Systematic review of predictive maintenance and digital twin technologies challenges, opportunities, and best practices

Identification

The process of identifying and selecting relevant studies consisted of four major steps. First, the keywords for each topic area were identified. Thesauruses, encyclopaedias, and prior investigations were utilised in order to discover suitable key phrases. Second, as provided in Table 3, in November 2023, search algorithms were developed using the keywords and characteristics of the eight databases, depending on the title of the articles. Numerous articles from various sites covering related search terms might be found. Among the keywords utilised in the choosing process are “equipment”, “machine”, “maintenance”, “smart maintenance”, “predictive”, “digital twin”, and “virtual model”.

Furthermore, this review emphasises the research publications to evaluate the methodology, data, and conclusions of the original studies that align with the purpose of this review. As a result, 278 articles from Scopus and 164 articles from Web of Science were discovered, and 5,242 items were determined in the six additional databases. Science Direct, IEEE Xplore, Emerald, Medline Complete, Dimensions, and Springer Link are the other six databases. Other methods, such as websites, organisations, and citation searches, were used to find pertinent studies. For the other methods of this study, the authors use the website Google Scholar. It is one of the other alternatives as a search engine, in addition to the eight databases used in this study. A total of 5,699 references, including articles and reports, were obtained after fifteen references were discovered by utilising similar keywords.

A software programme called VOSviewer is used to make maps from network data and to visualise and explore those maps. The VOSviewer produces a cluster, a collection of items contained within a map, while a network is a collection of items and the connections among the items in a map. Other than that, different colours denote the collection of items on a map and are used to distinguish the clusters. The boxes in the maps represent the object of interest, and the connections between two distinct objects are indicated by the link in the connection. A line’s width denotes the strength of a link, while a longer length denotes the availability of more publications online.

The visualisation map exhibits the terms or keywords present in the data file, utilising the clustering methodologies accessible in VOSviewer. The strategy is founded upon mapping and clustering, both of which operate on the identical fundamental principle (Bukar et al., 2023; Waltman, van Eck & Noyons, 2010). By examining the connections between the keywords in a variety of publications, co-occurrence analysis aims to identify timely issues and assist scholars in comprehending contemporary scientific concerns (Yin et al., 2022).

The crucial cluster or strong connection is indicated in red in Fig. 2, which provides an overview of PdM and DT research’s six distinct association clusters. The colour scheme is used to identify six distinct clusters: red denotes seven study subjects, green and blue for six topics, purple and yellow for five topics, and light blue for the smallest cluster, which has four themes. The primary subjects covered by the largest cluster in Fig. 2 include data analytics, Industry 4.0, maintenance, decision support, etc. Thus, these phrases will be used as keywords in this study’s investigation of these crucial subjects.

In Fig. 3 for the DT study, Cluster 1 in red consists of three items: data analytics, Industry 4.0, and manufacturing. Meanwhile, Cluster 2 in the purple colour indicates one item of ML. The following group of clusters, namely Cluster 3 in the colour yellow, with four different items, comprises decision-making, facility management, the Internet of Things, and artificial intelligence. Cluster 4 in light blue represents three items: real-time control, smart maintenance, and an information system. Cluster 5 in blue stands for five items: predictive maintenance, deep learning, anomaly detection, data-driven systems, and neural networks. Finally, Cluster 6 in green stands for two items: real-time fault diagnosis and remaining useful life.

The two items in Cluster 1 (highlighted in red) of Fig. 4 for the PdM study are manufacturing and Industry 4.0. Consequently, two items are indicated by Cluster 2, which is coloured purple: data mining and machine learning. Artificial intelligence, facility management, the Internet of Things, and decision-making are the four items in Cluster 3, which is coloured yellow. The single point represented by Cluster 4 in the colour light blue is the digital twin. On the other hand, six items depicted by Cluster 5 in blue are neural networks, data-driven, outlier detection, anomaly detection, and deep learning. Lastly, the four items represented by Cluster 6 in green are fault diagnosis, remaining useful life, LSTM, and k-means.

These results indicate a lot of information mapped under the PdM and DT. All these important details on a wide range of subjects available for thorough comprehension will be covered in more detail in the results section.

Screening

Among 5,699 references, 5,684 were obtained using databases and registries, while 15 were discovered using databases and other methods (see Fig. 5). Duplicate references and unrelated topics were separated from the obtained references. The databases contained 387 duplicate articles, and the other approaches yielded ten referrals. After deleting redundant references, 5,297 articles for the databases and five for the other approaches remained. Duplicate references were eliminated, and the remaining references were then carefully inspected by observing the title, keywords, and abstract.

Some criteria were also considered in other ways. First, the title and keywords included the general terms PdM, DT, and equipment. Second, a reference to the quantitative method for evaluating PdM performance using DT. From the 5,297 articles in the databases, 5,249 were removed due to inappropriate content to the themes, leaving 48 articles pertinent to the study. However, 16 of these 48 articles were eliminated due to inaccessible full-text articles in databases, and three of five were excluded using other methods. As a result, only 32 articles from databases and two reports on other approaches were discovered and chosen for the next phase.

Eligibility

The whole text of the publications was reviewed in this step to ensure that the 32 research articles and two reports were synthesised and analysed. The articles’ essential material was thoroughly evaluated to meet the inclusion and exclusion criteria. The study’s goal, methods, and research conclusions were all thoroughly assessed. As a result, no database articles or descriptions of other methodologies were included in these investigations. A hand search resulted in the inclusion of two additional pertinent articles. Consequently, 34 additional publications were included in this analysis.

