Generative AI in AI-Based Digital Twins for Fault Diagnosis for Predictive Maintenance in Industry 4.0/5.0

Featured Application

Potential applications of the work include more reliable dedicated AI-based predictive maintenance systems based on digital twins and generative AI.

Abstract

Generative AI (GenAI) is revolutionizing digital twins (DTs) for fault diagnosis and predictive maintenance in Industry 4.0 and 5.0 by enabling real-time simulation, data augmentation, and improved anomaly detection. DTs, virtual replicas of physical systems, already use generative models to simulate various failure scenarios and rare events, improving system resilience and failure prediction accuracy. They create synthetic datasets that improve training quality while addressing data scarcity and data imbalance. The aim of this paper was to present the current state of the art and perspectives for using AI-based generative DTs for fault diagnosis for predictive maintenance in Industry 4.0/5.0. With GenAI, DTs enable proactive maintenance and minimize downtime, and their latest implementations combine multimodal sensor data to generate more realistic and actionable insights into system performance. This provides realistic operational profiles, identifying potential failure scenarios that traditional methods may miss. New perspectives in this area include the incorporation of Explainable AI (XAI) to increase transparency in decision-making and improve reliability in key industries such as manufacturing, energy, and healthcare. As Industry 5.0 emphasizes a human-centric approach, AI-based generative DT can seamlessly integrate with human operators to support collaboration and decision-making. The implementation of edge computing increases the scalability and real-time capabilities of DTs in smart factories and industrial Internet of Things (IoT) systems. Future advances may include federated learning to ensure data privacy while enabling data exchange between enterprises for fault diagnostics, and the evolution of GenAI alongside industrial systems, ensuring their long-term validity. However, challenges remain in managing computational complexity, ensuring data security, and addressing ethical issues during implementation.

1. Introduction

The Industry 3.0 paradigm introduced automation and early diagnostics. Predictive maintenance is a method of forecasting and managing the condition of machines, devices, production lines based on historical data, mechanism models and domain knowledge that predicts equipment trends, and behavioral patterns and correlations using statistics or artificial intelligence (AI) models. This allows us to predict remaining useful life, upcoming failures, and other key indicators in advance, improving the decision-making process for maintenance activities, reducing the risks associated with failures and avoiding unnecessary equipment downtime [1,2]. Predictive maintenance began with the advent of automation in Industry 3.0, using sensors and condition monitoring systems to track equipment performance. In this phase, techniques such as vibration analysis and thermal imaging were used along with manual data logging for trend analysis, enabling basic failure prediction. Scientists and engineers developed statistical models to predict failures based on historical data, introducing an early form of predictive maintenance. The Industry 4.0 paradigm saw the transition to continuous, real-time data collection from interconnected devices and systems, powered by the Internet of Things (IoT) and Big Data. Machine learning (ML) algorithms began to analyze massive data sets collected from IoT devices to detect patterns, predict failures, and recommend maintenance actions. Digital twins (DTs) were introduced as virtual replicas of physical assets, integrating sensor data and simulations for real-time monitoring and fault diagnosis. Predictive maintenance is related to various areas of research and economic practice [3,4].

ML plays a crucial role in enhancing DT technology by enabling real-time data analysis, predictive modeling, and automation. DTs, which are virtual replicas of physical assets, leverage ML to process vast amounts of sensor data and detect anomalies [5]. By using predictive analytics, ML helps forecast potential failures and optimize maintenance schedules, reducing downtime and costs [6]. In manufacturing, DTs powered by ML enhance production efficiency by identifying inefficiencies and improving quality control [7]. Healthcare applications benefit from AI-driven DTs that simulate patient conditions, allowing personalized treatment planning [8]. In smart cities, ML-based DTs optimize traffic flow, energy usage, and infrastructure management [9]. The aerospace industry uses DTs to simulate aircraft performance under various conditions, ensuring better safety and efficiency [10]. ML also improves DTs in supply chain management by predicting demand fluctuations and optimizing logistics [11]. These AI-enhanced virtual models support sustainability by analyzing environmental impacts and optimizing resource usage. ML significantly enhances the capabilities of DTs, making them more accurate, intelligent, and beneficial across industries.

Industry 4.0 also saw the rise in edge computing, which enabled predictive models to process data locally, reducing latency and increasing responsiveness. In the Industry 5.0 paradigm, predictive maintenance has moved towards a collaborative human-AI approach, emphasizing user-friendly tools and sustainable maintenance practices [12]. GenAI has begun to enhance DTs by simulating complex scenarios, generating synthetic data, and improving fault detection and predictive models [13]. DTs based on GenAI in Industry 5.0 focus on optimizing asset performance while adapting to environmental goals, which is the latest advancement in predictive maintenance [14]. The future implications of GenAI, large language models (LLM) and search-augmented generation (RAG) will influence practices in such distant industrial sectors as construction, but also specialist education [15].

GenAI plays a significant transformational role in AI-based DTs for fault diagnostics, driving the advancement of predictive maintenance for Industry 5.0. It enables realistic, yet human-operator-level perceptual simulations of complex industrial systems by generating synthetic data and scenarios, helping to predict and diagnose faults with greater accuracy. This increases the ability of the operator—cyber-physical system to detect trends, upcoming anomalies, identify their root causes and predict failures before they occur, compared to solutions without GenAI [16]. GenAI also helps improve the quality and quantity of training data by addressing limitations in sensor data availability and improving the robustness of the ML model [17]. Additionally, GenAI supports real-time updates to DTs by synthesizing data from different sources, ensuring that the twin reflects the current state of the asset. It facilitates the analysis of alternative “what if” scenarios by generating multiple potential failure scenarios (instead of the most likely ones, as before), enabling proactive maintenance strategies [18]. The technology improves decision-making by providing actionable insights derived from patterns and correlations identified in the generated data. GenAI also helps optimize system performance by simulating the impact of different maintenance actions, allowing operators to select the most effective intervention [19]. As Industry 5.0 emphasizes human-centric and sustainable solutions, GenAI in DTs aligns with these goals, enabling smarter, more efficient, and less resource-intensive maintenance practices [20].

