AI-Driven Digital Twin Frameworks for Predictive Maintenance and Closed-Loop Remote Control in Smart Factories: A Structured Review

AI-Driven Digital Twin Frameworks for Predictive Maintenance and Closed-Loop Remote Control in Smart Factories: A Structured Review

Ranjitha Haleshappa Gowdru

School of Advanced Studies, GM University, Davangere, Karnataka, India

Neelambike

Program Director, GM School of Advanced Studies, GM University, Davangere, Karnataka, India

Abstract

The increasing connectivity of industrial machines has changed maintenance from a largely schedule-driven activ- ity into a continuous, data-driven process. Digital Twin (DT) technology supports this transition by providing a dynamic virtual representation of a physical asset that is continually updated using operational data. When articial i ntelligence ( AI) is e mbedded w ithin t he twin, the resulting system can move beyond visualization to support fault identication, anomaly detection, degrada- tion assessment, Remaining Useful Life (RUL) estimation, and maintenance decision-making. This paper presents a structured review of AI-enabled digital twin approaches for predictive maintenance in smart manufacturing, exam- ined end-to-end across physical sensing, data acquisition, communication, twin synchronization, intelligent analyt- ics, maintenance decision-making, and remote corrective control. Machine learning, deep learning, unsupervised anomaly detection, hybrid and physics-informed models, edge intelligence, federated learning, and generative AI (GenAI) are discussed with respect to their industrial suitability and their bearing on the accessibility of digital- twin maintenance for resource-constrained manufacturers. Particular attention is given to a limitation that remains evident throughout the literature: most reported systems detect or predict abnormal behavior but stop short of converting those predictions into safe, closed-loop actions. The review classies s ystems by their p rogression from monitoring and diagnosis to prediction, prescription, and autonomous control, and introduces a GenAI-Augmented Digital Twin Maturity Evaluation (GADT-ME) frame- work that scores a system jointly on functional matu- rity and generative-AI capability. Challenges relating to scarce failure data, latency, cybersecurity, synchroniza- tion, explainability, interoperability, computational cost, and deployment cost for small and medium enterprises (SMEs) are also examined. The review concludes that future research should prioritize hybrid edgecloud intelli- gence, secure actuation, multi-sensor fusion, interpretable and generative AI, federated learning, and self-healing, accessible industrial systems. Keywords: Articial In- telligence; Digital Twins; Predictive Maintenance; Smart Manufacturing; Closed-Loop Control.

1. INTRODUCTION

   Manufacturing systems are progressively becoming more connected, observable, and software-controlled.

   Modern production environments integrate sensors, pro- grammable controllers, industrial networks, edge

   devices, cloud platforms, robotics, and data-analytics services, together forming the technological foundation of

   Industry 4.0 and, increasingly, the human-centric and resilience-oriented goals of Industry 5.0 [1], [2].

   Maintenance remains one of the most consequential in- dustrial functions because equipment degradation directly affects p roduction c ontinuity, s afety, p roduct quality, energy consumption, and operating cost. An unexpected

   failure in a motor, bearing, conveyor, compressor, robotic arm, or turbine can interrupt an entire production line. Maintenance strategy has therefore evolved from reac-

   tive practice toward preventive, condition-based,

   predictive, and prescriptive approaches. Reactive main- tenance is initiated only after failure and can cause long, unplanned downtime; preventive maintenance reduces this risk through fixed servicing intervals but may service

   equipment that is still healthy; and predictive mainte- nance instead estimates equipment condition from mea- sured

   operating data, allowing degradation to be identified before a functional failure occurs [1].

   Digital Twin technology extends this idea by maintain- ing a virtual representation that stays continuously synchronized with its physical counterpart through data exchange. Measurements such as vibration, temperature,

   current, rotational speed, pressure, load, acoustic emis- sion, and energy consumption update the virtual

   representation in near real time [2], [5]. Artificial intel- ligence adds a further layer of functionality: machine- learning algorithms recognize fault patterns, estimate equipment health, and forecast degradation, while deep

   neural networks can learn temporal or spectral charac- teristics directly from raw sensor data. Most recently, generative AI (GenAI) including large language mod-els, diffusion models, a nd generative adversarial networks

   has begun to augment this pipeline by synthesizing rare-failure data, producing natural-language explana- tions

   of twin state, and drafting maintenance recommenda- tions, a direction discussed in detail in Section VII-A [26] [31].

   1. The Research Gap

      Despite this progress, the maturity of digital-twin-based maintenance systems remains uneven. Monitoring,

      diagnosis, and prediction have received substantial re- search attention, whereas closed-loop response is

      demonstrated far less often. Predicting that a bearing, motor, or production asset is approaching an unsafe state is useful, but prediction alone cannot prevent failure. A practical maintenance system must also determine what action should follow, how quickly that action must be

      executed, whether it can be performed safely, and increasingly whether an operator without deep AI

      expertise can trust and act on the twins output. Three specic gaps motivate this review:

      • Prediction-to-action gap: reported systems over- whelmingly stop at alarms or dashboards; validated end-

        to-end pipelines that carry a prediction through decision- making to safe remote actuation are

        comparatively rare (Section XIII-A).

      • Evaluation gap: existing maturity taxonomies classify digital twins by functional level (monitoring through

        autonomous control) but do not yet account for the rapidly growing role of generative AI within each level,

        making it dicult to compare a GenAI-assisted diag- nostic twin against a conventional one on common

        terms.

      • Accessibility gap: the literature on AI-enabled digital twins is dominated by large-enterprise deployments

      with substantial sensor, cloud, and engineering budgets; comparatively little synthesis connects predictive-

      maintenance capability with the low-cost, low-expertise constraints faced by small and medium

      manufacturing enterprises (SMEs) [32][34].

      The distinguishing perspective of this review is therefore to treat predictive maintenance as an end-to-end cyber- physical feedback process not solely as a machine- learning problem and to evaluate that process jointly

      on

      functional maturity, generative-AI contribution, and practical accessibility.

   2. Contributions

      The major contributions of this revised review are:

      • A structured classication of AI-enabled digital twin maintenance systems according to sensing modality,

        AI technique, deployment architecture, digital-twin capa- bility, and control level.

      • A comparative discussion f machine learning, deep learning, unsupervised anomaly detection, hybrid AI,

      physics-informed approaches, edge intelligence, and gen- erative AI.

      • A functional maturity taxonomy separating monitor- ing, diagnostic, predictive, prescriptive, and

        autonomous digital twins.

      • A new GenAI-Augmented Digital Twin Maturity Evaluation (GADT-ME) framework that scores sys- tems

        jointly on functional maturity and generative-AI capabil- ity (Section XV).

      • An examination of the transition from fault predic- tion to remote corrective action and closed-loop

        maintenance, including its practical implications for plant operators (Section XVI).

      • Identication of major technical barriers, including limited industrial failure data, latency, cybersecurity,

        explainability, synchronization, interoperability, computa- tional demand, and deployment cost for SMEs.

      • A synthesis of emerging research directions involving edgecloud intelligence, federated learning,

      explainable and generative AI, multi-sensor fusion, secure control, and self-healing, accessible systems.

2. EVOLUTION OF INDUSTRIAL MAINTENANCE

   Industrial maintenance strategies can be viewed as a progression from failure response toward intelligent

   intervention.

   1. Reactive Maintenance

      Reactive, or breakdown, maintenance is performed only after a component has failed. It requires little initial

      monitoring infrastructure but can produce unpre- dictable repair costs and prolonged equipment downtime.

   2. Preventive Maintenance

      Preventive maintenance schedules inspection or replace- ment according to time, operating cycles, or

      manufacturer recommendations. It reduces the proba- bility of sudden breakdown but may also result in

      premature servicing of components that remain healthy.

3. Condition-Based Maintenance

   Condition-based maintenance continuously observes equipment variables and triggers maintenance when

   measurements indicate deterioration. This approach depends strongly on sensor quality and on appropriate

   threshold selection.

   1. Predictive Maintenance

      Predictive maintenance estimates future machine condi- tion from historical and real-time observations. Statistical analysis, machine learning, signal processing, and deep learning can each be used to estimate fault probability or

      remaining useful life.

      Table 1: Comparison of Industrial Maintenance Strategies

      Strategy Operating Principle Primary Limita-

      1. Process Twin

         A process twin represents a manufacturing process or

         tion a sequence of operations rather than a single physical

         Reactive Repair after equipment failure

         Preventive Maintenance at prede-

         termined intervals

         High downtime and unpredictable repair cost

         Healthy compo- nents may be serviced unnecessar-

         component. It can describe relationships among process- ing stages and monitor variables such as production rate, temperature, pressure, energy consumption, machine uti- lization, processing time, and product quality. A process twin can therefore help identify how changes in one stage aect subsequent stages. For predictive maintenance,

         Condition-

         Maintenance based on

         ily

         reliable

         combining process information with equipment measure-

         based

         Requires

         monitored condition sensing and thresh-

         old selection

         ments can provide additional context when an abnormal operating condition is detected.

         Predictive Future degradation esti-

         mated using data mod- els

         Prescriptive System recommends a

         corrective maintenance action

         Autonomous Corrective response

         may be automatically

         Requires representa- tive historical or op- erational data Decision quality and explainability become critical Requires trustwor- thy AI and secure

      2. System Twin

      A system twin represents an interconnected manufacturing system, production line, or group of cooperating machines. It can integrate information from machines, robots, con- veyors, controllers, sensors, and other industrial assets. This system-level representation is important because

      executed closed-loop control degradation in one asset can inuence downstream or

      upstream operations. For example, deterioration of a con-

   2. Prescriptive and Autonomous Main- tenance

      Prescriptive maintenance extends prediction by recom- mending a specic response. Autonomous maintenance

      closes the feedback loop further still by executing an approved corrective action without requiring continuous

      human intervention.

4. DIGITAL TWIN TECHNOL- OGY

   Digital Twin (DT) technology provides a continuously updated virtual representation of a physical asset, pro- cess, or system. In smart manufacturing, measurements such as vibration, temperature, current, rotational speed, pressure, load, acoustic emission, and energy consump- tion can be transferred from the physical environment to the virtual representation. The resulting twin can support condition monitoring, fault identication, degra- dation assessment, predictive maintenance, and mainte- nance decision-making [2], [5]. The level of representation may range from an individual component to a complete manufacturing system.

   1. Component or Product Twin

      A component or product twin represents an individual physical asset, component, or manufactured product. Typ- ical examples include bearings, motors, gears, pumps, robotic arms, conveyor components, and turbine elements. The twin is updated using measurements from sensors at- tached to the corresponding physical object. For example, vibration, temperature, rotational speed, and load mea- surements from a motor bearing can be used to estimate its current health state and identify changes that may indicate degradation. In predictive maintenance, the com- ponent twin therefore provides an asset-level view of con- dition and supports early fault detection and remaining- useful-life estimation.

      veyor motor may change material ow and subsequently aect the operation of downstream equipment. A sys- tem twin can represent these dependencies and support system-level diagnosis, prediction, and maintenance plan- ning.

5. INDUSTRIAL IOT AND SEN- SOR INFRASTRUCTURE

   Industrial Internet of Things (IIoT) infrastructure pro- vides the sensing and communication foundation required to keep a Digital Twin synchronized with the physical environment. Sensors continuously collect operational measurements, while industrial communication networks transfer these observations to edge, local, or cloud process- ing resources. Typical measurements include vibration for rotating-component condition, temperature for ther- mal abnormalities, current for motor and load behavior, pressure for pumps and compressors, rotational speed for operating-state changes, acoustic emissions for mechani- cal damage, and energy consumption for eciency and equipment-state analysis.

   The resulting data ow can be represented as: physical machine sensors data acquisition communica- tion network Digital Twin AI analytics. Reliable sensing, synchronization, communication, and data qual- ity are therefore prerequisites for trustworthy AI-driven maintenance. Noise, missing observations, sensor drift, communication delay, and sensor failure can propagate uncertainty into the twin andsubsequently aect mainte- nance decisions.

6. ARTIFICIAL INTELLIGENCE FOR PREDICTIVE MAINTE- NANCE

   Articial Intelligence (AI) adds analytical capability to the Digital Twin by learning patterns from historical and real-time operational data. Depending on the objective

   and available data, AI can be used for anomaly detection, fault diagnosis, equipment-health assessment, degradation forecasting, Remaining Useful Life (RUL) estimation, and maintenance decision support. The principal approaches considered in this review are traditional machine learn- ing, deep learning, unsupervised anomaly detection, and hybrid or physics-informed models. Generative AI is dis- cussed separately as an emerging augmentation layer for explanation, data generation, and maintenance assistance.

