Chatter remains a persistent challenge in high-performance milling as it limits surface quality, tool life and productivity. Conventional detection methods based on external sensors or single-axis current signals often fail to capture the multidimensional dynamics of modern machining. This work advances the development toward sensorless chatter detection by incorporating bidirectional NC signal analysis in circular ramp milling. The new experiment enables simultaneous X- and Y-axis engagement, generating variable cutting force components and revealing cross-axis interactions that are not observable in uniaxial ramp tests. High-frequency NC signals from the position control loop are acquired via a Siemens Industrial Edge Device and evaluated using Principal Component Analysis (PCA) to identify dominant vibration modes and their axis-specific contributions. The results demonstrate that multiaxial motion enhances the observability of chatter phenomena, while PCA effectively compresses redundant information into a compact feature set that is suitable for future use at the periphery and thus for sensorless chatter detection. Beyond chatter detection, the approach provides deeper insight into dynamic tool-workpiece interactions and contributes to data-driven methods that support resilient, sustainable and AI-enabled manufacturing systems.
The arrangement of components within a robot cell is up to this point a mainly manual task. This paper presents a method for an automatic layout optimization of robot assembly cells in Visual Components using a genetic algorithm. The main goal of this contribution is to improve the cell layout of a robotic assembly cell automatically instead of manually changing it. In addition, a sensitivity analysis is conducted to evaluate different optimizers to present the most suitable parameter configuration for two differently packed robotic assembly cells. As expected, the quality and velocity of the optimization depend heavily on suitable parameters for the optimization method. While the quality of the results for the spacious cell remained virtually unchanged (~20% lower space utilisation, ~10% faster cycle time under the same conditions) and only the speed of approaching the optimum was improved by parameter optimisation, the difference was much more pronounced for the compact cell: the average optimisation was almost doubled.
Directed Energy Deposition (DED) as an additive manufacturing (AM) method offers great potential across various applications. However, challenges persist in controlling thermal history, grain structure, distortion, and residual stresses when manufacturing. Numerical simulation of the DED process enables prediction of these characteristics and supports first-time-right manufacturing. In most cases, computer-aided design (CAD) models of the post-machined component are used for toolpath planning and simulation. Since AM parts are produced layer by layer, their as-built geometry deviates from the idealized CAD model due to process-specific effects. This mismatch introduces errors in process simulations. To address this problem, a path-based geometry modeling tool is introduced that reconstructs a geometry closely approximating the as-built part, based on the actual toolpath and deposition bead dimensions. This model more accurately represents the final part’s volume and geometry. For thermal process simulation, material deposition is also modeled based on the same toolpath, establishing a unified workflow that uses standardized toolpath data for both geometry creation and thermal analysis. This ensures independence from specific machine types and enables early-stage simulation. The approach is validated through thermal simulations and experimental measurements, demonstrating high agreement and confirming that the proposed geometry modeling method improves the prediction of process parameters such as correct dwell times or interlayer temperatures.
Upper limb exoskeletons (ULEs) have shown significant potential in enhancing mobility and rehabilitation outcomes. However, pediatric applications pose unique challenges due to anatomical differences and the need for adaptable and comfortable components. This study develops an automated design-to-print optimization framework that combines simulation-informed design of experiments, physical testing, and artificial intelligence (AI) and machine learning (ML) to rapidly identify design and print parameters for 3D-printed exoskeleton joints. The workflow employs a two-stage design of experiments (DoE) approach: the first identifies optimal geometric parameters using finite element analysis (FEA), and the second optimizes print parameters through mechanical testing. The resulting data are used to train a multiple linear regression (MLR) model that predicts joint strength and print quality from design and manufacturing settings. Statistical analysis identified shaft diameter as the most significant design parameter and wall perimeter count as the dominant print factor influencing load capacity. The trained MLR model achieved an average prediction error below 0.10 MPa. This integrated workflow demonstrates how AI-assisted surrogate modeling can bridge virtual design and physical fabrication, enabling rapid, data-driven optimization of 3D-printed rehabilitation components.
