Research Review on Workshop Scheduling for Intelligent Manufacturing: Digital Twin Modeling, Optimization Algorithm, and System Architecture

Abstract

Digital twin, as a new generation of industrial intelligent technology, has become a key technology for achieving virtual-physical interaction and real-time optimization in intelligent manufacturing systems due to its capability for high-fidelity virtual mapping of physical systems. Production scheduling, as the core link in the operation of intelligent workshops, faces challenges such as frequent dynamic disturbances, rendering traditional static scheduling approaches inadequate to meet the real-time and flexibility requirements of backdrop operations. In this context, the significant potential of the deep integration of digital twin technology and workshop scheduling in enhancing scheduling real-time performance, agility, and robustness has increasingly been highlighted. This paper reviews the research progress in workshop digital twin scheduling technology over the past five years, focusing on the development paths and technical characteristics of typical workshop digital twin modeling techniques, intelligent scheduling algorithms, and system frameworks. Based on this, the paper proposes a conceptual framework for digital twin scheduling in complex manufacturing scenarios, providing theoretical references for developing highly real-time and robust intelligent manufacturing scheduling systems, and highlights future research directions and developmental trends.

1. Introduction

As the core pillar of the national economy, the manufacturing industry drives technological progress and enhances national competitiveness, promoting the development of related industries. The scale of manufacturing is increasing, and the complexity is gradually rising. With the entry of “intelligence” into manufacturing systems and the development and popularization of the next generation of information technology, the manufacturing industry is gradually moving towards digitization and intelligent development [1]. In this context, Germany proposed the “Vision 2030”, pointing out that autonomous, interoperable, and sustainable intelligent production and smart factories are the characteristics of future manufacturing. Meanwhile, the European Union emphasized a human-centered, sustainable, and resilient supply chain in “Industry 5.0” to ensure its global leading position in the manufacturing field [2]. In this ongoing “industrial revolution”, digital twin technology, which can dynamically interact with objects in the physical world and self-optimize and adjust through real-time data acquisition, has become a cutting-edge digital technology in manufacturing [3]. The most accepted concept of digital twin is the term “Digital Twin” proposed by NASA in 2010, defined as an integrated simulation process of multiple physical quantities, multiple scales, and multiple probabilities, completing the mapping of physical entities in virtual space and reflecting their entire life cycle process [4]. This technology allows for seamless bidirectional data flow between the virtual and physical worlds, adjusting and adapting according to changes in physical objects, and is widely applied in product design and development, fault detection, performance prediction, production scheduling, and other manufacturing scenarios. Thus, traditional modeling and simulation are comparatively inferior to the digitally interactive digital twin technology [5].

In the context of an increasingly intelligent industry, production scheduling models are gradually evolving from static to dynamic, requiring scheduling models to adjust production plans under continuously changing conditions and uncertain processing environments in a timely manner, thereby ensuring that the developed scheduling schemes can meet current workshop operations [6]. From the perspective of production scheduling, digital twins, as digital equivalents of the physical manufacturing environment, provide various decision-making supports, significantly enhancing the flexibility, adaptability, and efficiency of the manufacturing industry, while minimizing processing time, costs, and waste [7]. As this review illustrates, a digital twin can drive the transition of production scheduling towards dynamic optimization, thereby reducing and mitigating the chain reactions caused by interruptions and deviations from production goals [8].

This review aims to explore the latest developments and applications of digital twins in intelligent workshop scheduling and propose potential future research directions. By delving into the technologies and application cases of digital twins, this literature review systematically reflects the current status and future trends of digital twins, hoping to provide valuable references and insights for scholars and practitioners in related fields. Specifically, this review will address the following research questions:

RQ1: Which methods (such as algorithms and application programs) are referred to as digital twins in workshop dynamic scheduling?

RQ2: What modeling methods for digital twins are employed for workshop dynamic scheduling?

RQ3: What scheduling strategies based on digital twins are employed for workshop dynamic scheduling?

RQ4: What system technologies based on digital twins are employed for dynamic workshop scheduling in workshops?

RQ5: Which studies have focused on the architecture of digital twin technology in intelligent workshops within the context of dynamic scheduling?

RQ6: What are the requirements for applying digital twins in dynamic intelligent workshop scheduling, and what are their future developmental trends?

The structure of this paper is as follows: Section 2 provides an overview of the basic concepts of digital twins and production scheduling. Section 3 describes the methods and standards used for literature selection and the research strategy. In Section 4, a statistical analysis of the selected literature is conducted. Section 5 categorizes the included literature and provides a literature review. Section 6 introduces a description of the digital twin application framework found in the literature and proposes a conceptual overall framework for digital twins in intelligent workshops. This includes many manufacturing technologies that complement digital twins, thereby constructing a digital twin of the workshop to achieve dynamic scheduling and control. Section 7 summarizes the included literature and discusses future developmental trends and research gaps in the field of digital twins. Finally, Section 8 outlines the conclusions drawn from this review paper.

2. Background

To better apply digital twins in the dynamic production scheduling process, it is essential to thoroughly summarize and apply relevant knowledge about digital twins and production scheduling, maximizing their performance and multifunctionality in production scheduling [9]. This section briefly introduces the development and characteristics of digital twins and the methods and strategies for dynamic workshop scheduling.

2.1. Digital Twin

As the difficulty of workpiece processing gradually increases, it is necessary to carry out prediction and simulation during processing to ensure that the workpiece meets production requirements. Digital twins are one of the technologies that can meet these needs. With the rapid development of information technologies such as the Internet of Things and big data, digital twins have transitioned from theory to practice [10]. The essence of digital twin technology is to create high-precision models of physical entities through the “virtual-real integration” characteristic, reflecting and optimizing the operation of the physical world by collecting real-time data and performing intelligent analysis and predictions [11]. The term “digital twin” first appeared in David Gelernter’s book, The Mirror World, published in 1991. In its early application cases, NASA (National Aeronautics and Space Administration) constructed two identical spacecraft models for the Apollo Program, one for space missions and the other on Earth, simulating the behavior of the space model to predict its development. Since then, digital twins have gradually entered the public eye and have been increasingly applied in manufacturing [12]. The definition of digital twin has evolved from “a digital mirror of the entire product lifecycle” to “a real-time data-driven adaptive virtual-real symbiotic entity”. The early definition was proposed under NASA’s “twin machine mirror” concept and Grieves’ conceptual model of product lifecycle management, emphasizing high-fidelity digital models of physical entities and a full lifecycle perspective. Subsequently, driven by CPS/IoT, real-time data and continuous synchronization were gradually integrated into the core. In 2017, Tao, F first proposed the “Digital Twin Workshop (DTS)”, systematically elaborating on the four elements of physical workshop (PS), virtual workshop (VS), service system, and twin data, along with the dual-space interaction mechanism of sensors, data, and models, directly reflecting the idea of DT as the digital side of CPS, updated in real-time [13]. By 2018, Werner Kritzinger and others clarified “simulation ≠ DT” using the DM/DS/DT triad, defining the essence of DT as automatic decision-making based on current states and evolutionary predictions [14]. In 2021, Semeraro provided a consensus statement that DT is a collection of models driven by real-time data, adaptively updated throughout the entire lifecycle for simulation, prediction, and prescriptive decision-making, achieving bidirectional synchronization and closed-loop optimization [15]. To date, the definitions of DT have converged and formed a relatively stable theoretical consensus. Overall, a digital twin fuses physics-based modeling with data-driven methods to create a high-fidelity virtual counterpart that co-evolves in real-time with its physical asset and enables bidirectional interaction.

In the manufacturing sector, the main applications of digital twins are as follows: the first is to construct physical models during the design phase using digital twins, employing simulation analysis to evaluate different product design and development techniques, ensuring the accuracy of system models; real-time reflection of the operation status of the production line through digital twins, and formulating optimal process routes and production plans [16]. The second is to promptly identify issues during operation, record key production data, and complete rapid diagnostics, taking corrective actions in a timely manner to promote process optimization and closed-loop supervision. The third is that a digital twin utilizes its “forward-looking” mechanism to analyze historical and current data, predicting future evolutionary trends of the production system and identifying potential deviations in advance, thereby taking preventive measures before problems occur. Compared to post-event corrections, it aims to minimize lead time and costs, implementing optimal interventions before risks materialize, thereby enhancing stability, availability, and energy efficiency [17].

2.2. Dynamic Production Scheduling

Production scheduling is the process of managing, analyzing, and making decisions about production, designed to address the challenges posed by uncertainties and changes in the manufacturing environment in the workshop [18]. The goal is to optimize task allocation in the production process to ensure that the workshop can continue to operate efficiently in the face of unexpected situations. Under the demands of the Industry 4.0 era, manufacturing systems must gradually move towards “intelligence”, as traditional static scheduling has poor adaptability and struggles to respond promptly to the frequent dynamic events occurring in the workshop [19]. Compared to static scheduling, dynamic scheduling offers greater flexibility and adaptability. It can adjust production plans based on the actual conditions of the workshop and unforeseen processing demands, mitigating interruptions caused by uncertain events and making scheduling strategies more suited to the current workshop environment [20].