   1. Traditional Machine Learning

      Traditional machine-learning methods generally operate on structured or engineered features extracted from sensor measurements. Decision trees, random forests, support vector machines, and regression models can be used for fault classication, health-state classication, and predic- tive modelling. A typical pipeline converts measurements such as vibration, temperature, current, rotational speed, and load into features that are supplied to a trained model. The model then produces an estimated equipment state or fault class.

      These methods can be comparatively interpretable and can perform eectively with moderate-sized structured datasets. Random forests can provide nonlinear modelling and feature-importance information, while support vector machines can be eective for classication with appropri- ately selected feature sets. However, performance may depend strongly on feature engineering, data quality, and operating-condition coverage. The AI-method comparison in Table 2 summarizes these characteristics.

   2. Deep Learning

      Deep-learning models use multiple neural-network layers to learn complex representations from sensor and time- series data. Convolutional Neural Networks (CNNs) can process vibration signals or transformed representations such as spectrograms, while Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks are suited to sequential degradation patterns. Autoencoders can also learn compact representations of normal operat- ing behavior.

      For predictive maintenance, deep learning can learn temporal, spectral, or nonlinear relationships that may be dicult to capture through manually engineered features. For example, an LSTM or GRU model can use a sequence of historical measurements to estimate future equipment condition or RUL. The principal limitations are the need for representative training data, computational resources, careful model tuning, and additional techniques when interpretability is required.

   3. Unsupervised Anomaly Detection

      Unsupervised anomaly detection is particularly relevant when labelled failure examples are scarce. Industrial equipment may generate large volumes of normal oper- ating data while actual failures remain relatively rare. Instead of requiring a large set of labelled fault exam- ples, an unsupervised model can learn the characteristics of normal operation and ag observations that deviate substantially from that learned pattern.

      Autoencoders and Isolation Forests are examples consid- ered in this review. An autoencoder learns to reconstruct normal sensor patterns; a large reconstruction error can indicate a potential anomaly. Isolation Forests identify observations that are relatively easy to isolate from the ma- jority of the data. This approach can reduce dependence on labelled failure datasets, but the selection and calibra- tion of anomaly thresholds remain important challenges. An anomaly should therefore be interpreted together with operating context rather than treated automatically as a conrmed failure.

   4. Hybrid and Physics-Informed Models

      Hybrid and physics-informed approaches combine data- driven AI with engineering knowledge or physical models. A purely data-driven model learns relationships from observations, whereas a physics-informed approach can incorporate known relationships governing mechanical, thermal, electrical, or degradation behavior. The combi- nation is useful when industrial failure data are limited or when physically meaningful constraints are important. For example, a bearing-degradation model may com- bine vibration and temperature observations with engi- neering knowledge about load, friction, thermal eects, or degradation mechanisms. Such a model can be used for degradation assessment and RUL estimation while constraining predictions toward physically plausible be- havior. Hybrid approaches can therefore improve physical consistency and may support better generalization when training data do not cover all operating conditions. Their limitation is that an appropriate physical model or en- gineering relationship must be available and correctly

      specied.

   5. Generative AI

      Generative AI (GenAI) is an emerging augmentation layer for Digital Twin maintenance systems. Genera- tive models can support rare-failure data augmentation, natural-language explanation of twin states, maintenance- recommendation drafting, and operator assistance. In this role, GenAI complements rather than automatically replaces the predictive models responsible for detection or forecasting. Because generated outputs may contain errors or unsupported recommendations, guardrails, vali- dation, uncertainty information, and human or control- system oversight are required before generated content is allowed to inuence physical actuation [26][31].

7. REVIEW METHODOLOGY

   This study uses a structured evidence-mapping approach to synthesize research on articial intelligence (AI)- enabled digital twins (DTs) for predictive maintenance and closed-loop maintenance. The review is organized around predened research questions, an explicit eligibil- ity framework, thematic evidence mapping, bibliographic verication, and conservative maturity coding. It is in- tentionally reported as a structured review rather than a PRISMA-compliant systematic review because a con- temporaneous database-export and screening log was not retained for the present evidence set. Accordingly, no

   Table 2: Comparison of AI Methods for Predictive Mainte- nance

   Method Typical Appli- Strength Limitation

   maintenance and prognostics?

   1. RQ2: How are digital-twin systems evolving from condition monitoring and fault diagnosis toward pre-

      cation

      Decision Tree Fault classica- Simple and inter- May overt tion pretable noisy training

      diction, prescription, and autonomous maintenance?

      Random For- est

      Fault diagnosis, health classica- tion

      Robust nonlinear modeling, feature importance

      data

      Larger ensem- bles increase computational cost

   2. RQ3: Which sensing, communication, digital-twin, analytics, decision, and actuation components are required to support closed-loop predictive mainte-

      SVM Fault classica- tion

      CNN Signal/spectrogram analysis

      LSTM/GRU Temporal degra-

      dation, RUL pre- diction

      Autoencoder Unsupervised

      anomaly detec-

      Eective with moderate feature sets

      Automatic fea- ture learning

      Captures sequen- tial dependencies

      No large labelled fault dataset re-

      Training expen- sive for large datasets Requires su- cient labelled data

      Training cost

      and tuning complexity Anomaly thresh- old selection is

      nance?

   3. RQ4: What technical and operational barriers limit deployment, particularly with respect to data avail- ability, latency, cyberscurity, explainability, interop- erability, scalability, and cost?

   4. RQ5: What emerging opportunities, including

   Isolation For- est

   Physics- Informed AI

   Generative AI

   tion

   Tabular anomaly detection

   Degradation and RUL modeling

   Data augmenta- tion, explanation, recommendation drafting

   quired Computationally ecient

   Combines engi- neering knowl- edge and data Mitigates data scarcity; improves operator trust and accessibility

   dicult

   Limited tempo- ral representa- tion

   Requires an ap- propriate physi- cal model Hallucination risk; needs guardrails be- fore acting on

   edgecloud intelligence, federated learning, physics- informed AI, and generative AI, can strengthen digital-twin-based predictive maintenance and de- cision support?

   7.3 Information Sources and Search Pro- tocol

   output

   database retrieval, deduplication, screening, or PRISMA counts are inferred from the nal evidence matrix.

   1. Review Protocol and Reproducibility

      To improve methodological transparency, the review pro- tocol species the information sources, conceptual search terms, eligibility criteria, evidence categories, biblio- graphic verication procedure, and maturity-coding rules used in the evidence map. These elements dene a re- producible analytical framework without reconstructing unsupported retrieval or screening statistics. The original contemporaneous database-export and screening les were not retained; therefore, the manuscript does not report reconstructed numbers for database retrieval, duplicate removal, title/abstract screening, full-text exclusion, or – nal inclusion. Such numbers would not be auditable from the available evidence and are consequently not inferred. The reproducibility of the present study should there- fore be understood at the level of the documented protocol and coding framework. Another researcher can reproduce the intended search concepts, apply the stated eligibility rules, verify bibliographic records, and repeat the GADT- ME classication procedure. For a future update requiring formal systematic-review reporting, the searches should be re-executed and archived with database-specic search strings, execution dates, exported records, deduplication procedures, screening decisions, exclusion reasons, and the nal included-study set. This distinction is main- tained throughout the manuscript so that methodological transparency is strengthened without overstating the evi-

      dentiary basis of the current review.

   2. Research Questions

      The review is guided by ve questions:

      1. RQ1: Which AI and machine-learning techniques are being integrated with digital twins for predictive

      The review protocol was structured around ve biblio- graphic and scholarly information sources: Scopus, Web of Science Core Collection, IEEE Xplore, ScienceDirect, and SpringerLink. Publisher webpages and DOI/index records were used for bibliographic verication and DOI resolu- tion. The protocol was designed to be reproducible at the level of databases, concepts, eligibility rules, and coding variables. Importantly, the present manuscript does not retrospectively report database retrieval counts, duplicate counts, or screening counts because the contemporaneous database-export and screening log for the evidence set was not retained. No such counts are reconstructed from the nal reference list.

      The conceptual search was organized into four term blocks:

      1. Digital-twin block (A): digital twin, digital twins, virtual twin, digital replica;

      2. Maintenance block (B): predictive maintenance, condition monitoring, fault diagnosis, prognos- tics, remaining useful life, prescriptive mainte- nance;

      3. AI block (C): articial intelligence, machine learn- ing, deep learning, anomaly detection, physics- informed, reinforcement learning, generative AI;

      4. Control/deployment block (D): closed-loop, remote control, autonomous maintenance, edge AI, industrial IoT.

         The core conceptual expression was dened as

         (A) AND (B) AND (C), with block (D) used as an additional renement when the database supported suf- ciently specic control/deployment elds. To support reproducibility, the database adaptations were dened as follows:

         • Scopus: TITLE-ABS-KEY search across the four conceptual blocks.

           • Web of Science: Topic (TS) search using the same four blocks with database-specic eld syntax.

           • IEEE Xplore: Metadata/All Metadata search us- ing the same concepts, with phrase searching where supported.

           • ScienceDirect: Title, abstract, and keyword- oriented search using the same concepts.

           • SpringerLink: Article metadata/full-text search us- ing the same concepts, followed by relevance screen- ing.

      The evidence map covers literature from 2019 through 2026, reecting the period represented by the primary- study matrix, and prioritizes peer-reviewed journal arti- cles and conference papers. The intended language crite- rion is English. Records were eligible when they addressed AI-enabled digital twins in predictive maintenance, prog- nostics, fault diagnosis, maintenance optimization, or closed-loop maintenance/control, or when they provided a directly relevant enabling architecture or technology. Review and survey papers were retained for contextual synthesis but excluded from the primary-study maturity denominator.

      Because the original database execution logs were not retained, the database names, conceptual terms, eligibility rules, and coding protocol above should be interpreted as the documented review protocol rather than as a claim that a complete database export and screening sequence was reconstructed. This distinction is maintained throughout the manuscript to avoid introducing unsupported retrieval or screening numbers. If a future revision is required to meet a journals formal systematic-review reporting standard, the searches should be re-executed using the protocol above, with the exact execution dates, database- specic strings, retrieved records, deduplication, screening decisions, and nal included set archived as supplementary material.

      1. Eligibility and Study Classication

         The evidence map prioritizes peer-reviewed journal arti- cles and conference papers directly relevant to AI-enabled digital twins, predictive maintenance, prognostics, fault diagnosis, maintenance optimization, or closed-loop main- tenance. Records were retained when the title, bib- liographic identity (including DOI or publisher/index record), and topical relevance could be established. Re- view/context papers were retained for landscape synthe- sis but were not treated as empirical primary evidence. Framework/architecture papers were retained for concep- tual and system-design analysis, while adjacent/technical papers were retained only when they provided an enabling technology materially relevant to the review questions.

         Primary studies provide an empirical or implementation-oriented contribution. Frame- work/architecture studies describe reference architectures or methodologies. Adjacent/technical studies address enabling sensing, process-control, communication, or related infrastructure. Review/context studies synthesize previous literature. Bibliographic identication is kept separate from full-text evidence coding; records whose full-text capability could not be independently established

         are not presented as stronger evidence than the available source supports.

      2. Study Selection and Evidence Coding

         Selection was conducted at the evidence-map level us- ing the predened eligibility categories and bibliographic verication criteria. Each retained record was coded for publication year, title, venue, DOI, AI method, applica- tion domain, maintenance task, evidence type, and quality. Primary/implementation records were additionally coded on the GADT-ME maturity dimensions. The DOI veri- cation register records the bibliographic verication status and the evidence basis used for maturity coding.

         The evidence-coding workow followed ve stages: (1) identify a candidate record from the dened literature sources or cited literature; (2) verify title, venue, year, and DOI against a publisher or DOI/index record; (3) classify the record as primary/implementation, frame- work/architecture, adjacent/technical, or review/context;

         (4) extract the AI method, application domain, mainte- nance task, and evidence quality; and (5) assign GADT- ME maturity codes only where the available evidence sup- ports the classication. Where full-text evidence could not be independently established, the record was not upgraded beyond what the available publisher/index evi- dence supports.

         The primary evidence matrix and the separate re- view/survey matrix provide the audit trail for the syn- thesis. The DOI-Based Reference Verication Register provides a second bibliographic audit layer. These three representations are intentionally separated so that biblio- graphic completeness, evidence classication, and matu- rity coding are not conated.

      3. Literature Synthesis

         The literature was synthesized using qualitative thematic comparison and evidence mapping rather than statistical meta-analysis. Primary studies were compared across AI method, domain, maintenance task, architectural role, and maturity. Review and survey studies were analyzed separately so that secondary evidence did not enter the primary-study maturity calculations. The synthesis com- pares ve functional stages: monitoring, diagnosis, pre- diction/prognostics, prescription/optimization, and au- tonomous maintenance/control.