Cyber-Physical Systems (CPS) are central to modern manufacturing, yet their increasing interconnectivity exposes production environments to sophisticated cyberattacks. Existing resilience approaches often remain fragmented, focusing on isolated detection or prevention measures without providing actionable guidance for controller-level stabilization and recovery. This paper contributes a controller-level taxonomy for cybersecurity resilience in production CPS, integrating detection, mitigation, and recovery requirements directly into the control loop. The taxonomy is rooted in dynamic systems principles, and it structures resilience in terms of two classes of tuples, namely, attack tuple, specifying adversarial logic through vectors, targets, methods, effects, and informing outputs; and defence tuple, comprising detection mechanisms, defence actions, objectives, redundancy, secure estimation, and resilient controllers. To demonstrate how the taxonomy guides controller design, we instantiate it through a Model Predictive Control (MPC) formulation that embeds resilience objectives into predictive optimization, constraint handling, and trajectory reconfiguration. Comparative results with a Linear Quadratic Regulator (LQR) baseline illustrate how taxonomy-informed adaptations enable MPC to maintain stability and enforce constraints under adversarial conditions. Together, these contributions establish a foundation for embedding resilience into controller design, enabling CPS to anticipate, withstand, and recover from cyber threats in real time.
Accurate short-term forecasting of active (P), reactive (Q), and apparent power (S) at the production-cell level remains largely unexplored, despite increasing electrification and the prevalence of non-linear, inverter-driven loads in modern manufacturing. These systems exhibit frequency transients and state-dependent dynamics that are not observable in aggregated factory-level data. This work proposes a high-resolution forecasting framework that combines dynamic Working/Standby classification with a benchmarking of Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM), Transformers, and Light Gradient-Boosting Machine (LGBM). The results reveal a horizon-dependent performance crossover: TCNs achieve state-of-the-art accuracy for short-term horizons (5 min, R² = 0.88) by capturing deterministic kinematic patterns, while LGBM demonstrates superior robustness at medium horizons (15 min, R² = 0.68) due to its resilience against stochastic production effects. By forecasting not only P but also Q and S, the framework enables predictive power-quality management, including proactive voltage stabilization and dynamic power-factor correction, supporting more reliable and energy-efficient Industry 4.0 microgrids.
Atomic force microscopy (AFM) enables nanoscale surface characterization but is fundamentally constrained by slow scan speeds and sensitivity to noise. Increasing the scan rate introduces severe structured artefacts, notably line-scar effects : horizontal streaks and discontinuities arising from feedback instabilities, actuator dynamics, or intermittent tip–sample interactions. These artefacts mimic or obscure real topographic features and cannot be removed using standard filtering without compromising genuine nanoscale detail. We present a deep learning reconstruction framework that restores high-fidelity AFM images directly from fast, noisy measurements. The model maps line-scar–corrupted fast scans to their clean, slow-scan counterparts using a UNet-based architecture with skip connections to preserve fine spatial structure while suppressing structured noise. A curated paired dataset of noisy–clean AFM images is assembled, and results show substantial improvements in image quality, edge sharpness, and artefact removal across diverse morphologies. Quantitative evaluation via Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and visual assessment confirms recovery of nanoscale features obfuscated in raw fast-scan data. This framework enables reliable, high-throughput AFM imaging without hardware modification and provides a practical pathway toward real-time noise mitigation and artefact correction in advanced manufacturing and materials research.
This review explores the integration touch points for emerging Generative AI (GenAI) tools in various aspects of the Additive Manufacturing (AM) process chain, emphasizing application-specific design and optimization strategies. Over the past two decades, AM has evolved into a transformative manufacturing paradigm, enabling the realization of complex geometries and multi-material constructs with tailored functional properties. AM techniques such as Fused Filament Fabrication (FFF) have particularly matured to a point of democratizing AM methods for end users. Concurrently, the emergence of advanced sensing technologies and automation has enabled the adoption of AI-driven methodologies across the AM workflow. Despite these advancements, the underlying algorithmic frameworks, training protocols, and tunability for different target application functionalities exhibit substantial heterogeneity. This paper aims to consolidate the literature reports into a concurrent framework toward a function-driven design approach. Specifically, the authors identify various touch points for AI/ML approaches in design, manufacturing, and part performance within various stages of the AM workflow. Furthermore, potential opportunities for emerging GenAI approaches within the proposed concurrent framework are also presented.