Dynamic scheduling can be categorized into three major types based on the different ways of responding to and processing data. These include real-time scheduling, which adjusts scheduling plans in a timely manner according to real-time events; pre-reactive scheduling, which combines pre-planning and post-response to address uncertainties in the production process and is the most widely used dynamic scheduling approach in actual production applications; and robust scheduling, which fully considers future uncertainties when formulating scheduling plans and takes preventive measures to ensure smooth production process [21]. In real-time scheduling, no initial plan is generated in advance; instead, scheduling decisions are made based on the real-time production status of the workshop. Although this scheduling strategy can sense disturbances and respond, its scheduling response efficiency is relatively slow and its stability is weaker, making it prone to becoming stuck in local optima [22]. Robust scheduling predicts potential issues that may occur in future production processes, ensuring the stability of workshop operations. However, the scheduling strategy is overly conservative and cannot fully leverage the performance of workshop machines [23]. Pre-reaction scheduling combines real-time scheduling and robust scheduling, possessing the ability to predict anomalies in advance while also being able to adjust scheduling strategies in response to occurring anomalies [24]. To solve these three scheduling schemes, heuristic methods have been proposed. F. Zhao introduced a mixed-integer linear programming model and designed a population-based iterated greedy algorithm (PBIGA) to address workshop problems. To improve algorithm efficiency, an accelerated NR3 algorithm was proposed and high-quality initial populations were generated through heuristic algorithms. Meanwhile, local search and destruction construction mechanisms were designed for product and job sequences [25]. In meta-heuristics, Amir M proposed an online mixed-integer programming model to reduce workshop manufacturing time and limit energy consumption, employing two scheduling strategies; namely, prediction-reactive and proactive-reactive scheduling. The complexity of the model was managed using methods such as Lagrange relaxation and Benders decomposition, and large-scale instances were solved using simulated annealing and tabu search [26]. In hyper-heuristics, Emilio Singh et al. proposed an ant-based constructive hyper-heuristic method, applying it directly to the heuristic space while analyzing the impact of different pheromone maps (2D and 3D) on algorithm performance, automatically adjusting the usage strategies of various heuristics to enhance solving performance through intelligent selection or combination of heuristic algorithms [27].

3. Research Methodology

3.1. Document Search Strategies

The specific document retrieval strategies include keyword retrieval, database searching, and other measures. To ensure the high relevance and scientific rigor of the retrieved documents and content [28], this study develops a literature search strategy at the intersection of Digital Twin, Workshop/Dynamic Scheduling, and Intelligent Manufacturing. We first identify core search terms—“Digital Twin,” “Workshop Scheduling,” “Dynamic Scheduling,” “Intelligent Manufacturing,” “Digital Twin Modeling,” and “Intelligent Twin Framework”—and then refine the search queries using Boolean logic to balance relevance and coverage. For example, using AND restricts results to the intersection of topics (e.g., “Digital Twin AND Dynamic Scheduling”), thereby focusing on studies that address both; employing OR broadens the scope (e.g., “Workshop Scheduling OR Digital Twin”) to avoid missing potentially relevant work; and, when necessary, NOT excludes unrelated areas (e.g., “Digital Twin NOT Warehouse Scheduling”). The number of papers and research activities in this research direction have risen since 2015, and special attention was paid to relevant documents published in the past five years [29].

To balance rigor and coverage—and given the study’s interdisciplinary and international scope—we conducted parallel searches across multiple bibliographic databases. We used Web of Science to target high-quality journals and to conduct citation-network analyses, thereby identifying seminal works and research trajectories; its advanced search functions (e.g., keywords, authors, institutions) further enhanced retrieval precision. IEEE Xplore, with its emphasis on engineering and industrial applications, enabled the identification of technical pathways and case studies of digital-twin implementations in manufacturing and scheduling. ScienceDirect, with broad cross-disciplinary coverage, facilitated the capture of recent advances in intelligent manufacturing and the methodological evolution of dynamic scheduling [30]. CNKI complemented these sources by documenting developments in the Chinese context and reflecting domestic scholarly trends. Beyond these mainstream databases, we leveraged Google Scholar for natural-language queries and forward- and backward-citation chasing to surface influential works and gray literature (e.g., theses, conference papers). We also consulted proceedings from relevant international conferences (e.g., IEEE CASE, IEEM) to incorporate frontier and exploratory findings in a timely manner. Collectively, these sources provide authoritative and comprehensive coverage, effectively representing the current state of fundamental theory and applied technologies in the DT domain [31].

3.2. Document Screening and Evaluation Standards

In research on digital-twin-enabled dynamic workshop scheduling, the quality of the literature directly constrains the depth and breadth of inquiry; therefore, rigorous screening of the retrieved records is essential. We assess the literature along three dimensions—relevance, scholarly quality, and timeliness. First, for relevance, we verify whether the topic directly aligns with digital twins and dynamic scheduling in intelligent workshops, decisively excluding studies that deviate from manufacturing/scheduling scenarios. Second, for scholarly quality—treated as a key indicator of value—we prioritize publications indexed in authoritative journals, authored by recognized scholars, and subjected to rigorous peer review; we also consider quantitative indicators such as journal impact factors, citation counts, and position within citation networks, giving precedence to highly cited articles and those published in high-impact journals. Third, for timeliness, we focus on developments from the past five years while retaining a limited number of foundational early studies to establish the historical trajectory of theories and methods. Following this screening and quality assessment, the included literature is not only tightly coupled to the “digital twin–scheduling–intelligent manufacturing” theme but also meets requirements for academic rigor and timeliness, thereby laying a solid foundation for subsequent research.

The screening workflow follows a closed-loop principle of initial screening → relevance assessment → scholarly quality assessment → timeliness assessment. In the initial screening, titles, abstracts, and keywords are reviewed to remove records whose topics do not align or whose scenarios are misaligned. During the relevance assessment, full texts are examined to determine whether their research questions, methods, and conclusions closely correspond to digital-twin-driven dynamic scheduling; studies that touch only peripheral themes or adopt an overly broad scope are excluded. In the scholarly quality assessment, we prioritize publications in high-impact, peer-reviewed journals and major conferences, incorporating journal impact factors, citation counts, and citation network relationships into the evaluation. In the timeliness assessment, emphasis is placed on works from the last five years, while judiciously retaining a small number of milestone studies to ensure conceptual completeness. In addition, our quality appraisal considers the transparency and reproducibility of research design and methods, giving preference to studies that report adequately on data sources, model specifications, experimental/simulation design, and statistical analysis; where appropriate for the study type, we also draw on risk-of-bias assessment frameworks to identify and document factors that may affect the robustness of conclusions. On this basis, the retained literature supports the construction of the theoretical foundation, the development of technical roadmaps, and the formulation of subsequent methodological plans [32].

4. Bibliometric Analysis of Documents

Research on digital twin-based workshop scheduling attracts attention in a wide scope. Consequently, multiple representative databases were selected for retrieval, including Web of Science, IEEE Xplore, ScienceDirect, CNKI, etc. These databases cover a large number of documents within an extensive scope, ranging from fundamental theories to application technology, and comprehensively reflect the current research status. In this section, 136 documents are summarized, with a statistical overview provided in terms of publication types and periodical publishing houses. This study begins with the dimensions of Year, Authors, and Publication Title to further delineate the Research Subject. Building on this basis, it synthesizes the Algorithms and Techniques for Digital Twins, the Problems Addressed, the Methods Applied, and the Outcomes Achieved, while also compiling information on the Journal and on Databases and Publishers. Detailed data are provided in Appendix A. Analyzing these data can help extract key information, and based on the statistical results, researchers and journals with a significant influence on the study of digital twin-driven dynamic production scheduling can be identified, thus further deepening the research direction for scholars and students in the field [33]. Figure 1 illustrates the distribution of years of the publications selected, which reveals that the field attracts more attention in recent years. Over the past four years, journal articles have dominated among various types of published documents. Given that the current period is the beginning of 2025, with relatively limited research achievements that have been formally published and plenty of manuscripts still being reviewed, the number of articles is relatively small. Figure 2 primarily describes the types of publications selected, and Figure 3 presents the top eight periodical publishers. Figure 4 lists the top nine authors of publications. Figure 5 illustrates the major research orientations of the publications selected, and Figure 6 shows the research fields involved by the top seven publications. Figure 7 contains the types of workshops primarily studied in publications, and Figure 8 shows the word clouds generated from the keywords of the selected documents.

5. Literature Review

In smart workshops, digital twins abstract multi-entity systems such as equipment, workstations, and logistics into computable and predictable digital replicas through the integration of mechanisms, data, and simulation. Based on this, decision spaces that are assessable, comparable, and optimizable are constructed within the twin, generating executable scheduling plans around objectives such as delivery time, production capacity, and cost, using mathematical analysis, simulation collaboration, and intelligent algorithms. To support the feasibility and controllability of the twin model and scheduling strategy, data must be systematically collected in real-time, arranged in a hierarchical architecture from the data layer to the model layer, then to the decision layer, followed by the execution layer, and finally to the feedback layer [34]. Among them, modeling methods provide credible scenarios and predictions, scheduling strategies search and evaluate on the twin model, and system technologies close the loop of data-model-execution. The system architecture specifies interfaces, governance, and evolution paths, thus achieving maintainable, auditable, and scalable closed-loop optimization.

The analysis presented in the section and the literature selected herein are primarily aimed at addressing the following research questions:

RQ1: What approaches in dynamic workshop scheduling (such as algorithms and application programs) can be categorized as a Digital Twin?

RQ2: What modeling methods for digital twins are employed for workshop dynamic scheduling?

RQ3: What digital twin-based scheduling strategies are employed for dynamic workshop scheduling?

5.1. Digital Twin-Based Methods in Dynamic Workshop Production Scheduling

Digital twins serve as a new technical measure for dynamic scheduling in workshops. Through real-time data collection, simulation, and optimized decision-making, digital twins effectively enhance production efficiency and scheduling flexibility. Technical measures referred to as “digital twins” in workshop dynamic scheduling mainly include the following categories: model-based scheduling algorithms, which utilize digital modeling to simulate and optimize decisions for equipment, systems, production lines, or supply chains, with the most commonly used method being discrete event simulation (DES), which optimizes scheduling by simulating the sequence of events in the system (such as tasks, resource allocation, and equipment state changes) [35]; adaptive optimization scheduling algorithms, which combine digital twin models with optimization algorithms, such as genetic algorithms, particle swarm optimization (PSO), and simulated annealing, to dynamically optimize scheduling based on production objectives (such as minimizing total time and maximizing resource utilization) [36]; and data-driven dynamic scheduling methods, which utilize real-time data from IoT devices, sensors, and ERP/MES to perceive production status and dynamically adjust scheduling strategies. Virtual simulation and decision support system (DSS) use virtual simulation to model different scheduling schemes and optimize scheduling decisions; predictive scheduling and maintenance combines digital twins with predictive maintenance to anticipate equipment failures or production bottlenecks in advance and perform scheduling adjustments to reduce the impact of unexpected events [37].