Quality evaluation and data extraction

To ensure the publications’ suitability for research, the authors conducted a quality assessment of the ones they had chosen. The remaining articles were assessed using the qualitative analysis method. The research purpose, research technique, contribution, and article highlights were among the criteria used to evaluate the publications. The authors also used the information they excluded from the publications for additional synthesis and analysis. An organised table was created to compile the retrieved data. The authors then went over all the synthesised data. The synthesised data were divided into categories using thematic analysis (Braun & Clarke, 2006; Guest, Namey & Chen, 2020; Nowell et al., 2017). The authors extensively explored the classification of the input properties. The authors reached a consensus on any differences or contradictions prior to reaching an agreement.

Physical model

A PM is used to simulate physical qualities and loads (van Dinter, Tekinerdogan & Catal, 2022). The PMs are descriptive models. Descriptive models are primarily used for comprehension, prediction, and communication (Kaul, Bender & Sextro, 2019). According to Fig. 8, the percentage of PM in previous studies was 41%. The 14 studies involved in PM are Aivaliotis et al. (2023), Aivaliotis, Georgoulias & Chryssolouris (2019), Aivaliotis et al. (2019), Altun & Tavli (2019), Avornu et al. (2022), Centomo, Dall’Ora & Fummi (2020), Heim et al. (2020), Liu et al. (2019), Mubarak et al. (2022), Rajesh et al. (2019), Rossini et al. (2020), Siddiqui, Kahandawa & Hewawasam (2023), Singh et al. (2023) and Yakhni et al. (2022). PMs also have drawbacks, such as being time-consuming and expensive to destroy and restore.

The studies conducted by Aivaliotis et al. (2023), Aivaliotis, Georgoulias & Chryssolouris (2019), Aivaliotis et al. (2019) focused on forecasting the remaining useful life (RUL) of machines and methodology to enable dynamic DT and virtual model evolution in industrial robotics were studied. Yakhni et al. (2022) utilised the DT approach to monitor ventilation system conditions by incorporating fault prediction and maintenance methods (Liu et al., 2019).

Notably, this study indicated that eight studies applied ML techniques for PdM through data-driven modelling across various sectors such as aeroplane maintenance, automobile brake pads, and induction motors (Altun & Tavli, 2019; Avornu et al., 2022; Heim et al., 2020; Mubarak et al., 2022; Rajesh et al., 2019; Rossini et al., 2020; Siddiqui, Kahandawa & Hewawasam, 2023; Singh et al., 2023). Additionally, performance prediction studies have demonstrated the potential of ML as a useful technique.

Behavioural model

The behavioural model elucidates how external stimuli, such as driving forces or disruptive influences, impact the physical system (van Dinter, Tekinerdogan & Catal, 2022). These models serve a dual purpose, serving as descriptive models while enabling a comprehensive understanding of system component deterioration and their dynamic behaviour, specifically in reliability modelling. It is based on industry-standard models, such as Systems Modelling Language (SysML), Matrix laboratory (Matlab)/Simulation and link (Simulink), Dynamic modelling laboratory (Dymola), and MSC Automated Dynamic Analysis of Mechanical Systems (ADAMS), used in model-based development processes (Kaul, Bender & Sextro, 2019). In the context of behavioural models, as indicated in the previous studies (Bondoc, Tayefeh & Barari, 2022; Moghadam, de Rebouças & Nejad, 2021; Mourtzis, Tsoubou & Angelopoulos, 2023; Pillai, Shih & Roberts, 2021), these studies constitute approximately 12% of the field. However, an inherent limitation of these models is their reliance on inadequate theory, which occasionally leads to inaccuracies in behaviour prediction.

Within this field of interest, Mourtzis, Tsoubou & Angelopoulos (2023) elaborated on ways to identify and categorise the crucial component’s malfunctioning behaviour and for mechanical equipment vibration of a structural asset/machine (Bondoc, Tayefeh & Barari, 2022). For railway engineering, the field of interest of the behavioural model is switches and crossings (S&C) (Pillai, Shih & Roberts, 2021). This equipment uses PdM with DT using a proprietary DT model to improve the prediction accuracy of the results. In addition, PdM and DT, concerning marine engineering (Moghadam, de Rebouças & Nejad, 2021), gearboxes for floating offshore wind turbines.

Decision-making model

The decision-making model serves as a framework for the analysis, justification, and validation of the model, incorporating an algorithm, a defined set of restrictions and guidelines, and adjustable input. An intelligent data-driven modelling technique is the decision-making model. Decision-making techniques, such as ML algorithms, lower the time-to-market of DT while increasing accuracy slightly (van Dinter, Tekinerdogan & Catal, 2022). Despite its significance, studies employing decision-making models for DT development are relatively scarce, comprising only 6% (Feng et al., 2023; Mi et al., 2020). Meanwhile, the field of interest for the decision-making model for mechanical equipment and bearings for vertical mills was discussed in Mi et al. (2020). Additionally, Feng et al. (2023) highlighted the utilisation of DT in supporting multi-stage PdM for intelligent industrial systems, emphasizing its pivotal role in enhancing system reliability and efficiency.

Hybrid model

A hybrid model is another sort of representation. A PM and a data-driven model (DDM), for example, are two models that this model combines. DDMs can use this data to discover hidden patterns that can subsequently be used to model a physical asset. According to Errandonea, Beltrán & Arrizabalaga (2020), DT provides the most accurate synthetic data when employing a model-based approach. The PdM hybrid technique, superior to the single-strategy approach at each stage, can produce a more accurate RUL forecast with fewer errors (Luo et al., 2020). Several model representations of PdM using DT were developed. In total, 14 studies used such a hybrid model (Das et al., 2021; Harries et al., 2023; Hosamo et al., 2022, 2023; Hu et al., 2023a; Luo et al., 2020; Mrzyk et al., 2023; Panagou et al., 2022a, 2022b; Wang et al., 2022; Werner, Zimmermann & Lentes, 2019; Xiong et al., 2021; Yang et al., 2021; Zhang et al., 2022) with a percentage of 41% from these studies.