GenAI is a rapidly growing area of research in deep learning. GenAI algorithms generate new realistic data in various modalities (text, images, music, three-dimensional (3D) models) and multimodalities [21]. Key features of GenAI include the ability to synthesize data and learning distributions. GenAI systems can synthesize new data that has features from other elements of the data set [22]. Thus, a GenAI model trained on images of dogs can create new, realistic images of dogs that do not exist in the training set [23]. GenAI models probability distributions of data, from which it can sample new instances that follow learned patterns [24]. GenAI uses state-of-the-art deep learning architectures that are rapidly evolving. The most important of them are the following:

• Generative Adversarial Networks (GANs)—involve the interaction between two neural networks: a generator and a discriminator. The generator creates synthetic data, while the discriminator distinguishes real from fake data. Through iterative competition, GANs generate highly realistic results and are used for image synthesis (DeepArt, DeepFake), video content generation, and style transfer (painting, photography);
• Variational Autoencoders (VAE)—encode input data into a latent representation and decode it back, following a predefined distribution. This enables smooth interpolation between data points for image reconstruction, anomaly detection, or drug discovery;
• Transformer models—use self-attention mechanisms to generate coherent and contextually relevant sequences within language models (ChatGPT, BERT) or description-based image synthesis (DALL-E);
• Diffusion models—an alternative to GANs, learn to reverse the noise process applied to data, gradually producing realistic results in high-quality image synthesis or molecular structure generation [25,26,27].

GenAI allows us to create content that integrates multiple modalities, e.g., generating an image from a text description. It offers computationally efficient models that scale to large data sets. For the above reasons, GenAI in a sense expands creativity, providing a selection from a larger set of quickly generated data with similar characteristics [28]. This allows us to generate synthetic data (industrial, medical, etc.) for training AI models or DTs, including without concerns about privacy or security, creating unique narratives and visualizations, accelerating discoveries by generating molecular structures, optimizing engineering designs and simulating physical phenomena [29]. The biggest challenge is still to provide users with greater control over the generated content and make the data generation process more transparent (

Table 1. Research gaps observed in the state of the art in GenAI-based DTs for fault diagnosis for predictive maintenance in Industry 4.0/5.0 (own version).

Area	Identified
	Gap(s)	Possibilities of Closing Gap(s)
Real-time data fusion and processing	Current methods struggle to effectively integrate multimodal data (e.g., IoT sensor readings, historical records, and environmental factors) in real time to achieve predictive accuracy.	Explore scalable architectures for real-time data fusion using GenAI capabilities.
Explainability and interpretability	The “black box” nature of many GenAI models limits trust and adoption in industrial environments.	Develop interpretable GenAI models that can explain failure prediction decisions in a way that is understandable to human operators.
Generating synthetic data for rare faults	Industrial systems often lack sufficient labeled data for rare but critical fault types.	Explore GenAI techniques to generate high-quality synthetic datasets that mimic rare fault conditions for robust training.
Adaptive learning in changing environments	Current models are not suited for dynamic industrial environments where machine configurations and operating conditions change frequently.	Develop an adaptive GenAI framework that can learn continuously and update DTs without complete retraining.
Integration of domain knowledge	Many GenAI approaches neglect the integration of domain expert knowledge, leading to less reliable fault diagnosis.	Combine domain knowledge with generative models to improve fault diagnosis reliability and contextual validity.
Generalization across device(s) types	Existing GenAI models are often tailored to specific machines and lack cross-device generalization.	Design generalized GenAI-based DTs that can transfer knowledge across different device types and configurations.
Cybersecurity in GenAI-based DTs	Increased connectivity and dependency on GenAI increase cybersecurity vulnerability in DTs.	Develop secure GenAI frameworks that protect sensitive industrial data while maintaining predictive accuracy.
Low-power, edge-compatible solutions	Many GenAI models are computationally intensive, making them unsuitable for edge deployments in smart factories.	Optimize GenAI algorithms for resource-constrained environments, enabling deployment on edge devices.
Multi-twin collaboration	Collaboration between multiple DTs for complex systems or connected machines is underexplored.	Explore a framework for GenAI-enabled multi-twin ecosystems to improve fault diagnostics in connected environments.
DTs lifecycle management	Limited research addresses long-term DT lifecycle management, such as model updates or retirement of outdated twins.	Develop methods for continuous evolution and maintenance of GenAI-based DTs to ensure continued accuracy and relevance.

). GenAI expands the boundaries of what machines can create, including in cooperation with humans, both as a tool for analysis and as a partner in creation [30].

To fill these gaps, interdisciplinary approaches combining advances in AI, edge computing, industrial engineering, and cybersecurity are needed to fully leverage the potential of GenAI-based DT for predictive maintenance in Industry 4.0/5.0 [31].

The motivation behind integrating GenAI with AI-based DTs in fault diagnosis for predictive maintenance within Industry 4.0/5.0 stems from the need to enhance real-time decision-making and predictive capabilities. With industrial systems becoming increasingly complex and interconnected, traditional diagnostic methods struggle to cope with the sheer volume and variety of data generated. The primary challenges include accurately modeling dynamic systems, dealing with incomplete or noisy data, and providing timely predictions to avoid costly downtime. GenAI offers a way to simulate various fault scenarios and predict potential failures by creating realistic synthetic data that complement real-world data. This approach can improve the robustness of DTs by enabling them to adapt and learn from evolving system behaviors. A novel contribution of this work is leveraging GenAI to continuously update and refine digital twin models, ensuring they remain accurate and relevant. Additionally, integrating AI-driven anomaly detection with generative models helps in identifying previously unseen faults. This synergy fosters a proactive maintenance strategy, minimizing unexpected failures and optimizing maintenance schedules. By embedding generative capabilities, the DTs evolve beyond static representations, becoming adaptive tools capable of scenario analysis and prescriptive insights. Consequently, this advancement pushes the boundaries of what’s possible in predictive maintenance, aligning closely with the smart, automated vision of Industry 5.0 [32].