These methods and algorithms are closely integrated in the application of digital twins, enabling real-time monitoring of production status, optimizing scheduling strategies, and enhancing decision-making efficiency, thereby ensuring the efficient operation of intelligent workshops [38].

5.2. Exploratory Research on Digital Twin Modeling Methods for Dynamic Scheduling

A digital twin constructs a virtual model of physical entities through digital means, promoting the deep integration and interaction between the physical world and information world, serving as an important bridge for industrial information integration [39]. The model is the core of digital twin realization, which can be categorized based on different applications: product-level models targeting individual devices or components (such as aircraft engines) for design optimization, fault detection, and personalized customization; process-level models for manufacturing or business processes (like automobile production lines) to enhance production efficiency and quality management, as well as flexibly adjusting production efficiency, quality management, and production plans and resource allocation; and system-level models covering multiple interacting products or subsystems (such as wind power stations) for collaborative work, resource optimization, and risk assessment. From product-level to process-level and then to system-level, digital twin modeling continuously expands, driving production from a single product to multi-product collaboration, enhancing overall performance and efficiency [40]. Based on the surveyed literature, representative studies were extracted and synthesized to summarize and categorize method classes, principal characteristics, experimental data, baselines, and performance outcomes, as shown in

Table 1. Normalized Quantitative Evaluation of Modeling Approaches.

ID	Reference	Method Category	Key Features	Data/Scenario	Baseline/Comparator	Performance (Quant/Qual)
1	Liu et al., 2024 [41] (International Journal of Advanced Manufacturing Technology)—Fusion method for a production-line DT	Data modeling + multi-level fusion framework	Structured data modeling for real-time production data; full-factor semantic judgment; unit-to-system multi-level data–model fusion.	Micro-assembly workshop case study.	Qualitative comparisons with no-fusion/traditional integration.	Correctness/feasibility validated; supports multi-scale DT fusion.
2	Jiang et al., 2021 [42] Journal of Manufacturing Systems)—How to model and implement connections…	DT modeling framework/description model + cyber-physical connectivity	DES modeling; Seven Elements (controller/actuator/processor/buffer/flow entity/virtual service node/logistics path); service unit encapsulates I/O and control logic; virtual–physical mapping.	Real workshop DT application (connectivity validation).	No explicit numeric baseline; emphasis on practicality and engineering feasibility.	Faster modeling and lifecycle connectivity feasible for production use.
3	Zhou, 2022 [11] (Discrete Dynamics in Nature and Society)—Numerical Analysis aided by graph-theoretic optimization	Graph-theoretic combinatorial optimization + data preprocessing/clustering (AP)	AP clustering + graph theory; handles asynchronous/incomplete DT monitoring data; big-data-based DT; batching, outsourcing, and rolling scheduling framework.	Big-data platform examples; production line scheduling simulation.	Compared qualitatively with traditional methods (no unified public numeric baseline).	Streamlines numerical analysis and scheduling; lacks unified quantitative metrics.
4	Latsou et al., 2023 [43] (Journal of Manufacturing Systems)—DT-enabled anomaly detection and bottleneck identification	Multi-agent CPS + anomaly/bottleneck analytics	Extended 5C architecture; multi-level agents with a monitoring agent; real-time sensor data; automatic feedback to the physical system.	Real case in a cryogenic warehouse for cell and gene therapy; continuous sensing.	Before/after and process comparisons with manual monitoring/decision processes.	Average ~30% improvement in human resource utilization; decision and control efficiency improved.
5	Chua et al., 2022 [44] (Journal of Computing and Information Science in Engineering)—Surrogate model (MARS) for performance prediction	Surrogate/metamodel + production performance prediction	MARS surrogate; three input groups (system load/machine/product); uses DT’s real-time sync data to predict flow time, tardiness, and machine utilization.	Industrial wafer manufacturing case; random sampling with varying training sizes.	Compared with linear/non-linear models and feature-count variants.	Semi-quantitative: high correlations while substantially reducing input dimensionality.
6	Li et al., 2021 [45] (Journal of Manufacturing Systems)—Semantic modelling & resource recommendation	Ontology/semantics + manufacturing resource recommendation (MT&MR)	Ontology-based manufacturing task semantics; semantic indexing and retrieval; dynamic MR recommendation for DTS.	Workshop case validation.	Process/method comparisons; no unified numeric baseline.	Effective and feasible for fast production support.
7	Wang et al., 2024 [46] (International Journal of Advanced Manufacturing Technology)—Knowledge-driven multi-view XBOM reconfiguration	Knowledge-driven + XBOM reconfiguration/recommendation	Knowledge base for overhaul data; BiLSTM-CRF for entity recognition; interactive system for XBOM reconfiguration.	High-speed EMU bogie overhaul data; enterprise case.	Compared with manual/rule-based workflows (no unified numeric baseline).	Shorter cycle, higher efficiency/quality.
8	Xie et al., 2024 [47] (Journal of Manufacturing Systems)—MNOSE description model	Extended DT description model (MNOSE) + rapid reconfiguration	Material-node-oriented; captures both inter-equipment and equipment–material relations; enables fast (re)configuration and logical unification across DT services.	Modeling of typical production organizations and case studies.	Contrasted with the original Seven-Element concept.	Improves ease and speed of DT updates/reconfiguration.
9	Sun et al., 2022 [48] (Journal of Nanjing University of Aeronautics and Astronautics)—Modelling & application of DT for production process	System architecture + geometric and communication modeling	DT architecture for intelligent workshops; 3D geometry via point-cloud fitting; OPC UA information modeling; multi-source heterogeneous data integration.	Engine manufacturing workshop DT implementation.	Qualitative comparisons with non-DT/traditional integration.	Effective for visualized production control.
10	Luo et al., 2024 [49] (Proc. IMechE Part B: J. Eng. Manufacture)—Assembly-feature construction for equipment mesh models	Geometric/assembly features + workshop-level DT modeling efficiency	3D assembly information model; coarse-to-precise feature localization; precise mapping from info model to mesh models to speed assembly modeling.	Arc-welding and warehousing workshops; comparative experiments.	Compared with traditional geometric/mesh assembly workflows.	Improved modeling efficiency (no public numeric values).
11	Nejati et al., 2024 [50] (Computers & Industrial Engineering)—ML-based simulation metamodel (MLBSM)	ML metamodel + dynamic scheduling	SPBM vectorization of logs; multi-output AdaBoost regression; novel statistical risk assessment; bypasses multiple heavy simulation replications.	Synthetic MES data for a lithography workstation in semiconductor manufacturing.	Compared with discrete-event simulation and baseline predictors.	>80% recall for high-risk jobs; ≥70× faster than traditional simulation; consistent sensitivity analysis.
12	Wu et al., 2025 [51] (IEEE Internet Computing)—LLM-driven generative DT modelling	Generative modeling (LLM) + ontology/knowledge retrieval	LLM parses intent to produce hierarchical object structures; retrieval + dynamic prompting to create/fuse objects; builds structured, semantically rich DT models.	Workshop resource-scheduling example.	Compared with manual/ontology-only workflows.	Validated effectiveness; shows potential for automated modeling.

. The analysis indicates that modeling approaches for digital-twin manufacturing workshops can be broadly classified into three categories: classical simulation/graph-theoretic modeling methods, distributed intelligent modeling methods, and data- and knowledge-driven modeling methods. Each category exhibits distinct advantages, limitations, and application contexts. The following sections provide a detailed summary of these three classes of methods, outlining their respective strengths and weaknesses and highlighting typical application scenarios.

5.2.1. Classical Simulation and Graph Theory Modeling Methods

Research on modeling methods for digital twins has been continuously advancing across various manufacturing workshops, from semiconductor wafer workshops and aerospace assembly workshops to chemical fiber workshops, gene therapy cryogenic storage, and digitized intelligent factories [41]. Discrete event simulation (DES) and graph theory models are the most widely used classic foundational methods. DES formalizes discrete event-based modeling of systems, enabling the construction of virtual models of manufacturing systems and their connection to physical systems, thus achieving the simulation and optimization of production processes.

In the recent literature, Jiang Haifan et al. focus on complex discrete manufacturing workshops, proposing a digital twin modeling method based on DES theory and a virtual-physical interconnection mechanism to achieve the construction and application of digital twins throughout the lifecycle of manufacturing systems [42]. Wang, Yankai, and Wang Shilong, targeting hybrid flow digital workshops, proposed a green scheduling model that integrates a dynamic digital twin with Adaptive Multi-Objective Dynamic Harris Hawks Optimizer (AMODHHO), achieving dual optimization of total completion time and energy consumption [46]. In the application of graph theory models, Zhou Sujing proposed a big data-driven modeling approach that integrates digital twin with graph theoretic combinatorial optimization for discrete manufacturing workshops, constructing a real-time monitoring and scheduling system based on AP clustering and priority rules, addressing issues of data asynchrony, incompleteness, and scheduling uncertainty, significantly enhancing workshop flexibility and production efficiency [11]. Although graph theory models like discriminative graphs ensure rigorous understanding, their practicality may be insufficient as complexity increases and real-time requirements grow.

The core of this type of modeling lies in mathematical models and simulations, which deduce system behavior; its advantages include theoretical rigor, verifiability, and interpretability. DES supports the visualization, monitoring, and control of manufacturing systems, while graph theory models can formally verify complex structures to ensure correctness, making them easier to adopt in stable scenarios such as mass production and assembly lines. However, their limitations arise when systems are complex and require real-time responses, as the computational and modeling maintenance costs are high, and virtual-physical synchronization may lag; graph theory models lack flexibility under dynamic, uncertain events, making it difficult to meet real-time requirements. Such modeling methods are often used in large-scale discrete manufacturing, assembly lines, and large assembly lines where processes are relatively fixed and events are discrete; DES is commonly used as the main framework for scheduling and logistics simulation, combined with data analysis to assess bottlenecks and utilization; when processes and constraints are clear, graph theory models can also be used for combinatorial optimization of process paths and equipment layouts to enhance efficiency [52].