For the field of interest, CNC machine tools (CNCMT) (Luo et al., 2020) and hydraulic systems for hydraulic equipment (Wang et al., 2022; Zhang et al., 2022) offer a more practical solution method for integrated hydraulic systems for PdM using DT, made up of numerous interconnected parts and components. Meanwhile, Hosamo et al. (2022) state that the heart of a building’s heating, ventilation, and air conditioning (HVAC) systems is its air handling unit (AHU) system. AHU stability is essential for maintaining high efficiency and increasing the lifespan of HVAC systems. For railway engineering (Yang et al., 2021), a separate model DT is used for switch drives to enhance the outcomes’ forecasting accuracy. RUL estimation is used in seven studies for this hybrid model, which is of interest for various domains, such as flexible IT architectures, indoor climate, and others (Harries et al., 2023; Hosamo et al., 2023; Hu et al., 2023a; Luo et al., 2020; Mrzyk et al., 2023; Werner, Zimmermann & Lentes, 2019; Xiong et al., 2021). Other studies involving PdM and DT are in aerospace engineering, i.e., aircraft aero-engine (Xiong et al., 2021), marine engineering (Das et al., 2021), rolling mills (Panagou et al., 2022a, 2022b), and lithium-ion batteries for a fleet of vehicles.

Deep learning technique

DL algorithms are used to represent DT or for predictive analytics. DL is an important subset of AI and a subset of ML that demonstrates how machines can mimic humans in learning certain types of information and accumulating situational awareness (Akhtar et al., 2023; Brownlee, 2017; Huang et al., 2023; Mehrjardi et al., 2023; Rastall & Green, 2022; Tanveer et al., 2023; Usuga-Cadavid et al., 2022; Velichko et al., 2023; Zhao et al., 2020a; Zhou et al., 2023). DL techniques such as long short-term memory (LSTM), deep neural network (DNN), and other DL architectures were used (Neo et al., 2022). DNNs are formed by stacking these layered networks (Brownlee, 2017). In the review, four studies focus on DL, such as the LSTM approach (Hu et al., 2023a; Xiong et al., 2021; Yang et al., 2021). The issue of a series’ long-term reliance can be resolved with LSTM by analysing the inputs in time a series (Li, Wang & Li, 2022).

An example of a combined approach can be discovered in the study (Kamat, Sugandhi & Kumar, 2021). This study uses a combination of RUL, DL, and anomaly detection approaches. Machine supervisors must determine the RUL of rotating machinery based on deterioration data. DL models are popular and reliable techniques for predicting the breakdown of rotating machinery, such as bearings. Notably, RUL is not clearly defined when the machine is operating normally. This study suggests that anomalies be kept an eye on throughout RUL estimator training and used to overcome this problem. In addition, raw bearing vibration data are processed to extract relevant time-domain data, which is then processed using DL algorithms to look for anomalies. Consequently, this data triggers data-driven RUL estimation. Unsupervised clustering is utilised for anomaly trend analysis, and semi-supervised methods are employed for anomaly detection and RUL estimation. Estimating the equipment’s RUL and detecting anomalies are improved with DL-based methods (Kamat, Sugandhi & Kumar, 2021; Lin & Tao, 2019; Pang et al., 2021; Wen et al., 2021). DL algorithms have recently been proposed as a more effective approach for extracting text features for the classification of texts (Zhong et al., 2020). To automatically categorise occupational injuries, Zhang (2022) developed a structure-based DNN with a Word2Vec model. With an F1 score of 72%, the suggested hybrid architecture demonstrated good predictive power.

Machine learning technique

There are two types of ML, namely supervised and unsupervised learning. Supervised approaches, which may be used for both classification and regression issues, train a model based on known input and output data to predict future events. Unsupervised learning is typically applied to clustering issues and looks for hidden patterns or intrinsic structures in the input data (Abd-Elrazek et al., 2021). ML techniques such as naive Bayes, random forest, support vector machine (SVM), decision tree, and regression models are the most widely used AI approaches (Khairuddin et al., 2022, 2023; Neo et al., 2022), artificial neural network (ANN), and decision trees (Hosamo et al., 2022) and ML is also used for Autoregressive Integrated Moving Average (ARIMA) (Yang et al., 2021). From the review, 19 studies use ML approaches for PdM using DT (Altun & Tavli, 2019; Avornu et al., 2022; Das et al., 2021; Harries et al., 2023; Heim et al., 2020; Hosamo et al., 2022, 2023; Hu et al., 2023a; Luo et al., 2020; Mourtzis, Tsoubou & Angelopoulos, 2023; Mubarak et al., 2022; Panagou et al., 2022a, 2022b; Rajesh et al., 2019; Rossini et al., 2020; Siddiqui, Kahandawa & Hewawasam, 2023; Singh et al., 2023; Wang et al., 2022; Werner, Zimmermann & Lentes, 2019).

According to a preliminary study conducted by Wang & Wu (2023), the model extends the usual ML used in genetic studies to enable the discovery of epistatic effects. In Genome-Wide Association Studies (GWAS), it is the standard statistical method for estimating Single Nucleotide Polymorphism (SNP) interactions. With an accuracy of 99.42% for the maintenance preventative model and 99.80% for the replacement prioritisation model, SVM performs better than other algorithms when prioritising medical devices. The prioritisation mode was created by combining supervised and unsupervised ML techniques (Zamzam et al., 2021b).

Prognostics and health management

According to a preliminary study conducted by Toothman et al. (2023), PHM systems play a crucial role in diagnosing equipment health and anticipating potential issues. These systems offer valuable insights into the components likely to fail, expediting early detection and maintenance. Their efficacy lies in swiftly pinpointing machine parts requiring immediate attention, enhancing failure prediction. PHM aims to provide methods and tools for developing an appropriate maintenance policy for a given asset under its operating and degrading conditions to achieve high availability at the lowest possible cost. It can be perceived as a comprehensive approach to effective and efficient system health management (Fink et al., 2020). Utilising sensor technology and analytical capabilities, PHM continuously monitors machine health, detects degradation, and facilitates maintenance planning for various components. Sensors strategically placed on machines gather crucial raw data reflecting their condition, forming the backbone of the PHM system. It is crucial to emphasise the meticulous placement of these sensors to ensure accurate data collection. The collected data undergoes comprehensive analysis within PHM systems, enabling a deep understanding of the manufacturing system’s behaviour and the proactive anticipation of potential operational challenges, as elucidated by Ardila et al. (2020).

