Towards AI-assisted digital twins for smart railways: preliminary guideline and reference architecture

The Digital Twin is a dynamic and self-evolving virtual replica of a physical object and/or process, characterized by a bi-directional seamless communication that allows real-time data sharing between physical and digital world [7, 21, 63]. With DT technology, it is possible to implement several services (e.g., simulation, monitoring, prediction).

Despite DT informal introduction can be dated back to 2002 during Michael Grieves’s presentation about product lifecycle management (PLM), at present there is still a lack of a formalized definition and a standardized architecture for DT implementation [7, 38, 59]. Nonetheless, one of the most recurring definitions in academia and industry is the five-dimensional model [51, 59], reported in Fig. 1, according to which a Digital Twin Model (\(M_{DT}\)) is described by the following quintuple:where the Physical Space (PS) includes systems and/or processes and their internal and external interactions; the Virtual Space (VS) contains the digital replica fed with Digital Twin Data (DD), i.e., operational, environmental or historical data; the Services (Ss) are the applications that can be realized leveraging the DT technology (e.g., predictive maintenance); Finally, there are the Connections (CN) that create the communication link (data and control) between the four parties.

A DT is often depicted as a “mixture” of Industry 4.0 technologies, as it owes its industrial and academic diffusion to the massive exploitation of innovative information and communication technologies [49, 51]. For example, Internet of Things (IoT) and Industrial Internet of Things (IIoT) are among the enablers of the seamless communication between physical and virtual replicas: sensors capture events or environmental changes that can be further processed and analyses; actuators allow the physical system to respond to a set of inputs or commands received from the virtual replica. Data exchange is clearly achieved through communication technologies, such as MQTT, 5G/6G networks and so on.

Finally, Virtual and Augmented Reality technologies (VR/AR) better allow users to interface with the digital replica, thus enabling the human–machine interfacing with a computer-generated 3D visual environment.

As pointed out in Yokoyama [67], 21st-century innovation in terms of services for customers and technology supporting railway operations have enabled the shift from time-based maintenance to condition-based maintenance, leveraging large volumes of data directly acquired from monitoring to predict deterioration and failures of rolling stock and electrical equipment. The main maintenance paradigms and related actions can be distinguished in the following three categories [73]: corrective maintenance occurs when the fault is detected; preventive maintenance programs the repair actions at specific times according to past experiences; predictive maintenance (PdM) schedules the repairing activities predicting the failure occurrence through asset data analysis. According to Errandonea et al. [22], predictive maintenance also provides plans or strategies for taking action when the fault is predicted. Major benefits of PdM are maintenance costs and downtime decreasing, productivity and quality increasing.

The approaches used for prediction and thus for predictive maintenance are [22, 73]: (i) physical model-based, in which mathematical modelling reflexes component conditions; (ii) data-driven, relying on statistics-based, pattern recognition or AI models based on large amounts of data collected from the asset; (iii) knowledge-based approaches, aimed at reducing the complexity of physical modelling (typically used as a hybrid strategy).

The DT ecosystem provides a great opportunity for incorporating predictive models to evaluate the current state of a physical asset, analyse historical and operational data obtained from the sensors of the physical asset, and finally predict the degradation of a specific component.

Machine Learning (ML) applied to railway DTs mainly focuses on predictive maintenance [14, 40, 53] as a natural application of the former to the latter. Liu et al. [40] propose a framework enabling the application of industrial AI technologies for a Prognostic Health Management system for high-speed railways. The authors discuss the use of dynamic clustering of historical data for the identification of health conditions. Specifically, in a case study on traction motor condition monitoring, they discuss the use of an Artificial Neural Network for modelling the vibration between sensors that are close to the four wheels. Similarly, Consilvio et al. [14] propose a generic framework and build a Decision Support System by combining ML, data analysis, and simulation techniques for analysing real-time data to make automated decisions. The manuscript takes into account two case studies, one at the strategic level and the other at the operational level. At the strategic level, the authors describe a data-driven model based on the K-means clustering method, Petri net models, and Monte Carlo simulations for decision support in order to evaluate a sustainable mix of interventions, such as renewals and refurbishments. On the other hand, at the operation level, they describe a data-driven model based on a one-class support vector machine, the Bayesian network approach, and a linear programming model in order to improve maintenance operations and minimise costs caused by sudden failures and service disruptions.

