Digital Twin Incorporating Deep Learning and MBSE for Adaptive Manufacturing of Aerospace Parts

Abstract

With the growing demand for diverse and high-volume manufacturing of composite material parts in aerospace applications, traditional machining methods have faced significant challenges due to their low efficiency and inconsistent quality. To address these challenges, digital twin (DT) technology offers a promising solution for developing automated production systems by enabling optimal configuration of manufacturing parameters. However, despite its potential, the widespread adoption of DT in complex manufacturing systems remains hindered by inherent limitations in adaptability and inter-system collaboration. This paper proposes an integrated framework that combines Model-Based Systems Engineering (MBSE) with deep learning (DL) to develop a digital twin system capable of adaptive machining. The proposed system employs three core components: machine vision-based process quality inspection, cognition-driven reasoning mechanisms, and adaptive optimization modules. By emulating human-like cognitive error correction and learning capabilities, this system enables real-time adaptive optimization of aerospace manufacturing processes. Experimental validation demonstrates that the cognition-driven DT framework achieves a defect recognition accuracy of 99.59% in aircraft cable fairing machining tasks. The system autonomously adapts to dynamic manufacturing conditions with minimal human intervention, significantly outperforming conventional processes in both efficiency and quality consistency. This work underscores the potential of integrating MBSE with DL to enhance the adaptability and robustness of digital twin systems in complex manufacturing environments.

1. Introduction

High-quality manufacturing of aerospace products is pivotal to advancing national defense security and economic progress. Advances in high-performance materials have spurred the rapid adoption of composite material systems, which exhibit distinct advantages over traditional metallic components. These include superior specific strength-to-weight ratios and elastic modulus, enhanced wear and corrosion resistance, as well as improved fatigue performance. Furthermore, composite materials provide critical lightweight design benefits without compromising structural integrity. Consequently, they are fast emerging as the preferred material for an expanding range of components in advanced aerospace applications [1,2]. The complex multistep manufacturing processes required for composite materials, despite their superior material and mechanical properties (e.g., high strength-to-weight ratios), pose significant technical challenges characterized by excessive labor intensity and heavy reliance on manual processes [3,4]. To address these challenges, smart manufacturing systems now leverage advanced technologies such as the Industrial Internet of Things (IIoT), real-time communication networks, cutting edge precision manufacturing techniques, and data-driven information processing algorithms. This integration fundamentally transforms conventional industrial production paradigms, enabling adaptive, flexible manufacturing of aerospace components while simultaneously promoting enterprise sustainability through optimized resource utilization and workflow automation [5].

The DT, a technology grounded in cyber–physical systems that enables real-time monitoring and analysis of physical processes by linking virtual models with sensor data, has emerged as a cornerstone in advancing smart manufacturing [6]. In response to the complex challenges inherent in composite material part production—characterized by large-scale manufacturing demands, highly variable product specifications, manual machining processes, and reliance on human inspection practices—this study proposes the development of an adaptive manufacturing system for aerospace parts based on DT technology. This system aims to address key industry pain points, including low production efficiency, elevated labor demands, and inconsistent part quality. While recent years have seen extensive theoretical exploration of DT applications in manufacturing, empirical adoption and implementation across enterprises remain uneven, underscoring the need for practical solutions tailored to aerospace production requirements [7]. Scholars have concentrated on modeling frameworks for DT applications in real-time monitoring, human–machine collaboration/interaction, and production planning [8]. The machining of composite aerospace parts constitutes a complex system involving multiple elements (e.g., personnel, machinery, materials, methods, and environment) and interconnected processes such as technical workflows, business procedures, material logistics, and control flows [9].

While prior studies have explored DT applications in manufacturing, three critical gaps persist:

(1)

System adaptability deficit: Current DT implementations predominantly focus on static system mirroring rather than cognitive adaptation [7,8];

(2)

Integration methodology gap: The synergistic potential between MBSE’s structural rigor and DL’s dynamic learning remains underexplored [10];

(3)

Due to a lack of holistically designing and evolving DT-based manufacturing systems through modular, structured frameworks frequently result in low reusability and system autonomy [11].

Although recent works by Hoang demonstrated DL-enhanced process monitoring, their reliance on predefined system architectures limits adaptability [12]. Conversely, Liu’s MBSE framework achieves structural clarity but lacks learning capabilities [11]. Our integration approach uniquely bridges these disconnected paradigms by constructing DT in machining systems, as shown in Figure 1. The key contributions of this research are outlined below:

(1)

A synergistic framework integrating MBSE-driven formal modeling with DL is proposed to address the unresolved tension between system rigidity and environmental adaptability in existing DT research;

(2)

A cognitive-driven adaptive feedback mechanism is developed to transcend the limitations of conventional DT systems constrained by passive one-to-one mappings;

(3)

The proposed DT system achieves high-precision machining of composite aerospace components and enables adaptive workflow execution, thereby facilitating a paradigm shift from laboratory-scale validation to industrial applicability in DT systems.