Fault diagnosis

In studies verified using data from controlled experiments, test beds, or numerical simulations, PdM confronts various failure diagnosis and prognosis concerns that are typically overlooked (Fernandes, Corchado & Marreiros, 2022). Correspondingly, Xu et al. (2019) proposed a fault diagnosis method that uses DT technology to move fault information from a virtual entity to its physical counterpart. Therefore, fault detection and diagnostics significantly reduce downtime and unexpected failure of increasingly sophisticated industrial machinery and equipment (Sun et al., 2017). In contrast to traditional fault diagnosis, which relies on engineers’ extensive human expertise to determine the relationship between monitored data and machine health, intelligent fault diagnosis applies ML theories to fault diagnosis, automating the fault detection and classification process (Lei et al., 2020).

Anomaly detection

Anomaly detection is the process of identifying significant deviations from the rest of the data (López et al., 2023; Wang et al., 2021). Failures can be made, and the operator is notified in time to save the system from failing by spotting anomalies (He et al., 2023). Thus, extensive study has gone into the field of log anomaly detection, which attempts to automatically identify any aberrant logs from log data to allow operators to swiftly address problems and ensure system stability (Bertero et al., 2017; Zhao et al., 2021). A machine intelligence method provides an anomaly detection strategy that captures the temporal and spatial link between several battery characteristics to increase battery life and system safety (Das et al., 2021). Any divergence from the nominal torque signal corresponding to the current task (defined by a specific set of waypoints—position signals) would indicate an anomaly in the robot’s performance (Aivaliotis et al., 2023). Additionally, IoT devices continuously gather energy and meteorological data, which the model carefully processes and accurately examines in real-time to find any potential irregularities in energy consumption (Malki, Atlam & Gad, 2022).

Remaining useful life estimation

The RUL is the time frame from the present until the expiration of the useful life (Cai et al., 2022; Han, Li & Chen, 2023; Hu et al., 2023b; Mitici et al., 2023; Shaheen, Kocsis & Németh, 2023; Wang et al., 2023; Yang et al., 2023; Yousuf, Khan & Khursheed, 2022; Zhu et al., 2023). RUL has been used in numerous industries, such as rotating equipment, batteries, and aerospace, to name a few, to warn operators of early failures (Zhao et al., 2020b). According to a preliminary study conducted by Kang, Catal & Tekinerdogan (2021), the RUL is primarily utilised as a risk indication. It suggests how long a machine may be used properly before breaking down. Run-to-failure data from machine operations must be collected for RUL modelling. However, this task is challenging. Knowledge-based models (KBM), PM, DDM, and DL are possible areas for study on RUL prediction. In KBM, experts define the rule sets and evaluate the equipment’s state based on historical failures, occasionally leading to disparities (Garga et al., 2001).

PdM using DT for medical equipment maintenance management

In the ever-evolving realm of contemporary healthcare, the proficient administration of medical equipment maintenance emerges as a pivotal element in guaranteeing the highest standard of patient care and operational effectiveness. Integrating PdM and DT technologies offers a paradigm-shifting approach to improve maintenance practices for medical equipment. Moreover, the integration of PdM and DT allows for proactive problem prediction, enhances the effectiveness of troubleshooting, optimises maintenance scheduling, and enables decision-making based on data analysis. Drawing upon the findings obtained from the systematic evaluation of relevant literature, a fully integrated framework is proposed, which combines the capabilities of PdM and DT technologies, as illustrated in Fig. 9. This framework functions as a crucial blueprint that coordinates the smooth integration of diverse components to enable effective management of medical equipment maintenance in healthcare environments. Furthermore, the core of the integrated framework is comprised of a precisely crafted system architecture. The architectural design functions as a central point of connection, facilitating the integration of many components essential to the maintenance management process. The system is continuously supplied with a flow of data streams, each contributing to a comprehensive comprehension of the status and operation of the medical equipment. However, the most essential data source for this integrated system is the physical medical equipment.

The aforementioned entity serves as the fundamental component of the maintenance ecosystem, playing a crucial role in supporting the overall structure. In addition to the physical equipment, there are data streams created by sensors. The real-time data inputs are obtained via a network of embedded sensors that are strategically placed on the medical equipment. These inputs provide a live view of the operating condition of the equipment. The framework further incorporates a DT model, serving as a virtual depiction of the medical equipment. The digital replica has diverse attributes, qualities, and actions that closely resemble those of its tangible counterpart. The virtual model functions as a dynamic representation that has the ability to simulate real-world circumstances and forecast future outcomes. Data technology plays a crucial role in enabling the use of predictive analytics and real-time monitoring. In addition, DL algorithms can be used to represent DT or predictive analysis, including Health Index (HI), predictive model, and fault diagnosis or prediction of RUL.

The architecture for PdM of medical equipment, known as the DT architecture, is constructed based on actual physical medical equipment and their corresponding operating environment. Various kinds of sensors are often installed on physical medical equipment. These sensors are crucial in providing data for data-driven operations such as RUL prediction and component failure detection. Sensors are used to gather real-time operating data for medical equipment, which is then employed to develop a DT model or virtual model of the item using simulation data. Furthermore, the use of sensor technologies for the collection and transmission of data from medical equipment to a central system enables the instantaneous remote monitoring of the operational data of such equipment. This technology facilitates the remote monitoring of medical equipment by healthcare practitioners, allowing for tracking usage trends and timely detection of possible issues or failures. Through the real-time monitoring of operational data, healthcare practitioners have the ability to proactively detect and address issues before they escalate, reducing the probability of medical equipment failure or outage.