Sahal et al. [53] introduce a conceptual framework to fulfil the DTs collaboration requirements and to enhance an intelligent DT based on blockchain technology and predictive analysis. By using distributed ledger technology, several DTs can be connected through the blockchain network and provide data analytics and management across the railway sector. To minimise maintenance costs, the authors explain that the framework consists of predictive models that use data-driven ledger-based historical DT operational data and a distributed decision-making algorithm for improving the DTs collaboration. The authors argue that the framework can predict potential failures from rail-based operational data by applying ML algorithms and training the offline predictive model with this data. Then, an online predictive model can be evaluated to predict the failures in railways. The outcome of these models serves as input for the consensus algorithm to make the best decision for data anomalies or to help decision-makers.

One of the most important issues in DTs development is their design and generation. Ariyachandra and Brilakis [4] argue that generating a DT of an existing railway from its Point Cloud Data (PCD) is a time-consuming and tedious task. Automation processes of this task face the challenge of detecting masts from air-borne LiDAR data since masts when scanned from above, are visible as thin lines. Thus, it is difficult to differentiate objects of similar shapes, such as signal poles, from masts. The authors propose a method that begins with tools that clean the PCD and identify its positioning, then, the resulting data sets are processed and masts are detected using the RANSAC algorithm. To summarise, the article focuses on DT generation to save time and modelling costs and proposes a method to enable the rapid adoption of a geometric DT.

Lastly, Zhou et al. [71] introduce a DT framework for automatic train regulation and control with the goal of reducing train delay times and energy consumption. The authors achieve this goal by using a Convolutional Neural Network (CNN) to map the relationship between the train running time and the energy consumption of the speed profile and switch points.

Jiang et al. [34] discuss the railway industry’s demand for on-site data collection, control, and management. The authors propose a monitoring platform design and architecture for intelligent high-speed railways. Then, they also discuss the DT architecture including aspects such as the collection of real-time data from the sensors and the integration of sensor monitoring data to provide optimal solutions for failure handling. The aim of the DT is to develop a high-intelligence stage to intervene during abnormal working conditions through automatic monitoring and to prevent accidents.

Similarly, Errandonea et al. [23] propose an IoT approach for intelligent data acquisition to generate DTs in the railway industry with the goal of creating an onboard system for maintenance prediction in trains. The authors detail the architecture of the approach in three aspects: (1) The communication module for near-real-time communication and batch information transfer; (2) The functional design of the system, where the processing and integration of the sensor data are done in modules so that the system remains independent; (3) Data ingestion technologies, where the authors make use of Apache NIFI since it allows different data sources and parallel executions.

Du et al. [20] introduce a DT framework and its implementation method for urban rail transit. The proposed implementation method for the DT consists of four components, where: (1) the sensor network collects data; (2) onboard data is saved to a ground data storage device; (3) an analysis process is performed for mining and visualising data; and (4) AI algorithms are used for data mining, evaluating and predicting faults. The authors point out the necessity to pre-process the data by cleaning, integrating, transforming and reducing the data before the actual data mining process. For the analysis, the authors suggest using k-means when the labels belong to an unknown category, and for the classification of known data, they suggest using KNN, regression classification, neural networks, or naive Bayes.

Avizzano et al. [5] introduce an algorithm for reconstructing rolling stocks from a sequence of images into an image model. The aim is to be able to generate DTs for complete rolling-stock vessels to serve for monitoring, diagnosis and failure prediction. The authors combine Kalman Filters, Gaussian mixture models and specific algorithms to track images of a train, which were taken by a fixed camera, so to provide a light and robust tracking system for image stitching.

To enable DT to analyze and exchange data from heterogeneous systems in real time, interoperability is essential. Dimitrova and Tomov [17] argue that current railway networks are managed by different tools and devices, resulting in too heterogeneous and non-integrated data, which makes interoperability and decision-making especially challenging. Sahal et al. [53] discuss the DT interoperability challenge and propose a data-driven ledger-based predictive model, which ensures intelligent and secure interoperability.

Despite the large adoption of IoT sensors in different sectors, connectivity challenges still exist especially when dealing with real-time monitoring. When trying to connect all of the sensors in a railway system at the same time, it could face some obstacles and difficulties in obtaining data due to the variety and mutability of (harsh) environmental conditions including, among others, storms, freezing, and extremely high temperature which may disrupt connectivity and/or cause loss of data. To tackle this obstacle, Sahal et al. [53] discuss communication technologies such as Beyond Fifth Generation (B5G) and 6G as opportunities for handling connectivity among the increasing number of smart devices in intelligent transportation systems.