The remainder of the paper is organized as follows: Section 2 reviews existing work on deep learning-driven machining and DT. Section 3 details the agile design and modeling framework for DT using MBSE methodology. In addition, Section 4 presents a deep learning-driven adaptive digital twin and its evaluation methodology. Section 5 provides case studies demonstrating the benefits of the proposed framework in composite aerospace manufacturing systems, analyzing its performance metrics. Finally, Section 6 summarizes this research and outlines future directions.

2. Related Works

2.1. Machining Technology of Composite Material Parts

Composite material parts are characterized by large dimensions, low stiffness, and complex geometries, which inevitably induce residual stresses during manufacturing processes. These stresses can lead to part distortion or even component scrapping; thus, simulation plays a crucial role in composite material processing. For instance, Lupo et al. [13] specifically examined the parameter calibration of particle–particle interactions in Discrete Element Method (DEM) models. By adjusting model parameters and time steps, simulations enable the evaluation of adhesive forces between particles. In current workshop practices, manual trial-and-error machining is deemed the best approach to address such challenges [13]. However, this method becomes prohibitively expensive and time-consuming when scaling to mass production of multiple variants. Shifting toward automated manufacturing with real-time monitoring of workpiece status changes and predictive deformation analysis could mitigate investments in labor-intensive processes. Over time, the research focus has shifted toward applying artificial intelligence (AI) technologies in industrial settings [14]. Studies indicate that these approaches are key catalysts for transitioning traditional manufacturing systems to Industry 4.0. For instance, Liao et al. [15] proposed a method for manufacturing process monitoring by integrating time–frequency analysis with a deep neural network (DNN). During turning operations, acoustic emission signals were captured at various spindle speeds, feed rates, and cutting depths. Hoang et al. employed deep learning for defect detection through a three-step framework: data collection, feature extraction, and classification [12].

Beyond material-related causes, geometric variations in machining are further exacerbated by factors such as positioning errors, clamping methods, tool selection, and manufacturing processes. This complexity necessitates more robust manufacturing systems to counteract challenges [16]. Enhancing monitoring capabilities in production processes represents a significant step toward achieving cost-effective development and manufacturing of composite materials.

2.2. MBSE and Construction of Digital Twin

System Engineering (SE) is an organizational management technology encompassing both text-based SE and model-based SE (MBSE) [17]. MBSE represents an evolution from traditional text-based systems engineering, describing and modeling systems in digital and graphical forms rather than textual descriptions [18]. Leveraging MBSE to decouple complex multi-level relationships and address diverse adjustments is feasible. The V-model summarizes the workflow of MBSE, employing a top-down approach for forward design and bottom-up validation. By efficiently realizing complex system architectures, MBSE provides a foundation for constructing a digital twin [10]. Simultaneously, DT offers a virtual environment for SE models to enable faster modeling and simulation, thereby reducing system design costs [11]. A shared objective between DT and MBSE is to automate the transition of virtual models into unified representations [19]. Tong et al. demonstrated that MBSE enables interconnected and co-evolving DT across different application scenarios in manufacturing systems, implemented in a hierarchical, structured, and modular manner [20]. MBSE builds DT through clear requirement descriptions and intuitive design diagrams, which not only avoids errors in the top-level design phase but also establishes a standardized repository of DT models.

Table 1. A summary of the MBSE carried out in DT.

Classification	Representative Works	Key Techniques	Objectives
System optimization-oriented	[20,21]	SysML, Enterprise Architect	Facilitate cross-departmental collaboration
Unity of representation-oriented	[19,22]	SysML, system design, MBD	Solve the problem of inconsistency between actual and DT theoretical simulation
System development-oriented	[10,23,24]	SysML, MagicGrid, transient simulation	Create hierarchical, structured, and modular DT systems to reduce design costs

summarizes the work of performing MBSE in DT.

The MBSE methodology comprises three core components: modeling language, methodologies, and tools. SysML is an object-oriented paradigm extended from a subset of UML 2.0 [25]. It visualizes system structures, behaviors, requirements, and mathematical models applicable to SE modeling for systems of any scale incorporating hardware, human factors, and equipment. The MagicGrid [26], a novel methodology for MBSE, is tool-agnostic and fully compatible with SysML. Liu et al. proposed a modular development solution based on MagicGrid to construct workshop DT tailored to customized requirements [11]. Zhang et al. [27] adopted MagicGrid to propose a standardized, hierarchical, modular, and generic architecture for describing industrial robot DT comprehensively and dynamically. The methodology facilitates rapid system development of DT in a hierarchical, structured, and modular manner across diverse application scenarios. However, most MBSE-based DT applications focus on preliminary implementation of the system rather than deeper integration of modeling and real-time adaptive capabilities.