      4. Maturity Coding Protocol

         For the GADT-ME analysis, each pri- mary/implementation record was assigned a functional maturity code and a generative-AI capability code only when the available evidence supported the clas- sication. The functional axis is M1 (monitoring), M2 (diagnostic/anomaly identication), M3 (predic- tion/prognostics), M4 (prescriptive decision-making or maintenance optimization), and M5 (validated autonomous maintenance/control). The generative-AI axis is G0 (no demonstrated generative-AI capability), G1 (generative-AI augmentation), G2 (generative-AI co-pilot), and G3 (generative-AI-enabled actuation). Records without a demonstrated maturity stage are coded NR and excluded from the maturity denominator.

         The coding is conservative: prediction is not upgraded to M4 unless a maintenance recommendation, optimiza- tion, or scheduling function is demonstrated; an architec- ture is not upgraded to M5 unless autonomous corrective action is demonstrated; and conventional neural networks or autoencoders are not treated as GenAI merely because they are generative models in the broad machine-learning sense. The 2025 Generative and Predictive AI for digital twin systems in manufacturing study is coded M3/G1 be- cause the published study demonstrates predictive/quality capabilities together with generative-AI augmentation, while agentic autonomous control is described as future work. The Investigating determinants of digital twins for predictive maintenance study is coded NR/NR be- cause it evaluates implementation determinants rather than demonstrating a maintenance maturity stage. The 2025 Industry 5.0 ML study is coded M3/G0 because its documented contribution is predictive maintenance using ML without demonstrated GenAI capability.

         Of the 35 records in the primary matrix, 34 provide a maturity-stage code and one is NR. The resulting matu- rity distribution is M1: 2 (5.9%), M2: 6 (17.6%), M3: 23

         (67.6%), M4: 3 (8.8%), and M5: 0. The GenAI distribu-

         tion is G0: 33 (97.1%), G1: 1 (2.9%), G2: 0, and G3: 0.

         These percentages describe the evidence map and are not meta-analytic eect estimates.

      5. Generative AI in Digital Twins (2025

      2026 developments)

      Generative AI (GenAI) is emerging as an augmentation layer for digital-twin systems rather than a replacement for established predictive-maintenance models. Recent work considers large language models and other gener- ative techniques for synthesizing scarce or rare-failure data, explaining equipment condition in natural language, assisting maintenance planning, and supporting operator interaction with complex twin states. In a manufacturing setting, a GenAI component can translate model outputs such as anomaly scores, health indices, or predicted failure probabilities into structured maintenance explanations and recommendations while the underlying numerical pre- diction remains the responsibility of validated analytical models [26][31].

      The evidence mapped in this review indicates that GenAI capability is still at an early stage relative to conventional AI-enabled predictive maintenance. Most identied systems remain focused on monitoring, diag- nosis, or prediction, while validated GenAI-enabled ac- tuation is not yet demonstrated in the primary evidence set. Accordingly, the GADT-ME framework treats GenAI augmentation, co-pilot functionality, and GenAI-enabled actuation as distinct capability levels. This separation is important because natural-language generation or data synthesis alone should not be interpreted as autonomous industrial control. Safety constraints, human oversight, verication, and secure interfaces remain necessary be- fore generated recommendations can inuence physical equipment.

8. TAXONOMY OF AI-ENABLED DIGITAL TWIN SYSTEMS

   1. Classication by Sensor Modality

      AI-enabled digital twins can be classied according to the type and combination of sensor information used to update the virtual representation. Common modalities include vibration, temperature, current, pressure, rota- tional speed, acoustic emission, energy consumption, and process-quality measurements. Single-modality twins fo- cus on one dominant signal, whereas multi-sensor twins combine complementary observations to obtain a more complete representation of equipment condition.

      Multi-sensor integration is particularly useful when dif- ferent failure mechanisms produce dierent observable signatures. For example, vibration can reveal mechani- cal abnormalities while temperature and current provide additional thermal and electrical context. Sensor fusion can therefore improve the robustness of condition assess- ment, although synchronization, missing values, noise, and sensor reliability must be managed explicitly.

   2. Classication by Intelligence

      Digital-twin systems can also be classied according to the level of intelligence applied to the synchronized data. At the basic level, the twin supports monitoring and visu- alization. Higher levels introduce anomaly detection and diagnosis, predictive models for degradation and RUL, prescriptive optimization, and eventually autonomous maintenance or control. AI methods may include tra- ditional machine learning, deep learning, unsupervised detection, hybrid models, physics-informed learning, rein- forcement learning, and generative AI.

      This classication complements the GADT-ME matu- rity model used in this review. The distinction is im- portant because the presence of an AI algorithm does not by itself indicate that a twin performs predictive or autonomous maintenance. The functional role of the AI within the end-to-end system must be established.

9. Classication by Twin Function

   Digital twins can be dierentiated by the primary function they perform in the maintenance workow. A component or product twin focuses on an individual asset; a process twin represents a manufacturing process; and a system twin represents interacting machines and production re- sources. From a maintenance perspective, the functional role may further progress from monitoring and diagnosis to prediction, prescription, and autonomous control.

   This functional classication provides a bridge between the physical scope of the twin and the intelligence ap- plied to it. It also supports comparison of studies that use dierent manufacturing domains but address similar maintenance functions.

   1. Classication by Deployment

      Deployment architecture describes where sensing, Digital Twin synchronization, AI inference, and decision-making are executed. Cloud-based deployment provides scalable

      computing and centralized data management, whereas edge deployment places processing close to the physical equipment to reduce communication delay. Hybrid edge cloud architectures distribute tasks between local and remote resources.

      For closed-loop maintenance, deployment location di- rectly aects latency, connectivity, privacy, and compu- tational constraints. Time-critical anomaly detection or safety-related responses may require local processing, while model training, historical analysis, and large-scale eet analytics can be performed in cloud infrastructure.

10. REFERENCE ARCHITEC- TURE FOR AI-DRIVEN PRE- DICTIVE MAINTENANCE

    1. Physical Asset Layer

       The physical asset layer contains the machines, compo- nents, production lines, robots, actuators, and other equip- ment whose condition is represented by the Digital Twin. This layer is the source of the physical state that the virtual model seeks to represent. Typical assets include motors, bearings, pumps, conveyors, robotic arms, tur- bines, and manufacturing machines.

       Maintenance actions ultimately aect this layer. Con- sequently, any closed-loop architecture must maintain a clear relationship between predicted equipment state, au- thorized maintenance decisions, and the physical action applied to the asset.

    2. Sensing Layer

       The sensing layer captures measurements from the phys- ical asset using industrial sensors and data-acquisition devices. Vibration, temperature, current, pressure, speed, load, acoustic emission, and energy measurements pro- vide observations of operating condition. The quality of these observations determines the reliability of subsequent Digital Twin and AI functions.

       Sensor calibration, synchronization, redundancy, missing-data handling, and fault detection are therefore essential. A sensor failure or systematic drift can other- wise be interpreted incorrectly as equipment degradation.

11. Communication Layer

    The communication layer transfers measurements and control information between physical equipment, edge devices, Digital Twin services, AI models, and actuation systems. Industrial networks and messaging infrastruc- tures provide the connectivity required for continuous synchronization.

    Communication latency and packet loss are particularly important in closed-loop predictive maintenance. A tech- nically accurate prediction may still be operationally inef- fective if the information arrives too late for the required intervention. Communication security is also essential be- cause unauthorized modication of sensor data or control commands can aect physical equipment.

    1. Digital Twin Layer

       The Digital Twin layer maintains the virtual representa- tion of the physical asset or system and updates it as new observations arrive. It provides the state representation used by downstream analytics and can maintain historical information needed to understand degradation trends.

       The twin should preserve synchronization between phys- ical and virtual states while representing relevant operat- ing context. In predictive maintenance, this layer connects raw industrial observations with higher-level asset-health and maintenance information.

    2. AI Analytics Layer

       The AI analytics layer transforms synchronized Digital Twin data into maintenance-related intelligence. Depend- ing on the application, it can perform anomaly detection, fault diagnosis, health estimation, degradation prediction, RUL estimation, or maintenance optimization.

       Dierent models can be combined within this layer. For example, an unsupervised detector may identify an abnormal operating condition, followed by a supervised or deep-learning model that estimates the likely fault or future degradation. Hybrid and physics-informed models can provide additional engineering constraints.

    3. Decision and Actuation Layer

       The decision and actuation layer converts analytical re- sults into maintenance actions. At lower maturity levels, the output may be an alarm or operator notication. At higher levels, the system may recommend inspection, maintenance scheduling, parameter adjustment, or an- other corrective response. Autonomous operation requires an additional level of validation and secure authorization before commands reach physical equipment.

       The distinction between prediction and action is central to this review. A prediction becomes part of a closed-loop system only when its output is connected to a validated decision process and, where appropriate, a controlled physical response.

12. MATHEMATICAL FOUNDA- TIONS

    1. Digital Twin State

       The Digital Twin state can be represented by a vector of variables describing the current condition and oper- ating context of the physical system. A generic state representation is

       xt = [x1,t, x2,t,…, xn,t]T , (1)

       where each component may represent a measured or es- timated quantity such as temperature, vibration, speed, load, health index, or other relevant state variable at time t. The state is updated as new sensor observations become available.

       The quality of this state representation depends on sensor coverage, synchronization, preprocessing, and the ability of the twin to retain relevant temporal information. AI models use the resulting state or a sequence of states as input for diagnosis, prediction, or decision support.

       1. Remaining Useful Life

          Remaining Useful Life (RUL) represents the estimated amount of useful operating time remaining before an as- set reaches a dened failure or maintenance threshold. An AI model can estimate RUL from historical degrada- tion trajectories, current operating state, and contextual variables.

          RUL estimation is particularly useful for maintenance planning because it connects condition prediction with scheduling decisions. However, the estimate is uncertain and should be interpreted together with condence in- formation and operating context rather than as an exact failure time.

       2. Failure Probability

          Failure probability represents the estimated likelihood that an asset will enter a dened failure state within a specied prediction horizon. A generic formulation can be written as

          Figure 1: Closed-loop architecture of an AI-driven digital twin maintenance system. Each stage feeds forward into the next; outcome verication closes the loop back to the physical asset.

    2. Feature Normalization

       Sensor variables can have very dierent numerical ranges and units. Feature normalization reduces scale dierences before model training or inference. A common minmax transformation is

       xi = x xmin , (2)

       P (Ft+h = 1 | x1:t), (3)

       where t+h denotes failure within the future horizon and

       x1:t represents observations available up to time t.

       This measure supports risk-based maintenance deci- sions. The prediction horizon and failure denition must be stated clearly because a probability without a time window is dicult to interpret operationally.

       1. Z-Score Anomaly Detection

          The z-score measures how far an observation is from the mean in units of standard deviation. It can be used as a simple statistical anomaly indicator when the monitored variable has a reasonably stable reference distribution.

          For an observation x, mean , and standard deviation

          , the standardized score is

          x

          xmax xmin

          which maps a feature to a dened range when the mini- mum and maximum values are known. Standardization can alternatively be used when a zero-mean, unit-variance representation is appropriate.

          Consistent preprocessing is important because changes in scaling between training and deployment can alter model behavior. Normalization parameters should there- fore be derived from the training procedure and applied consistently during operational inference.

13. Health Index

    A health index (HI) provides a compact representation of equipment condition derived from one or more sensor variables or model outputs. It can combine vibration, temperature, current, operating load, and other indicators into a value that changes as equipment condition evolves.

    z = . (4)

    Large positive or negative values indicate observations that are far from the reference behavior. In industrial applications, the threshold should be calibrated using operating context because normal changes in load or speed can also produce large standardized values.

    1. Autoencoder Anomaly Score

       An autoencoder learns to reconstruct input observations. When trained primarily on normal operating data, it should reconstruct familiar patterns with relatively small error. The reconstruction error can therefore be used as an anomaly score.

       For an input vector x and reconstructed vector x, a mean-squared reconstruction score can be expressed as

       X

       n

       The HI is useful for Digital Twin maintenance because it provides a common state variable for visualization, trend analysis, degradation modelling, and RUL estimation. Its

       E = 1 (x

       n i

       i=1

       xi)2. (5)

       interpretation should remain consistent across operating conditions; otherwise changes in load or environment may be mistaken for degradation.

       A suciently large value of E can trigger an anomaly ag. Threshold calibration and changing operating conditions remain important limitations.