To be efficient, DTs require well-defined frameworks taking into account the IoT paradigm, data analytics techniques, and many other enablers. Furthermore, within all forms of DT development, also the modelling seems to be still an open issue given that, to the best of our knowledge, there is no standardized approach to model DTs. Standardized procedures enable domain and user comprehension, as well as information flow between phases of a DT’s development and implementation. In a such view, Gan et al. [30] discuss the main challenges to developing standards for Industry 4.0 and achieving agreements with the goal to connect the entire supply chain and stakeholders. In Borjigin et al. [10], the authors discuss the evolution of BIM towards an international standard. BIM is described as a method using a shared digital representation of a built asset, which progresses to become a DT. In that context, the authors stress the need for collaboration and shared information among different stakeholders at various stages, so that more accurate decisions can be made.

In the railway sector, DTs can improve operational efficiency, reduce costs, and enhance the passenger experience. However, the lack of standardized frameworks and guidelines for implementing DTs in this sector can make it challenging for railway operators and manufacturers to adopt and implement DT solutions. Therefore, it is crucial to develop standardized frameworks and guidelines to ensure consistency, interoperability, and security when implementing DT solutions in the railway sector. Krishna et al. [36] tackle the lack of frameworks by introducing a train-track multi-body simulation model to assess long-term rail surface damage evolution in accordance with European standards. The framework can be turned into a DT and provide guidance for infrastructure managers regarding the condition of the rail surface. Moreover, Shift2Rail³ and CENELEC WG⁴ are initiatives that recognize the significance of DT in this sector and aim to promote the development and implementation of innovative solutions in a coordinated and standardized way. Shift2Rail focuses on integrating new technologies and solutions into the railway sector, including those related to DT, while CENELEC WG has developed standards related to railway applications, signalling, and telecommunications, providing a framework for DT solution development and implementation. Collaborating with these initiatives to develop and implement DT solutions in accordance with established standards can help ensure interoperability, consistency, and effectiveness in the adoption and implementation of DT solutions in the railway sector.

Data privacy and security are other relevant challenges when it comes to railway systems, especially safety-critical ones. The main reasons are related, but not limited, to the fact that data exchanged (e.g., between trains and a zone controller) are extremely critical and any intrusion could potentially cause considerable damages both in terms of costs and human lives. Hence, to overcome this issue, DT-enabling technologies must meet existing security and privacy policies and legislation. Along this line, Boockmeyer et al. [9] point out the importance of IT security and suggest the use of appropriate encryption techniques and software update strategy.

Scalability refers to the possibility of integrating and managing multiple DTs at the same time since (within manufacturing systems, as an example), it may happen that there are multiple DTs deployed each of which virtually represents a specific machine or workstation. In this context, Sahal et al. [53] identify in the blockchain technology a suitable solution for scalability since large amounts of DTs can work together in a hierarchical and granular manner.

The first task lays the groundwork for the design and the implementation of a DT, providing the primary information upon which the rest of the steps will be based. First of all, it is needed to identify for which industry the DT will be developed [50] and this is significant as some industries (e.g., railway, aerospace) have to follow some regulations. Step two aims to extract DT functional requirements and thus to define its purpose, such as optimization, prediction, monitoring, security assessment, development improvement and so on [56]. The following step is the identification of the process or asset to digitally replicate: it may occur that the Physical Twin (PT) does not already exist; in this case, a prototype of the PT can be built to guide the construction of the actual DT [56]. For example, in the railway context, in the design phase engineers can start the creation of the DT of a train to produce an accurate computer model of every single part (inside and out), but also to carefully plan the sensing layer (e.g. sensors, IoT devices, and asset and facility management system). In general, using a DT in the asset’s design stage helps designers and engineers to model, simulate, and conduct what-if scenarios analysis to improve and optimise the asset design. Once the physical counterpart leaves production, the DT can also help engineers to monitor it, increasing not only production management and work performance but also the health, safety, and wellbeing of workers and materials [56].

Once the global goal of the DT-based application has been identified, it is needed to define what kind of use the DT is expected to deal with [50]. Indeed, we should at least define the DT usage frequency, i.e., if the DT will be used when needed or continuously or even in hybrid mode, and the DT nature, i.e., static or dynamic, and thus if the DT will adapt to initial conditions alterations or not. In fact, even if the DT definition that we presented in Sect. 2.1 describes the DT as a dynamic and self-evolving virtual replica, a DT can execute in two modes:Further steps are the identification of technologies needed for DT development and specification of input and output parameters of the PT and so of the DT, according to each use case. This will increase the DT understandability by a broader audience and the re-usability of DT in other scenarios [50].