3. Building an MBSE-Based Digital Twin for Adaptive Manufacturing

The modeling methodology of MBSE stems from the distillation of human knowledge in systems engineering, representing critical foundational principles for constructing complex systems. In this study, MagicGrid, EA (Enterprise Architect), and SysML are adopted as the methodology, tool, and modeling language framework for MBSE-driven DT modeling. These are leveraged to support systematic activities, including requirements definition, design, analysis, verification, and validation. Since the feasibility of DT’s incremental modeling from the problem domain to the solution domain has been well established in the prior literature, this study selectively focuses on critical modeling steps within the workflow. The emphasis is placed specifically on adaptability- and collaboration-related model construction, which constitutes both the core research contribution of this paper and a current gap in the existing literature.

3.1. Creating a Digital Twin Requirements Model

The construction process of the DT requirements model is divided into two stages. The first stage involves analyzing interactions between the system and specific stakeholders, gradually finalizing the preliminary design of its internal composition and functional architecture. Subsequent work focuses on further detailed analysis and design to elaborate the precise expected behaviors and conceptual structures of DT. As shown in Table 1, this phased approach ensures systematic progression from high-level requirements to detailed specifications.

Table 2. DT stakeholder needs.

1. Technological needs
1.1. High-precision data synchronization
	Requirements background: Composite parts processing involves positioning (±0.1 mm), processing (±0.1 mm), measurement (±0.07 mm), and other links the system needs to ensure that the data of each subsystem are consistent in real time.
	Stakeholders: Process engineers (quality consistency), quality control department (reduce rework).
	Specific requirements: Cross-subsystem data delay ≤ 1 ms, key indicators (e.g., positioning accuracy) error thresholds can be configured.
1.2. Dynamic adaptive
	Requirements background: Tool wear, environmental vibration, and other factors during machining require a system to adjust parameters in real time.
	Stakeholders: Production supervisor (efficiency improvement), equipment maintenance team (reduce downtime).
	Specific requirements: Adaptive adjustment response time ≤ 0.5 s, support at least 10 kinds of abnormal working conditions automatic processing.
2. Business needs
2.1. Scalability
	Background: The system needs to be quickly adapted to new part models or machining processes in the future.
	Stakeholders: R&D department (new technology integration), marketing department (shorten time-to-market).
	Specific requirements: Integration cycle for new part models or machining processes ≤ 2 weeks, API interface standardization.
2.2. Human–computer interaction
	Requirement background: The system needs to provide a visual monitoring and operating interface.
	Stakeholders: Process engineers and frontline operators (ease of operation).
	Specific requirements: The system displays product status online, and the system recommends adjustments for manual confirmation (e.g., emergency shutdown privileges).
3. Management needs
3.1. Cross-sectoral collaborative support
	Requirement background: Manufacturing, design, and QA departments need to share digital twin data.
	Stakeholders: Cross-departmental managers (process optimization), IT department (data security).
	Specific requirements: Hierarchical management of permissions (design data can only be modified by R&D department), audit log retention ≥ 6 months.
3.2. Data security
	Requirement background: Processing parameters and design drawings involve trade secrets.
	Stakeholders: Legal department (intellectual property protection), information security team.
	Specific requirements: Data transmission encryption (AES-256), two-factor authentication for access rights.

illustrates the needs of the DT stakeholders.

Through the identification and refinement of stakeholder requirements for manufacturing systems, a clear and consistent functional description of the DT system is established. Utilizing requirement diagrams within MBSE enables a user-friendly approach to building the DT system’s requirements model, as demonstrated in Figure 2.

3.2. System Context for Creating Digital Twin

The contextual framework of the DT system is then established, encompassing four key dimensions: manufacturing plant (work orders), manufacturing environment (users, physical conditions, and regulatory constraints), business operations (subsystems and processes), and adaptive manufacturing. The system environment is modeled using block definition diagrams (BDDs), as illustrated in Figure 3.

3.3. Digital Twin Use Cases

To meet the demands of mass-personalized manufacturing for material parts, the processing system must first possess part characteristic recognition and classification capabilities. Subsequently, based on these requirements, the digital twin system’s adaptive manufacturing capability is established. Within Model-Based Systems Engineering (MBSE), system use case diagrams are leveraged to model the specific scenarios of each use case, as illustrated in Figure 4.

To enable data-driven operations in the physical manufacturing system, data acquisition devices are employed to collect status monitoring data. After data are transmitted via an object stream to the learning unit, they are immediately stored. Subsequently, an artificial intelligence algorithm model reads the status data and outputs model inference results, which are then utilized by the manufacturing equipment control program to update its operational parameters. If no special conditions exist in the digital twin system, after the predefined Data Reading Rate time expires, the system will execute the next iterative cycle of data collection, parsing, allocation, and updating for scenario actions (Figure 5). Conversely, if the “Close” condition evaluates to “true”, the behavior-synchronized use case terminates.

To capture the dynamic interactions between the physical system and its digital twin, system dynamics modeling is applied to represent key feedback loops and time-dependent behaviors. To formally model these interactions, activity diagrams are adopted to represent the workflow of each use case scenario. Specifically, Figure 5 illustrates the formalization of the “State Transition” use case through this approach, ensuring precise synchronization between physical and virtual system operations.