The model is founded upon empirical facts, with the functioning of medical equipment yielding a substantial volume of data. The use of a data-driven approach mandates that prior sensor data undergo a series of procedures prior to its application in prediction tasks. These procedures include noise reduction, data pre-processing, feature extraction, and state detection. Creating hybrid simulation models that strike a compromise between computational economy and accuracy by fusing data-driven and physics-based modelling. These are some methods for coming up with creative techniques to maximise the computational effectiveness of real-time DT simulations. By processing data closer to the origination point, edge computing lowers the latency in transmitting data to a central server. Lower latency allows for faster analysis of sensor data and quicker reaction to possible maintenance issues, which is crucial in PdM, where rapid insights are required.

The integrated system effectively collects real-time data from physical equipment and sensors, including performance measurements, operational parameters, and environmental factors. The data stream is further processed using sophisticated analytical techniques to derive a HI. It serves as a measurable indicator of the equipment’s comprehensive health and operational soundness (Yin et al., 2020). Note that the state of medical equipment is tracked using a HI for PdM to determine when maintenance is required. A HI is usually created by analysing data obtained from sensors and other monitoring equipment that record numerous factors. First, the equipment’s health factors were identified, and its state of health was assessed. The HI-based prediction model uses the input data and labels it generates (Zheng et al., 2021). DL methods are employed, including ANN, LSTM, and DNN, amongst others (Neo et al., 2022). DL models can be trained to forecast the need for maintenance, enabling maintenance personnel to plan replacements or repairs before a significant breakdown occurs. The DT model/virtual model, driven by DT for PdM of medical equipment, has assessment, optimisation, prediction, and other capabilities, such as fault detection and RUL prediction of medical equipment, as well as a conceptual model. Organisations can facilitate the efficient transfer of knowledge for applying PdM using DT in the medical industry through various strategies. For example, developing comprehensive training programmes for employees at all levels, including engineers, data scientists, medical professionals, and maintenance staff. Other than that, providing practical training that covers the principles of PdM, the development and use of DT, and the specific requirements and intricacies of the medical industry. Personnel involved in using PdM insights using DT for decision-making in the medical industry should have a combination of technical, domain-specific, and operational skills. Personnel must comprehensively understand the medical devices and systems to be monitored. This includes knowing how the devices work, their critical parameters, and potential failure modes. Training in collaboration and communication between different disciplines is essential. Personnel should be able to effectively communicate findings and recommendations to medical professionals, engineers, and decision-makers.

ML and DL were utilised, each showcasing unique advantages. A comparative analysis was performed, revealing ML’s superiority over DL in accuracy, precision, specificity, and F1 score. Although DL has a shorter training time, its accuracy is lower than that of ML (Rahman et al., 2023). However, according to the preliminary studies by Zhai & Qiao (2020), DL’s excellent performance is primarily due to a massive amount of training data and a deep network topology. The proposed framework employs DT technology to detect and diagnose problems and predict medical equipment components, enabling maintenance personnel to make more informed decisions at the appropriate moment. This will enhance the models to create the ideal strategies for future medical equipment in the healthcare industry. Several technical obstacles and challenges can arise when seamlessly integrating these technologies, such as medical facilities being subject to constant changes such as equipment upgrades and replacements and workflow changes.

To ensure that DT remains accurate and current in a dynamic medical environment, a robust mechanism for continuous updating and adaptation is required. In addition, collaboration between healthcare providers, vendors, medical device manufacturers, and regulatory bodies is essential for successfully implementing PdM and DT. Coordinating the efforts and expectations of these different stakeholders can be challenging due to differing priorities, timelines, and business models.

The significant outputs of the DT model will maximise operating costs within budgeted costs by improving medical equipment maintenance operations through the implementation of effective maintenance management practices. Moreover, accidents, financial losses, and fatalities can be avoided by scheduling maintenance in advance when it is acknowledged that medical equipment will fail. Medical equipment significantly contributes to the effectiveness of healthcare quality (Zamzam et al., 2021a). In addition, the primary goals of maintenance are dependability, maintainability, availability, and safety. Hence, medical equipment should not fail frequently and should be repaired as soon as a flaw is discovered (Abd Rahman et al., 2023).

This study has some limitations despite producing good performance prediction results. Most studies only evaluate monitoring data without fully examining the equipment’s real-time state, leading to biased results and low real-time forecast accuracy. As a result, the relationship between the actual operational state and the virtual simulation must be actualised using real-time monitoring data. It will aid in comprehensively forecasting equipment health conditions (Yang et al., 2021). Thus, the proposed framework can be embedded with the existing national standard of handling medical equipment.

For example, the concept of data-driven DT-PdM can be adopted in any healthcare institution, such as an imaging department. DT can simulate the behaviour of imaging equipment, including Magnetic Resonance Imaging (MRI) machines, Computed Tomography (CT) scanners, and ultrasound machines. Analysing real-time data from these medical types of equipment makes it possible to develop PdM models that optimise maintenance schedules and prevent unscheduled downtime.

PdM using DT for smart water management

As medical equipment health indices paved the way for timely interventions, intelligent maintenance strategies can monitor and optimise the urban equivalent, from water and energy distribution to transportation and waste management. Due to urbanisation, the world has recently confronted numerous issues, including urban poverty, exorbitant prices, traffic congestion, lack of shelter, absence of financial backing, increased crime, environmental degradation, and inequality (O’Brien, Pike & Tomaney, 2019). To meet the predicted growth in urbanisation, it is becoming increasingly difficult to design, build, run, and maintain urban infrastructure systems such as water distribution networks (WDN) with sufficient sustainability and resilience (Wu et al., 2023). The adoption of a Smart Water Grid (SWG) is an effective technique for ensuring adequate WDN sustainability and resilience (Public Utilities Board Singapore, 2016).