GenAI is being integrated with AI-based DTs by enhancing their ability to simulate, predict, and adapt to real-world conditions in industrial environments. This integration involves using generative models such as VAEs and GANs to create synthetic datasets that complement real-world sensor data, improving fault detection even when historical failure data are sparse. Device states are described by feature vectors (at points in time) or feature matrices (time-varying feature vectors) in edge computing and include all relevant features (states, parameters) of both normal operation and impending wear and tear, failure, and attack, for example. GenAI also enables anomaly detection by learning the normal behavior of the system and generating deviations that signal potential failures before they become critical failures. GenAI-powered DTs can run multiple failure simulations to predict the impact of different faults, allowing industries to proactively optimize maintenance strategies. One of the major issues they address is the challenge of incomplete or noisy data, as GenAI can fill in missing information and remove signal noise to increase diagnostic accuracy. The integration also mitigates the high costs and risks associated with physical fault testing by creating realistic virtual failure scenarios for analysis. Another problem it addresses is the inability of traditional AI models to generalize across machines and environments, as generative models can adapt to changes in operating conditions. By continuously updating DTs with new synthetic and live data, the system remains dynamic and responsive to evolving industrial processes. This approach enhances predictive maintenance by reducing false positives and false negatives, ensuring that maintenance actions are taken only when necessary. In this way, generative AI empowers DTs, transforming them from static models into self-learning, adaptive systems that provide more accurate and timely fault diagnosis in Industry 4.0/5.0 [33].

This paper presents the current state of the art and prospects for using generative AI-based DTs to diagnose faults for predictive maintenance in Industry 4.0/5.0.

2. Materials and Methods

2.1. Data Set

Our bibliometric analysis aimed to investigate the research landscape and current knowledge and practices related to planning and implementing GenAI-based DTs for fault diagnosis in predictive maintenance within the framework of Industry 4.0/5.0 paradigms. To achieve this, we used bibliometric methods to examine scientific publications by defining research questions to identify key aspects, including the current state of the field, the origin and evolution of research topics, sources of publications (institutions, countries, and funding mechanisms), and the most influential authors and research teams. This methodology provides a comprehensive overview of current research and industry trends in GenAI-based DTs for fault diagnosis in predictive maintenance. By analyzing bibliometric data, this study contributes to the ongoing discussions and helps to establish a solid foundation for future research, identifying priority directions and research teams to follow in the coming years.

2.2. Methods

The study used the bibliographic databases Web of Science (WoS), Scopus, and dblp, selected for their extensive research collection and rich citation data, which facilitated a comprehensive bibliometric analysis of GenAI DT for failure diagnosis in predictive maintenance under Industry 4.0/5.0 (

Table 2. Bibliometric analysis procedure (own approach).

Stage	Name	Tasks
1	Defining research objectives	Defining goals of the bibliometric analysis
2	Selecting databases and data collections	Choosing appropriate data set(s) and developing research queries according to the study goals
3	Data preprocessing	Cleaning the collected date to remove duplicates and irrelevant records
4	Bibliometric software selection	Choosing suitable bibliometric software tools for analysis
5	Data analysis	Description, author, journal, area/topics, institution/country, etc.
6	Visualization (where possible)	Visualizing the analysis results to present insights
7	Interpretation and discussion	Interpreting findings in the context of the research goals

). To ensure greater relevance of the results, filters were applied to focus on original articles in English. Each of the selected articles was then manually reviewed to confirm its compliance with the inclusion criteria, refining the final set of analyzed articles. Key features of the data set were then examined, including prominent authors, research groups, institutions, countries, topic areas and emerging trends. This analysis helped to trace the evolution of key terminology and major research advances in the field. Furthermore, where possible, temporal trends were analyzed to track changes in research coverage over time, and publications were categorized into topic areas to discover relationships among different clusters of them. This process ultimately highlighted significant themes and subfields within the overall domain of study.

This study followed specific elements of the PRISMA 2020 guidelines [34] for bibliographic reviews (Supplementary Materials), focusing on key aspects such as rationale (item 3), objectives (item 4), eligibility criteria (item 5), information sources (item 6), search strategy (item 7), selection process (item 8), data collection (item 9), synthesis methods (item 13a), synthesis results (item 20b), and discussion (item 23a). Bibliometric analysis was performed using tools available in the Web of Science (WoS), Scopus, and dblp databases, as well as the Biblioshiny tool from the Bibliometrix v.4.1.3 package. The results are presented in a table, which allows for flexible analysis and visualization. Considering the interdisciplinary nature and complexity of the topic, the most important results of the review are summarized in a concise table for clarity.

3. Results

3.1. Data Sources

To refine the search, advanced filtering techniques were used, limiting the results to articles in English. In WoS, searches were performed using the “Subject” field, which includes title, abstract, keyword plus, and additional keywords. In Scopus, searches were performed using article title, abstract, and keywords, while in dblp, manual keyword selection was used. Databases were searched using keywords such as “generative artificial intelligence”, “digital twin”, and “Industry 4.0” or “Industry 5.0” (

Table 3. Detail search query over databases.

Parameter	Description
Inclusion criteria	Articles (original, reviews, communication, editorials) and chapters, including conference proceedings, in English
Exclusion criteria	Books older than 10 years, letters, conference abstracts without full text, other languages than English
Keywords used	Artificial intelligence, generative AI, digital twin, predictive maintenance, Industry 4.0, Industry 5.0
Used field codes (WoS)	“Subject” field (consisting of title, abstract, keyword plus, and other keywords)
Used field codes (Sopus)	Article title, abstract, and keywords
Used field codes (dblp)	Manually
Boolean operators used	Yes, e.g., “digital twin” AND (“Industry 4.0” OR “Industry 5.0”) AND rehabilitation
Applied filters	Results refined by publication year, document type (e.g., articles, reviews), and subject area (e.g., industry, engineering)
Iteration and validation options	Query run iteratively, refinement based on the results, and validation by ensuring relevant articles appear among the top hits
Leverage truncation and wildcards used	Used symbols like * for word variations (e.g., “digital twin *”) and ? for alternative spellings (e.g., “Industry ?.0”)

).

The selected set of publications was then further refined (see Figure 1) by manually re-reviewing the article and removing irrelevant items and duplicates, which allowed us to determine the final sample size.

The summary of the bibliographic analysis results is presented in

Table 4. Summary of results of bibliographic analysis (WoS, Scopus, dblp).

Parameter/Feature	Value
Leading types of publication	Conference review (50.0%), article (16.7%), conference paper (33.3%)
Leading areas of science	Computer science (50.0%), Engineering (20.0%), Mathematics (20.0%), Materials Science (10.0%)
Leading topics	Industrial: Design and Manufacturing
Leading countries	Bulgaria, Germany
Leading scientists	Mateev, M., Jazdi, N., Weyrich, M., Xia, Y., Xiao, Z.
Leading affiliations	University of Architecture, Civil Engineering and Geodesy, Sofia, Bulgaria, Universitat Stuttgart, Germany
Leading funders (where information available)	None
Sustainable development goals	Industry Innovation and Infrastructure, Responsible Consumption and Production

. The review included 21 articles (2023–2024) published in the last two years (no older ones were included).