Functioning as a hybrid-driven simulation framework, the DT model enables the prediction of system-level behaviors. This integration leverages Model-Based Design tools and open-source libraries (e.g., MATLAB and MWorks), incorporating internal scripting and external function interfaces to implement methods and simulate static models. Scripts written in C or Python are executed within software environments, while external function interfaces define code-based linkages between the software and external files. A complementary model repository may further enhance simulation reliability by providing validated modular components.

4. Deep Learning-Driven Adaptive Systems

In this section, hybrid data–physical-driven technologies are further integrated into the framework established in the previous section to enhance the autonomy of the DT system, particularly in access control, real-time monitoring, and change analysis. First, the DT system incorporates a data-driven deep learning methodology to achieve functionalities such as object identification, spatial localization, and process quality perception. Subsequently, leveraging the advantages of a physics-driven approach, a method is developed to extract geometric features of worn cutting tools, and a set of geometric parameters representative of tool wear states is proposed. These wear characteristic parameters are then combined with machining and structural tool specifications to improve process quality, thereby driving system optimization through adaptive adjustments.

4.1. Adaptive Modeling

The physical manufacturing system requires not only fundamental machining capabilities but also the ability to swiftly identify product quality during production. This enables adaptive response to tool wear caused by frequent, high-intensity material fragmentation (a common consequence of intensive machining processes). By invoking corresponding machining programs and rapidly locating product machining references, the system can shorten alignment time and enhance overall production efficiency. As illustrated in Figure 6, this approach is embedded within the logical architecture of the DT system.

4.2. Deep Learning-Based Cognitive Reasoning

4.2.1. Identification and Localization

To achieve more efficient feature extraction, a YOLOv8-based algorithm is employed to leverage both spatial information from low-level features and semantic information from high-level features. For part boundary box prediction, an anchor-free object detection method is adopted, directly predicting the center coordinates, width, and height of target bounding boxes. Specifically, for each grid cell on the feature map, the model predicts box attributes—center coordinates (x, y), width w, and height h. Predictions are transformed via activation functions into values within the range (0, 1). For center coordinate prediction,

$$\widehat{x}=\sigma \left(p_x\right)+i,\widehat{y}=\sigma \left(p_y\right)+j$$

(1)

where σ denotes the activation function. Next, predicted width and height values are scaled using the following:

$$\widehat{w}=e^{p_w}\ast s_w,\widehat{h}=e^{p_h}\ast s_h$$

(2)

where p_w, p_h are raw predictions and s_w, s_h are scaling factors. A task-aligned assignment strategy is introduced to synergize classification and regression tasks during training. This strategy evaluates bounding box quality via the Anchor Alignment Metric, computed as follows:

$$M_{aligm}=C_{score}\ast IoU\left(p_{box},g_{box}\right)$$

(3)

where C_score is the classification confidence score, p_box is the predicted box, and g_box represents the ground-truth bounding box. The total loss function Total_loss is defined as follows:

$$Total\_loss=VFL\_loss+{\lambda }_1\ast CIoU\_loss+{\lambda }_2\ast DFL\_loss$$

(4)

where λ₁ and λ₂ are hyperparameters balancing the contributions of classification and regression losses, minimized via backpropagation to optimize detection accuracy. Where VFL_Loss is as follows:

$$L_{vfl}=-{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^C\left[y_{i,j}\log \left(\frac{{\displaystyle {\sum }_{k\in G}{w}_k{y}_{i,j,k}}}{{\displaystyle {\sum }_{k\in G}{w}_k}}\right)\right]}}$$

(5)

where n is the batch size, C denotes class count, y is the one-hot encoded ground-truth label, G represents participating nodes in federated learning, and w_k is their weights. The prediction y is as follows:

$${\widehat{y}}_{i,j,k}={\displaystyle \frac{\exp \left(z_{i,j,k}/\tau \right)}{{\displaystyle {\sum }_{l\in G}\exp \left(z_{i,j,l}/\tau \right)}}}$$

(6)

where z is the model’s output and τ is a temperature parameter. CIoU_loss is the improved IoU loss function with the following detailed expression:

$$\begin{array}{c}CIoU\_loss={\displaystyle \frac{Pre{d}^{adjusted}}{IoU}}+\alpha \cdot v\\ Pre{d}_{area}^{adjusted}=Pre{d}_w^{adjusted}\times Pre{d}_h^{adjusted}\\ IoU={\displaystyle \frac{Pre{d}_{area}\cap G{t}_{area}}{Pre{d}_{area}\cup G{t}_{area}}}\\ v={\displaystyle \frac{4}{{\pi }^2}}{\left(\arctan {\displaystyle \frac{w_g}{h_g}}-\arctan {\displaystyle \frac{w}{h}}\right)}^2\\ \alpha ={\displaystyle \frac{v}{1-IoU+v}}\end{array}$$

(7)

where Pred is the model-predicted bounding box, Gt is the aspect ratio of ground-truth boxes, and v is the compensation term for IoU. The parameter α balances area overlap and penalization terms. The DFL_loss is as follows:

$$DFL\_loss={\displaystyle \frac{1}{N}}{\displaystyle {\sum }_{i=1}^{N}f\left(y_i,{\widehat{y}}_i\right)}$$

(8)

where N is the number of samples, y_i and hat{y_i} and denote the true and predicted labels of the ith sample, respectively, and f (·) is the dynamic focal function, which is expressed as follows:

$$f\left(y_i,{\widehat{y}}_i\right)={\left(1-{\widehat{y}}_i\right)}^{\gamma }{y}_i\log \left({\widehat{y}}_i\right)+{\widehat{y}}_i^{\gamma }\left(1-y_i\right)\log \left(1-{\widehat{y}}_i\right)$$

(9)

where γ is the dynamic focal factor.