       Table 3: Comparative Analysis of Predictive Maintenance Approaches

       1. CHALLENGES AND RE- SEARCH GAPS

          Approach Data

          InterpretabilityReal-Time

          Main Limita-

          Rule-Based

          Need

          Suitability

          tion

          The evidence map identies several barriers to depend-

          Low High Very High Cannot learn

          complex degra-

          able AI-driven Digital Twin maintenance. These include

          Traditional ML Medium Medium

          High

          High

          dation behavior Depends strongly on engineered features

          limited validation of closed-loop actuation, scarcity of real failure data, domain shift, latency, sensor reliability, cyber- security, explainability, interoperability, and deployment

          Deep Learning High Low Medium Large datasets

          and compute re- quired

          cost. These challenges aect not only model accuracy but

          also the ability to move from laboratory demonstrations

          Unsupervised Detection

          Cloud AI Digi- tal Twin

          Edge AI Digital Twin

          Medium Medium High Threshold cali-

          bration can be dicult

          High Medium Medium Network

          latency, connec- tivity, privacy concerns

          Medium Medium Very High Limited edge

          memory and processing

          to reliable industrial operation.

          16-.1 Limited Closed-Loop Validation

          A major gap is the limited number of studies that demon- strate an end-to-end path from fault prediction to safe physical response. Many systems generate an alarm,

          Hybrid Edge Cloud Twin

          Medium High

          Medium High

          capacity Greater orches- tration and synchroniza-

          health estimate, or RUL value without validating the subsequent maintenance action. Future studies should therefore report decision logic, actuation constraints, re-

          tion complexity

    2. Total Closed-Loop Latency

    Closed-loop latency is the elapsed time between sens- ing a relevant physical change and completing the corre- sponding validated response. It can be considered as the combined delay of sensing, data transfer, Digital Twin synchronization, AI inference, decision-making, communi- cation, and actuation.

    A simplied representation is

    Ttotal = Tsense+Tcomm+Ttwin+TAI +Tdecision+Tactuation.

    (6)

    Reducing this latency is particularly important when equipment condition can deteriorate rapidly. Edge or hybrid architectures can reduce communication and in- ference delays, but they introduce additional resource- management and synchronization requirements.

14. COMPARATIVE ANALYSIS

15. CLOSED-LOOP PREDICTIVE MAINTENANCE

Closed-loop predictive maintenance extends conventional predictive maintenance by connecting prediction to a validated decision and, where authorized, to corrective ac- tion. The overall sequence can be represented as physical observation sensing Digital Twin synchronization

AI prediction decision actuation outcome verication.

Most predictive-maintenance systems reported in the evidence map stop at monitoring, diagnosis, prediction, or an operator alarm. A closed-loop system must addition- ally determine what action should follow, verify that the action is safe and appropriate, and observe the resulting equipment state. This prediction-to-action transition is therefore a central research gap identied by the review. Human approval can remain in the loop for maintenance actions that require additional operational or safety au- thorization.

sponse latency, safety mechanisms, and outcome verica- tion rather than evaluating prediction accuracy alone.

16.2 Scarcity of Industrial Failure Data

Real industrial failures are relatively rare and may be costly or unsafe to reproduce. Consequently, many datasets contain abundant normal observations but com- paratively few labelled failure events. Synthetic data generation, transfer learning, self-supervised learning, fed- erated learning, and carefully validated generative meth- ods can help address this limitation, but synthetic obser- vations should not be assumed to represent real failure mechanisms without validation.

1. Domain Shift

   AI models trained under one machine conguration, load prole, environment, or operating regime may perform dif- ferently when deployed under another. This domain-shift problem is especially relevant to Digital Twins because the physical system may change over time and because twins can be transferred across machines or production environ- ments. Continual learning, domain adaptation, transfer learning, and explicit operating-context representation can help maintain model reliability.

   1. Latency

      Prediction and intervention are time-sensitive in many industrial processes. Communication delay, cloud round trips, model inference time, and actuation delay can collec- tively reduce the usefulness of a prediction. Edge AI and hybrid edgecloud architectures can place time-critical functions closer to the equipent while retaining cloud resources for training and eet-level analysis.

   2. Sensor Reliability

      Sensor noise, missing values, drift, calibration errors, and sensor failure can degrade Digital Twin synchronization and AI performance. Redundant sensing, sensor-health monitoring, multi-sensor fusion, robust preprocessing, and

      uncertainty estimation can reduce the eect of unreliable measurements.

      Table 4: Major Research Gaps and Potential Research Direc- tions

      Research Current Problem Potential Direction

   3. Cybersecurity

      Gap

      A Digital Twin connected to physical equipment creates security requirements at both the data and control levels. Manipulated sensor measurements can cause incorrect

      Closed-loop control

      Limited la- belled failures

      Prediction often ends with an alarm

      Real industrial failure data are scarce

      Integrate constrained decision-making and vali- dated remote actuation Synthetic/GenAI data, transfer learning, self- supervised and federated

      predictions, while unauthorized control commands can directly aect equipment. Authentication, authorization,

      Cloud latency Network communica-

      tion delays interven-

      learning

      Edge AI and hybrid edge cloud architectures

      secure communication, signed commands, monitoring, and intrusion-detection mechanisms are therefore important for closed-loop deployment.

      Model general- ization

      Cybersecurity

      tion

      Performance decreases across machines or op- erating conditions Manipulated data or commands can aect physical equipment

      Domain adaptation and continual learning

      Secure communication, au- thentication, signed com- mands, intrusion detection

   4. Explainability

      Maintenance personnel may need to understand why an AI system generated an alarm or recommendation before acting on it. Interpretable models, feature-importance methods, condence estimates, and validated natural- language explanations can improve transparency. Gen-

      Explainability Operators may not

      trust opaque predic- tions

      Sensor quality Noise, drift, missing

      values, sensor failure

      SHAP, LIME, inter- pretable models, validated GenAI rationales, con- dence estimates

      Sensor fusion, redundancy, health monitoring, uncer- tainty estimation

      erative explanations should remain grounded in veried model outputs and should not be treated as independent evidence of equipment condition.

   5. Interoperability

      Industrial Digital Twins must exchange information across sensors, controllers, enterprise systems, AI services, and heterogeneous machines. Dierences in data models, com- munication protocols, semantic representations, and ven- dor interfaces can make integration dicult. Standardized interfaces and semantically consistent data representations are therefore important for scalable deployment.

2. Cost and SME Accessibility

   The deployment of AI-enabled Digital Twins can be dif- cult for small and medium enterprises (SMEs) because the complete solution may require sensors, industrial gate- ways, networking infrastructure, computing resources, software integration, data engineering, cybersecurity con- trols, and personnel with specialized expertise. A techni- cally eective predictive-maintenance architecture there- fore also needs to be evaluated in terms of acquisition cost, operating cost, maintenance eort, scalability, and required technical skills.

   For SMEs, a practical deployment can begin with a limited number of high-value assets and a focused sensor set rather than instrumenting the entire production facil- ity. Edge computing can reduce continuous cloud-transfer requirements, while hybrid edgecloud architectures can reserve cloud resources for model training, historical anal- ysis, eet-level learning, and long-term storage. Open interfaces and interoperable data representations can also reduce dependence on a single vendor. The objective is not simply to minimize the initial investment, but to achieve an acceptable balance among predictive-maintenance ben- et, computational cost, latency, reliability, and opera- tional complexity.

   The accessibility problem is also related to human ex- pertise. Maintenance personnel may need actionable ex- planations rather than raw model outputs. Carefully

   constrained GenAI interfaces may assist with natural- language explanations, maintenance documentation, and operator interaction, but recommendations aecting phys- ical equipment should remain grounded in veried sensor observations, validated model outputs, and appropriate safety controls.

3. FUTURE RESEARCH DIREC- TIONS

   The reviewed evidence indicates that future Digital Twin research needs to move from isolated prediction models to- ward deployable, secure, interoperable, and continuously improving cyberphysical systems. The following direc- tions are particularly relevant to predictive maintenance and closed-loop manufacturing.

   1. Edge AI

      Edge AI places selected analytics and inference functions close to the physical equipment. This reduces dependence on continuous cloud connectivity and can shorten the time between sensing, inference, and intervention. Edge deployment is particularly useful for vibration monitoring, anomaly detection, safety-related alarms, and other appli- cations where latency is important. Resource constraints remain a challenge because industrial edge devices may have limited memory, processing capability, and energy budgets. Model compression, lightweight architectures, hardware acceleration, and adaptive workload placement can therefore support practical deployment.

   2. Hybrid EdgeCloud Digital Twins

      A hybrid edgecloud Digital Twin can divide computation according to latency, data-volume, privacy, and computa- tional requirements. Time-critical anomaly detection and local control can be performed at the edge, while com- putationally intensive model training, historical analysis, eet-level comparison, and long-term storage can be per- formed in the cloud. Such architectures require reliable

      synchronization, workload orchestration, secure communi- cation, and mechanisms for maintaining consistent twin state across distributed computing locations.

4. Federated Learning

   Federated learning provides a potential approach for learn- ing from data distributed across multiple factories, ma- chines, or organizations without requiring all raw opera- tional data to be transferred to a central repository. Local models can be trained using local data and model updates can then be aggregated. For industrial Digital Twins, this can support collaborative learning while reducing direct sharing of sensitive production data. Important challenges include heterogeneous data distributions, communication overhead, model poisoning, privacy protection, and main- taining consistent model performance across participating sites.

   1. Multi-Sensor Fusion

      Predictive maintenance can benet from combining com- plementary measurements such as vibration, temperature, current, pressure, speed, load, and acoustic signals. Multi- sensor fusion can improve observability because a failure mechanism may not be clearly represented in a single sensor stream. Fusion can occur at the data, feature, or decision level. Future systems should also estimate sensor reliability so that a failed or degraded sensor does not dis- proportionately inuence the Digital Twin or maintenance decision.

   2. Explainable and Generative, Uncertainty-Aware AI

      Industrial maintenance ecisions require more than a prediction label. Operators may need to know which measurements contributed to an alert, how condent the system is, and what uncertainty surrounds a forecast. Ex- plainable AI can provide feature importance, local expla- nations, or interpretable model outputs, while uncertainty estimation can communicate the reliability of predictions. Generative AI can provide natural-language explanations and operator assistance, but generated content should remain grounded in veried data and model outputs. For safety-critical applications, uncertainty and validation information should accompany any AI-generated recom- mendation.

   3. Self-Healing Industrial Systems

      Self-healing systems aim to detect abnormal conditions, determine an appropriate response, execute an approved corrective action, and verify the outcome. In the context of Digital Twins, this represents a progression from pre- diction toward closed-loop operation. A safe self-healing architecture should include constrained decision-making, authorization, fail-safe behavior, action verication, and human override. Autonomous actuation should therefore be treated as a controlled engineering function rather than as an unrestricted AI capability.

      Figure 2: Proposed GenAI-Augmented Digital Twin Maturity Evaluation (GADT-ME) framework. Shading is illustrative of maturity progression.

   4. Semantic and Interoperable Digital Twins

      Industrial Digital Twins often operate across heteroge- neous machines, controllers, databases, communication protocols, and enterprise applications. Semantic interop- erability can help dierent components exchange informa- tion with consistent meanings rather than merely exchang- ing raw values. Future work should therefore investigate standardized information models, machine-readable asset descriptions, common semantic representations, and in- teroperable interfaces that allow twins to be integrated across production environments.

   5. Accessible, GenAI-Assisted Digital Twins for SMEs

      GenAI may make Digital Twin systems more accessi- ble by providing natural-language interfaces for querying equipment state, summarizing alarms, explaining model outputs, generating maintenance documentation, and as- sisting with troubleshooting. For SMEs, such interfaces could reduce the expertise required to interact with com- plex analytics platforms. However, generated explana- tions and recommendations should be constrained by trusted industrial data, model outputs, access controls, and safety policies. Research should therefore focus on af- fordable architectures in which GenAI complementsnot replacesvalidated engineering and AI components.

5. A PROPOSED EVALUATION FRAMEWORK: GENAI- AUGMENTED DIGITAL TWIN MATURITY

   The GenAI-Augmented Digital Twin Maturity Evaluation (GADT-ME) framework proposed in this review evaluates Digital Twin systems along two complementary dimen- sions: functional maturity and generative-AI capability. The framework is intended as an analytical taxonomy for comparing systems rather than as a claim that all reviewed studies implement every level.