Shipbuilding is characterized by constant adjustments to the product during the production phase, which has already begun. The reasons for this range from changing technical specifications to the implementation of customer requirements. Due to the parallelization of design, planning, and manufacturing, schedules are often based on assumptions and rough estimates. Deadlines are therefore gradually specified and adjusted. This necessitates active production process control. The paper shows the dependencies between the various production areas and the challenges involved in data acquisition and ensuring the necessary data quality. Furthermore, the potential of digitally networked production is demonstrated using production-specific key figures. Practical examples are used to illustrate measures that result in a leveling of capacity utilization. At the same time, the complexity of holistic production control in shipbuilding becomes apparent.
Accurate demand forecasting is essential for effective inventory management, production planning, and supply chain decision making. Customer reviews have become an important data source in the past decade, offering additional signals that can improve forecast accuracy. However, most existing forecasting studies treat all reviews as equally informative and do not distinguish between verified and unverified purchases. This distinction is increasingly important, as online retailers now display the Verified Purchase indicator, and customers actively rely on it when assessing review credibility. This study evaluates whether Verified Purchase reviews, as an indicator of credible reviews, provide a more reliable basis for demand forecasting. Using Amazon review data for a kitchen appliance category, two XGBoost models were developed. The first model incorporates all reviews together with historical sales, while the second model includes only Verified Purchase reviews with historical sales. In addition, a Prophet model was built to forecast demand using only historical sales data. The XGBoost models outperform the Prophet model, confirming the value of incorporating review information. The model based solely on Verified Purchase reviews shows substantially lower forecasting error compared to the model that incorporates all reviews. To quantify the operational impact of this improvement, a discrete event simulation of an (R, Q) inventory system was used to estimate holding and shortage costs under each forecasting approach. The results show that models using Verified Purchase reviews consistently deliver higher forecast accuracy, leading to significantly lower inventory costs. Overall, the findings demonstrate that incorporating review credibility improves forecasting accuracy and provides meaningful economic benefits for inventory management.
Effective personnel allocation is a crucial factor for improving productivity in back-office operations. It is essential to consider both ‘employer satisfaction’ from an organizational perspective and individual ‘employee satisfaction.’ It is also necessary to consider both objective indicators, such as employee aptitude and proficiency, and subjective indicators, such as their placement preferences. Furthermore, employee proficiency improves through experience, and considering this factor makes it possible to plan for employee promotions. This study proposes a personnel allocation planning method that considers employee proficiency. In the proposed method, the multi-objective optimization problem, which aims to maximize both ‘employer satisfaction’ and ‘employee satisfaction’, is formulated as a single-objective function. The effectiveness of the proposed method is verified through computer experiments’ results.
The transition towards circular production challenges the assumptions of traditional production planning and control systems, which were designed for stable, linear supply chains. By contrast, remanufacturing is characterised by irregular product returns, uncertain quality conditions, and variable process routes. Such dynamics necessitate planning mechanisms that are adaptive and continuously updated. This paper presents a Digital Twin–enabled framework for capacity planning in remanufacturing production systems. By connecting tactical and operational planning horizons, the framework maintains synchronisation between planning and execution through information integration and event-triggered orchestration. The paper proposes a multi-time-frame digital twin architecture to transform classical horizon planning into a feedback-driven system that can adapt to stochastic disturbances in real time. This concept establishes the basis for data-driven, event-responsive capacity planning, thereby supporting the development of resilient, resource-efficient remanufacturing operations.
The diversity of tools and processes in modern turn-mill centers makes it difficult to deploy consistent process control across all processes. This paper presents an IoT-based system architecture for process monitoring through sensorized clamping devices, exploiting the clamping system as a process-agnostic measurement location. The architecture comprises a newly developed sensorized live center, an MQTT-based communication infrastructure unifying heterogeneous sensor nodes, and flexible machine control integration via digital I/O gateways or industrial edge devices. Edge computing on the sensor nodes reduces data volume. Experimental investigations demonstrate the system’s capability and show excellent temporal stability of the wireless data acquisition and, thus, confrming reliable operation in the challenging machine tool environment. The proposed architecture provides a low-threshold entry point for SMEs seeking to adopt sensor-based process monitoring without extensive IT infrastructure investment.