4.2.2. Data-Driven Process Quality Perception

Process quality perception forms the foundational function of a cognitive system. Such perception generally involves collecting multi-source, heterogeneous, and time-series data via sensors. For the extracted target part contours, deep learning-based feature recognition and classification are performed. This paper leverages process quality perception to rapidly infer tool wear severity and autonomously update the selection of optimal cutting tools.

A tool wear detection model for machining was first constructed using VGG19, with the network architecture shown in Figure 7. The neural network extracts features through six convolutional layers, stabilizes the training process using batch normalization layers, and reduces dimensionality via max pooling layers to extract primary feature information from images. Subsequently, the network implemented 10 alternating convolutional and batch normalization layers while maintaining feature map spatial dimensions. The architecture gradually deepened to extract higher-order features step by step.

During Backward Propagation in the neural network, the momentum ratio was set to α, allowing current gradient updates to appropriately reference prior velocity. This effectively prevents drastic changes in weights w:

$$\Delta w_{t+1}=-\eta {\displaystyle \frac{\partial E}{\partial w_t}}+\alpha \Delta w_t$$

(10)

where η denotes the learning rate, and E represents the loss function. For regularization, Weight Decay was adopted to mitigate overfitting in deep neural networks. Assuming the original loss function as E₀, the regularized loss function E incorporating L₂ regularization is as follows:

$$E=E_0+{\displaystyle \frac{\lambda }{2}}{\displaystyle \sum w_i^2}$$

(11)

where λ is the penalty factor. Taking gradients of this equation yields the following:

$$\Delta w_{t+1}=-\eta {\displaystyle \frac{\partial E_0}{\partial w}}-\lambda \eta w$$

(12)

This shows that L₂ regularization tends to shrink weights toward zero, promoting equitable treatment of inputs and preventing excessive amplification of any single feature. When combining both momentum and Weight Decay, the gradient becomes the following:

$$\Delta w_{t+1}=-\eta {\displaystyle \frac{\partial E}{\partial w_t}}+\alpha \Delta w_t-\lambda \eta w_t$$

(13)

Following these layers, three fully connected layers (FC) were connected, and dropout regularization was applied three times, as illustrated in Figure 7. The FC layers map convolutional features to the classification result space, while dropout randomly discards a portion of neurons during training. This enhances generalization by reducing overfitting, ensuring the model performs effectively on unseen data. The final output provides the category information of the target part in the input image.

4.2.3. Physically Driven Wear Mechanism of Machining Tools

During material removal, machining tools undergo continuous impact loads that induce micro-chipping and other wear mechanisms. These processes cause significant fluctuations in the blunted radius. Meanwhile, the outer edge of the cutting tool’s chisel edge experiences less fluctuation in blunted radius compared to its central region due to differential cutting speeds and the abrasive action of hard particles (e.g., fibers) on the edge. Initially, the cutting edge radius (CER) increases as the cutting edge blunts until it loses sharpness and cannot effectively sever fibers. Unsevered fibers then contribute to abrasive wear on the flank surface, forming a pronounced wear land (flank wear band). This process is analogous to re-sharpening the cutting edge, resulting in a subsequent decrease in the CER (Figure 8).

The CER and flank wear width (VB) of drill bits were evaluated to characterize the blunting degree of the cutting edge via CER and flank wear extent via VB, as shown in Figure 9. These parameters exhibited corresponding variation patterns, revealing the mechanism by which drill bits transition from blunted edges to re-sharpened states after re-sharpening.

The radial changes in the cutting edge’s CER and the maximum flank wear width (VBmax) are thus commonly used as quantitative metrics for milling tool wear. During CFRP drilling, primary cutting edge wear predominantly occurs at the cutting edge and rake face. The volumetric wear rate ν induced by an individual carbon fiber in the workpiece on the primary cutting edge can be expressed as the superposition of normal stresses acting on both the cutting edge and the rake face. Furthermore, based on the contact conditions between carbon fibers in the workpiece and the primary cutting edge, the number of carbon fibers contacting the primary cutting edge depends on its length. Consequently, the total volumetric wear rate ν of the primary cutting edge caused by all carbon fibers in the workpiece can be modeled as follows:

$$v=(k_1{p}_1+k_2{p}_2)L$$

(14)

where k₁ and k₂ denote the wear coefficients for the cutting edge and rake face, respectively, p₁ and p₂ represent the normal stresses acting on the cutting edge and rake face, respectively, and L is the length of the primary cutting edge. By integrating experimental studies on cutting forces and tool wear morphology, a correlation between tool wear rate and geometric parameters of worn tools (e.g., edge rounding, flank wear width) was established. This relationship will be leveraged to optimize tool geometry and cutting parameters for enhanced durability in CFRP drilling.

