. A materials degradation assessment DT contained a CAD model to retrieve geometry and material data; it relied on ThermoStudio to process images. The results were analyzed in the degradation characterization module and transformed into JSON files to be sent to the learning module for predicting degradation.

3.4.1. BIM approaches.

3.4.2. Physics-based modeling approaches.

3.4.3. SM approaches.

Throughout the literature, it is well established that BIM has revolutionized the design and construction phases of an infrastructure's lifecycle (Chen and Luo 2014). Recently, BIM has been explored as a data-rich repository for extending the monitoring process beyond the construction phases. The operation and maintenance phase occupies up to 80% of a structure's lifecycle (Cholakis and Hunt 2017) and has become an increasing topic of interest, especially with the introduction of DTs to the industry. Within the reviewed case implementations, (Grosso et al 2017, Boddupalli et al 2019, O'Shea and Murphy 2020, Angelosanti et al 2021, Oreto et al 2021), to state some, have successfully mitigated the limitations of the IFC schema to integrate data-driven sensory information into the static database. However, long-term SHM data generates unwieldy data sets, such as several thousands of data points within time–history intervals. The studies still rely on manual data extraction from offline and external databases, files, and hyperlinks to feed the BIM repository with limited visualization capabilities within the IFC format, obstructing the full functionality of a DT.

A crucial characteristic of a DT is the ability to transfer data into an online cloud server accessible by several parties regardless of the time and location of login. Such permissibility is granted and eased with IoT technology. In many cases (Davila Delgado et al 2018, Tygesen et al 2018, Karve et al 2020, Liu et al 2020, Ye et al 2020, Yin et al 2020, Lu et al 2020b), the use of IoT is disregarded, and the data collected to feed DTs remain accessed from offline local servers and PCs. Some studies (Wang et al 2017, Cheng et al 2020, Hou et al 2021, Mahmoodian et al 2022, Scianna et al 2022) integrated IoT in their framework and have dealt with the issues of automated data transfer through a centralized hub. However, SHM data extends beyond the high-speed transfer of raw data. The processing units present on portable IoT sensors face limitations when dealing with the transfer of data with minimal latency, especially when introducing multiple sensors that exploit the wireless transmission bandwidth capabilities. The studies in the literature focused on integrating IoT without tackling these limitations for achieving a near real-time monitoring process. In addition, the data collected ought to undergo several types of data processing such as data filtering, modal analysis, and frequency-domain analysis in real-time specifically along with the simultaneous transmittal of raw data.

The ultimate goal of a DT platform focuses on granting the necessary information regarding the physical structure's behavior and integrity within the immediate request by the end user. In the literature, IoT technology and sensory data integration within a static data repository are being explored. However, the algorithms and procedures chosen for processing the data for outputting fundamental system identification play crucial roles. For instance, the phase of data processing consumes time: either the run of the algorithm itself or the wait on gathering enough input data as time-histories to feed the algorithm. In both cases, real-time monitoring through the DT is affected, and data latency is introduced. In addition, a DT-based SHM application is automatable, however, to a certain extent. Throughout the literature, continuous data acquisition through sensors and initial analysis undertook successful forms of automation, ensuring a constant flow of information about the structural conditions. ML algorithms play a role in automating the detection of patterns and anomalies in the collected data. However, aspects like model updating and certain decision-making processes still require human intervention, particularly for more complex adjustments and critical judgments related to structural integrity. The process has not yet achieved a fully automatable characteristic and requires multiple skills and expertise from various industries of software engineering, structural engineering, engineering management, equipment specialists, and inspection to attain its full potential.

Several articles in the review dealt solely with the theoretical approach to create a DT platform for SHM and remain not working cases (D'Amico et al 2019, Ye et al 2019, Wagg et al 2020, Sanfilippo et al 2022, Tian et al 2021, Wenner et al 2021). Propositions must become case studies while stating the limitations that accompany the use of novel technologies. It is important to note that a significant number of advancements in the field often remain in the theoretical realm due to data privacy concerns, hindering their public sharing. The intricate and proprietary nature of algorithms, modeling techniques, and sensor integration strategies employed in creating DTs for SHM often leads companies to safeguard these innovations as closely held intellectual property. Fear of compromising competitive advantages or exposing proprietary methodologies can discourage entities from openly sharing their practical DT applications. Consequently, the field faces a challenge in realizing its full potential, as the reluctance to disclose key details impedes the widespread dissemination of proven, real-world implementations. Striking a balance between protecting intellectual assets and fostering collaborative efforts is crucial to overcoming such barriers and facilitating the broader adoption of practical applications for the benefit of the industry as a whole.