The proposed combined PdM and DT framework consists of a water distribution system (WDS), sensors, data collection and management from sensors and PdM using the DT model (see Fig. 10). The purpose is to track drinking water quality and pipeline leaks and exchange this data in order to save water and enhance the present WDS. The DT architecture for PdM is built on real physical WDS for pipe monitoring or the PipeSense system of a physical layer. In physical WDS, numerous types of sensors have been placed in water treatment facilities as well as around the pipe monitoring or PipeSense system, and these sensors then provide information for data-driven analysis. The sensors displayed in Fig. 10 are water quality sensors and pH and flow sensors that serve as measurement objects of the WDS layer. The idea of numerous networked sensors on equipment, individuals, and products, along with intelligent monitors, is becoming more attainable due to the advancements in communication technology (Olsen & Tomlin, 2020).

The sensor data layer of this structure includes sensor data collecting and management to analyse the acquired data. This involves network data transfer as well as server storage and database administration. Sensors collect operational data in real-time for WDS, and simulation data is used to generate a PdM using the DT model. Furthermore, implementing real-time remote monitoring will allow for more efficient decision-making for PdM using DT. These technologies could serve as the basis for the PdM using DT of the city and help in the sustainable design of smart cities. With this technology, facility management staff or users can remotely monitor water system operation, examine usage patterns, and spot potential faults or malfunctions in real-time. The model is based on data, and the operation of the water network generates a large amount of data. In the data-driven method, the previous sensor data must go through a series of steps before it can be used for prediction, including cloud data storage, communication, data analysis, and processing.

Businesses can use a range of approaches to facilitate knowledge transfer that will guarantee the effective use of DT for PdM in smart water management. Other than that, encourage the sharing of knowledge through unofficial channels so team members with less experience can benefit from their mentors’ real-world expertise. In order to properly analyse insights from PdM and DT for smart water management decision-making, staff members need to be trained in problem-solving and critical thinking. Hence, they will gain communication skills to communicate technical findings effectively and clearly to stakeholders who are not technical.

Optimising the computational efficiency of real-time simulations with DT in large-scale and complex industrial environments, especially in smart water management, requires the consideration of several innovative approaches and frameworks. Therefore, integrating ML algorithms aid in predicting system behaviour and optimise simulation parameters. Train models using historical data to make predictions, facilitating real-time adjustment of simulation parameters. This enhances computational efficiency for improved performance. With edge computing and distributed processing, real-time analytics for PdM is made possible in large, geographically dispersed businesses like smart water management. By processing data closer to the source, edge computing lowers the delay in sending data to a central server. In smart water management, where prompt problem-solving and prevention are paramount, reduced latency guarantees prompt sensor data processing for PdM.

The application level is the highest level in the water management system’s framework. Data from sensors can be enhanced by models to create useful PdM using DT for better maintenance decision-making. To produce a more precise forecast for pipe monitoring or the PipeSense system, the anticipated circumstances are then updated using the observational data collected by the system to produce an accurate prediction result.

An effective strategy to ensure PdM using DT according to the standard of the proposed approach can be incorporated into the current national standard for maintaining water treatment assets for Smart Cities System maintenance. Therefore, urban drainage terminology and actions to establish a DT through constructing digital ecosystems and open data standards format are crucial in the water sector (Pedersen et al., 2021). The integration of DT for PdM in smart water management is promising. Nevertheless, companies may encounter unforeseen technological hurdles and difficulties. These difficulties include accurate prediction depending on the availability and quality of real-time data from sensors and other devices. This data can sometimes be delayed, erroneous, or incomplete. Hence, the protection of confidential information is an important issue, especially when it comes to vital infrastructure such as water management. When using DT, users can transmit sensitive information. Furthermore, several rules and regulations apply to the smart water business, and adhering to them might be difficult. Establishing unambiguous communication channels with regulatory bodies, keeping up a continuous dialogue to comprehend and fulfil compliance obligations, and actively engaging in industry standardisation initiatives are all crucial. The successful deployment of PdM and DT frequently necessitates cooperation between several entities, such as regulators, technology vendors, and water utilities. Encourage a culture of cooperation, form alliances and consortia, and take part in industry coalitions state expectations for cooperation as well as roles and responsibilities.

A limitation of this study is that a large data set must be collected, which requires pre-processing to determine if the data can be used. Data pre-processing involves checking the data for errors such as missing or duplicate records. In addition, DT necessitates collecting enormous amounts of data from various endpoints, each providing a possible point of vulnerability (Wu et al., 2023).

A practical and affordable alternative, an integrated sensor platform, is offered by PipeSense. Monitoring the WDS and locating leakage through real-time monitoring data or DT is possible. The idea of data-driven DT using PdM, for instance, can be used in any system for smart cities, particularly in the field of WDS, for example, real-time monitoring of asset conditions. This proposed model contributes to innovative maintenance management using the concept of PdM with DT. The advantages that DT offers in implementing PdM in the water industry in terms of digital transformation and improving system performance are substantial and immediate impacts on customer satisfaction in water services, cost efficiency in asset maintenance, and ensuring greater upkeep and environmental safety. In addition, DT enables estimation of utilisation during periods of high demand, improves water flow and pressure control, and significantly increases system dependability and flexibility during operation (Ramos et al., 2022).