As the DT is typically a virtual replica of only a part of the physical process/asset [7, 38], the purpose of process planning is to determine the functionalities and properties of the physical twin to be modelled in the DT. In some cases, the DT may be an identical copy of the PT, but this one-on-one replication could most probably bring more complexity and redundancy [56].

In this task, the other two objectives are: i) the definition of the comprehensive set of data that suitably describes the physical twin [56] (e.g., environmental, historical, real-time or near real-time operational data) and that are used to tune, validate and eventually update the DT; and ii) the design of the DT and PT communication infrastructure. Depending on requirement specifications, for example, a 5G, B5G, or 6G communication technology could be taken into account.

As pointed out in Sect. 4, there is no consensus about DT components and corresponding architecture, but software architectures are cornerstones for the set-up and engineering of software complex systems such as DTs [15, 26]. For this reason, it is crucial to perform the architectural design task by focusing on both the architectural pattern choice and DT components identification steps.

Although there is a lack of a widely accepted architectural solution for DTs, according to Ferko et al. [26], the majority of the academic and industrial proposals are built using the layered architectural pattern or the combination of layered and service-oriented patterns. Furthermore, regardless of DT-based application, it is possible to identify a set of strictly needed DT components [55]: i) the Models that digitally represent the physical process/asset; ii) an Even Source whose objective is to send information and/or commands to the PT; iii) a set of algorithms (e.g., AI) provides the reasoning capability that aims to extract useful information which feeds DT models and the event source element.

Finally, if DT nature is dynamic, the architecture is expected to evolve over time to integrate system changes (e.g., new components), to adapt the system behaviour, and also to support the interconnection with other DTs [18, 56].

This task is entirely dedicated to the modelling activity since PT models are the core elements of the virtual replica. It is mandatory to reproduce the physical properties, geometry, behaviours, and rules of the physical system as faithfully as possible. According to Segovia and Garcia-Alfaro [56] and Schroeder et al. [55], it is possible to distinguish two modelling techniques:

The objective of this task is the set-up of the whole DT infrastructure, starting with the generation, tuning and validation of PT models [50, 56]. Indeed, in order to obtain a comprehensive view of the physical system, DTs are characterized by different kinds of models previously listed. Thus, in this task, it is required to integrate hierarchically or collaboratively models and other DT components (which and how many depend on the use case). It is worth underlining that deciding whether to automate the building of the DT or rather to construct it manually are both feasible choices: for example,Ariyachandra and Brilakis [4] generates a geometric DT (gDT) from its Point Cloud Data.

This task aims to design a service that leverages DT technology. Thus, after DT design, in our work, we suggest defining the predictive maintenance model and choosing whether to perform it with supervised or unsupervised learning. Notable implementations relying on supervised learning algorithms exploit Neural Networks [71], Support Vector Machines or Naive-Bayes [14]. Similarly, unsupervised learning (such as the K-means algorithm) has been used for anomaly detection [14]. Regardless of the chosen learning algorithms, this task will lay the basis for the implementation of the DT-based predictive maintenance service.

Once the set-up of DT is completed, the virtual replica has to be connected to the physical twin through the seamless bidirectional connection that describes the DT itself, with respect to the communication infrastructure previously designed. Thus, the sixth task aims to create the data flow, i.e., the stream of environmental and operational data coming from the physical world combined with historical data. For example, in a railway context, operational data are the position of the railway [20], the temperature of the motor [11], or the vibrations [17], environmental data are the different weather conditions or humidity. All data will be collected and stored using one or more storage solutions, such as cloud services or a local server with Internet access.

The objective of the conversion task is to process, clean and eventually fuse and reduce data due to their multi-temporal, multi-dimensional, multi-source and heterogeneous nature [7, 15, 39].

After collecting data from the physical system for the ML-based problem (i.e., predictive maintenance) and properly treating them with AI techniques, it is possible to build the predictive model training and deploying it. At the time of writing this paper, the most used DT platforms for predictive maintenance applications are OpenModelica and MathWorks Simulink [61].

This task operates when ML-based predictive maintenance model has been deployed. The algorithm executes using real-time data continuously coming from the physical twin and a visualisation of the prediction.

Once the insight is obtained and a fault is predicted, the decision process task designs the set of actions to be performed on the physical twin. The DT can be used to support human decisions, perform automatic repair actions, or assign tasks to the right personnel [40].