Successful establishment of DT requires high fidelity virtual modeling and strong information interactions. GenAI can use advanced AI algorithms to automatically create, manipulate, and modify the desired sparse, correct, and diverse data. However, the implementation of this technology faces numerous challenges and perhaps should be implemented more quickly in specific areas of Industry 4.0/5.0 [35].

3.2. IIoT Background and the Potential of GenAI-Driven DT

IIoT is revolutionizing industrial sectors by connecting machines, sensors, and devices through networks, enabling real-time data acquisition and intelligent decision-making. Industrial IoT can be used for process monitoring [36,37]. It emphasizes predictive maintenance, process optimization, and asset management, leveraging advanced analytics on massive data sets collected from industrial operations. DTs, acting as virtual replicas of physical systems, leverage IIoT data to model, simulate, and monitor performance [38,39,40,41]. By integrating IIoT and GenAI, enterprises can unlock new, higher levels of operational intelligence and decision-making agility. GenAI introduces a new dimension to DTs by enabling them to learn, predict, and generate new data scenarios, significantly increasing their accuracy and usability. Unlike traditional DTs that rely on fixed data patterns, generative DTs powered by AI can model complex, nonlinear systems and propose innovative solutions. This capability is especially transformative in industries such as manufacturing, energy, and transportation, where high variability and system complexity require dynamic adaptation. GenAI can simulate how machines will operate under unprecedented conditions, reducing downtime and mitigating risk. It also enables real-time scenario analysis, helping industries to adapt to disruptions very quickly or optimize resource allocation. Additionally, these intelligent DTs facilitate sustainable practices by modeling energy efficiency and predicting the carbon footprint of industrial operations. This convergence may redefine industry standards, supporting innovation and resilience in an era of rapid technological advancement, not only within the Industry 4.0 and Industry 5.0 paradigms but also their next generations [42].

Recent publications on GenAI in AI-based DTs for fault diagnosis and predictive maintenance in Industry 4.0/5.0 show the growing focus on increasing predictive accuracy and real-time decision-making. Researchers are increasingly exploring the integration of GenAI with IoT, edge computing, and cloud-based architectures to improve system responsiveness. Many studies emphasize the use of synthetic data generation to address the problem of limited failure data in industrial environments. There is a noticeable shift towards explainable AI (XAI) to improve transparency and trust in AI-based maintenance decisions. The most progressive topics include deep learning-based anomaly detection, reinforcement learning for adaptive maintenance strategies, and hybrid AI models combining physics-based and data-based approaches. Publications also emphasize the role of GenAI in enabling self-learning DTs that evolve with operational data. This trend indicates a shift towards more autonomous, scalable, and human-centric AI-based maintenance solutions in smart industries [43].

3.3. Basic Methods of Generative AI-Driven DTs

GenAI-driven DTs leverage advanced ML techniques (such as GANs and VAEs) to simulate and generate realistic data scenarios. These models learn patterns from IIoT data, allowing them to predict, optimize, and replicate complex behaviors of real-world systems. Reinforcement learning (RL) further enables DTs to adapt to dynamic environments by testing and refining decision-making strategies. Natural language processing (NLP) techniques allow these DTs to interpret and integrate textual data such as maintenance logs, operator reports, and technical documentation into their analyses. Additionally, transfer learning helps leverage pre-trained models to accelerate the development and deployment of AI-driven generative DTs in various industrial contexts [44].

GANs enhance AI-based DTs by generating synthetic sensor data that replicates real-world failure conditions, helping predictive maintenance models detect rare and complex failure modes in Industry 4.0/5.0 environments. GANs create different failure scenarios by learning statistical patterns of normal and faulty operating states, enabling digital twins to improve fault diagnostic accuracy even when historical failure data are limited. The generator synthesizes realistic fault signals, while the discriminator improves its ability to distinguish between real and artificial faults, strengthening the fault detection and anomaly recognition capabilities of the digital twin. By continuously learning from live sensor data, GAN-based digital twins detect deviations from normal operation in real time, enabling early failure prediction and reducing unplanned downtime. When combined with reinforcement learning, GANs optimize maintenance decisions based on synthetic fault simulations, while federated learning ensures secure model training across multiple industrial sites without sharing sensitive operational data [45].

VAEs enhance AI-based DTs by learning the underlying distribution of normal operating data, enabling the detection of deviations that signal potential failures in predictive maintenance for Industry 4.0/5.0. VAEs encode high-dimensional sensor data into a lower-dimensional latent space, capturing salient features of machine behavior and facilitating the identification of anomalies that indicate emerging failures. VAEs reconstruct sensor signals from compressed latent representations, and when the reconstruction error exceeds a threshold, it suggests the presence of an unknown or abnormal fault condition. Unlike traditional deterministic models, VAEs generate probabilistic results, enabling digital twins to assess the uncertainty of failure predictions and reduce false positives in predictive maintenance. VAE can be combined with RL to optimize maintenance decisions based on detected anomalies, while federated learning enables secure, decentralized model training across multiple industrial facilities without the need to share raw sensor data [46].

Transformers, such as Bidirectional Encoder Representations from Transformers (BERT), improve predictive maintenance by processing sensor data in large time series, capturing complex relationships in industrial systems, and improving fault diagnostics in Industry 4.0/5.0. Transformers’ self-driving mechanism enables AI-based digital twins to analyze long-term relationships in sensor data, identifying subtle fault patterns that traditional machine learning models may miss. Transformer models pre-trained on extensive industrial data sets can be tuned to specific device types, enabling more accurate fault detection and predictive maintenance tailored to unique operating conditions. Transformers analyze streaming data in real time, learning about normal machine behavior and flagging anomalies that indicate early signs of failure, reducing unplanned downtime in manufacturing and industrial processes. Combined with generative AI, transformers improve fault simulation and synthetic data generation, while RL optimizes maintenance strategies by continuously adapting fault diagnosis models based on real-time feedback from DTs [47].