   1. Framework Axes

      The rst axis represents functional maturity. Five levels are used: M1, monitoring; M2, diagnosis; M3, prediction or prognostics; M4, prescriptive decision-making or op- timization; and M5, autonomous operation. The second axis represents the role of generative AI: G0, no demon- strated GenAI capability; G1, GenAI augmentation; G2, GenAI co-pilot functionality; and G3, GenAI-enabled ac- tuation. The two axes are intentionally separated because a system may have a high predictive maturity without using GenAI, or it may use GenAI for explanation while remaining at a diagnostic or predictive level.

   2. Reading the Framework

      A system should be assigned a maturity level according to the highest capability demonstrated by the evidence available for that system. For example, a Digital Twin that detects faults but does not forecast future degrada- tion is classied at M2, whereas a system that estimates RUL is classied at M3. A system that recommends a maintenance action based on a prediction may reach M4, while validated autonomous execution of a corrective ac- tion is required for M5. GenAI coding follows the same evidence-based principle: natural-language explanation or data augmentation corresponds to G1, operator co- pilot functionality corresponds to G2, and GenAI directly participating in authorized actuation is required for G3.

6. Illustrative Scoring Procedure

   The GADT-ME framework is qualitative and evidence- based rather than a universal numerical benchmark. For an individual study, the reviewer rst identies the highest demonstrated functional maturity (M1M5) and then assigns the highest demonstrated GenAI capability (G0 G3). The resulting pair, such as M3/G1, summarizes the systems position in the framework. If a study does not demonstrate a maturity stagefor example, because it focuses on implementation determinants rather than a functioning maintenance systemthe relevant axis can be recorded as not reported (NR) rather than inferred. This conservative procedure prevents a study from receiving a higher maturity level solely because a technology is mentioned without evidence of implementation.

   The framework can subsequently be used to compare groups of studies by domain, sensor modality, deployment architecture, or maintenance task. Such comparisons should report the number of classiable studies and should distinguish bibliographic or abstract-level evidence from independently veried full-text evidence.

7. DISCUSSION

   The evidence synthesized in this review indicates that AI- enabled Digital Twins are increasingly capable of support- ing monitoring, diagnosis, and predictive maintenance, but the transition from prediction to validated closed-loop action remains comparatively challenging. The technical pipeline requires reliable sensing, synchronized twin state, appropriate AI models, decision logic, secure communica- tion, and controlled actuation. A weakness at any layer

   can reduce the reliability of the overall system.

   The analysis also shows that no single AI approach is sucient for every manufacturing condition. Traditional machine learning can provide useful and comparatively interpretable models for structured sensor features; deep learning can capture complex temporal or signal patterns; unsupervised methods can address the scarcity of labelled failures; and hybrid or physics-informed methods can incorporate engineering knowledge. GenAI adds a dif- ferent capability by supporting interaction, explanation, data augmentation, and recommendation drafting, but it requires grounding and safeguards before its outputs inuence physical operations.

   1. Practical Implications

      For manufacturing practitioners, the ndings support an incremental deployment strategy. Organizations can begin with reliable sensing and condition monitoring, es- tablish a synchronized Digital Twin, and then add diag- nosis, prediction, prescriptive analytics, and controlled actuation as evidence and validation increase. Edge com- puting can be used where latency is critical, while cloud resources can support training and eet-level analysis. SMEs may particularly benet from modular architec- tures, open interfaces, selective sensor deployment, and operator-oriented AI interfaces that reduce the expertise required to interpret system outputs.

      For researchers, evaluation should extend beyond pre- dictive accuracy. Important measures include false-alarm behavior, RUL error, inference and end-to-end latency, synchronization quality, robustness to sensor degradation and domain shift, cybersecurity, explainability, uncer- tainty, and the safety andreliability of any closed-loop action. These measures better reect the requirements of an industrial cyberphysical system than model accuracy alone.

8. CONCLUSION

This structured review examined the integration of Ar- ticial Intelligence (AI), Digital Twin (DT) technology, and predictive maintenance in smart manufacturing, with particular emphasis on the transition from condition mon- itoring and prediction toward prescriptive and closed- loop maintenance. The reviewed literature indicates that Digital Twins provide a cyberphysical representation of manufacturing assets by connecting physical equipment, sensor infrastructure, communication systems, virtual models, and analytical services. When AI is embedded within this architecture, Digital Twins can support fault diagnosis, anomaly detection, degradation assessment, Re- maining Useful Life (RUL) estimation, and maintenance decision-making.

The analysis shows that traditional machine learning, deep learning, unsupervised anomaly detection, hybrid models, and physics-informed approaches provide comple- mentary capabilities for predictive maintenance. Tradi- tional machine-learning methods are useful for structured sensor data and fault classication, while deep-learning approaches can capture complex temporal and signal pat- terns. Unsupervised methods are particularly relevant where labelled industrial failure data are scarce. Hybrid

and physics-informed approaches provide a means of com- bining data-driven learning with engineering knowledge and physical constraints, thereby supporting more physi- cally consistent degradation and RUL modelling.

A major nding of this review is the continuing prediction-to-action gap. Many reported Digital Twin sys- tems identify or predict abnormal equipment behaviour but stop at an alarm, dashboard, or maintenance rec- ommendation rather than demonstrating a complete and validated transition from prediction to safe corrective action. Moving toward closed-loop predictive mainte- nance therefore requires reliable decision-making, low- latency communication, secure actuation, accurate twin synchronization, uncertainty assessment, and appropriate human oversight. The transition from prediction to ac- tion should consequently be evaluated as an end-to-end cyberphysical process rather than as a machine-learning problem alone.

The review also identies persistent challenges involv- ing limited industrial failure data, domain shift, sensor reliability, latency, cybersecurity, explainability, interoper- ability, computational requirements, and deployment cost for small and medium-sized enterprises (SMEs). These challenges indicate that increasing model complexity alone is insucient for industrial deployment. Practical systems require dependable sensing, robust data pipelines, edge cloud coordination, secure communication, interpretable outputs, uncertainty awareness, and architectures that can be maintained under changing operating conditions. The proposed GenAI-Augmented Digital Twin Maturity Evaluation (GADT-ME) framework provides a structured perspective for assessing Digital Twin systems according to functional maturity and generative-AI capability. The framework distinguishes progression from monitoring and diagnosis through prediction and prescription toward au- tonomous operation, while separately representing the contribution of generative AI. This perspective can sup- port comparison of emerging systems that use GenAI for maintenance knowledge assistance, natural-language interaction, data augmentation, explanation, and recom- mendation generation. Because GenAI-enabled industrial actuation introduces additional safety and reliability con- cerns, generated recommendations should remain subject to validation, engineering constraints, and appropriate human or supervisory control before physical actions are

executed.

Future research should therefore focus on hybrid edge cloud Digital Twins, federated learning, multi-sensor fu- sion, explainable and uncertainty-aware AI, secure closed- loop actuation, self-healing industrial systems, semantic interoperability, and accessible GenAI-assisted Digital Twins for SMEs. Evaluation should extend beyond pre- dictive accuracy to include false-alarm behaviour, RUL error, inference and end-to-end latency, synchronization quality, robustness to sensor degradation and domain shift, cybersecurity, explainability, uncertainty, and the safety and reliability of closed-loop actions. These mea- sures better represent the requirements of an industrial cyberphysical system than model accuracy alone.

Overall, the literature demonstrates a continuing tran- sition from detect and predict toward understand, de- cide, and safely act. Realizing this transition will require coordinated advances in AI modelling, Digital Twin syn- chronization, sensing and communication infrastructure,

secure control, and human-centred deployment. For man- ufacturing organizations, an incremental pathway begin- ning with reliable sensing and monitoring, followed by diagnosis, prediction, prescriptive analytics, and validated actuation, provides a practical basis for increasing system capability while maintaining control over operational risk. For researchers, the remaining gaps provide opportunities to develop reproducible evaluation protocols, interoper- able architectures, trustworthy GenAI integration, and validated closed-loop Digital Twin systems for realistic industrial environments.

1. PRIMARY-STUDY EVIDENCE MATRIX

   Table 5: Primary-study evidence matrix for AI-enabled digital twins and predictive maintenance.

   Year	Title	Venue	DOI	AI Method	Domain	Task	M	G	Quality
   2019	Digital Twin for Machining Tool	Procedia CIRP	10.1016/j.procir.2019.04.049	Deep	Machining/CNC	Tool condition	M3	G0	Moderate
   	Condition Prediction			learning/DSGRU		prediction
   2019	The use of Digital Twin for	International Journal	10.1080/0951192X.2019.1686173	Physics-based DT	Manufacturing	RUL/predictive	M3	G0	Moderate
   	predictive maintenance in	of Computer			robot	maintenance
   	manufacturing	Integrated
   		Manufacturing
   2019	Methodology for enabling Digital	Procedia CIRP	10.1016/j.procir.2019.03.072	Physics-based	Industrial	Predictive	M1	G0	Moderate
   	Twin using advanced			modelling	robot	maintenance
   	physics-based modelling in
   	predictive maintenance
   2020	A hybrid predictive maintenance	Robotics and	10.1016/j.rcim.2020.101974	Hybrid DT +	CNC	RUL/prediction	M3	<>G0	High
   	approach for CNC machine tool	Computer-Integrated		particle ltering
   	driven by Digital Twin	Manufacturing
   2020	Approach for a Holistic Predictive Maintenance Strategy	Procedia 10.1016/j.promfg.2020.01.265 Manufacturing		Simulation/statistical DT	Production machine	Predictive maintenance	M3	G0	Moderate
   	by Incorporating a Digital Twin
   2021	Digital twin modeling for	Forschungsingenieurwesen10.1007/s10010-021-00468-9		DT degradation	Wind turbine	Degradation/RUL	M3	G0	Moderate
   	predictive maintenance of			model	gearbox
   	gearboxes in oating oshore
   	wind turbine drivetrains
   2021	Prediction maintenance	Journal of 10.1016/j.jmsy.2020.08.001		Decision/optimization	Industrial	Maintenance	M4	G0	Moderate
   	integrated decision-making	Manufacturing			maintenance	decision support
   	approach supported by digital	Systems
   	twin-driven cooperative
   	awareness and interconnection
   	framework
   2022	Digital Twin Enabled Industry	IEEE 10.1109/ICEEICT53079.2022.9768590		ML/DT	Industry 4.0	Predictive	M3	G0	Moderate
   	4.0 Predictive Maintenance					maintenance
   	Under Reliability-Centred
   	Strategy
   2022	A Digital Twin Design for	Procedia CIRP 10.1016/j.procir.2022.05.268		Degradation/RUL	Milling	Maintenance	M4	G0	Moderate
   	Maintenance Optimization			model	machine	optimization
   2022	Feature investigation with	IFAC-PapersOnLine 10.1016/j.ifacol.2022.04.182		Feature	Steel	PdM features	M2	G0	Moderate
   	Digital Twin for predictive			selection/ML	production
   	maintenance following a machine
   	learning approach
   2023	Digital twin-enhanced predictive	Energy and Buildings 10.1016/j.enbuild.2023.113738		LSTM	Building	Failure prediction	M3	G0	Moderate
   	maintenance for indoor climate			autoencoder	systems
   2023	A digital twin framework for	Computers in 10.1016/j.compind.2023.103948		Data-driven +	Manufacturing	PHM	M2	G0	Moderate
   	prognostics and health	Industry		physics	equipment
   	management
   2023	Digital Twins for predictive	Procedia CIRP 10.1016/j.procir.2023.03.087		IoT/service	Industrial asset	PdM case	M3	G0	Moderate
   	maintenance: A case study for a			architecture
   	exible IT-architecture
   2023	Digital Twins for Predictive	Procedia CIRP 10.1016/j.procir.2023.06.053		Statistical + ML	Virtual factory	PdM	M3	G0	Moderate
   	Maintenance
   2023	Experience from implementing	Journal of Intelligent 10.1007/s10845-023-02078-4		DT +	Industrial	Maintenance	M3	G0	High
   	digital twins for maintenance in	Manufacturing		anomaly/RUL	process	analytics
   	industrial processes			analytics
   2023	Predictive maintenance of	Reliability 10.1016/j.ress.2023.109859		AI regression	Steel	Failure prediction	M3	G0	High
   	standalone steel industrial	Engineering & System			components
   	components powered by a	Safety
   	dynamic reliability digital twin
   	model with articial intelligence