The last task is the evaluation which can be done through the generation of a proof-of-concept. Indeed, although model validation metrics for AI-based models are overall accuracy, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and so on, different works [7, 26] have demonstrated that evaluation of DT-based solutions must almost always be carried out through examples. This is due to the lack of DT-specific evaluation measures because of the heterogeneity of its components and its dependency on the physical system.

This layer interconnects the physical asset and its virtual counterpart (the DTs) by means of suitable sensing and actuating capabilities. The (IoT) sensors monitoring the physical asset collect different kinds of operational data which are then sent to the Data Layer for preliminary elaborations. Actuators, on the other hand, physically execute the decisions received from the upper layers.

The Data Layer is in charge of managing both the row data collected through sensors at the PL, and the information retrieved through the elaborations performed at the upper layers, i.e., the Virtual Layer (VL) and Service Layer (SL). This is achieved through a fundamental knowledge management capability that leverages shared knowledge. At the PL, depending on the asset and the choices made within the Requirement Specification phase of the guidelines proposed in Fig. 3, sensors may collect many heterogeneous data (e.g., vibration, acceleration, video, audio). We can subdivide these data into Primary Data (PD) and Secondary Data (SD). By PD we mean those data that, after some basic elaboration that does not involve any AI application (e.g., cleaning, transformation), can be directly adopted at the upper layers to model the DT and perform the necessary analyses. Differently, by SD, we mean data that need to undergo an “advanced” pre-processing to be used at the upper layers. As an example, data captured by a camera may be processed by suitable AI approaches to extract motion features (i.e., new data) that can be used in the VL and SL for modelling and/or analysis. Capabilities that enable to extract both PD and SD are referred to as basic elaboration and AI pre-processing, respectively. It is worth noting that the integration of AI pre-processing capabilities is very helpful in a context such as a railway system. Indeed, most railway assets are safety-critical systems which must be compliant with strict standards and regulations to be deployed; intrusive modifications (e.g., the physical installation of new sensors on the deployed system) may lead to time-consuming and expensive (re)certification processes [16]. New assets, recently deployed, may already include (by design) all the sensors needed to produce their digitalized version (i.e., a DT); however, issues may arise in the case of legacy assets for which sensorization and digitalization might not be viable. The possibility to adopt AI pre-processing capabilities to elaborate data from non-intrusive sensors (e.g., cameras, microphones, LiDARs) and use these data to potentially digitalize any kind of railway asset would avoid the aforementioned issues. To conclude, the DL includes a presentation capability related to the Human Machine Interface (HMI), devoted to displaying information to human operators.

This layer includes the core capabilities required to create and manage the virtual representation of a physical asset. The core capability within the VL is modelling, which typically encompasses the set-up and management of different models, characterising the structure and/or behaviour of the different parts of the asset. Besides “traditional models” (e.g., geometric, mathematical), AI can be adopted to build a behavioural model leveraging learning processes capable of extracting hidden information from data. In particular, system models may be built from historical data; for example, taking into account an ML approach, a model may be trained, validated, and tested on historical data and then used (without any modification) for simulations or predictions. An alternative approach may adopt, for instance, online machine learning [32] to allow ML models to evolve in real-time with the on-line stream of data. The latter approach, despite being potentially more appropriate to keep the digital model constantly updated, is quite more challenging especially when it comes to Deep Learning applications most of which, notwithstanding their potential, are opaque (i.e., it is not always possible to understand their functioning and how the mapping between input and output has been performed) and unstable (i.e., to slight variations of the input may correspond heavy variations on the output). It is worth noting that it is already extremely challenging to demonstrate the dependability and trustworthiness of Deep Learning models [28] built starting from a specific and, if possible, reliable set of historical data. On-line learning would introduce additional complexity that would make this demonstration process even more complicated to address. A suitable compromise would be to periodically update the AI models (manually or automatically) so as to have the possibility to properly integrate new data and re-validate and re-test the new versions of the models.

This layer implements the forecasting and decision-making processes that can be attributed to DTs. In our proposal, it includes some capabilities that are specific to predictive maintenance, while additional capabilities may be considered depending on the specific application (e.g., simulation, monitoring, anomaly detection, etc.). The inference/prediction capability represents the forecasting logic based on the selected prediction models. In our conceptual architecture, we have explicitly considered a RUL Prediction Modelling capability to set-up the models for the estimation of the remaining useful life (RUL) of an asset, as it is a key indicator identified by the research community to determine the maintenance time [54]. Based on the run-time prediction and other information stored in the shared knowledge, the maintenance scheduling capability is responsible for optimizing the maintenance operation scheduling. Finally, alerting capabilities must be typically considered to support maintenance.