RL enables AI-based DTs to optimize predictive maintenance by continuously learning from real-time equipment data and adapting maintenance strategies based on Industry 4.0/5.0 fault diagnostics results. RL agents interact with digital twins to simulate different maintenance actions, receiving rewards based on reduced downtime, increased fault detection accuracy, and minimized operating costs. Unlike traditional rule-based approaches, RL-based digital twins dynamically refine fault diagnostics and maintenance strategies, improving decision-making as more operational data are processed over time. By combining RL with generative AI, digital twins can simulate rare fault conditions, enabling the RL model to learn from different failure scenarios and make more accurate predictive maintenance recommendations. In complex manufacturing environments, multi-agent RL enables multiple AI-powered digital twins to collaborate, optimizing fault diagnosis and predictive maintenance strategies for interconnected industrial assets [48].

Federated learning enables multiple industrial plants to jointly train predictive maintenance models while maintaining data privacy, and generative AI enhances this process by generating synthetic fault data to increase model robustness in Industry 4.0/5.0. Instead of sharing raw sensor data, federated learning enables AI-based digital twins to exchange model updates, keeping sensitive operational data secure while leveraging collective intelligence across manufacturing plants. Generative models such as GANs and VAEs generate realistic fault scenarios to enrich federated learning models, compensating for imbalanced data sets where certain failure conditions may be underrepresented. AI-based digital twins within a federated learning network continuously update fault diagnostic models using locally generated synthetic data, enabling real-time adaptation to unique operating conditions in different industrial environments. By integrating federated learning with generative AI, industrial systems develop more accurate, context-aware predictive maintenance strategies that reduce downtime and improve fault detection without compromising data security or requiring centralized data storage [49]. The generation of extended data to improve error detection in images or time series signals has been proposed in [50,51].

Numerical comparisons of different generative AI methods in AI-based DTs for fault diagnosis and predictive maintenance reveal distinct advantages based on accuracy, efficiency, and computational cost. GANs typically improve fault classification accuracy by 5–15% by generating realistic failure data for training models. VAEs achieve anomaly detection precision of 85–95% on average, depending on system complexity and available data. Transformer-based models such as BERT time series variations outperform traditional recursive models by 10–20% in predictive accuracy due to their ability to capture long-range dependencies. Reinforcement learning (RL) approaches reduce maintenance costs by 20–40% by optimizing predictive scheduling strategies compared to rule-based systems. Physics-based generative models reduce false positive rates by 30–50% because they combine physical simulations with AI-based analyses to provide more reliable fault diagnosis. Federated learning with Generative AI improves model generalization across multiple industrial sites while maintaining accuracy at 95%+, although it may require 20–30% more computational resources due to decentralized processing. These comparisons highlight the trade-offs between accuracy, computational cost, and scalability, which is the basis for selecting the right Generative AI techniques for different industrial applications [52,53].

The increasing understanding of AI algorithms brings about a deeper understanding and classification of them by researchers and practitioners, who can apply the appropriate ones to obtain optimal results in the shortest possible time with less effort for their specific application area problems in a novel and significant way [54].

Here are some important frameworks that can help understand generative AI for fault diagnosis in AI-based digital twins for predictive maintenance in Industry 4.0/5.0:

• Conceptual diagram of GenAI in DTs:
  • Shows interaction between real-world industrial systems, IoT sensors, AI-based digital twins, and predictive maintenance systems;
  • Highlights how real-time data are collected, processed by GenAI models, and used to predict faults.
• Comparison chart of GenAI methods comparing GANs, VAEs, transformer-based models, reinforcement learning, and physics-informed models based on accuracy, computational cost, scalability, and fault detection capabilities.
• Workflow of GenAI-Based fault diagnosis, i.e., a step-by-step process illustrating the following:

  ▪

  Data collection from IoT sensors;

  ▪

  Preprocessing and feature extraction;

  ▪

  Model training using GenAI;

  ▪

  Fault detection and predictive maintenance actions.

• Illustration of GANs vs. VAEs for anomaly detection showing how GANs generate synthetic failure data while VAEs reconstruct normal operation data to detect anomalies.
• Transformer-based time-series prediction showing how transformer models analyze historical machine data and predict failures with multi-step forecasting.
• Impact of GenAI on maintenance costs and downtime comparing reactive, preventive, and predictive maintenance strategies, highlighting the efficiency gains achieved by Generative AI.
• Federated learning for cross-site fault diagnosis shows how decentralized AI models share insights across multiple industrial sites while preserving data privacy [55,56].

The digitization of data throughout the product life cycle provided by DTs in cyber physical systems enables a rapid transition from current industrial solutions to intelligent and adaptive solutions. GenAI promotes the construction, modernization, and updating of data in DTs to increase the predictive accuracy and ensure differentiated smart manufacturing by detecting IIoT devices and sharing data. The problem here is the adverse selection caused by information asymmetry. Here, a contract theory model based on a balanced soft actor-critic algorithm based on diffusion is proposed. It provides the identification of the optimal feasible contract and also reduces the number of actor network parameters through the dynamic structural pruning technique [57].

It does this by analyzing huge amounts of data from various sources, such as sensors, smart meters, and historical production and energy consumption patterns, and AI algorithms can identify patterns and anomalies that are difficult for humans to detect. This enables the development of predictive models that optimize production and consumption.

The reliability of data generated by Generative AI in AI-based DTs for fault diagnosis and predictive maintenance is critical to ensuring accurate decision-making in Industry 4.0/5.0. One way to establish reliability is through rigorous validation techniques, such as comparing generated data with real-world sensor readings to assess accuracy. Generative models can be trained using high-quality, diverse data sets to reduce bias and improve their ability to generalize across industrial environments. Another approach is uncertainty quantification, where confidence levels are assigned to generated results, helping maintenance teams assess the reliability of AI-generated predictions. Hybrid AI models that combine generative approaches with physics-based simulations increase data reliability by ensuring that the patterns generated are consistent with known system behaviors. Domain adaptation techniques can further refine generative models to match specific machine characteristics, increasing the relevance of synthetic data for fault diagnosis. Continuous learning mechanisms enable generative models to be updated based on real-time feedback, ensuring they evolve with changing system conditions and avoiding outdated or misleading predictions. Implementing explainability techniques such as attention mechanisms or feature attribution can increase transparency and help engineers understand how the model generates and uses synthetic data. Cross-validation with expert knowledge ensures that AI-generated insights align with human expertise, reducing the risk of incorrect error predictions. Finally, regulatory compliance and adherence to industry standards help build trust in generative AI-powered DTs, ensuring that the generated data meets the reliability and safety requirements for industrial applications.