4.2.4. Cognitive Reasoning of Wear Analysis

Combining the wear mechanisms described above, correlations between tool wear rate and geometric parameters (e.g., CER, VB) were derived. The tool wear rate was calculated using a predictive model that incorporates geometric parameters, workpiece material properties, and calibration coefficients. This cognition not only achieves precise wear prediction but also enables analysis of geometric parameter effects on tool wear, thereby providing guidelines for optimizing tool geometry (e.g., selecting angles and radii) to minimize wear. To capture the temporal evolution of system states, a state machine diagram is integrated into the MBSE framework. The system transitions between “Data Acquisition”, “Model Inference”, and “Emergency Brake” modes based on real-time confidence scores from the AI model.

Figure 10 depicts the SysML-compliant state machine governing the tool parameter optimization loop. When tool wear confidence drops below 0.8 (calculated via Equation (14)), the system enters DiagnosticMode to perform wear-cause analysis and parameter adjustment, ensuring continuous quality improvement through physics-guided optimization.

4.3. System Effectiveness Metrics

In systems engineering, the Measure of Effectiveness (MoE) quantitatively characterizes a system’s performance through numerical values. In this study, MoE was modeled as a parameter of the DT, enabling system evaluation from both functional and performance perspectives to measure critical indicators, as depicted in Figure 11.

(1)

Experiment Setup: An adaptive manufacturing testbed for composite material cable cover production was established;

(2)

Validation of Critical Metrics:

(a)

Flexible fixture repeatability positioning accuracy: ±0.1 mm;

(b)

Machining efficiency improvement via multi-spindle heads: over 50% enhancement;

(c)

Digital measurement accuracy in hole spacing: ±0.1 mm.

(3)

Cognitive Collaboration Effect: The system demonstrated adaptive optimization capability in dynamic operational environments.

5. Case Study

Aircraft cables are typically mounted along the outer surfaces of airframe components and subjected to external aerodynamic forces and thermal loads during flight. To ensure their reliable operation and prevent electrical system failures, cable covers are employed to protect onboard cables. In this case study, engineering experiments were conducted to validate a DT implementation workflow developed. This methodology was applied to enhance the manufacturing process of cable cover assemblies, demonstrating its effectiveness in improving design-to-production integration and reducing defects through real-time digital feedback.

5.1. System Integration

The system integration encompasses software integration and hardware integration. In the software domain, the DT model and physical processing system were integrated with SysML models via a communication mechanism that unifies SysML, deep learning algorithms, and external function interfaces. An intermediate database was implemented to store communication data, enabling seamless integration by linking the interfaces of these platforms through this centralized repository.

For hardware integration, a LAN-based communication network was established to enable data exchange from diverse devices. Clients connected to servers using the OPC protocol for real-time interoperability. As depicted in Figure 12, hardware–software communication occurs as follows: sensor data are first stored in a buffer zone, undergo preliminary preprocessing, and are then transmitted to the DT. Concurrently, a subset of this processed data is stored as operational logs in the database for traceability and analysis.

As demonstrated in the construction process details presented earlier, within the modeling tool, engineers first preview and validate the proposed design. After verifying the completeness of the DT, a SysML model is generated to systematically encapsulate system requirements, design specifications, analysis parameters, verification protocols, and validation workflows. Subsequently, based on the defined architecture, corresponding modules were identified in MWorks and Unity3D, which were then retrieved from the model library, configured via drag-and-drop operations, and assembled into a virtual simulation scenario. Three-dimensional models were positioned precisely. Engineers assigned physical attributes (e.g., material properties) to components and implemented componentized scripts for modular functionality. Additionally, user interfaces were customized to enable seamless interaction with the virtual environment.

The workpiece is a thin-walled component with a thickness of 1.3–1.6 mm. Post-molding, it undergoes specialized machining processes (drilling and edge milling) to meet assembly and connectivity requirements. To achieve this, a multi-spindle machining experimental platform was developed, as illustrated in Figure 13. This setup ensures precise fabrication while accommodating the demands of complex geometries and material behaviors in thin-walled components.

In the DT system, a C#-based class library was developed to encapsulate encoding functions. The integration of Unity3D, MWorks, and SysML models is facilitated through an intermediate database, enabling seamless interoperability. The interfaces and data types were designed in accordance with predefined specifications within the overarching system architectural framework, ensuring consistency between theoretical design and implementation. These models exchange events, states, and data via message-passing protocols through the database—specifically, reading/writing messages to synchronize information.