PdM using DT for smart farm maintenance management

Population growth has caused a sharp increase in food consumption, and agricultural mechanisation has become a key strategy for enhancing grain yield and quality (Mantoam et al., 2020). The machinery of agriculture is becoming a fundamental and important equipment in modern farming (Han et al., 2020). Regular maintenance is required to avoid losses caused by failures, and firms deploy maintenance vehicles to assure stability in agricultural machinery operations (Wang, Hu & Ren, 2021). Modern agriculture is not possible without accurate and current information regarding the farming process. Nevertheless, farms are being forced to concentrate a growing amount on digital technology, like measuring and monitoring tools, sophisticated data analysis, and smart machinery (Verdouw et al., 2021). As a result of the quick growth of technologies like computing in the cloud, big data, the IoT, ML, virtual reality, and the field of robotics, agricultural production is quickly shifting towards smart farming technologies (Kamilaris & Prenafeta-Boldú, 2018; Zhai et al., 2020).

The proposed architecture of PdM with DT for smart farm equipment is divided into three layers: the physical, communication protocol, and application layers. Each layer is important for obtaining a good model with accurate results (see Fig. 11).

At the physical level, various sensors are installed in the smart agriculture field, where data is collected using sensor applications. The proposed framework has two types of sensors: environmental and camera. Camera sensors work to detect the farm condition, while environmental sensors are smart farm characteristics such as temperature, humidity, air velocity, light, and ventilation (Sung & Kim, 2022). Multiple types of sensors are strategically implemented into agricultural machinery, generating a vast amount of real-time operating data that serves as the foundation for data-driven analysis (Javaid, Haleem & Suman, 2023). The sensors are used to generate real-time data for smart farming and store it in the database for the next shift process. Furthermore, real-time remote monitoring will enable more efficient decision-making for PdM using DT. These developments could serve as the basis for PdM using DT of farm equipment and help develop sustainable farm equipment. Hence, users or farmers can use this technology to monitor farm equipment remotely, investigate consumption trends, and detect potential defects or malfunctions in real-time.

Optimising the computational efficiency of real-time DT simulations in large-scale and complex industrial environments, especially in the smart farm industry, requires the consideration of innovative approaches and frameworks. Thus, real-time data compression techniques are implemented to reduce the amount of data transferred between sensors and the central system. The use of intelligent filters to transfer only the information is essential for running simulations and reduces the computing load. In sectors like smart farms with extensive and dispersed infrastructures, edge computing and distributed processing are essential to providing real-time analytics for PdM. It keeps private data inside the local infrastructure, which helps alleviate privacy worries. Consequently, encryption and secure communication protocols guarantee the confidentiality and integrity of information shared among dispersed nodes.

Next is the communication protocol layer, where data is transferred to the communication protocol to analyse and visualise the pre-process. In this process, observation of missing data and data cleaning is conducted. In addition, normalisation of the data was necessary to guarantee that the gap between every characteristic value was equally weighted (Zamzam et al., 2021b). This preliminary process will also send this data to the data analytics layer. A PdM using the DT model is created at the application level from the analysed, integrated, and visualised data for improved maintenance decisions. The model can provide accurate outcome predictions for smart farm equipment monitoring. In addition, the authors propose to call this model the model for intelligence, efficiency, and visibility.

The contribution to the creation of this model in the agricultural field is that it can help the farmer or user increase crop yield by increasing equipment uptime. PdM using DT can help predict equipment damage that occurs and reduce maintenance costs. Organisations that adopt the DT can gain major advantages such as improved processes, reduced time to the marketplace, and innovation in goods and services (Ante, 2021; Javaid et al., 2022; Vitorino et al., 2019). Businesses can guarantee that staff members possess the skills and knowledge required to apply PdM using DT in the smart farm sector by employing various approaches, including cross-disciplinary workshops, external training opportunities, internal workshops and seminars, and collaborative learning platforms. Maintaining the workforce’s current knowledge of cutting-edge technology and industry best practices requires ongoing training and upskilling. Therefore, workers in smart farms must possess crucial knowledge and experience requirements, including familiarity with agricultural operations, maintenance procedures, and the ability to convert PdM knowledge into actionable decisions.

The usage of sensors in individual equipment to gather data has helped IoT applications advance. Nevertheless, discrete sensor and internet covering of every location resulted in substantial facility management expenses. Moreover, adherence, safety, and trustworthiness standards are becoming more stringent and subject to rapid change (Masmoudi et al., 2016).

Using applications for real-time monitoring equipment, the model can be integrated with PdM, utilising DT to facilitate decision-making. The user and farmer can comprehend the data without difficulty. The concept of DT in smart agriculture is still in its earliest stages. At the beginning demonstrating stage, numerous farmers are exploring incorporating smart technology and methods that enhance the effectiveness of agricultural processes (Verdouw et al., 2021). Agricultural mechanisation can save labour time and other resources, enhance labour efficiency, and lower production costs (Mu et al., 2018). In addition, the proposed technique can be implemented in the existing national standard for agricultural machinery maintenance as an efficient method to ensure PdM with DT in accordance with the standard. DT integration of PdM in the smart agricultural sector could encounter several technological challenges. Variations in the weather, plant health, and pest infestations are only a few examples of the dynamic and unpredictable variations that affect agriculture the creation of adaptable DT models that can be modified on the fly. Hence, it establishes feedback and real-time monitoring systems to inform the DT about the changing agricultural landscape. In addition, effective implementation of PdM and DT often requires collaboration between different organisations, including farmers, technology providers, and regulators. These organisations can foster a culture of collaboration, establish partnerships, and participate in industry consortia or alliances. Clearly define roles, responsibilities, and expectations for collaboration.