A lightweight automatic data augmentation framework (ALADA) is proposed to optimize data augmentation rules and industrial defect detection solutions. It provides more efficient augmentation and generation of augmented images for joint optimization, hyperparameter tuning for retraining with searched rules, and also reduces the risk of defect failure in four situations: textured background, non-uniform brightness, low contrast, and intra-class difference, which is validated on three industrial defect detection datasets, namely Tianchi-TILE, GC10-DET, and NEU-DET [58].

Effective communication is the backbone of GenAI-driven DTs, enabling robust, adaptive, real-time industrial operations. Fast and error-free communication in AI-driven DTs is essential for seamless interaction between physical and virtual systems. Data integration is achieved through IIoT devices, where sensors and edge devices transmit real-time data to DTs for analysis and simulation. Advanced protocols such as MQTT, OPC-UA, and 5G networks facilitate low-latency and secure data transfer, ensuring synchronized operations between physical and digital twins. It is worth nothing that GenAI-driven DTs leverage cloud and edge computing for scalable and efficient communication, enabling rapid processing of complex data streams. Interoperability standards and application programming interfaces (APIs) are key, enabling DTs to interact with diverse systems, applications, and stakeholders in industrial ecosystems. Bidirectional communication ensures that insights or decisions generated by DT can be fed back to physical systems for execution or intervention. GenAI enhances communication by interpreting unstructured data sources (such as natural language maintenance reports, image data) and integrating them into the DT model. User interfaces, including dashboards and voice command systems, enable intuitive communication with human operators, making insights accessible and actionable. Collaboration capabilities allow multiple GenAI-based DTs or systems to share insights, enabling optimized performance at the network or organizational level.

Data management in GenAI-powered DTs is the foundation of their functionality and effectiveness. It starts with robust data acquisition from IIoT devices, collecting real-time information such as sensor readings, operational parameters, and environmental conditions. These data are then preprocessed using techniques such as normalization, filtering, and outlier detection to ensure quality and relevance. Centralized or distributed data architectures, often leveraging cloud and edge computing, are used to store and process massive amounts of structured and unstructured data. Advanced data integration frameworks enable the merging of heterogeneous data sources such as machine logs, video feeds, and maintenance records into a unified platform, and multi-modal data are increasingly being integrated. GenAI algorithms analyze and synthesize these data, generating predictive insights, simulations, or new system optimization scenarios. Metadata management (data about data) is also critical, ensuring that data provenance, context, and relationships are well documented to support the interpretability/explainability of GenAI. Security and compliance protocols that protect sensitive industrial data implement encryption, access controls, and compliance with regulations such as the General Data Protection Regulation (GDPR) or industry standards. Scalable storage solutions and intelligent indexing facilitate efficient retrieval and manipulation of data in real time. Periodic data collection and lifecycle management ensure that DT operates on accurate, timely, and meaningful data sets. By combining advanced analytics with disciplined data practices, AI-powered generative DT can drive transformational improvements in industrial processes.

A Weighted Extreme Learning Machine (WELM) is proposed to provide balanced class distribution and reduce data complexity by generating new samples and removing overlapping noisy samples at class boundaries. The effectiveness of the above solution is demonstrated by allocating the most efficient resources to the most urgent orders to avoid delays in the supply chain [59].

GenAI extends DTs beyond their current capabilities into more dynamic, predictive, and interactive tools that simulate complex scenarios and predict future conditions with remarkable accuracy. Depending on the level of GenAI integration into DTs, DTs can be extended to varying degrees to generate synthetic data sets, simulate events/scenarios that have no previous equivalents (e.g., isolated failures), and provide second opinions for decision-making based on LLM agent networks. This has varying implications for operational efficiency, innovation, and decision-making processes [60]. Different GenAI models allow for DT state emulation, function abstraction, and decision-making based on the interaction between GAI-based and model-driven data processing [61]. Three approaches have been proposed for network management:

• Light model weighting;
• Adaptive model selection;
• Data model-driven management [61].

Modeling in the absence of data, e.g., higher resolution photovoltaic (PV) systems (down to individual households or hourly), is a huge obstacle to making informed and accurate decisions. This requires new methods to generate detailed realistic data sets—such as integrated ML models identifying PV users—and methods to augment data using explainable AI techniques based on key features and their interactions and to generate hourly solar energy production at the household level using an analytical model. The synthetic data sets obtained by the above method are validated against real-world data for DTs for further modeling tasks [62]. Depending on the type and method of providing input data, predictive analysis within GenAI-based DT can be based on different LLM: GPT, DALL-E, DAVINCI, or WHISPER.

3.4. Typical Applications

Typical applications include adaptive monitoring and diagnostics, and adaptive response. Dynamic, evolving features of the physical world require a huge amount of data transmission/exchange to ensure synchronization between the physical world and its virtual image. Such a communication framework can be based on the “look only once” (YOLO) principle. The YOLOv7-X object detector in the case of an apple orchard was used to extract semantic information from captured images of edge devices, reducing the amount of data needed to be transmitted. The meaning of each piece of semantic information is determined based on the trust generated by the object detector. Two resource allocation schemes are proposed:

• Trust-based scheme;
• AI-generated scheme acrlong.

Diffusion models generate an optimal allocation scheme that outperforms the results obtained from the schemes used separately. An additional improvement is provided by the attention modules of the ELAN-H and SimAM layer aggregation network that reduce model parameters and computational complexity when using edge devices with limited performance [63]. The complexity of the supply chain poses a particular challenge due to its lack of transparency, generally accepted standards, and regulations. A blockchain-based process data management solution for recycling and reuse of used electronic devices is proposed, combining DT and GenAI to solve the blockchain performance bottleneck by predicting future data flows. This improves the adaptability and throughput of the system, as well as traceability, prediction accuracy, and efficiency throughout the process [64]. GenAI-supported cellular network DT is also proposed to learn complex network data distribution (environmental, user, and service) from samples from the distribution [65]. GenAI uses these data to generate different scenarios, improving flexibility and practically solving network optimization [66,67]. The integration of GenAI and urban DTs is used to address challenges in the planning and management of built environments, including various urban subsystems (transportation, energy, water, and construction and infrastructure) [68].