Cold spray is an emerging additive manufacturing technology with great potential in the solid freeform fabrication of parts with complex geometries. While cold spray manufacturing with conventional converging-diverging nozzles has had success with softer metals, it has struggled to produce quality films and coatings from hard metals or ceramics. Vented nozzle designs, recently developed for micro-cold spray applications, offer a potential solution to this issue, as they allow for significantly less abrasive, finer powders (e.g., 1-5 µm) to be used. Vented nozzles incorporate pressure relief channels that enhance particle deposition velocity and enable deposition at comparatively lower temperatures than converging-diverging nozzles (i.e., from 450↑C to 200↑C). However, a comprehensive profile characterization of vented nozzle depositions in cold spray is lacking in the literature. As such, this work fills this gap by presenting the first comprehensive investigation of the profile characteristics of vented nozzles using high-purity niobium powder. To study the profile of vented nozzle depositions, a full factorial design of experiments was employed to evaluate the effects of the number of passes and standoff distance, two critical features in cold spray shape control analysis. After depositing track lines for each set of process parameters, the cross-sectional profiles were characterized using optical profilometry and fit using exponential-family distributions. Results demonstrate that vented nozzle depositions exhibit characteristic Gaussian profile distributions consistent with conventional cold spray, suggesting that established modeling frameworks remain applicable despite fundamentally altered gas flow dynamics. These preliminary findings are promising as they suggest that vented nozzles can be more readily integrated into cold spray additive manufacturing applications, leveraging lower-temperature deposition while providing a pathway into broader material compatibility with hard metals and ceramics.
Virtual commissioning (VC) is a method to verify and validate manufacturing systems by combining simulations, emulations, and hardware components. In robotic bin-picking applications, machine vision systems are key components but are often excluded from the VC process. The challenges are the complexity of creating virtual models of vision systems and integrating them with the overall simulation. As a result, the virtual commissioning of bin-picking systems relies on trial-and-error approaches. In this paper a method has been presented for virtual commissioning of robot vision system in bin picking applications. Digital twin of the robot vision system has been created in Tecnomatix Process Simulate Software. In this work, the digital twin represents a high-fidelity simulation model of the robotic vision system used for virtual commissioning. A virtual camera was mounted on the robot to create virtual vision for the robot. Deep learning algorithm was then applied on the virtual images captured by the virtual camera for detection of parts. The successful implementation of this method may help in virtual commissioning of the robot vision system for bin-picking application. This will help in robot pose estimation and path planning as well as error detection prior to physical commissioning.
Large metal plates, such as those utilized in the shipbuilding industry, frequently exhibit undesirable deformations during manufacturing steps or the storage process due to their substantial self-weight. These deformations can, on occasion, be so significant that seamless further processing of the plates becomes unfeasible. Such cases require complex work steps, resulting in additional costs due to labor and delays. Ideal support or storage conditions can mitigate these negative effects. However, they cannot always be realized due to the large number and dimensions of the plates. Consequently, numerical analyses can be employed to predict the deformations that occur in relation to specific support conditions. These analyses can then provide an optimal storage recommendation that can be implemented in accordance with the actual on-site possibilities. The geometry of the investigated plates was captured using a three-dimensional (3D) laser scanning device. After adequate preprocessing steps and reverse engineering of the data, the geometries were integrated into a deformation analysis framework based on Isogeometric Analysis (IGA). The intention of using IGA is to exploit its advantages when modeling the plate geometry, such as its accurate and smooth representation. While IGA as well as classical and ML-assisted optimization are established individually, there has been no systematic exploration of their combined application to support condition optimization. The influence of different support conditions on the deformation of the plates is investigated and demonstrated.