Five review articles are included in this discussion and discuss PdM using DT. There are different aims for all the papers. The first review article aims to redefine the concept of next-generation-Digital-Twin (nexDT). Furthermore, it introduces a specialised term for electrical machines manufacturing, PdM, and control, summarising the majority of pertinent work in the process. Other than that, it offers a new definition unique to PdM and serves as a basis for future efforts (Falekas & Karlis, 2021). This study reviews the most recent descriptions found in the broader literature and addresses the open challenge given by the indirect usage of an established Digital Twin Framework (DTF) in industry, specifically system security and risk assessment. Aspects such as goals, application areas, platforms, representation types, approaches, abstraction levels, patterns of design, protocols for communication, twin parameters, challenges, and solutions are highlighted in the second review article on PdM through the use of DT (van Dinter, Tekinerdogan & Catal, 2022). There are 42 primary studies that have been evaluated. The study revealed that the computational effort, the variety of data, and the complexity of the models, systems, or components represent the greatest challenges in the development of these models. Consequently, a presentation is given on using a DT framework in PdM and its expansions via physics-based modelling and ML. Planning for repair, logistics for deployment, and performance assessment measures are all included in fleet management (Kunzer, Berges & Dubrawski, 2022). This study attempts to clarify the definition of the DT by looking at the term’s history and original context in asset maintenance and fleet management, planning, operations and product lifecycle management. The challenges for this study are safety protocols, the adoption of the DT framework in the workplace, the robustness of the sensors, missing data, poor quality data, and offline sensors. Review four examines the modelling techniques, the application framework, and the relationship between the virtual image and the physical object as recent developments (You et al., 2022). This study aims to describe the benefits of the PdM using the DT paradigm that are thought to exist and examine the methods and applications for each category. A total of 30 primary studies were examined. Four critical challenges for DT-driven PdM discussed in this study are the standardisation of the framework, holistic evaluation methods, the need for a digital model with high fidelity, and a multi-level model and multi-component. Note that the creation and assessment of reference architectures created with the use of well-known software architecture techniques are the topics of the final review article (van Dinter, Tekinerdogan & Catal, 2023). The study revealed three viewpoints for DT-based PdM systems. The authors developed a context diagram for the user view, and for the structural view, they created a package diagram. The authors also created an application architecture for each case study based on the features of the study using each reference architecture view, a layered view to illustrate the system’s breakdown into layers, and a deployment view to depict the hardware, software, and surrounding environment. There are 42 selected primary studies for which a trait model was developed. This study is about the continuous consideration of challenges in production, such as changes in data distribution, skewing of training data, or problems with data quality.

From all the highlighted review articles, this study differs by focusing on three areas of PdM-based DT implementation, namely healthcare, utilities (smart water management), and agriculture (smart farm), and fulfilling the objectives of this study. This work analysed 34 articles from 2018 to 2023 and discussed various challenging PdM with DT. In addition, an explanation of the use of sensor types for the relevant industries, and finally, the implementation of maintenance management, which focuses on the diversity of maintenance management methods and strategies, issues proposed in the study, PdM and DT methods, studies conducted in real-time monitoring, and the importance of research.

The integration of DT and PdM technologies has resulted in industry-altering advancements. This integration enables proactive equipment maintenance in healthcare management by simulating real-time conditions and anticipating potential malfunctions. DT and PdM optimise the operation of infrastructure in utilities (smart water management) by integrating real-time sensor data with predictive analytics to anticipate leakage and system inefficiencies. Similarly, in agriculture (smart farm), the combination of DT and PdM aids in the surveillance of apparatus in real-time, while predictive insights guide opportune productivity-enhancing interventions. These applications demonstrate the value of DT and PdM in revolutionising maintenance practises across industries, thereby paving the way for enhanced operational efficiency, resource optimisation, and overall performance.

However, implementing PdM and DT to maintain equipment or systems can present various challenges. Here are some of the major considerations to address prior to adopting the concept:

1. Data collection and management: Collecting and managing the enormous quantities of data required for PdM and DT can be difficult, particularly for aged equipment or systems not necessarily designed to collect and transmit data. Facilities must invest in automated data acquisition and management tools to ensure accurate, trustworthy, and secure data. To achieve a balance between optimising network maintenance through PdM and DT and meeting stringent regulatory requirements, meticulous planning and implementation are required.

2. Data integration: It can be difficult to integrate data from multiple sources, such as equipment sensors, electronic medical records, and patient feedback. A data integration strategy must be developed by facilities to ensure that all data sources are integrated and accessible for analysis.

3. Model development: Developing accurate PdM models and DT requires an in-depth comprehension of the behaviour and performance of equipment or systems. To develop and evaluate these models, facilities must invest in expertise and resources.

4. Compliance with regulations: PdM and DT technologies must comply with the country’s own regulations, which can present additional challenges for facilities. It is essential to ensure that these technologies adhere to regulatory requirements and that data is kept secure. In healthcare, utilities (smart water management), and agriculture (smart farm), the need to adhere to sector-specific rules can confound the integration of PdM and DT. Utilising these technologies in healthcare, patient data privacy statutes such as the Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR) must be adhered to. Similarly, water networks must comply with water quality regulations and data security. In agriculture, machinery optimisation must consider environmental regulations and data privacy. It is essential for the safe and effective implementation of PdM and DT methods in each field that they adhere to these regulations.

5. Cost: Implementing PdM and DT technologies can be costly, necessitating substantial investments in hardware, software, and personnel. The costs and benefits of these technologies must be thoroughly evaluated by facilities to determine if they are worth the investment.

6. Changes to organisational cultures: Adopting new technologies and modifying maintenance practises can be difficult for organisations, necessitating a shift in organisational culture and perspective. Facilities must invest in change management and training to ensure their employees are prepared to implement these technologies and procedures.

Across all industries, integrating PdM and DT technologies presents obstacles. This involves managing immense quantities of data, integrating diverse data sources, creating accurate models, and ensuring compliance with sector-specific regulations such as HIPAA, GDPR, water quality standards, and agricultural regulations. Implementing PdM and DT can be expensive, necessitating a thorough cost-benefit analysis and a cultural transformation in organisations, necessitating change management and training for successful adoption. Therefore, implementing PdM and a DT to maintain three potential applications for equipment or systems requires meticulous planning, investment, and specialised knowledge. Facilities must be prepared to surmount these obstacles To reap the benefits of these technologies, which include improved equipment or system uptime, reduced costs, and increased patient or worker safety.