Upon system initialization, the virtual experimental platform synchronizes with the physical entity’s operational states (e.g., hole machining positions and tool wear levels). Corresponding synchronization messages are logged into the intermediate database, while SysML models maintain the flexibility to incorporate additional specification details in future iterations.

5.2. Digital Twin System Development

To address the demand for automatic workpiece identification, a cable cover model recognition software was developed using PyQt5. The software interface is divided into three primary sections:

(1)

Camera Feed Display Area: Real-time visualization of the live camera feed;

(2)

Function Menu Bar: Controls for image capture, cropping, and recognition initiation;

(3)

Result Display Section: Outputs classification outcomes (see Figure 14 for a detailed interface layout).

Operational Workflow:

(1)

Image Acquisition: After placing the cable cover on the platform and launching the software, clicking the Capture Image button establishes real-time communication with the camera. The live feed is displayed in the Camera Feed Display Area;

(2)

Image Cropping: Once the workpiece is positioned within the red rectangular frame on-screen, clicking Crop Image captures the region of interest (ROI), generating a processed image for analysis;

(3)

Recognition Execution: Initiating the Start Recognition command loads both the preprocessed image and the trained neural network model. The software classifies the input, outputting the result to the Result Display Section in real time.

The software’s interface design and workflow are illustrated in Figure 14, demonstrating the seamless integration of user interaction with automated recognition processes.

5.3. Quality Perception for Cable Cover Machining

A total of 17 different types of cable covers were selected for the experiment. A Hikvision DS-E14a image acquisition device was used to capture images of cable covers. After acquiring the imagery, grayscale distribution histograms were analyzed to study how pixel intensity varied with lighting conditions. An optimal threshold value was selected based on grayscale distributions to segment the images. Subsequent morphological transformations, including hole filling and noise suppression, were applied to refine the initial segmented images. Edge detection was then performed using the Canny operator to extract target contours (see Figure 15 for comparative analysis).

The extracted cable cover contours were used to construct a contour recognition dataset for further processing. A VGG neural network was trained using this dataset, with parameters configured as

Table 3. Training parameter settings.

Name	Value
Input image size	640 × 480
Training rounds	100
Batch size	16
Initial learning rate	0.01
Periodic learning rate	0.01
Learning rate momentum	0.937
Weight Decay factor	0.0005

.

Figure 16 presents the parameter evolution during training, including curves for bounding box loss, segmentation loss, class loss, and confidence loss on both the training and validation datasets.

The results demonstrate that the class loss curve achieves optimal convergence, indicating successful model training in capturing subtle workpiece features. The iterative optimization process ensures high classification accuracy, enabling precise identification of manufacturing components.

5.4. Results and Discussion

5.4.1. Tool Wear-Adaptive Optimization

The vision-based quality monitoring system demonstrated robust convergence characteristics, with training loss stabilizing at 0.0072 (MSE) and classification accuracy reaching 99.59% by Epoch 40 (Figure 17). This temporal resolution proved sufficient for detecting critical burr formation events as small as 42 ± 8 μm, with detection sensitivity outperforming laser profilometry in sub-surface defect identification.

To compute the tool wear rate, geometric parameters of the cutting tool, workpiece material properties, and calibration coefficients were substituted into the predictive model. The accuracy of this model was validated through three experimental configurations, where rake angles, clearance angles, and apex angles of the cutting tool were systematically varied. These experiments were conducted during a 72 h continuous production run of 125 CFRP components, as illustrated in Figure 18, which compares the model’s calculated wear rates with experimentally measured values.

Beyond accurate wear prediction, this model enables analysis of the correlation between tool geometrical parameters and wear generation, thereby guiding engineers to select optimal geometrical configurations that minimize tool wear. Building on the theoretical research foundation, the tool structural parameters were optimized by integrating a micro-blade controllable arrangement milling cutter design methodology and the reverse shear drilling concept, leading to the development of specialized tools for drilling and milling operations.

The study systematically investigated variations in cutting forces, machining quality, spindle speed, and feed rate across different tool configurations. Key influencing factors affecting drilling quality were identified to establish optimal process parameters enabling high-quality machining of CFRP composites. Furthermore, process parameter optimization was extended to address variations in fiber orientation and resin matrix systems, with the reverse shear drilling concept achieving a 38% reduction in fiber pull-out. This advancement enables consistent, high-quality machining across diverse CFRP configurations.

To evaluate performance, continuous machining experiments were conducted using the developed tools in conjunction with optimized process parameters, and their performance was benchmarked against two internationally advanced milling tools.

The experimental results demonstrated significant improvements. The optimized tools exhibited a lifespan increase of over 1.8 times compared to their international counterparts. Achieved high-quality machining of CFRP with reduced material delamination defects, validating their suitability for industrial applications (

Table 4. Export quality of four tools drilling CFRP.

Workpiece Materials	Carbon Fiber Reinforced Polymer (CFRP)
Tool Type
Quality
Life of Cutting	4.5 m	37.5 m	82.5 m

summarizes quantitative comparisons). This breakthrough enables high-efficiency, long-life machining of CFRP components, which is critical for aerospace industries where precision and durability are paramount.