5.2.2. Distributed Intelligent Modeling Methods

To overcome the limitations of centralized models in complex dynamic environments, distributed and intelligent modeling approaches have emerged. These methods abstract machines, workpieces, and people in manufacturing systems as intelligent agents, achieving autonomous control and optimized scheduling of workshops through the collaboration and game theory of agents [53]. Latsou, Christina proposed a new method integrating digital twin with multi-agent CPS for low-temperature workshops in the cell and gene therapy industry, introducing “monitoring agents” to achieve micro-level anomaly detection and bottleneck identification, enhancing the utilization of human resources and decision-making capabilities in complex manufacturing systems [43]. In flexible job shop scenarios, the decentralized GMAS scheduling model utilizes multi-agent centralized learning and local execution to achieve real-time bottleneck detection and global performance improvement. Pu Yu, Fang Li et al. addressed modeling problems in flexible job shops for personalized manufacturing, proposing a centralized learning decentralized execution multi-agent scheduling method (GMAS) based on graph convolution network, constructing a directed acyclic graph probabilistic model and global action modeling to address the complexity and dynamics of Flexible Job Shop Scheduling (FJSP), achieving intelligent optimization of scheduling strategies [54].

Multi-agent modeling has the advantages of distributed autonomy and real-time responsiveness. Agents can perceive local states, make decisions in real-time, and collaborate to achieve global objectives, thus being able to timely detect and alleviate bottlenecks and adapt to dynamic changes in production, suitable for dynamic changes such as interleaving and system failures; for instance, “monitoring agents” can automatically detect anomalies and self-optimize human resource reallocation, while GMAS combines reinforcement learning to enable agents to autonomously learn scheduling strategies, making them more agile and robust in high-variability scenarios [55]. The trade-off is an increase in modeling and coordination complexity, with larger scales prone to instability and convergence issues; improper collaboration mechanisms can lead to process conflicts. Additionally, there is a strong reliance on data, with significant cold start as well as training costs, and safety and interpretability must also be considered in industrial real-time applications. Therefore, Multi-Agent Systems (MAS) are more suitable for flexible, customized manufacturing scenarios including complex workshops with heterogeneous coupling, such as flexible assembly lines, FJSP, and cell gene therapy; as well as scenarios requiring real-time bottleneck detection and autonomous correction in intelligent warehousing/AGV scheduling. When systems are highly dynamic and centralized control is inadequate, distributed intelligent agent modeling is particularly effective [56].

5.2.3. Data-Driven and Knowledge-Driven Modeling Methods

With the acceleration of workshop digitalization and data accumulation, machine learning (ML) and data-driven technologies are becoming mainstream solutions for digital twin modeling. These methods utilize big data and intelligent algorithms from the workshop to train models, predict performance, and optimize decisions, thereby partially replacing and enhancing mechanism-based models. Firstly, agent models represented by MARS can quickly approximate the workshop response surface using real-time data and provide lightweight decision support for scheduling [57]. Chua, Ping Chong, et al. proposed a proxy modeling method based on Multivariate Adaptive Regression Splines (MARS), integrating system load, equipment parameters, and product parameters to efficiently predict key production performance indicators and provide lightweight decision support for scheduling optimization [44]. Secondly, in relational structure learning, Graph Neural Networks (GNN) excel at capturing the complex relationships within manufacturing systems, representing equipment, tasks, etc., as graph nodes to predict bottlenecks [58]. Liu Tingyu, Hong Qing, and Sun Yifeng proposed a production action recognition method based on Attention Graph Convolutional Networks (GCN), utilizing topological graph structures to extract digital twin features, effectively supporting standardized process construction and digital twin model building [59]. Jwo, Jung-Sing addressed the challenges of complex manufacturing processes and virtual modeling in workshops by simplifying digital twin modeling with machine learning methods and designing Data Twin Service (DTS) and Cyber-Physical Factory (CPF), achieving collaborative integration of real and simulated data through physical and network operations [60]. Ladj, Asma integrated unsupervised learning and knowledge reasoning into digital shadows and digital twin models aimed at bottleneck identification to achieve micro-behavior-driven system modeling and intelligent decision optimization [61].

In the knowledge-driven direction, modeling utilizes structured semantics (such as knowledge graphs) to integrate heterogeneous data and support model interoperability and reconfigurability. Li, Xixing, Lei Wang, and Chuanjun constructed an ontology-based semantic modeling and indexing mechanism to achieve efficient matching and recommendation of production resources, enhancing the accuracy and timeliness of scheduling optimization in DTS [45]. Liu, Xiaojun, and Chongxin Wang proposed a structured data modeling and multi-level fusion framework, achieving deep integration of real-time production data and multi-dimensional digital twin models through a full-factor semantic mechanism, addressing the issue of insufficient multi-scale feature modeling [62]. Wang, Yunrui, and Yaodong Wang proposed a knowledge-driven XBOM reconstruction modeling method based on knowledge bases and the BiLSTM-CRF algorithm for maintenance scenarios of complex products like EMU bogies in digital twin workshops, effectively improving the efficiency and quality of multi-view bill of materials reconstruction [63]. Xie, Jiaxiang, and Haifan Jiang proposed an improved material node-oriented SevenElements (MNOSE) description model for discrete manufacturing workshops, enhancing the expressiveness of relationships between equipment and equipment-material, achieving efficient construction and rapid reconfiguration of workshop digital twins [47].

To simultaneously meet high fidelity and real-time synchronization [64], modeling methods such as point cloud + OPC UA + multi-source integration [48], multi-dimensional modeling methods combined with anomaly handling services [65], dynamic fidelity reconstruction [66], and assembly feature construction based on equipment grid models [49] have emerged. In high-precision, strongly constrained workshop modeling—for example, in semiconductor photolithography workshops—Machine Learning-Based Simulation Models (MLBSM) combine data vectorization and multi-output adaptive regressors with risk assessment to replace traditional simulation bottlenecks, achieving rapid prediction and scheduling support for high-risk operations [50]. Additionally, large language models (LLMs) automatically generate hierarchical object structures through intent parsing and domain knowledge extraction, shortening modeling cycles and enhancing the semantic expression and intelligence level of models.

Data-driven and knowledge-driven modeling centers around real-time data. The former learns input-output mappings from a large amount of historical data through proxy models, predicting complex workshop behaviors in milliseconds and presenting them with far lower computational costs than fine simulations; machine learning excels at mining high-dimensional non-linear patterns, and graph learning models like GNN can eliminate information silos. The latter ensures semantic consistency and reconfigurability through ontologies and knowledge graphs, spanning design, processes, and operations while enhancing decision reusability and interpretability [67]. However, challenges are also prominent: data-driven approaches depend on data quality and coverage, easily leading to underfitting due to insufficient historical data; the cost of deploying computational power raises the threshold for implementation; on the knowledge side, there are human resource costs and update lags in constructing and maintaining ontologies and knowledge bases; at the same time, there are weaknesses in generalization ability, the need for online adaptation, and security risks such as data poisoning, so caution should be exercised in data-scarce and security-sensitive scenarios [68]. Particular benefits can be achieved in the context of highly digitalized factories and data-intensive decision-making; for instance, in semiconductor manufacturing, machine learning replaces part of the simulation to achieve rapid prediction and scheduling of high-risk processes, customized production supports rolling plans with real-time proxies to reduce delay rates, and complex assembly and maintenance workshops leverage knowledge graphs to integrate cross-link information to support collaborative decision-making, while three-dimensional visual monitoring integrates point clouds, sensors, and OPC UA to achieve high-fidelity virtual-real synchronization [51]. Overall, when data is abundant and rapid iteration is needed, data-driven approaches are prioritized to gain real-time and intelligent dividends; when systems are heterogeneous and cross-domain complexity arises, knowledge-driven approaches are introduced to bridge semantics and hierarchies, constructing a scalable and reconfigurable digital twin system.

5.3. Strategies and Methods Implemented Based on Workshop Scheduling in Digital Twin Models

Section 2.2 outlines three strategies for production scheduling: real-time, robustness, and proactive-reactive scheduling. In the field of intelligent scheduling workshops oriented towards digital twins, intelligent scheduling methods have been validated and applied in various types of digital twin manufacturing workshops. Flexible job shops need to cope with flexible scheduling across multiple processes and devices, assembly workshops emphasize process coordination and timely response, semiconductor workshops involve reentrant processes and AGV logistics scheduling, while wood furniture workshops require efficiency in customized production [69]. In response to the uncertainties arising in workshop scheduling, research focus has expanded from static scheduling to dynamic rescheduling, robust optimization, and production maintenance collaboration, and production maintenance collaboration. Researchers have deeply integrated digital twins with real-time feedback, event-driven rolling rescheduling, logistics path planning, and predictive maintenance, constructing robust optimization frameworks to collaboratively schedule processing equipment, AGVs, and maintenance resources, significantly enhancing the adaptability and overall reliability of workshop production systems to random disturbances [70]. Based on the literature search, representative studies were selected and systematically synthesized across the dimensions of method category, key characteristics, experimental datasets, comparative baselines, and performance metrics; the findings are presented in

Table 2. Normalized Quantitative Evaluation of Scheduling Algorithms.