Adaptive Response

Enabling sustainable development and efficient use of resources, especially expensive energy, is becoming critical for today’s economy. This is due to a number of factors: climate change, resource depletion, and the need for decarbonization and increased innovation in solutions. GenAI, DTs, and big data can help the energy sector achieve greater efficiency, optimize operations, and facilitate decision-making to optimize energy use and reduce waste [69]. The confusion in construction stems from the need to differentiate between two technologies: building information modeling (BIM) and DT, which differ in terms of technologies, maturity levels, data layers, enablers, and functionalities. The research emphasis here is on the convergence of BIM and DT, data integrity, their integration and transmission, bidirectional interoperability, non-technical factors, and data security [12]. Asset Administration Shell (AAS) is a digital twin model in the context of Industry 4.0, assuming that semantic-based communication and meaningful textual data generation are directly related and that these processes are equivalent. An LLM-supported system is implemented to generate standard DT models as instances from raw textual data collected automatically from data sheets describing technical assets. The achievable effective generation rate was 62–79%. The resulting AAS model can be integrated with compliant DT software for data exchange and DT communication and interoperability in industrial applications [70]. OpenAI GPT-4 Turbo with Vision LLM can interpret images and provide textual answers to queries about those images by combining natural language processing and visual understanding. This allows for intelligent extraction of key metadata from images and videos to assess the state of real-world systems and propose sustainability measures. This helps to implement efficient image analysis and prediction models and optimize the cost of the solution using a hybrid approach. GenAI in data analysis increasingly offers efficient and cost-effective solutions for predictive analysis based on vector search and other data analysis methods, including image analysis, case decomposition, hybrid search, and generation of self-adaptive models to find trends and offer preventive actions, even for smaller companies [71].

4. Discussion

AI is based on replicating human intelligence in machine control systems, enabling them to perform tasks that require human cognitive abilities (perception, learning, reasoning, and problem-solving). AI encompasses various methodologies and technologies such as ML, natural language processing, computer vision, and robotics, and GenAI further extends these capabilities with rapid creativity previously unavailable to humans [72].

Generative AI development in AI-based DTs for predictive maintenance fault diagnosis benefits the community by increasing industrial efficiency, reducing operational costs and improving safety. By enabling early fault detection, it minimizes unexpected equipment failures, leading to fewer disruptions in key sectors such as manufacturing, energy, and transportation. The technology supports sustainability efforts by optimizing resource utilization, reducing waste, and extending machine life. Small- and medium-sized enterprises benefit from cost-effective predictive maintenance solutions that were previously available only to large corporations. The workforce also benefits from AI-assisted maintenance by reducing unsafe manual checks and freeing skilled professionals to focus on higher-value tasks. From a managerial perspective, integrating Generative AI with DTs requires a shift to data-driven decision-making, which requires investment in AI knowledge and infrastructure. Managers must ensure ethical AI implementation, balancing automation with human oversight to maintain accountability and transparency. Real-time insights from DTs enable proactive maintenance planning, improved asset utilization, and reduced downtime. Additionally, industries adopting this technology gain competitive advantage by improving service reliability and customer satisfaction. Overall, the advancement of Generative AI in DTs aligns with the Industry 5.0 vision of human-centric, sustainable, and resilient industrial ecosystems [73].

Introducing Generative AI into AI-based DTs for fault diagnosis and predictive maintenance in Industry 4.0/5.0 raises several difficulties and scientific questions beyond the usual technical AI challenges. One major issue is the interpretability of AI-generated insights—how can engineers and decision-makers trust and understand the synthetic fault scenarios generated by AI models? There is also a concern about data authenticity and bias, as synthetic data might reinforce existing biases in training datasets, leading to skewed predictions. A key scientific question is how to balance the trade-off between real-world physics-based models and AI-generated simulations to ensure reliability without excessive computational costs. The integration of Generative AI with human decision-making poses an epistemological challenge: to what extent should AI-driven insights override human expertise in maintenance decisions? Ethical considerations arise when AI is used to automate critical fault diagnosis, particularly regarding accountability in cases of incorrect predictions leading to safety risks or economic losses. Standardization remains an unresolved issue—how can industries establish universal guidelines for using Generative AI in DTs across different sectors? Organizational resistance is another difficulty, as industries must overcome skepticism from stakeholders who may not fully trust AI-generated diagnostics. The scalability of this approach in highly diverse industrial settings raises scientific questions about adaptability—how can generative models generalize across different machine types and operational environments? Additionally, regulatory and compliance concerns present challenges, as industries must ensure AI-driven fault diagnosis meets safety and legal requirements. The economic implications of adopting Generative AI for predictive maintenance need further exploration—how can businesses quantify the long-term cost savings and return on investment from implementing these advanced AI-driven systems [74]?

Improved fault diagnostics and data augmentation for robust models enables faster and more reliable detection of anomalies in real time based on live sensor data, identifying patterns that indicate, for example, early signs of equipment degradation. This allows maintenance schedules to be dynamically adjusted based on real-time information, on the one hand avoiding downtime and on the other reducing unnecessary maintenance. Clear, explainable GenAI models with easily interpretable results/hints are key to earning operator trust and ensuring that generative AI-based fault predictions are consistent with expert knowledge. Combining AI-generated insights with expert judgment improves decision-making, ensuring that maintenance personnel can verify and refine GenAI-based fault diagnosis recommendations. For these reasons, protecting sensor communications, databases, and GenAI algorithms from attacks and data manipulation is essential to maintaining reliable predictions [75].

4.1. Limitations of Current Solutions and Concepts

The modernity of technologies using GenAI in AI-based DTs for fault diagnostics brings with it a number of limitations that must be taken into account when planning, building, operating, modernizing, and decommissioning/replacing such systems. They are presented in

Table 5. Limitations of AI in AI-based DTs for fault diagnostics within Industry 4.0/5.0 paradigm (own version).