5.4.2. Discussion

To evaluate the relative merits of the proposed MBSE-DL framework, we benchmarked against three industry-relevant baselines:

(1) 

Baseline 1 (Manual System): Two senior process engineers independently diagnosed tool wear and adjusted parameters;

(2) 

Baseline 2 (Pure DL Model): A ResNet-34 trained end-to-end on raw vision data without MBSE constraints;

(3) 

Baseline 3 (Rule-Based System): A Drools engine implementing 127 decision rules extracted from historical process manuals.

All methods were tested on the same 125 CFRP cable covers under identical CNC conditions. Compared to baseline methods (

Table 5. Comparison of key metrics with baseline methods.

Metric	Proposed (MBSE-DL)	Pure DL	Rule-Based	Manual System
False Negative Rate	4.2%	42.0%	28.5%	51.7%
Parameter Adjustment Time	2.3 s	1.8 s	6.7 s	18.4 min
Uptime (72 h)	99.6%	95.1%	92.4%	89.3%

), the MBSE-DL framework significantly outperforms baselines in reliability (uptime) and safety (false negatives), albeit with slight latency trade-offs versus pure DL. This demonstrates the critical role of MBSE in balancing adaptability and traceability.

Unlike conventional DT systems that treat sensory data and physical models as separate layers [23], our cognitive–physical fusion enables causality-aware adaptation (see

Table 6. Comparison of key features with baseline methods.

Comparison Dimension	Manual Systems	Rule-Based Systems	Pure DL Models	Our
Mechanism	Phenomenon-driven	Rules-driven	Data-driven	Cognitive–physical synergy
Tool Optimization Paradigm	Direct replacement based on monitoring	Direct replacement based on threshold	Predictive replacement based on data	Pre-optimization based on process quality
Industrial Scalability	Difficult	Limited expansion (similarity rule)	Limited expansion (data sets)	Easily extendable

). The vision system’s burr morphology characterization directly informs the physical wear model, creating a bidirectional feedback loop that closes the ‘sense-decide-act’ gap in autonomous machining.

(1)

Cognitive–Physical Synergy Mechanism. The observed performance enhancements stem from the unique integration of visual cognition and physical modeling. The vision system’s high temporal–spatial resolution enables real-time burr morphology characterization, which correlates with tool wear state through our developed DT system;

(2)

Tool Optimization Paradigm Shift. Quality analysis of machining (Table 4) reveals that the tool structural parameters were optimized through the integration of a micro-blade controllable arrangement milling cutter design method and a reverse shear drilling tool design concept. This yielded specialized milling tools tailored for complex material processing, particularly critical for aerospace components [15];

(3)

Industrial Scalability Considerations. While demonstrating 99.6% uptime in controlled environments, field deployments revealed two key challenges:

• Ambient lighting variations caused 12% false positives in burr detection (mitigated through active IR illumination);
• When transitioning between 17 distinct CFRP components, the MBSE kernel maintained structural coherence while system constraints were automatically transferred through product family architecture mapping;
• During continuous 72 h production of 1250 CFRPs, the DT architecture demonstrated predictive replacement triggered wear threshold, eliminating sudden failures.

The MBSE kernel’s product family architecture enables constrained parameter inheritance across geometrically dissimilar components, as 62% of machining constraints (e.g., max spindle torque, minimum tool engagement) were successfully transferred from cable covers to unmanned aerial vehicle (UAV) battery hatches during pilot tests. This suggests that the framework’s core ontology can capture domain-invariant manufacturing knowledge, reducing reconfiguration efforts for new components. However, geometry-specific constraints (e.g., five-axis tool accessibility for deep cavity parts) still require case-by-case calibration.

6. Conclusions

This study establishes a DT framework that synergizes MBSE with deep learning, addressing three fundamental challenges in adaptive aerospace manufacturing:

(1)

Industrial Scalability: The MBSE kernel achieved automatic constraint transfer across 17 CFRP variants, resolving the “one-twin-per-variant” bottleneck in traditional DT systems;

(2)

Integration Methodology: By correlating vision-based burr morphology with tool wear physics models, the experimental results demonstrated significant improvements. The optimized tools exhibited a lifespan increase of over 1.8 times compared to purely data-driven benchmarks;

(3)

System Adaptability: In 72 h continuous production of 125 CFRP components, the DT framework maintained 99.6% uptime through predictive replacement triggered by wear thresholds via minimized unplanned downtime.

These findings advance DT theory by embedding first-principle physical causality into data-driven adaptation loops, offering a replicable template for high-mix manufacturing. While the current implementation focuses on aerospace composites, the constraint can be extended to other material systems. In future research, we will focus on a hybrid digital twin architecture that tightly couples physics-based simulations (e.g., FEM for stress distribution analysis, CFD for thermal behavior prediction) with data-driven models. Additionally, cloud manufacturing services will be leveraged to enable distributed collaboration on physics-data fusion models.