ID	Reference	Method Category	Key Features	Data/Scenario	Baseline/Comparator	Performance (Quant/Qual)
1	Gao et al., 2024 [71] (Robotics and Computer-Integrated Manufacturing)—DT-driven dynamic scheduling with worker allocation	DT + dynamic scheduling + multi-objective optimization (improved NSGA-II/IMOEA)	Monitors disruptive events (insertion, cancelation, absenteeism, rotation); trigger-based rescheduling; multi-skill/multi-level worker assignment; integer programming; three population-initialization rules with tuned parameters.	Complex-product assembly workshop; DT system constructed for validation.	Compared with vanilla NSGA-II and no-trigger strategies (qualitative).	Qualitative: balances efficiency (makespan) and stability (time deviation); no unified public numbers.
2	Huang et al., 2021 [72] (Journal of Shandong University: Engineering Science)—Lion swarm algorithm for DT job-shop scheduling	DT + metaheuristic (lion swarm) + flexible job shop	Real-time digital–physical interaction; equipment-utilization-oriented optimization; addresses machine failures; LSA used for initial plan generation and improvement.	Real machining shop-floor data.	Compared with traditional heuristics/rules.	Qualitative: stronger search and faster speed across scales; improved overall system performance.
3	Zhang et al., 2023 [73] (Flexible Services and Manufacturing Journal)—DT-driven flexible scheduling via hierarchical RL in a human–machine collaborative workshop	DT + hierarchical RL + human–robot collaboration	Parallel lines as communities; community-level flow optimization and H–M participation ratio tuning; improves flexibility and load balance.	Ventilator assembly case.	Compared with traditional scheduling/fixed participation ratios.	Qualitative: stronger adaptability to demand and line changes.
4	Wang et al, 2025 [74] (Process Safety and Environmental Protection)—End-to-end scheduling DT for multistage batch plants with safety	End-to-end DT + deep reinforcement learning + safety constraints	Multi-product, multistage batching; robust to interruptions/maintenance insertions; minimal retraining for different scales.	Comparative tests on multistage batch processes.	Contrasted with mathematical programming/metaheuristics/hybrid models (qualitative).	Qualitative: fast interaction and strong solution quality with robustness.
5	Xia et al., 2021 [75] (Journal of Manufacturing Systems)—A DT to train deep RL agents: environment, interfaces, and intelligence	DT + deep Q-learning + industrial control	Training environment and interface network; near-synchronous digital–physical control; ‘digital engine’ for process knowledge and task orchestration.	DQN training case study.	Compared with traditional control/no-DRL.	Qualitative: potential to improve robustness and efficiency via DRL within system-level DT.
6	Yuan et al., 2023 [76] (Advanced Engineering Informatics)—Multi-agent double DQN with state machine & event stream for FJSP	Multi-agent RL (double DQN) + event-driven MDP + flexible job shop	Decouples event-driven environment and decision; job and machine agents; Boltzmann exploration to avoid local optima; real-time capability.	Large-scale numerical experiments.	Compared with traditional scheduling methods (heuristics/rules).	Qualitative: outperforms traditional methods on large instances; real-time decisions.
7	He et al., 2023 [77] (Computer Integrated Manufacturing Systems)—Dynamic scheduling for textile dyeing via multi-agent recurrent PPO	Multi-agent RL (recurrent PPO) + LSTM + batching/vat scheduling	Batching and vat agents; LSTM for dynamics; inter-agent interaction for global optimization; minimizes total tardiness.	Real textile-dyeing enterprise; multiple scales.	Compared against strong heuristic rules.	Qualitative: markedly reduces total tardiness; improves on-time delivery.
8	Fang et al., 2023 [78] (Journal of Computing and Information Science in Engineering)—Adaptive job-shop scheduling via RL in a DT environment	DT + distributed RL + disturbance-adaptive rescheduling	Monitors state and deviation; defines triggers; distributed RL senses dynamics and applies corrective actions; closed-loop feedback.	Real shop-floor case and DT prototype.	Compared with predictive/static scheduling processes.	Qualitative: effective in deployment; more timely disturbance handling.
9	Xu et al, 2021 [79] (Journal of Information Science and Engineering)—Dynamic DT job-shop scheduling based on edge computing	DT architecture + edge computing + data-mining-based scheduling	DT inserted between business and execution layers; remote monitoring/analysis/management; data acquisition model + multiple scheduling-knowledge models.	Simulation validation on a real shop-floor setup.	Compared with traditional/no-edge architectures (qualitative).	Qualitative: supports dynamic scheduling; improves interaction and response speed.
10	Ma et al., 2025 [80] (Expert Systems with Applications)—Data-driven scheduling via DT with GAN-based sample expansion	Data-driven scheduling + DT data fusion + GAN augmentation	Fuses physical and digital shop-floor data; model-level data-fusion mechanism; GAN-augmented samples + multilayer feed-forward network for decisions.	Semiconductor workshop experiments.	Compared with physical-data-only/conventional training pipelines.	Qualitative: improves training efficiency and scheduling performance.
11	Yue et al., 2024 [81] (International Journal of Advanced Manufacturing Technology)—Disturbance evaluation for DT-based scheduling	DT + disturbance evaluation (CFC + CNN) + adaptive mechanism	Causal-factor graph and CNN to assess disturbance impact; selects scheduling mechanisms by impact level; DT provides data and model support.	Experimental validation (cases and comparisons).	Compared with frequent rescheduling/no evaluation.	Quantitative: avoided two unnecessary reschedulings; disturbance-handling time reduced by 66.3%.
12	Chen et al., 2024 [82] (International Journal of Industrial Engineering—Theory, Applications and Practice)—DT-oriented collaborative optimization of process planning and scheduling	DT + co-optimization of process planning and scheduling (enhanced GA + hybrid PSO)	Enhanced GA generates near-optimal routes with four-level encoding; hybrid PSO with multiple neighborhoods for better local search.	Production instance simulation.	Compared with GA and PSO baselines.	Qualitative: faster convergence, shorter runtime, higher accuracy; optimizes maximum manufacturing span.
13	Guo et al., 2023 [83] (Journal of Manufacturing Systems)—Joint multi-objective dynamic scheduling of machine tools and vehicles via DT	DT + joint scheduling (machines + vehicles) + predictive maintenance/quality	Incorporates machine failure, tool wear, quality monitoring, and energy cost; objectives: minimize makespan and flexibly manage energy consumption; multi-factor service system.	Marine diesel engine workshop; key parts machining.	Compared with separated scheduling/single-factor approaches.	Qualitative: improved timeliness and predictability; superior to comparators.
14	Zhou et al., 2024 [84] (International Journal of Computer Integrated Manufacturing)—DT-based job-shop multi-objective scheduling model and strategy	DT + multi-objective model (makespan/tardiness/energy) + improved NSGA-II	Cloud–edge decision framework; precise processing-time estimation via spatiotemporal compression ratio; data comparison and anomaly detection; multi-mode crossover, random mutation, variable-ratio elitism.	Standard datasets and real processing problems.	Compared against standard NSGA-II and others.	Qualitative: validated strategy and algorithm (better Pareto fronts and stability).

. Building on this synthesis—and focusing on strategies and methods for digital-twin-driven shop-floor scheduling—we further distill five categories of technical strategies and propose two deep-integration pathways.

5.3.1. Five Technical Strategies for Digital Twin-Driven Workshop Scheduling

In the strategy of deepening the combination of classical algorithms, mature classical optimization algorithms (such as Particle Swarm Optimization (PSO), Genetic Algorithms (GA), etc.) are deeply integrated with digital twin technology, driving algorithmic iteration through real-time simulation and feedback from DT, thereby improving scheduling efficiency and adaptability to dynamic changes. Weng et al. deeply coupled DT with competitive PSO and introduced a twin network for perception of production disturbances, achieving rapid convergence and strong anti-interference capability [85]. Javaid et al. constructed a simulation-optimization model (SBOM) based on Simio, integrating digital twin real-time data with hybrid PSO to compensate for the shortcomings of traditional PPS in uncertainty and system integration [86]. Chen et al. embedded hybrid PSO (H-PSO) into DT to solve fuzzy multi-objective scheduling, maintaining population diversity through learning factor adaptation and fuzzy membership-driven crowding distance, suppressing premature convergence [87]. Liu et al. constructed a workshop digital twin and event-driven rolling window rescheduling mechanism for island layout flexible assembly workshops, achieving dynamic optimization in conjunction with GA [88]. Gao et al. nested improved NSGA-II and integer programming within the joint simulation framework, DT-Simio, hierarchically solving rescheduling trigger thresholds and dual-objective trade-offs [71]. Zhou et al. proposed a two-stage GA decentralized scheduling for multi-agent discrete manufacturing, expanding the search space through agent crossover and block insertion, and generating high social welfare consensus schedules through non-dominating sorting and grey relational analysis, enhancing efficiency in large-scale multi-agent scenarios [89]. Within the strategies of integration and expansion of swarm intelligence and bionic algorithms, the in-depth coupling of bionic swarm intelligence with digital twins opens new pathways for enhancing optimization performance. Huang et al. mapped the “leader-cooperation-wandering” phase of LOA to three layers of search subspaces for process selection, equipment allocation, and fault recovery, dynamically adjusting the wandering radius based on the equipment health returned by the digital twin [72]. Huo et al. introduced attraction-repulsion potential field functions in the bacterial foraging algorithm (IBFOA), achieving adaptive trade-offs between machine load and completion time [90].

In the optimization scheduling strategy of bionic swarm intelligence in digital twins, relying on the characteristics of bionic algorithms with few parameters and easy parallelization, online evolution in the real-time simulation evaluation closed loop of digital twins significantly accelerates algorithm convergence, while integrating real-time data feedback and the advantages of algorithm heterogeneity, effectively providing a more efficient and robust solution for complex workshop scheduling problems. Liu, A. Y. proposed a multi-agent scheduling method that integrates attention mechanisms, constructing an optimized simulation model considering factors such as failure rates, route selection, and layout, enhancing the efficiency and robustness of scheduling in complex systems under intelligent manufacturing environments [91].