A two-way data flow focuses on achieving the mutual benefit of feeding real-time information to the DT and also inducing countermeasures to enhance the structure's physical aspect based on the data. In the literature, reactive maintenance, which deals with the notion of fixing when broken, and preventive maintenance, which deals with scheduled fixing to prevent breakdowns, have been achieved using a DT. However, predictive maintenance which deals with predicting structural behavior and health, is still feeble in implementation. For example, the researchers (Anil et al 2016, Valinejadshoubi et al 2017, 2018, 2019) successfully managed and queried physical information necessary for a DT's success; however lacked the elements of post-processing the data for future model updating, learning, and insightful decision-making. Other researchers (Wang et al 2017, Scianna et al 2022) made use of threshold values to generate warnings using the BIM model, yet the model does not self-update and does not predict the future behavior of the physical structure, a fundamental characteristic for a platform to be classified as a full DT.

As shown previously, the amount of software development under the framework of DT applications is vast. However, most software adopted in the literature remains custom-developed with the use of APIs in trial modes, inaccessible to the public, requires skillful programming expertise, and is limited to specific applications. Institutes such as Unity and Microsoft Azure have embarked on unlocking the technological tools for achieving future DT platforms and standards for users to follow. However, such propositions are not adequate to address SHM applications using DTs, and significant knowledge of adopting existing software to specific objectives is needed.

As illustrated in section 4, the latency of DT with near real-time information remains a challenge. Initiatives to explore data processing techniques and system identification ought to be undertaken while exploiting short-duration data inputs and fast processing algorithms to minimize the twining period of physical to digital assets.

The metaverse refers to a collective virtual shared space where people can interact with digital objects in real time. It is often portrayed as an immersive, multi-dimensional space that combines elements of VR, AR, and the internet (Huang et al 2022). The metaverse encompasses a vast network of interconnected digital representations, virtual environments, and user experiences. The potential integration of DTs into the metaverse as interactive entities within the virtual space creates a smart monitoring platform interconnecting multiple DT structures with their environment, realizing the full capabilities of studying entire city behaviors.

State-of-the-art IoT technology has successfully dealt with transmitting near real-time data into DT models. However, future research must focus on developing multi-modal sensors that can capture various structural parameters, exploring energy-efficient IoT solutions for SHM and prolonging sensor lifespans, enabling real-time analysis and decision-making at the edge of the network via edge computing, and enhancing the reliability, scalability, and connectivity of IoT networks in SHM through new communication protocols, network topologies, and data transmission methods.

The emergence of more case studies leads the industry to create a unified consent on the standards and protocols a DT should follow. Based on multiple trials, failures, and successes, DT applications prosper and highlight the gaps. An abundance of examples and implementations fixates the researchers' focus on tackling the weaknesses in the field of interest.

Research can dive into developing practices, standards, and guidelines for implementing DT, addressing concerns related to data integration, interoperability, security, and privacy. Collaborative research endeavors involving academia, industry, and policymakers can play a pivotal role in driving the necessary mindset shift, creating a conducive environment for embracing DTs as a powerful tool for the future of industries. This endeavor requires efforts to educate and raise awareness among industry stakeholders about the transformative capabilities and benefits of DT-based SHM applications. Research must focus on showcasing successful case studies and demonstrating tangible value across the field.

The trifecta of high data quality, precise sensor calibration, and meticulous consideration of measurement uncertainty stands as a linchpin for the reliability and efficacy of decision-making processes. Given that sensors are often deployed for extended periods on critical structures for long-term monitoring, the financial and safety implications of decisions based on the collected data are substantial. Therefore, ensuring a balance between deployment costs and high data quality is paramount. Precise sensor calibration guarantees accurate measurements, contributing to the fidelity of the DT's representation of the physical structure. Equally critical is addressing measurement uncertainty and acknowledging the inherent variability in sensor readings—a phase often absent in case implementations. Decision-makers must have confidence that the data is not only accurate but also sufficiently robust for the specific task to be executed.