Cooperative 3D printing has recently emerged as a scalable extension of traditional additive manufacturing, enabling substantial gains in both makespan and build volume through the use of multiple robots. However, coordinating several printers within a shared workspace introduces the critical challenge of preventing inter-robot collisions. While prior work has largely focused on computationally intensive methods, such as optimal partitioning and coordinated toolpath generation, these approaches can be difficult to deploy in real-time. We introduce a lightweight, threat field–based method for generating minimum-risk trajectories when two printers require repositioning (e.g., homing to avoid collisions). The proposed algorithm integrates naturally with existing constraint-based strategies that command robots to retreat to safe configurations during hazardous interactions, providing a rapid, maximum-clearance motion solution. This paper offers a formal analysis of the underlying framework and numerical characteristics of such online planners, with particular emphasis on minimum-risk motion sequences for SCARA-based systems. Our numerical study examines how different distance metrics and workspace configurations affect computation time, threat-field reconstruction, and the accuracy of the resulting trajectories, as well as a simulated success rates.
The increased complexity of modern manufacturing systems demands accurate and adaptive supervision to support system-level decision making. Digital Twin technologies address this need by enabling “what-if” analysis and predictive insights through simulations performed on digital replicas of physical systems. However, existing automated production system modelling approaches are constrained by existing control policies, and the resulting models are weakly integrated with Manufacturing Execution Systems (MES). This paper proposes a graph-based production system modelling method based on labelled directed graphs. By keeping clear correspondence with production system configuration, the method can enable smooth dynamic evolution of the system model when combined with model generation techniques. Building on this modelling method, an MES-driven Discrete Event Simulation (DES) architecture is developed by connecting an MES instance to a DES simulation engine. By integrating MES into simulation, the control logic of the actual system can be applied to simulated production, yielding more realistic and finegrained predictions of event sequences. Experiments have been conducted on two different production systems, demonstrating that the MES-driven DES architecture produces policy-consistent behaviour and validating the effectiveness of the proposed method.
Artificial intelligence (AI) offers significant potential for optimizing manufacturing through improved decision-making, shorter cycle times, and improved product quality. However, many AI implementations fail to generate measurable business value due to organizational and skills-related barriers. A key obstacle is the lack of skills development among employees. While the EU Artificial Intelligence Act provides a legal framework for mandatory training, current approaches are predominantly general basic training that do not address the specific skills, job profiles, and responsibilities required in a manufacturing context. This leads to a paradoxical situation in which companies with mature lean production systems (LPS), which offer the ideal structural and cultural conditions for AI integration, struggle to leverage these advantages due to their employees’ insufficient AI-specific competences. This paper addresses this gap and develops a comprehensive framework for employee-oriented AI training that combines LPS principles with the development of AI competencies. It serves as a structured and methodical frame that organizes and manages employee-specific training for AI. The framework comprises a multidimensional requirements analysis, derives five AI Competence Dimensions in manufacturing, demonstrates a role-specific competence matrix with room for individual development paths, and an outlook on a modular training concept enabling individualized learning. This approach enables differentiated, role-specific competence development that supports successful and responsible AI adoption while ensuring regulatory compliance and active employee involvement in a digital transformation era.
All manufacturing systems encounter diverse and often unknown forms of variability, including changing operating conditions, configuration shifts, external disturbances, equipment wear, and evolving task requirements. Traditional strategies for handling such variability typically involve re-tuning controller gains when feasible, explicitly modeling the variability and re-deriving the control law, or replacing the controller with more sophisticated laws. These approaches are costly, require downtime, and often interfere with legacy control systems that are already tested, validated, and certified. This paper proposes a plug-and-play, variability-agnostic Adaptive Control Augmentation (ACA) method that attaches externally to the legacy controller without modifying it, requires no variability models or plant re-identification, and compensates variability using only tracking-error feedback. A Lyapunov-based formulation provides conditions under which tracking remains bounded for admissible variability levels, offering explicit stability guarantees while preserving the original control logic. The approach is validated on a 2-DoF gantry equipped with a PID-LQR legacy controller under multiple intrinsic and extrinsic variability modes. Across all cases, ACA achieves steady-state tracking performance comparable to individually re-tuned controllers while using a single fixed gain set and no prior knowledge of variability. These results position ACA as a practical foundation for variability-agnostic, legacy-preserving control in intelligent manufacturing systems.