   Year	Title	Venue	DOI	AI Mehod	Domain	Task	M	G	Quality
   2024	Machine learning-based digital	International Journal	10.1007/s00170-024-14097-3	Machine learning	Conveyor	PdM	M3	G0	High
   	twin of a conveyor belt for	of Advanced
   	predictive maintenance	Manufacturing
   		Technology
   2024	Incremental learning with	Engineering	10.1016/j.engappai.2024.108212	Incremental/deep	Bearing	Fault diagnosis	M2	G0	High
   	multi-delity information fusion	Applications of		learning
   	for digital twin-driven bearing	Articial Intelligence
   	fault diagnosis
   2024	Twin data multimode	Engineering	10.1016/j.engappai.2024.109142	Transfer learning	Bearing	Failure diagnosis	M2	G0	High
   	collaborative transfer learning	Applications of
   	for bearing failure diagnosis	Articial Intelligence
   2024	Deep learning-based digital twin	Journal of Intelligent	10.3233/JIFS-233863	Deep learning	Rapier loom	PdM	M3	G0	Moderate
   	for intelligent predictive	& Fuzzy Systems
   	maintenance of rapier loom
   2024	Digital Twin-based Predictive	Procedia Computer	10.1016/j.procs.2024.01.050	Hybrid	Sheet-metal	PdM	M3	G0	Moderate
   	Maintenance for Sheet Metal	Science		physics/data	bending
   	Bending
   2024	Digital twin (DT)-based	Procedia Computer	10.1016/j.procs.2024.06.058	AI/ML	6G network	PdM	M3	G0	Moderate
   	predictive maintenance of a 6G	Science
   	communication network
   2024	Real-world prototyping and	Journal of Building	10.1016/j.jobe.2024.110890	DT platform	Facilities	Predictive	M3	G0	High
   	evaluation of digital twins for	Engineering		evaluation		maintenance
   	predictive facility maintenance
   2024	IoT and Digital Twin: a new	Procedia Structural	10.1016/j.prostr.2024.09.334	IoT/DT	Cultural	Predictive	M3	G0	Moderate
   	perspective for Cultural Heritage	Integrity			heritage	maintenance
   	predictive maintenance
   2024	From Sensors to Digital Twins toward an Iterative Approach for	Sensors	10.3390/s24051434	Sensor/DT integration	Manufacturing systems	Monitoring/integration M1		G0	Moderate
   	Existing Manufacturing Systems
   2025	Data-Driven Digital Twin	Machines	10.3390/machines13060481	ML comparison	CNC turning	PdM	M3	G0	High
   	Framework for Predictive
   	Maintenance of Smart
   	Manufacturing Systems
   2025	Rolling bearings remaining	Engineering	10.1016/j.engappai.2025.111070	Physics-informed	Rolling bearing	RUL	M3	G0	High
   	useful life estimation using	Applications of		NN
   	digital twin and	Articial Intelligence
   	physics-informed methods with
   	uncertainty quantication
   2025	Bridging gray-box modeling and	Measurement:	10.1016/j.meadig.2025.100018	Gray-box + ML	Refrigeration	Identication/PdM	M3	G0	Moderate
   	machine learning: a digital twin	Digitalization
   	approach to refrigeration system
   	identication and predictive
   	maintenance
   2025	Digital twin integration in	Frontiers in	10.3389/fmech.2025.1655565	Datafusion/deep	Metalworking	PdM	M3	G0	Moderate
   	metalworking: enhancing	Mechanical		learning
   	eciency and predictive	Engineering
   	maintenance
   2025	A simulation-based digital twin	Flexible Services and	10.1007/s10696-024-09579-1	Actor-critic RL	Production line	Maintenance	M3	G0	High
   	for data-driven maintenance	Manufacturing				scheduling
   	scheduling of risk-prone	Journal
   	production lines via actor critics
   2025	Generative and Predictive AI for digital twin systems in	Frontiers in Articial Intelligence	10.3389/frai.2025.1655470	GenAI + Manufacturing/w predictive AI		eldQinugality/PdM assistance	M3	G1	Moderate
   	manufacturing
   2025	Investigating determinants of	Journal of Quality in	10.1108/JQME-05-2024-0047	MCDM/DEMATEL/ISIMndustrial im-		PdM determinants	NR	NR	Moderate
   	digital twins for predictive	Maintenance		plementation
   	maintenance	Engineering
   2025	Digital Twin-Driven Predictive	IEEE CISES	10.1109/CISES66934.2025.11265639	ML Industry 5.0		PdM	M3	G0	Moderate
   	Maintenance Using ML Models
   	in Industry 5.0 Environments

   Year	Title	Venue	DOI	AI Method	Domain	Task	M	G	Quality
   2026	System and system-of-systems	Journal of Intelligent	10.1007/s10845-025-02789-w	Real-time	Ring rolling	Root-cause	M2	G0	High
   	digital twins for predictive	Manufacturing		analytics/root	process	analysis/PdM
   	maintenance and root cause			cause
   	analysis
   2026	Physics-informed auto-encoder	Engineering	10.1016/j.engappai.2025.113017	Physics-informed	Rolling bearing	Fault diagnosis	M2	G0	High
   	based on digital twin for rolling	Applications of		autoencoder
   	bearing fault diagnosis under	Articial Intelligence
   	imbalanced sample conditions

   Note: The matrix contains the empirical primary/implementation records currently represented in the evidence map. M/G codes were assigned using the conservative GADT-ME coding protocol described in Section 9.4. NR denotes no demonstrated maturity stage and is excluded from the maturity denominator.

2. REVIEW / SURVEY EVIDENCE MATRIX

   Table 6: Review and survey evidence matrix used for landscape and contextual synthesis.

   Year	Title	Venue	DOI	AI Method	Domain	Task	Quality
   2020	Digital Twin for maintenance: A litera-	Computers in Industry	10.1016/j.compind.2020.103316	Literature review	Industrial mainte-	Review	High
   	ture review				nance
   2020	Digital Twin Models in Industrial Oper-	Procedia Manufacturing	10.1016/j.promfg.2020.02.084	Systematic review	Industrial opera-	Review	High
   	ations: A Systematic Literature Review				tions
   2022	Predictive maintenance using digital	Information and Software	10.1016/j.infsof.2022.107008	Systematic review	Industrial PdM	Review of 42 primary	High
   	twins: A systematic literature review	Technology				studies
   2022	Digital Twin-Driven Machine Condition	Journal of Sensors	10.1155/2022/6129995	Literature review	Condition moni-	Review	High
   	Monitoring: A Literature Review				toring
   2022	Advances of Digital Twins for Predictive	Procedia Computer Sci-	10.1016/j.procs.2022.01.348	Review	PdM	Review	Moderate
   	Maintenance	ence
   2022	Digital Twin for Integration of Design-	Chinese Journal of Me-	10.1186/s10033-022-00760-x	Overview	Manufacturing	Review	High
   	Manufacturing-Maintenance: An	chanical Engineering
   	Overview
   2023	Overview of predictive maintenance	Heliyon	10.1016/j.heliyon.2023.e14534	Review	PdM	Review	High
   	based on digital twin technology
   2023	The advance of digital twin for predic-	Journal of Manufacturing	10.1016/j.jmsy.2023.10.010	Review	PdM/ML	Review	High
   	tive maintenance: The role and function	Systems
   	of machine learning
   2024	Articial intelligence in digital twins	Data & Knowledge Engi-	10.1016/j.datak.2024.102304	Systematic review	AI + DT	Review of 149 studies	High
   	A systematic literature review	neering
   2024	Systematic review of predictive main-	PeerJ Computer Science	10.7717/peerj-cs.1943	Systematic review	PdM + DT	Review	High
   	tenance and digital twin technologies
   	challenges, opportunities, and best prac-
   	tices

   Year	Title	Venue	DOI	AI Method	Domain	Task	Quality
   2024	Predictive maintenance in Industry 4.0:	CIRP Journal of Manufac-	10.1016/j.cirpj.2024.02.003	Systematic mapping	Industry 4.0 PdM	Review	High
   	A systematic multi-sector mapping	turing Science and Tech-
   		nology
   2025	AI-enhanced digital twins in main-	Journal of Manufacturing	10.1016/j.jmsy.2025.07.006	SLR + industry inter-	AI-DT mainte-	Review of 51 cases	High
   	tenance: Systematic review, indus-	Systems		views	nance
   	trial challenges, and bridging research
   	practice gaps
   2025	Systematic review of predictive main-	Intelligent Systems with	10.1016/j.iswa.2025.200501	Systematic review	Manufacturing	Review	High
   	tenance practices in the manufacturing	Applications			PdM
   	sector
   2025	Generative AI-Empowered Digital	IEEE Transactions on In-	10.1109/TII.2025.3540473	Survey/taxonomy	GenAI + DT	Review	High
   	Twin: A Comprehensive Survey With	dustrial Informatics
   	Taxonomy
   2025	Generative AI in AI-Based Digital	Applied Sciences	10.3390/app15063166	State-of-the-	GenAI + DT +	Review	High
   	Twins for Fault Diagnosis for Predic-			art/review	PdM
   	tive Maintenance in Industry 4.0/5.0
   2025	Electronic Articial Intelligence and	Machines	10.3390/machines13090755	Systematic review	AI + DT + Indus-	Review	High
   	Digital Twins in Industry 5.0: A Sys-				try 5.0
   	tematic Review and Perspectives
   2025	Digital Twins, Extended Reality, and	Sustainability	10.3390/su17052318	Systematic review	Manufacturing	Review	High
   	Articial Intelligence in Sustainability
   	Manufacturing Reconguration: A Sys-
   	tematic Literature Review
   2025	Digital twin technology in wind turbine	Intelligent Systems with	10.1016/j.iswa.2025.200535	Review	Wind energy DT	Review	High
   	components: A review	Applications
   2025	Digital twins in manufacturing: a uni-	Annual Reviews in Control	10.1016/j.arcontrol.2025.101031	Conceptual review	Manufacturing	Review	Moderate
   	ed conceptual framework				DT
   2026	AI-driven predictive maintenance for in-	International Journal of	10.1007/s00170-026-17531-w	Systematic review	AI + PdM	Review	High
   	dustry 4.0: a systematic review of mod-	Advanced Manufacturing
   	els, methods, and challenges	Technology
   2026	AI-Driven Digital Twins for Manufac-	Sensors	10.3390/s26010124	Review	AI + DT manu-	Review	High
   	turing: A Review Across Hierarchical				facturing
   	Manufacturing System Levels
   2026	Generative AI for Predictive Mainte-	Procedia CIRP	10.1016/j.procir.2025.09.012	Survey	GenAI + PdM	Review	Moderate
   	nance for Manufacturing: A Survey
   2026	Exploring AI in Digital Twin: trends,	Digital Twin	10.1080/27525783.2026.2704992	Review	AI + DT	Review	High
   	challenges, enets, and contributions
   2026	Digitalised Predictive Maintenance in	Infrastructures	10.3390/infrastructures11030087	PRISMA SLR	Railways	Review of 73 articles	High
   	Railways: A Systematic Review of AI,
   	BIM, and Digital Twins
   2025	Digital twin innovations for induction	Next Research	10.1016/j.nexres.2025.100969	Review	Induction motor	Review of 60 studies	High
   	motor monitoring and predictive main-
   	tenance: A 20202025 review

   Year Title Venue DOI AI Method Domain Task Quality

   2025	A tutorial on digital twins for predictive Woodhead Publishing	10.1016/B978-0-443-15476-	Tutorial/review	PdM	Tutorial	Moderate
   	maintenance	8.00005-8

   Note: Review/survey records provide landscape and contextual evidence and are not counted as empirical maturity demonstrations.

3. DOI-BASED REFERENCE VERIFICATION REGISTER

   Table 7: DOI-based reference verication register for the evidence map.