In terms of scheduling approaches integrating machine learning and digital twins, Zhang abstracts parallel production lines as “production communities”, constructs a digital twin community model, and dynamically balances loads using hierarchical reinforcement learning, verifying its high adaptability to demand fluctuations and line adjustments [73]; Wang proposes an end-to-end solution DRL plus DT solution for multi-product multi-stage batch processing plants, allowing for quick reordering under events such as equipment failures and maintenance insertions without extensive retraining, demonstrating robustness and efficiency superior to traditional algorithms [74]; Xia introduces data-driven digital twin, mapping manufacturing units to a virtual environment for pre-training before validating DRL, and then deploying to achieve autonomous and robust control [75]; Yuan builds an event-driven DT environment, transforming FJSP into MDP, proposing a multi-agent double DQN where job agents and machine agents collaborate, balancing real-time performance and robustness, suitable for large-scale scenarios [76]; Serrano-Ruiz implements flexible autonomous scheduling in a real workshop DT using PPO combined with multi-feature observation and multi-objective rewards [92]; Park first introduces RL for reentrant job workshops before constructing DT to improve control and learning effects [93]. He Junjie and Zhang Jie achieve fully reactive global optimization for dynamically arriving orders using MA-RPPO + LSTM [77]. Fang integrates distributed RL into DT, triggering feedback based on real-time perception of disturbance deviations, automatically correcting and self-learning for optimization [78]. Heik quickly balances energy consumption, project duration, and resource utilization on an IIoT testing platform using PPO [94]. Gu designs a two-stage adaptive dynamic scheduling strategy with PPO and achieves good results [95]. Overall, DRL plus DT, through event-driven and hierarchical systems combined with high-frequency closed-loop evaluations of DT, enables rapid re-tuning in highly uncertain environments, cross-objective trade-offs, and deployable real-time decision-making.

In the research on hybrid improved strategy optimization, Gan, XueMei proposed an adaptive scheduling method called Explicit Exploration and Asynchronous Update Proximal Policy Optimization Algorithm (E2APPO), achieving self-regulation and self-learning under dynamic interaction between virtual and physical workshops [96]; Li, Yuxin and Xinyu Li proposed a dynamic reconfigurable workshop scheduling method based on multi-agent deep reinforcement learning, effectively optimizing tardiness costs and enhancing the scheduling robustness and intelligence of complex workshops in dynamic environments [97]. Lv Lingling proposed a type-aware multi-agent deep reinforcement learning (MADRL) method, addressing the frequent machine failures in flexible job shop environments, realizing real-time schedule recovery through heterogeneous graph modeling and meta-path-type-aware mechanisms [98]. Hu Liang and Liu Zhenyu adopted the method of the deep Q network (DQN) to handle scenarios with shared resources, flexible paths, and random arrivals, approximating the state-action value function with a GCN-based Petri-net Convolution (PNC) layer, leveraging input/output matrix propagation features and parameter reduction for stability, outperforming traditional algorithms [99].

Classic algorithms are deepened and combined based on instant evaluation, featuring low cost, controllable parameters and processes, and fast convergence, making them suitable for scenarios with clear rules, moderate dynamics, and few insertions. However, they are parameter-sensitive, prone to premature convergence, and heavily reliant on high-precision digital twins (DT), facing global and response insufficiencies in highly dynamic situations. Bionic swarm intelligence enables parallel exploration and online adaptation, capable of handling frequent insertions and vast solution spaces, but essentially optimizes heuristic algorithms, lacking global optimality and relying on high-fidelity low-latency DT. Swarm intelligence optimization achieves rapid trial-and-error and multi-objective collaboration through high-frequency feedback from DT, being more robust to disturbances, but requires complex reward mechanisms and stability governance, with high demands on DT concurrency and real-time performance, and still lacks optimal guarantees [100]. Reinforcement Learning (RL) combined with DT models scheduling as a Markov Decision Process (MDP), enabling millisecond-level decision-making, handling high-dimensional non-linear and multi-objective (reward shaping), significantly reducing reliance on expert rules, but incurs high training/simulation costs, weak interpretability, and generalization, and faces challenges in interdisciplinary modeling. DT combined with hybrid improved algorithms achieves global-to-local coupling, such as EGA→PSO, embedding integer programming in swarm intelligence and improving NSGA-II for more robust multi-objective trade-offs, customizable for dynamic reconfigurable and frequently failing scenarios, but with complex integration, heavy parameter tuning, high computational overhead, and high team capability requirements [101]. In practical applications, stable process lines and processing workshops prioritize the application of DT and classic or swarm intelligence; in highly dynamic and uncertain scenarios (such as semiconductors and custom assembly), RL combined with DT or multi-agent swarm intelligence is preferred; extreme performance or system-level collaboration is best achieved through hybrid improved strategies, with DT leading simulation verification and gradual implementation to control risks.

5.3.2. Two Deep Integration Paths for Digital Twin-Driven Workshop Scheduling

In the research on the cloud-edge-end collaborative integration path, Xu Li-Zhang proposed a dynamic scheduling method for digital twin job workshops (DTJ) based on edge computing to address the shortcomings of traditional scheduling models in real-time interaction between information and physical spaces. A DT architecture is constructed at the management execution layer, building a scheduling model through data mining, integrating data collection and multi-scheduling knowledge to achieve dynamic scheduling optimization and remote monitoring and management of the manufacturing process [79]. Li Juan and Tian Xianghong proposed a dynamic data scheduling method that integrates digital twin and a cloud genetic algorithm (CGA) to address the premature convergence and yield decline issues in flexible industrial workshop scheduling, achieving efficient collaborative scheduling and stable output among multiple production lines through the construction of a comprehensive information integration model and the introduction of a chaotic particle swarm optimization algorithm [102]. Wang Jin and Liu Yang proposed a real-time digital twin flexible job shop scheduling method (R-DTFJSS) based on edge computing to address the limitations of traditional dynamic scheduling methods (TDSM) in coping with frequent abnormal disturbances. By achieving real-time interaction between the physical and virtual workshops, efficient job allocation is realized, and an improved Hungarian algorithm is introduced to optimize scheduling results, enhancing accuracy and robustness [103]. Liu Zhifeng constructed a hyper-network model of feature-process-machine through the integration of digital twin and hyper-network technology, achieving the association and clustering of multi-source data, making intelligent scheduling more efficient [104]. Ma Yumin and Li Luyao integrated the data of physical workshops and digital workshops, combined them with the sample expansion mechanism of the generative adversarial network and the multi-layer forward neural network to perform high-quality training of the scheduling model, and effectively enhanced the scheduling efficiency and precision [80]. Yue Pengjun proposed a scheduling mechanism disturbance evaluation method based on DT technology to address the continuous rescheduling issues caused by frequent disturbances. By evaluating the disturbance impact degree through causal factor graphs (CFC) and convolutional neural networks (CNN), combined with the full-process data and model support offered by DT, targeted scheduling response strategies are formulated [81].

In the integrated technological fusion path for multi-task multi-objective optimization, relying on explicit exploration entropy, parallel evolution pruning, and online transfer learning, combined with graph theory models, hyper-network structures, and data augmentation, collaborative optimization of production scheduling, logistics path planning, and equipment maintenance is achieved [105]. Serrano-Ruiz Julio C. proposed a job shop intelligent manufacturing scheduling method (JSSMS) based on the Markov decision process (MDP) and deep reinforcement learning (DRL), effectively enhancing the flexibility and autonomy of the scheduling system in complex and uncertain environments by constructing a digital twin model, designing a multi-feature observation space and multi-objective reward mechanism, and optimizing scheduling strategies with the PPO algorithm [92]. Yan Qi and Wang Hongfeng proposed a dynamic scheduling method that integrates Digital Twin (DT) and double-layer Q-learning (DLQL) to address uncertainties in the production process, using DT for real-time monitoring of virtual-physical differences and triggering rescheduling, while DLQL synchronously learns job and equipment-related decision-making strategies, thus improving the efficiency of scheduling responses [106]. Liu Zhifeng integrated digital twin and super-network technology to construct a feature-process-machine tool super-network technology to realize efficient association and clustering of multi-source heterogeneous data [107]. Yan Jun and Liu Zhifeng proposed an enhanced genetic algorithm that introduces redundant three-layer encoding and correction decoding mechanisms to address the often-overlooked transportation constraints in flexible job shop scheduling problems (FJSP), effectively enhancing the feasibility of scheduling results and the docking capability with digital twin (DT) systems [108]. Meanwhile, the multi-algorithm coupling strategy also demonstrates outstanding performance. For instance, Chen ZM, et al. applied an enhanced genetic algorithm (EGA) and particle swarm optimization (PSO) in series, obtained a near-optimal solution through the rapid convergence of EGA, and then conducted local and fine retrieval based on PSO, which significantly shortened the maximum completion time in flexible workshops [82]. For green manufacturing goals, most studies have expanded on traditional multi-objective evolutionary algorithms, focusing on optimizing control of factors such as carbon emission and energy consumption [83]. For example, by integrating reliability maintenance duration (RMD) and reliability processing coefficients (RPC) into workshop monitoring systems, a multi-objective scheduling model reflecting dynamic equipment reliability is constructed [109]. Wu Jiawei expanded the flexible manufacturing scheduling model, proposing a multi-objective dynamic partial reentrant mixed flow green workshop scheduling problem (MDPR-HFSP), and introduced a multi-agent proximal policy optimization algorithm (MMAPPO) to achieve dynamic resource and job scheduling through routing agents and sorting agents, integrating multiple scheduling rules and the Wasserstein distance restriction mechanism to enhance the stability of policy optimization and multi-objective adaptability [110]. Li Zhi proposed a flexible job shop scheduling strategy that integrates workers’ multi-memory behaviors (learning and forgetting) and digital twin technology, optimizing manufacturing span, carbon emissions, costs, and quality in dynamic situations such as machine failures, achieving green and efficient scheduling [111]. Zhou Zhuo and Xu Liyun proposed a scheduling strategy based on digital twin to address issues such as information opacity, response lag, and insufficient optimization capabilities in traditional scheduling, relying on a cloud-edge collaborative architecture to comprehensively optimize energy consumption, completion time, and delay metrics, and improving NSGA-II through multi-mode crossover and variable ratio elite retention mechanisms, validating the effectiveness and practicality of the method in standard datasets and actual workshop environments [84].