Limitation	Description
Dependence on data quality	GenAI models rely heavily on the quality and diversity of training data, so incomplete, uncertain, or biased data can lead to inaccurate simulations and fault diagnoses.
Computational complexity	The computational power required to train and deploy GenAI models can be significantly higher, making it challenging for real-time applications in resource-constrained environments and cost- and energy-intensive. Regular updates and retraining of GenAI models are necessary to keep them current, which increases operational costs.
Scenarios validity	Generated data or scenarios may not always reflect realistic or physically plausible conditions, which can lead to misleading conclusions. This often requires consulting experts.
Model interpretability/explainability	GenAI models, especially those using DL, are often black boxes, making it difficult to understand or undermining the trust in the decisions they generate.
Integration challenges	Integrating GenAI into existing DT frameworks can be complex and require significant AI and domain-specific expertise.
Risk of overfitting	Generative models can overfit to specific patterns in training data, reducing their ability to generalize to unseen error conditions.
Lack of domain-specific context	Without sufficient domain expertise incorporated into the AI model, generative AI may not account for the nuances of operational behavior of specific industrial systems.
Dependence on AI expertise	Successful implementation requires skilled AI practitioners who understand generative models and digital twin technologies, which can be a limiting factor in many industries.
Ethics and security concerns	Data generated by GenAI could potentially raise ethical issues or be used maliciously, such as by creating misleading error scenarios.

.

Overcoming the above limitations, even partially, will increase the effectiveness of the discussed group of systems and accelerate their full implementation [76,77].

4.2. Directions of Further Research

Complementing the overcoming of the fundamental limitations described above are the most promising directions for further research on GenAI in AI-based DTs for fault diagnosis for predictive maintenance in Industry 4.0/5.0. Advances in generating high-fidelity synthetic data that closely reflect real-world conditions can address data scarcity and improve model performance [78]. Research into GenAI techniques that enable real-time updates and learning can make DTs more dynamic and responsive to changing system conditions. Developing hybrid models that combine GenAI with physics/chemistry/mechanics-based simulations can increase the realism and reliability of DTs for fault diagnosis [79]. Research into interpretable/explainable GenAI (XAI) models can help build trust and make it easier to understand the insights provided by AI-based DTs. Adaptively tailoring GenAI architectures to specific industries or assets can improve their accuracy and relevance in predictive maintenance applications [80]. Furthermore, research into lightweight GenAI models can enable scalability and real-time processing in resource-constrained industrial environments. Exploring how GenAI can seamlessly integrate with IoT sensors and edge computing can enhance data collection and error detection capabilities [81]. Collaborative interdisciplinary research involving AI experts and industry practitioners can ensure that GenAI solutions are practical and tailored to real-world needs [82]. Focusing on the secure implementation of GenAI in DTs can prevent vulnerabilities related to data manipulation and model exploitation [83]. Exploring how GenAI can optimize resource utilization and minimize energy consumption aligns with Industry 5.0’s emphasis on sustainability and human-centric approaches [84,85].

Future work on generative AI in AI-based DTs for fault diagnosis and predictive maintenance in Industry 4.0/5.0 will focus on increasing model adaptability, scalability, and real-time decision-making capabilities. One key direction is to integrate multimodal data sources, such as sensor data, historical maintenance logs, and expert knowledge, to increase the accuracy and robustness of fault predictions. Advanced reinforcement learning techniques can be combined with generative AI to enable self-learning DTs that continuously evolve without human intervention. Future developments may also include federated learning to ensure data privacy and enable collaborative intelligence across multiple industrial sites. The use of quantum computing can further accelerate training and inference of generative models, enabling more complex simulations and faster fault diagnosis. Another promising extension is the implementation of AI-based edge computing, where generative models run on localized devices to provide immediate fault predictions without relying on cloud infrastructure. Generative AI can also enable synthetic data augmentation, improving model generalization for rare or invisible fault conditions. As AI-based DTs become more advanced, they can integrate with augmented reality and virtual reality systems to provide immersive diagnostics and training for maintenance personnel as part of Industry 5.0 [86,87]. Additionally, human-in-the-loop AI frameworks will be key to maintaining the interpretability and trust of automated fault diagnostic systems. Future research should also consider regulatory compliance, ethical issues, and standardization of AI-based predictive maintenance technologies to ensure safe and responsible implementation across industries [88,89].

5. Conclusions

DT technologies, including those based on GenAI, enable early detection and correct diagnosis of faults, which will facilitate corrective actions to replace predicted damaged components before failures occur. The number of publications on GenAI-based DTs is not large in relation to the needs, nor does it cover all the observed research gaps, which is why there should be more emphasis on interdisciplinary scientific and economic cooperation in this area. This applies to both collected and generated data, as well as entire environments for their processing. This will not only allow for maintaining control over the development of this group of solutions, but also for their standardization and synchronization of development with the consideration of Explainable AI (XAI).

With respect to the previously observed knowledge and experience gaps, it can be said that Generative AI plays a key role in AI-based DTs for fault diagnosis and predictive maintenance in Industry 4.0 and 5.0. By creating realistic simulations, it enables accurate modeling of machine behavior under different conditions. These AI-driven DTs continuously learn from real-time data, enhancing the ability to detect and predict faults. Generative AI helps in synthesizing missing or sparse failure data, improving diagnostic accuracy. It also enables adaptive maintenance strategies by predicting potential failures before they occur. This reduces downtime, minimizes maintenance costs, and optimizes resource utilization. Moreover, the integration of Generative AI with IoT and edge computing improves real-time monitoring and decision-making. The synergy between AI and digital twins facilitates a proactive, data-driven approach to maintenance in smart industries. As Industry 5.0 emphasizes human-AI collaboration, Generative AI increases interpretability and decision support for human operators. Using Generative AI in DTs transforms predictive maintenance, ensuring reliability, efficiency, and sustainability in industrial operations.

In predictive maintenance, GenAI DTs enable realistic operational profiles, identifying potential failure modes that traditional methods may miss. New opportunities in GenAI-based DTs include:

• Incorporating XAI to enhance decision-making clarity and improve reliability in key industries such as manufacturing, energy, and even employee healthcare as part of preventive medicine;
• Emphasizing a human-centric approach, so GenAI-based DTs can better integrate with human operators to support collaboration and decision-making;
• Implementation of edge AI and distributed computing further increases the scalability and real-time capabilities of DT, and federated learning ensures data privacy.

In this way, increasingly effective DT technologies will increasingly cooperate with operators within the Industry 5.0 paradigm. Challenges remain in managing computational complexity, ensuring data security, and addressing ethical issues during implementation.