   Year	Title	Venue		DOI	Evidence Type; Biblio- graphic Verication	Full-Text Coding / Status
   2019 2019	Digital Twin for Machining Tool Condition Prediction The use of Digital Twin for predic-	Procedia CIRP International	Journal	10.1016/j.procir.2019.04.049 10.1080/0951192X.2019.1686173	Primary; DOI/publisher- index record identied Primary; DOI/publisher-	Bibliographically veried; maturity coding based on available publisher/index evidence Bibliographically veried; maturity coding based on

   tive maintenance in manufacturing

   of Computer Integrated Manufacturing

   index record identied

   available publisher/index evidence

   2019 Methodology for enabling Digital Twin using advanced physics-based modelling in predictive maintenance

   Procedia CIRP 10.1016/j.procir.2019.03.072 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2020 A hybrid predictive maintenance ap- proach for CNC machine tool driven by Digital Twin

   Robotics and Computer- Integrated Manufactur- ing

   10.1016/j.rcim.2020.101974 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2020 Approach for a Holistic Predictive Maintenance Strategy by Incorpo- rating a Digital Twin

   Procedia Manufacturing 10.1016/j.promfg.2020.01.265 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2020 On the requirements of digital twin- driven autonomous maintenance

   Annual Reviews in Con- trol

   10.1016/j.arcontrol.2020.08.003 Framework/architecture;

   DOI/publisher-index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2021 Digital twin modeling for predictive maintenance of gearboxes in oating oshore wind turbine drivetrains

   Forschungsingenieurwesen 10.1007/s10010-021-00468-9 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2021 Prediction maintenance integrated decision-making approach sup- ported by digital twin-driven cooperative awareness and intercon- nection framework

   Journal of Manufacturing Systems

   10.1016/j.jmsy.2020.08.001 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   2021 Holistic Concept Towards a Refer- ence Architecture Model for Predic- tive Maintenance

   2021 Low-cost real-time multivariate pro- cess control from large sensor arrays (DOLPHINS)

   2022 Digital Twin Enabled Industry

   4.0 Predictive Maintenance Under Reliability-Centred Strategy

   2022 A Digital Twin Design for Mainte- nance Optimization

   2022 Feature investigation with Digital Twin for predictive maintenance fol- lowing a machine learning approach

   Procedia CIRP 10.1016/j.procir.2021.11.241 Framework/architecture;

   DOI/publisher-index record identied

   Frontiers in Chemistry 10.3389/fchem.2021.734132 Adjacent/technical;

   DOI/publisher-index record identied

   IEEE 10.1109/ICEEICT53079.2022.9768590 Primary; DOI/publisher-

   index record identied

   Procedia CIRP 10.1016/j.procir.2022.05.268 Primary; DOI/publisher-

   index record identied IFAC-PapersOnLine 10.1016/j.ifacol.2022.04.182 Primary; DOI/publisher-

   index record identied

   Bibliographically veried; maturity coding based on available publisher/index evidence

   Bibliographically veried; maturity coding based on available publisher/index evidence

   Bibliographically veried; maturity coding based on available publisher/index evidence

   Bibliographically veried; maturity coding based on available publisher/index evidence Bibliographically veried; maturity coding based on available publisher/index evidence

   Year	Title	Venue	DOI	Evidence Type; Biblio- graphic Verication	Full-Text Coding / Status
   2023	Digital twin-enhanced predictive	Energy and Buildings	10.1016/j.enbuild.2023.113738	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	maintenance for indoor climate			index record identied	available publisher/index evidence
   2023	A digital twin framework for prog-	Computers in Industry	10.1016/j.compind.2023.103948	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	nostics and health management			index record identied	available publisher/index evidence
   2023	Reference architecture for digital	Computers & Industrial	10.1016/j.cie.2023.109099	Framework/architecture;	Bibliographically veried; maturity coding based on
   	twin-based predictive maintenance	Engineering		DOI/publisher-index	available publisher/index evidence
   	systems			record identied
   2023	Digital Twins for predictive mainte-	Procedia CIRP	10.1016/j.procir.2023.03.087	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	nance: A case study for a exible				index record identied	available publisher/index evidence
   	IT-architecture
   2023	Digital Twins for Predictive Mainte-	Procedia CIRP	10.1016/j.procir.2023.06.053	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	nance			index record identied	available publisher/index evidence
   2023	Experience from implementing digi-	Journal of Intelligent	10.1007/s10845-023-02078-4	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	tal twins for maintenance in indus-	Manufacturing		index record identied	available publisher/index evidence
   	trial processes
   2023	Predictive maintenance of stan-	Reliability Engineering &	10.1016/j.ress.2023.109859	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	dalone steel industrial components	System Safety		index record identied	available publisher/index evidence
   	powered by a dynamic reliability dig-
   	ital twin model with articial intel-
   	ligence
   2024	Machine learning- based digital twin	International Journal of	10.1007/s00170-024-14097-3	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	of a conveyor belt for predictive	Advanced Manufacturing		index record identied	available publisher/index evidence
   	maintenance	Technology
   2024	Incremental learning with multi-	Engineering Applications	10.1016/j.engappai.2024.108212	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	delity information fusion for digital	of Articial Intelligence		index record identied	available publisher/index evidence
   	twin-driven bearing fault diagnosis
   2024	Twin data multimode collaborative	Engineering Applications	10.1016/j.engappai.2024.109142	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	transfer learning for bearing failure	of Articial Intelligence		index record identied	available publisher/index evidence
   	diagnosis
   2024	Deep learning-based digital twin for	Journal of Intelligent &	10.3233/JIFS-233863	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	intelligent predictive maintenance of	Fuzzy Systems		index record identied	available publisher/index evidence
   	rapier loom
   2024	Digital Twin-based Predictive Main-	Procedia Computer Sci-	10.1016/j.procs.2024.01.050	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	tenance for Sheet Metal Bending	ence		index record identied	available publisher/index evidence
   2024	Digital twin (DT)- based predictive	Procedia Computer Sci-	10.1016/j.procs.2024.06.058	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	maintenance of a 6G communication	ence		index record identied	available publisher/index evidence
   	network
   2024	Real-world prototyping and evalua-	Journal of Building Engi-	10.1016/j.jobe.2024.110890	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	tion of digital twins for predictive	neering		index record identied	available publisher/index evidence
   	facility maintenance
   2024	IoT and Digital Twin: a new per-	Procedia Structural In-	10.1016/j.prostr.2024.09.334	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	spective for Cultural Heritage pre-	tegrity		index record identied	available publisher/index evidence
   	dictive maintenance

   Year	Title	Venue	DOI	Evidence Type; Biblio- graphic Verication	Full-Text Coding / Status
   2024	From Sensors to Digital Twins to-	Sensors	10.3390/s24051434	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	ward an Iterative Approach for Ex-			index record identied	available publisher/index evidence
   	isting Manufacturing Systems
   2024	State-of-the-art review and synthe-	Advanced Engineering In-	10.1016/j.aei.2024.102800	Framework/architecture;	Bibliographically veried; maturity coding based on
   	sis: A requirement-based roadmap	formatics		DOI/publisher-index	available publisher/index evidence
   	for standardized predictive mainte-			record identied
   	nance automation using digital twin
   	technologies
   2025	Data-Driven Digital Twin Frame-	Machines	10.3390/machines13060481	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	work for Predictive Maintenance of			index record identied	available publisher/index evidence
   	Smat Manufacturing Systems
   2025	Rolling bearings remaining useful	Engineering Applications	10.1016/j.engappai.2025.111070	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	life estimation using digital twin and	of Articial Intelligence		index record identied	available publisher/index evidence
   	physics-informed methods with un-
   	certainty quantication
   2025	Bridging gray-box modeling and	Measurement: Digitaliza-	10.1016/j.meadig.2025.100018	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	machine learning: a digital twin	tion		index record identied	available publisher/index evidence
   	approach to refrigeration system
   	identication and predictive mainte-
   	nance
   2025	Digital twin integration in metal-	Frontiers in Mechanical	10.3389/fmech.2025.1655565	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	working: enhancing eciency and	Engineering		index record identied	available publisher/index evidence
   	predictive maintenance
   2025	A simulation-based digital twin for	Flexible Services and	10.1007/s10696-024-09579-1	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	data- driven maintenance scheduling	Manufacturing Journal		index record identied	available publisher/index evidence
   	of risk- prone production lines via
   	actor critics
   2025	Method for the development and ap-	Production Engineering	10.1007/s11740-025-01346-x	Framework/architecture;	Bibliographically veried; maturity coding based on
   	plication of digital twins in manu-			DOI/publisher-index	available publisher/index evidence
   	facturing			record identied
   2025	Generative and Predictive AI for dig-	Frontiers in Articial In-	10.3389/frai.2025.1655470	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	ital twin systems in manufacturing	telligence		index record identied	available publisher/index evidence
   2025	Investigating determinants of digital	Journal of Quality in	10.1108/JQME-05-2024-0047	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	twins for predictive maintenance	Maintenance Engineering		index record identied	available publisher/index evidence
   2025	Digital Twin-Driven Predictive	IEEE CISES	10.1109/CISES66934.2025.112656399	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	Maintenance Using ML Models in			index record identied	available publisher/index evidence
   	Industry 5.0 Environments
   2026	System and system- of-systems digi-	Journal of Intelligent	10.1007/s10845-025-02789-w	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	tal twins for predictive maintenance	Manufacturing		index record identied	available publisher/index evidence
   	and root cause analysis
   2026	Physics-informed auto-encoder	Engineering Applications	10.1016/j.engappai.2025.113017	Primary; DOI/publisher-	Bibliographically veried; maturity coding based on
   	based on digital twin for rolling	of Articial Intelligence		index record identied	available publisher/index evidence
   	bearing fault diagnosis under
   	imbalanced sample conditions

   Year	Title	Venue	DOI	Evidence Type; Biblio- graphic Verication	Full-Text Coding / Status
   2026	A Critical Analysis and Roadmap	Applied System Innova-	10.3390/asi9060123	Framework/architecture;	Bibliographically veried; maturity coding based on
   	for the Development and Implemen-	tion		DOI/publisher-index	available publisher/index evidence
   	tation of Predictive Maintenance			record identied
   	Digital Twins
   2020	Digital Twin for maintenance: A lit-	Computers in Industry	10.1016/j.compind.2020.103316	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	erature review			index record identied	mary M/G percentages
   2020	Digital Twin Models in Industrial	Procedia Manufacturing	10.1016/j.promfg.2020.02.084	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	Operations: A Systematic Litera-			index record identied	mary M/G percentages
   	ture Review
   2022	Predictive maintenance using digi-	Information and Software	10.1016/j.infsof.2022.107008	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	tal twins: A systematic literature	Technology		index record identied	mary M/G percentages
   	review
   2022	Digital Twin-Driven Machine Con-	Journal of Sensors	10.1155/2022/6129995	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	dition Monitoring: A Literature Re-			index record identied	mary M/G percentages
   	view
   2022	Advances of Digital Twins for Pre-	Procedia Computer Sci-	10.1016/j.procs.2022.01.348	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	dictive Maintenance	ence		index record identied	mary M/G percentages
   2022	Digital Twin for Integration of	Chinese Journal of Me-	10.1186/s10033-022-00760-x	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	Design- Manufacturing- Mainte-	chanical Engineering		index record identied	mary M/G percentages
   	nance: An Overview
   2023	Overview of predictive maintenance	Heliyon	10.1016/j.heliyon.2023.e14534	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	based on digital twin technology			index record identied	mary M/G percentages
   2023	The advance of digital twin for pre-	Journal of Manufacturing	10.1016/j.jmsy.2023.10.010	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	dictive maintenance: The role and	Systems		index record identied	mary M/G percentages
   	function of machine learning
   2024	Articial intelligence in digital	Data & Knowledge Engi-	10.1016/j.datak.2024.102304	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	twinsA systematic literature re-	neering		index record identied	mary M/G percentages
   	view
   2024	Systematic review of predictive	PeerJ Computer Science	10.7717/peerj-cs.1943	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	maintenance and digital twin tech-			index record identied	mary M/G percentages
   	nologies challenges,
   2024	Predictive maintenance in Industry	CIRP Journal of Man-	10.1016/j.cirpj.2024.02.003	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	4.0: A systematic multi- sector map-	ufacturing Science and		index record identied	mary M/G percentages
   	ping	Technology
   2025	AI-enhanced digital twins in mainte-	Journal of Manufacturing	10.1016/j.jmsy.2025.07.006	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	nance: Systematic review, industrial	Systems		index record identied	mary M/G percentages
   	challenges, and bridging research
   	practice gaps
   2025	Systematic review of predictive	Intelligent Systems with	10.1016/j.iswa.2025.200501	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	maintenance practices in the manu-	Applications		index record identied	mary M/G percentages
   	facturing sector
   2025	Generative AI- Empowered Digi-	IEEE Transactions on In-	10.1109/TII.2025.3540473	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	tal Twin: A Comprehensive Survey	dustrial Informatics		index record identied	mary M/G percentages
   	With Taxonomy