In the cloud-edge-end collaborative path, the cloud is responsible for global optimization and deep learning training, thus processing massive data, while the edge is close to the site and achieves millisecond-level control with low-latency autonomy. DT runs through the cloud and edge to ensure virtual-physical synchronization and closed-loop decision-making, thus combining high scalability with rapid response. The main challenges in this context include the high demands for network and computing infrastructure, latency, data consistency governance, and security protection. Applicable scenarios include smart factories with multiple production lines and equipment density requiring real-time optimization, achieving adaptive operation through “cloud computing and edge control”. The complementary multi-task multi-objective integrated path interacts within the same DT framework to link production, logistics, maintenance, energy/carbon emissions, and quality, achieving cross-functional consistency and real constraint integration; its difficulties arise from the solving and modeling challenges posed by problem scale and coupling complexity, as well as engineering implementation costs. This path is suitable for complex manufacturing (such as aerospace and lighthouse factories) that pursues system-level optimality and sustainability, achieving global trade-offs under multiple KPIs (efficiency, energy, delays, quality). Overall, the former addresses the issue of concentrated computing power combined with edge-side execution in seconds, while the latter resolves the problem of “multi-objective multi-link collaboration” [112]. In the actual production process, each algorithm has its unique advantages and applicable scenarios, and the most suitable algorithm should be chosen according to actual needs and system complexity.

6. Technology Adopted and Application Framework

This section reports on the selected literature primarily to address the following research questions:

RQ4: What system technologies based on digital twins are employed for dynamic workshop scheduling in workshops?

RQ5: What content is studied in research on the architecture of digital twin technology in intelligent plants?

6.1. What System Techniques Are Applied to Digital Twin?

In the digital twin system of a workshop, manufacturing perception technology is responsible for the real-time collection, transmission, and analysis of information such as manufacturing processes, equipment status, and environmental conditions in the physical world [113]. Sensor technology is the core of manufacturing perception, including environmental sensors, position and motion sensors, equipment status sensors, and image and video sensors. Through manufacturing perception technology, the digital twin system can perceive changes in the physical world in real-time [114]. IoT technology serves as a bridge for real-time data transmission, connecting sensors, devices, production systems, and information platforms to the network for intelligent data collection and transmission. Through IoT technology, operational data from factory equipment can be transmitted in real-time to the digital twin model, ensuring dynamic optimization of production scheduling [115]. Artificial intelligence (AI) technology and machine learning (ML) technology possess strong adaptive capabilities, allowing them to learn from historical data and optimize production scheduling, reducing manual intervention. AI can predict while analyzing potential production issues, assisting in scheduling decisions and enabling dynamic adjustment to optimize production efficiency [116]. Edge computing technology brings data processing capabilities closer to the device side, reducing cloud computing latency to improve response speed. Cloud computing and big data analysis technologies have powerful computing and storage capabilities, allowing for the simultaneous processing of massive amounts of data, enabling large-scale production optimization. For instance, the cloud can integrate digital twin models from multiple factories for global production scheduling and analysis. Virtual reality (VR) and augmented reality (AR) technologies provide intuitive visualization capabilities, helping managers and operators to view the production line status in real-time [117]. Programmable Logic Controllers (PLC) and Supervisory Control and Data Acquisition (SCADA) systems achieve automated management and scheduling of production line equipment [118]. Wireless communication technology ensures smooth data transmission, especially with 5G technology which, due to its high bandwidth and low latency characteristics, can support efficient data synchronization of large-scale devices and sensors, thus improving the flexibility and precision of production scheduling [119].

6.2. Digital Twin Technology Architecture

The core architecture of digital twin technology in smart factories connects physical devices with virtual models, collecting and processing data in real-time through sensor technology to construct virtual factory models and simulate production processes, equipment behavior, and data flows [120]. This architecture integrates multiple technology modules (such as IoT, cloud computing, big data, AI, automatic control, etc.) for unified management systems, thereby possessing cross-system and cross-platform data integration capabilities, providing intelligent decision support for production scheduling in digital twin workshops [121].

6.2.1. Research and Application Methods of Digital Twin Technology Architecture

Summarizing the recent research and literature, this paper categorizes digital twin architectures into two main types: the digital twin-driven dynamic scheduling workshop optimization framework and the digital twin-integrated intelligent manufacturing ecosystem framework.

In the digital twin-driven dynamic scheduling workshop optimization framework, this type of architecture focuses on achieving real-time scheduling optimization and operational collaboration in complex workshop environments. Liu Weiran, Zou Xiaofu et al. have constructed a DTPLSS framework for multi-level production logistics synchronization satellite large-scale assembly workshops using digital twins, integrating resource workstation workshop-level twin bodies with distributed control strategies to realize real-time coupling of assembly rhythm and material delivery, as well as disturbance self-adaptation [122]. For aircraft overhaul workshops, Liu Mengnan, et al. have integrated real-time work hour data with improved KDE methods to dynamically evaluate workshop capacity, providing a decision framework for maintenance planning and resource allocation [123]. At the scheduling level, Villalonga Alberto, et al. proposed a distributed decision framework that integrates local and global digital twins combined with fuzzy reasoning, in order to perceive and optimize equipment status and system capacity in response to the challenges of dynamic scheduling in cyber-physical production systems [124]. Pan, Jianguo proposed a digital twin-based hanging workshop management framework, combining graph neural networks and deep reinforcement learning (PPO) to effectively enhance scheduling decision-making capabilities and workshop operational performance [113]. Gu Wenbin and Duan Lianshui proposed a three-layer real-time scheduling framework based on digital twins combined with deep reinforcement learning (PPO) to realize adaptive dynamic scheduling optimization in mixed-flow workshops, enhancing decision autonomy [125]. Li, Yibing developed an anomaly detection and dynamic scheduling framework based on digital twins, integrating multi-level monitoring models and improved gray wolf optimization algorithms to achieve real-time optimization and deviation correction of scheduling plans [126]. Pandhare and others incorporated predictive health management (PHM) into twin scheduling, optimizing plans with genetic algorithms to maintain robust production rhythms even in equipment degradation scenarios [127]. The digital twin-driven dynamic scheduling workshop optimization framework is suitable for highly complex, frequently changing tasks and variable processes that need to respond in real-time to unexpected events in discrete manufacturing and assembly workshops, such as satellite batch assembly lines, aircraft overhaul workshops, or high-end equipment manufacturing production lines. Meanwhile, the implementation of this method is also challenging, requiring a complete perception network and high-precision digital model support, with algorithms such as multi-agent or deep reinforcement learning increasing system complexity and demanding high data quality and computing power [128]. In other words, deploying such frameworks in small factories with many customized devices and low data standardization may face integration difficulties; however, in complex workshops requiring flexible scheduling, real-time capacity assessment, and logistics synchronization, it can significantly enhance resource utilization and production resilience [129].

The digital twin-integrated intelligent manufacturing ecosystem framework aims to construct a unified digital twin ecosystem across processes and systems to achieve global optimization and lifecycle management, focusing on building a unified digital twin ecosystem [115]. Nie, Qingwei, and Dunbing Tang deployed a multi-agent cloud-edge collaborative digital twin framework at the virtual layer, allowing resource agents to coordinate in the cloud and respond quickly at the edge, achieving distributed production control [130]. Guo, Daqiang proposed the GiMS framework based on IIoT and digital twins, achieving synchronization of manufacturing and logistics at the information, decision, and execution levels through mixed-integer programming, effectively enhancing workshop operational coordination and decision-making efficiency [131]. In intelligent manufacturing workshops, targeting general intelligent manufacturing workshops across processes and lifecycles, Sun, Mengke proposed the digital twin intelligent manufacturing system (DT-IMS) framework based on multi-layer information system integration, constructing high-fidelity multi-dimensional twin models to achieve real-time collaboration of Intelligent Workshop processes, planning, and scheduling, effectively enhancing management efficiency and reducing workshop complexity, achieving a digitized closed loop of all elements including people, machines, materials, approaches, and environments [132]. Xia Luyao proposed the concept of a digital twin manufacturing ecosystem (DTME), establishing a cross-domain, multi-model intelligent factory DTS architecture, and validating its intelligent improvements in inventory reduction and delivery acceleration in hydraulic cylinder workshops [133]. At the same time, some reference models, such as the HexaSFDT six-dimensional workshop twin model and the hierarchical SDT model, provide standardized methods for constructing workshop twin frameworks from dimensional and hierarchical perspectives, emphasizing key elements such as information model standardization, high-performance data processing, and system interoperability [34]. Such system frameworks are suitable for large intelligent manufacturing enterprises or scenarios requiring cross-departmental collaboration, such as those needing to connect data across design, production, and operation phases or implement collaborative manufacturing across multiple factories and business units. In these environments [134], the digital twin ecosystem can serve as a key enabler, providing comprehensive perception and predictive analysis to ensure consistency in planning and execution under uncertain disturbances. Its disadvantages include complex implementation and high investment, requiring a complete industrial IoT and IT architecture support, the need to connect data standards from different source systems, and considerations for network security and data privacy issues [135]. Additionally, establishing high-precision all-element twin models is itself time-consuming and challenging, making it difficult for small- and medium-sized enterprises with weak information foundations to reap immediate benefits. However, overall, in smart factory scenarios pursuing global optimization and lifecycle digitization, such digital twin architectures can significantly enhance decision-making efficiency, resource allocation optimization levels, and rapid response capabilities to environmental changes [136].