   Year	Title	Venue	DOI	Evidence Type; Biblio- graphic Verication	Full-Text Coding / Status
   2025	Generative AI in AI- Based Digital	Applied Sciences	10.3390/app15063166	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	Twins for Fault Diagnosis for Predic-			index record identied	mary M/G percentages
   	tive Maintenance in Industry 4.0/5.0
   2025	Electronic Articial Intelligence and	Machines	10.3390/machines13090755	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	Digital Twins in Industry 5.0: A			index record identied	mary M/G percentages
   	Systematic Review and Perspectives
   2025	Digital Twins, Extended Reality,	Sustainability	10.3390/su17052318	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	and Articial Intelligence in Man-			index record identied	mary M/G percentages
   	ufacturing Reconguration: A Sys-
   	tematic Literature Review
   2025	Digital twin technology in wind tur-	Intelligent Systems with	10.1016/j.iswa.2025.200535	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	bine components: A review	Applications		index record identied	mary M/G percentages
   2025	Digital twins in manufacturing: A	Annual Reviews in Con-	10.1016/j.arcontrol.2025.101031	Review; DOI/publisher-	Veried: review/context rcord; not used for pri-
   	unied conceptual framework	trol		index record identied	mary M/G percentages
   2026	AI-driven predictive maintenance	International Journal of	10.1007/s00170-026-17531-w	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	for industry 4.0: a systematic review	Advanced Manufacturing		index record identied	mary M/G percentages
   	of models, methods, and challenges	Technology
   2026	AI-Driven Digital Twins for Manu-	Sensors	10.3390/s26010124	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	facturing: A Review Across Hierar-			index record identied	mary M/G percentages
   	chical Manufacturing System Levels
   2026	Generative AI for Predictive Mainte-	Procedia CIRP	10.1016/j.procir.2025.09.012	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	nance for Manufacturing: A Survey			index record identied	mary M/G percentages
   2026	Exploring AI in Digital Twin:	Digital Twin	10.1080/27525783.2026.2704992	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	trends, challenges, benets, and con-			index record identied	mary M/G percentages
   	tributions
   2026	Digitalised Predictive Maintenance	Infrastructures	10.3390/infrastructures11030087	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	in Railways: A Systematic Review			index record identied	mary M/G percentages
   	of AI, BIM, and Digital Twins
   2025	Digital twin innovations for induc-	Next Research	10.1016/j.nexres.2025.100969	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	tion motor monitoring and predic-			index record identied	mary M/G percentages
   	tive maintenance: A 2020 2025 re-
   	view
   2025	A tutorial on digital twins for pre-	Woodhead Publishing	10.1016/B978-0-443-15476-8.00005-	Review; DOI/publisher-	Veried: review/context record; not used for pri-
   	dictive maintenance		8	index record identied	mary M/G percentages

   Note: Bibliographic verication records identify the DOI/publisher-index record. Primary-study maturity codes use the conservative GADT-ME protocol and available publisher/index evidence; review/context records are separated from primary-study maturity calculations.

   Declarations

   Funding

   Not applicable. This research received no specic grant from any funding agency in the public, commercial, or not-for-prot sectors.

   Conicts of interest/Competing interests

   The authors declare that they have no conicts of interest or competing interests that could have inuenced the work reported in this manuscript.

   Authors contributions

   All authors contributed to the conception and design of the study, literature search and analysis, interpretation of the ndings, and preparation and revision of the manuscript. All authors have read and approved the nal version of the manuscript.

   References

   1. M. Achouch, M. Dimitrova, K. Ziane, S. S. Karganroudi, R. Dhouib, H. Ibrahim, and M. Adda, On predictive maintenance in Industry 4.0: Overview, models, and challenges, Applied Sciences, vol. 12, no. 16, p. 8081, 2022, doi: 10.3390/app12168081.

   2. R. van Dinter, B. Tekinerdogan, and C. Catal, Predictive maintenance using digital twins: A systematic literature review, Information and Software Technology, vol. 151, p. 107008, 2022, doi: 10.1016/j.infsof.2022.107008.

   3. R. Stark, C. Fresemann, and K. Lindow, Development and operation of Digital Twins for technical systems and services, CIRP Annals, vol. 68, no. 1, pp. 129132, 2019, doi: 10.1016/j.cirp.2019.04.024.

   4. B. Kunzer, M. Berges, and A. Dubrawski, The Digital Twin Landscape at the Crossroads of Predictive Maintenance, Machine Learning and Physics Based Modeling, arXiv:2206.10462, 2022.

   5. M. Hassan, M. Svadling, and N. Björsell, Experience from implementing digital twins for maintenance in industrial processes, Journal of Intelligent Manufacturing, vol. 35, pp. 875884, 2024, doi: 10.1007/s10845-023-02078-4.

   6. T. Kreuzer, P. Papapetrou, and J. Zdravkovic, Articial intelligence in digital twinsA systematic literature review, Data & Knowledge Engineering, vol. 151, p. 102304, 2024, doi: 10.1016/j.datak.2024.102304.

   7. V. Pulcini and G. Modoni, Machine learning-based digital twin of a conveyor belt for predictive maintenance, International Journal of Advanced Manufacturing Technology, vol. 133, pp. 60956110, 2024, doi: 10.1007/s00170- 024-14097-3.

   8. S. Chen, E. Turanoglu Bekar, J. Bokrantz, and A. Skoogh, AI-enhanced digital twins in maintenance: Systematic review, industrial challenges, and bridging researchpractice gaps, Journal of Manufacturing Systems, vol. 82, pp. 678699, 2025, doi: 10.1016/j.jmsy.2025.07.006.

   9. A. Benhania, Z. Ben Cheikh, P. Moura Oliveira, A. Valente, and J. Lima, Systematic review of predictive maintenance practices in the manufacturing sector, Intelligent Systems with Applications, vol. 26, p. 200501, 2025, doi: 10.1016/j.iswa.2025.200501.

   10. S. Singh, S. Sethi, R. Sharma, D. Vaibhavi, and A. Tiwari, AI-powered CNC digital twin for predictive maintenance, in Proc. 4th Int. Conf. Power, Control and Computing Technologies (ICPC2T), pp. 892897, 2025, doi: 10.1109/ICPC2T63847.2025.10958573.

   11. T. Khan, U. Khan, A. Khan, C. Mollan, I. Morkvenaite-Vilkonciene, and V. Pandey, Data-driven digital twin framework for predictive maintenance of smart manufacturing systems, Machines, vol. 13, no. 6, p. 481, 2025, doi: 10.3390/machines13060481.

   12. R. Kerkeni, A. Mhalla, and K. Bouzrara, Unsupervised learning and digital twin applied to predictive main- tenance for Industry 4.0, Journal of Electrical and Computer Engineering, vol. 2025, p. 3295799, 2025, doi: 10.1155/jece/3295799.

   13. Z. Huang, Y. Shen, J. Li, M. Fey, and C. Brecher, A survey on AI-driven digital twins in Industry 4.0: Smart manufacturing and advanced robotics, Sensors, vol. 21, no. 19, p. 6340, 2021, doi: 10.3390/s21196340.

   14. Y. Hu, O. W. Taha, and K. Yang, Fault detection in active magnetic bearings using digital twin technology,

       Applied Sciences, vol. 14, no. 4, p. 1384, 2024, doi: 10.3390/app14041384.

   15. A. M. Barbosa, T. V. N. Ngo, E. Jafarigol, T. B. Trafalis, and E. P. Ojoboh, Using federated machine learning in predictive maintenance of jet engines, Journal of Machine Learning and Applications, vol. 1, no. 1, pp. 1724, 2025, doi: 10.61577/jmla.2025.100007.

   16. A. L. Hananto, A. Tirta, S. G. Herawan, M. Idris, M. E. M. Soudagar, D. W. Djamari, and I. Veza, Digital Twin and 3D Digital Twin: Concepts, Applications, and Challenges in Industry 4.0 for Digital Twin, Computers, vol. 13, no. 4, p. 100, 2024, doi: 10.3390/computers13040100.

   17. T. Mukherjee and T. DebRoy, A digital twin for rapid qualication of 3D printed metallic components, Applied Materials Today, vol. 14, pp. 5965, 2019, doi: 10.1016/j.apmt.2018.11.003.

   18. C. Zhang, C. Gupta, A. K. Farahat, and K. Ristovski, Equipment health indicator learning using deep reinforcement learning, in Machine Learning and Knowledge Discovery in Databases, pp. 488504, 2019, doi: 10.1007/978-3-030- 10997-4_30.

   19. M. El-Hajj, T. Itäpelto, and T. Gebremariam, Systematic literature review: Digital twins role in enhancing security for Industry 4.0 applications, Security and Privacy, vol. 7, no. 5, p. e396, 2024, doi: 10.1002/spy2.396.

   20. J. Michael, J. Pfeier, B. Rumpe, and A. Wortmann, Integration challenges for digital twin systems-of-systems, in

       Proc. 10th IEEE/ACM Int. Workshop on Software Engineering for Systems-of-Systems and Software Ecosystems,

       pp. 912, 2022, doi: 10.1145/3528229.3529384.

   21. F. Ramzan and D. Reforgiato Recupero, A literature review on enhancing predictive maintenance in smart manufacturing industries: Fostering human-technology collaboration and overcoming data scarcity limitations with advanced AI models, Operations Research Forum, vol. 6, p. 181, 2025, doi: 10.1007/s43069-025-00584-0.

   22. S. M. M. Sajadieh and S. D. Noh, From simulation to autonomy: Reviews of the integration of articial intelligence and digital twins, International Journal of Precision Engineering and Manufacturing-Green Technology, vol. 12,

       pp. 15971628, 2025, doi: 10.1007/s40684-025-00750-z.

   23. L. Nele, G. Mattera, E. W. Yap, M. Vozza, and S. Vespoli, Towards the application of machine learning in digital twin technology: A multi-scale review, Discover Applied Sciences, vol. 6, p. 502, 2024, doi: 10.1007/s42452-024- 06206-4.

   24. M. R. Kabir and S. Ray, Digital Twin Tools for Smart Manufacturing: A Paradigm Shift for Industry 4.0, IEEE Open Journal of the Industrial Electronics Society, vol. 6, pp. 17561770, 2025, doi: 10.1109/OJIES.2025.3628531.

   25. Y. Shu, A. Hellicar, W. Wang, and A. Rahman, Semantic modelling of AI-integrated digital twins for enhanced transparency, Journal of Intelligent Information Systems, 2026, doi: 10.1007/s10844-026-01065-4.

   26. E. Mikoajewska, D. Mikoajewski, T. Mikoajczyk, and T. Paczkowski, Generative AI in AI-based digital twins for fault diagnosis for predictive maintenance in Industry 4.0/5.0, Applied Sciences, vol. 15, no. 6, p. 3166, 2025, doi: 10.3390/app15063166.

   27. D. Dai, B. Zhao, Z. Yu, P. Franciosa, and D. Ceglarek, Generative and Predictive AI for digital twin systems in manufacturing, Frontiers in Articial Intelligence, vol. 8, p. 1655470, 2025, doi: 10.3389/frai.2025.1655470.

   28. F. Naeem, M. Ali, G. Kaddoum, Y. Faheem, Y. Zhang, M. Debbah, and C. Yuen, A survey on GenAI-driven digital twins: Toward intelligent 6G networks and metaverse systems, IEEE Open Journal of the Communications Society, vol. 6, pp. 1036510402, 2025, doi: 10.1109/OJCOMS.2025.3641307.

   29. D. Chiaro, P. Qi, A. Pescapé, and F. Piccialli, Generative AI-empowered digital twin: A comprehensive survey with taxonomy, IEEE Transactions on Industrial Informatics, vol. 21, no. 6, pp. 42874295, 2025, doi: 10.1109/TII.2025.3540473.

   30. L. Ismail, A. Abdelmoti, A. Basu, A. D. E. Berini, and M. Naouss, A systematic review of digital twin-driven predictive maintenance in industrial engineering: Taxonomy, architectural elements, and future research directions, arXiv:2509.24443, 2025, doi: 10.48550/arXiv.2509.24443.

   31. P. Nguyen, M. Kim, E. Nichols, and H.-S. Yoon, AI-driven digital twins for manufacturing: A review across hierarchical manufacturing system levels, Sensors, vol. 26, no. 1, p. 124, 2026, doi: 10.3390/s26010124.

   32. M. Nasirinejad, H. Afshari, and S. Sampalli, Implementing Digital Twin for Maintenance 4.0 in SMEs: A Framework for Aordable and Secure Solutions, IFAC-PapersOnLine, vol. 59, no. 10, pp. 130135, 2025, doi: 10.1016/j.ifacol.2025.09.024.

   33. S. de Larmelina, A. L. da Silva, and L. A. Risso, A technology readiness assessment approach for Digital Twin implementation in SMEs, International Journal of Advanced Manufacturing Technology, vol. 140, nos. 56, pp. 27772796, 2025, doi: 10.1007/s00170-025-16370-5.

   34. J. Molter, A. Ibrahim, M. Eichenwald, and R. Müller, Implementing Digital Twins in SMEs: A step-by-step method- ology for digital production twins, Procedia CIRP, vol. 136, pp. 7984, 2025, doi: 10.1016/j.procir.2025.08.016.