6.2.2. Ideal Framework for Digital Twin Workshop Scheduling Based on Reinforcement Learning

Building upon prior research and incorporating the experimentally validated integration of dynamic scheduling and multi-agent reinforcement learning reported in “Multi-agent reinforcement learning based textile dyeing workshop dynamic scheduling method [77],” this paper proposes and develops a reinforcement-learning-based digital-twin shop-floor scheduling framework, as illustrated in Figure 9, which systematically displays its structure and functional relationships [137]. This digital twin ideal framework is an intelligent manufacturing architecture that integrates digital twin technology with reinforcement learning algorithms, presenting a dynamic collaboration mechanism from production instruction input to physical execution feedback and then to policy optimization in a closed loop [138]. The system first receives production orders, objective restrictions, and task resource allocation information from the manufacturing execution system (MES) or advanced planning and scheduling system (APS). The underlying physical layer is equipped with various industrial devices and sensors (such as PLCs, RFID, industrial cameras, visual recognition, and environmental monitoring sensors), achieving real-time data collection and feedback from the production site through communication modes, including 5G, Wi-Fi, Bluetooth, and industrial buses, covering multi-dimensional information such as equipment status, workpiece location, energy consumption information, and operator behavior [139]. The collected data undergoes preliminary pre-processing (data cleaning, noise reduction, format conversion, etc.) before being transmitted to the virtual-physical interaction layer, where the overall simulation and optimization module, based on the virtual factory, conducts manufacturing system modeling, production simulation, and scheduling optimization calculations [130]; meanwhile, the visualization module constructs a digital workshop view through 3D modeling and real-time rendering technology, achieving visual display of equipment operation, material circulation, and task execution processes, assisting decision-making and management intervention [140]. The core of the system is the digital twin and reinforcement learning integration layer. This layer constructs a multi-agent digital twin environment simulator, where multiple agents (such as scheduling agents, path planning agents, resource allocation agents, etc.) perceive, interact, and select strategies based on the current system state [141]. The system employs deep reinforcement learning algorithms for training and iterative updates of scheduling strategies and resource optimization strategies to maximize production efficiency, resource utilization, and response robustness [142]. During training, the system corrects the reward function and optimizes strategies based on actual feedback while continuously updating the precision of the twin model, achieving “perception, decision-making, execution, and feedback” closed-loop control [143]. Finally, the scheduling strategies generated by reinforcement learning are dispatched to the physical execution layer through the virtual-physical interaction layer, where controllers and devices complete manufacturing tasks, and the execution results are used as feedback for further training of the model and updating of the system state [144]. The system integrates multiple key techniques, including Industrial Internet of Things (IIoT), Edge Computing, 3D visualization, Digital Twin modeling, deep reinforcement learning, multi-agent collaboration, and adaptive strategy optimization, effectively addressing the challenges of task complexity, resource variability, and environmental uncertainty in discrete manufacturing processes, contributing to the development of an efficient, intelligent, and flexible dynamic manufacturing system [145].

7. Discussion

This section reports on the selected literature primarily to address the following research questions:

RQ6: What are the requirements for applying digital twins in dynamic intelligent workshop scheduling, and what are their future developmental trends?

7.1. The Role of Humans in the Digital Twin System

Humans play a crucial role in digital twin systems. Although digital twins possess high levels of automation and data-driven decision-making capabilities, human involvement is indispensable [146].

In the production process, human decision-making and planning involve decision-makers utilizing digital twin systems to obtain data and analyze results, thereby formulating long-term strategies [147]; for instance, predicting market development trends, optimizing resource allocation, and adjusting business objectives through digital twin models. In production planning and scheduling, humans manage production plans and adjust production line configurations through the digital twin interface to respond to unexpected situations or changes in demand [148]. Meanwhile, with the predictive maintenance capabilities of digital twins, personnel can recognize potential faults in advance and utilize real-time data for remote diagnostics and repairs [149]. Remote experts can also assist on-site operators in diagnosing and repairing complex equipment through augmented reality (AR) or virtual reality (VR) features of the digital twin platform. In production monitoring, humans interact in real-time through graphical interfaces, touch screens, and AR glasses to monitor production processes, view alerts, and data streams, ensuring stability in the production process. In quality control, personnel can analyze data streams using the digital twin system to identify process defects and propose optimization suggestions, thereby enhancing production quality [150]. Furthermore, the integrated platform of the digital twin system promotes cross-department collaboration, enabling data and information sharing, and facilitating coordinated work among production, maintenance, research and development, and supply chain departments [151]. In terms of engineering ethics, humans can ensure that factory operations comply with safety, environmental, and labor regulations through the digital twin platform and utilize simulations to ensure operational compliance. As digital twin systems become increasingly intelligent and automated, balancing technical development with ethical issues (such as privacy protection and accountability in artificial intelligence) will become an important task. In production safety, personnel also assume the role of risk management [152].

In digital twin systems, humans are not only operators and supervisors but also the “decision-makers” and “feedback providers” of the system [153]. Human intuition, judgment, and creativity are vital for the smooth operation of digital twin systems [154]. By collaborating with the digital twin platform, personnel can achieve refined management, optimize production processes, and improve work efficiency while ensuring safety and ethical compliance, thus enabling rapid responses to changes in complex and dynamic production environments [155].

7.2. Prospect of Future Development

The role of digital twins in the dynamic scheduling of smart factories is gradually shifting from a “simulation tool” to a “decision-making hub”. The modeling of manufacturing workshops is accelerating towards hybridization and intelligence, deeply coupling with specific workshop forms and problem scenarios. In the future, flexible assembly lines and operational workshops will be more suitable for multi-agent systems (MAS) and reinforcement learning; for knowledge-intensive tasks, data-driven strategies such as semantic models, GNN, and ML will help break down information silos and achieve cross-domain integration [156]; large-scale discrete manufacturing will still primarily rely on discrete event simulation (DES) supplemented by data analysis [157]. In scenarios with high system heterogeneity and modeling complexity, new methods such as dynamic fidelity reconstruction, low-carbon footprint assessment, green scheduling optimization [158], multi-scale micro-assembly modeling [159], mixed reality visualization scheduling [143], multi-layer aggregation models trained with truncated normal distribution [160], visual-ultrasound-3D collaborative twins [161], heterogeneous AAS three-dimensional fusion [162], and edge-fog-cloud three-tier twins [163] will significantly enhance the real-time performance and interpretability of models, promoting the transition of manufacturing modeling from isolated, static simulation to an integrated, intelligent, and highly adaptive digital twin system [164].

In terms of strategy and method selection, future emphasis should be placed on the matching of “scenarios and methods”. In scheduling with stable processes and clear constraints, classical optimization or heuristic algorithms should be prioritized for offline and rolling production scheduling to obtain optimal solutions; whereas, in highly dynamic scenarios, strategies with adaptive and online rescheduling capabilities should be adopted due to the ineffectiveness of static plans [60]. When traditional rules and deterministic algorithms struggle to address issues such as random arrivals and equipment failures, it is advisable to introduce swarm intelligence optimization and MAS to achieve decentralized collaboration and rapid response, combined with deep reinforcement learning (DRL) to enhance robustness and flexibility. In terms of integration paths, the combination of digital twins (DT) and DRL will significantly enhance the dynamic optimization capabilities of flexible workshops through virtual workshop training and real-time decision-making in actual factories, in conjunction with GNN/PPO; the combination of DT and MAS will rely on twins to conduct global assessments and visual interventions to improve yield and identify and alleviate bottlenecks early [151].

At the technical and architectural levels of DT systems, “cloud-edge-end” collaboration is leading the future direction, with the front end connecting through industrial IoT sensors/PLCs and IIoT gateways, combined with high-precision timing and streaming acquisition using OPC UA/Modbus and PTP/TSN, establishing a continuous, reliable, and traceable data foundation to support millisecond-level scheduling [165]; the cloud side aggregates data from multiple workshops and factories for global optimization and resource orchestration, with microservices + containerization + event bus achieving low-latency, scalable computing scheduling; under the support of 5G/industrial wireless, high concurrency and low-latency data backhaul and remote operation and maintenance are realized, supporting real-time collaboration across production lines and factories; simultaneously, in human–computer interaction, AR is used for on-the-job decision-making and visual guidance, while VR is used for panoramic simulation, production scheduling simulation, and training, with 3D visualization and interpretable analysis enhancing transparency and collaboration efficiency [166].

Looking ahead, the key focus of digital twin workshop scheduling should be on compressing decision-making delays, explainable reinforcement learning (RL), and strategy visualization, while incorporating energy consumption and carbon emissions into normalized optimization. This should be supported by high-fidelity digital twin models and strict data collection to ensure generalization and scalable implementation. At the same time, digital twin (DT) technology will continue to serve as a central platform integrating RL, deep reinforcement learning (DRL), multi-agent systems (MAS), bionic swarm intelligence, and cloud-edge collaboration, supporting more real-time, efficient, transparent, and green workshop operations [167].

8. Conclusions

This study provides a systematic review of scientific publications, focusing on the relationship between Industry 4.0, workshop production scheduling, and digital twins, particularly in relation to intelligent workshop scheduling models.

The findings of this review indicate a significant increase in academic research on dynamic production scheduling driven by digital twins. The number of related publications from 2019 to 2024 has risen year by year, with particular attention to digital twin-based job shop scheduling and green scheduling. The study also reveals that most of the relevant literature employs case studies to verify the proposed frameworks, with cross-sector studies being particularly common. Nevertheless, most studies lack validation in real industrial environments and assume that the implementation and real-time synchronization capabilities of digital twins are guaranteed.

Another notable finding is the variation in the definition of digital twins in the literature: some studies view digital twins as models covering the entire scheduling system, while others define them as independent real-time monitoring platforms. This discrepancy has led to a lack of unified understanding of the concept of digital twins.

This study further clarifies the various applications of digital twins in production scheduling, providing effective references for subsequent theoretical research and proposing several future research directions. Specific contributions include: (1) analyzing the application algorithms, modeling approaches, and application frameworks of digital twins in dynamic production scheduling; (2) conducting a systematic review of recent literature on digital twins, summarizing their innovations; (3) providing a research road map for dynamic production scheduling driven by digital twins for academia and practitioners; (4) discussing relevant manufacturing technologies that can collaboratively build workshop digital twins; (5) underscoring more innovative applications of digital twins in dynamic scheduling within intelligent factories, aiming to promote understanding and profound studies on digital twins and dynamic scheduling.

Despite achieving significant results, this study still has certain limitations: (1) the research scope focuses solely on production scheduling and does not address other planning issues in manufacturing; (2) the research materials are limited to articles published in indexed journals, papers, and conference proceedings; (3) research on scheduling technology that has real-time capabilities but is not explicitly identified as a digital twin is not included; (4) only literature from a specific time period was reviewed, and subsequent new research was not covered.

Despite the aforementioned limitations, this study followed a clear methodology and successfully achieved its intended goals.