This paper proposes a data readiness framework to evaluate whether a manufacturing company can realistically proceed with an Artificial Intelligence (AI) investment. The methodology assesses the organization’s data infrastructure through a two-phase approach. Phase 1 (Problem Framing and Data Awareness) supports in defining operational objective and translating it into a measurable Key Performance Indicator (KPI), formalizing plausible causes of the performance gap, mapping available datasets across the production process and inspecting the selected datasets from multiple perspectives. Phase 2 (Feasibility Assessment) combines the evidence generated in Phase 1 with the results of an Analytic Hierarchy Process (AHP)-based pairwise weighting into a decision-oriented Data Readiness Gate. The gate evaluates whether the company’s available datasets satisfy the criteria prioritized by the organization and whether the corresponding process stages are adequately covered. This enables an explicit assessment of feasibility at both the dataset level and the cross-dataset level, ultimately yielding a transparent go/no-go recommendation for the intended AI investment. The framework is demonstrated through a case study of a large Printed Circuit Board Assembly (PCBA) manufacturing company, where the dataset mapping step revealed that more than 80% of the identified sources were not directly usable for the intended investigation. A subsequent inspection of the remaining datasets highlighted structural constraints that still prevent end-to-end, board-level analysis. As a result, the framework indicated that, under current conditions, it is not advisable to proceed with AI investments. Overall, the study shows that early data readiness assessment can prevent premature AI investments by making feasibility constraints explicit and by producing actionable outputs that should guide future data-structure and integration improvements.

Control cabinet manufacturing is a representative engineer-to-order domain in which high product variety, small batch sizes, and inconsistent data availability limit the automation of assembly processes. In this setting, semantic segmentation of component geometry is not only a perception problem but also a prerequisite for deriving positional information that can be consumed by robotic assembly, wiring, and digital assistance systems. This paper presents an approach on using graph neural networks for semantic segmentation of CAD files for retrieving positional information. Four GNN architectures are compared for vertex-wise classification of five assembly-relevant classes. The experiments are based on a shared PyTorch geometric pipeline with automated hyperparameter optimization. GraphSAGE yields the strongest and most stable performance, reaching a mean weighted intersection-over-union (IoU) of about 0.72 during the main comparison and 0.80 after focused optimization. Beyond reporting the benchmark, the paper argues that the industrial value of mesh segmentation lies in the conversion of semantic masks into positionally explicit assembly information such as centroids, surface normals, target regions, and local action frames. These outputs are required for robotic automation as well as for worker assistance systems. The study therefore positions graph-based mesh perception as an enabling layer between vendor-independent CAD models and robotics-oriented assembly automation.

Discrete-event simulation (DES) is a powerful method for analyzing and improving production and material flow systems, yet its practical adoption remains difficult in engineer-to-order (ETO)-oriented small and medium-sized enterprises (SMEs). In these contexts, high product variety, limited standardization, fragmented data, and low internal simulation expertise create substantial barriers to simulation use. While prior studies and reviews have discussed DES applications in terms of industries, modelling approaches, optimization objectives, or software environments, they provide only limited insight into the conditions under which DES becomes practically accessible in constrained production settings. This paper addresses that gap through a systematic literature review with a barrier-oriented perspective. We analyze how the literature reflects obstacles related to data, model construction, abstraction, verification and validation, and organizational adoption, and we synthesize enabling mechanisms proposed to reduce these obstacles. Based on this synthesis, we derive five design requirements for making DES more accessible in ETO-oriented SMEs. The review shows that existing DES literature remains largely expert-centric and frequently presupposes comparatively structured, data-rich environments. In contrast, ETO-oriented SMEs face conditions under which simulation accessibility becomes a central concern. The paper contributes a structured account of DES access barriers, a synthesis of enabling mechanisms, and a design-oriented foundation for future methodological work on simulation approaches tailored to resource-constrained, high-variance production environments.

The inspection of components with complex and variable geometries remains a critical challenge in industrial quality control, particularly under limited availability of labeled defect data. This paper presents a multimodal robotic visual inspection framework for anomaly detection in variable geometry panels under limited defect data availability. The system combines three-dimensional (3D) geometric sensing with two-dimensional (2D) imaging through a geometry-aware acquisition strategy, where point cloud slicing and surface normal estimation guide robotic positioning to ensure perpendicular, consistent image capture across complex surfaces. A synthetic data generation pipeline using physically-based rendering and generative image editing addresses data scarcity. For anomaly detection, a modified SuperSimpleNet architecture with an enlarged 5×5 neighbourhood pooling kernel is proposed, adapted for spatially extended defects in wood-textured surfaces. The proposed variant achieves an image-level AUROC of 93.7% and a pixel-level AUPRO of 85.4%, improving image-level detection over the baseline configuration on domain specific (wood panels). The integrated system has been validated at Technology Readiness Level 5 (TRL 5) in a laboratory environment.

The Fourth Industrial Revolution (4IR) is reshaping manufacturing through technologies such as Artificial Intelligence, Big Data, and the Internet of Things. For South African manufacturing SMEs, the ability to adopt and leverage these technologies is increasingly linked to competitiveness and long-term resilience. However, skills gaps and fragmented support systems continue to limit meaningful 4IR adoption. This study proposes a structured skills development framework to support manufacturing SMEs in building 4IR-relevant capabilities using a Systems Engineering methodology. The framework was developed through four stages: input identification and concept development (based on literature, policy documents, and training models), requirement analysis (translating stakeholder needs into functional and user requirements), functional analysis (mapping requirements into a logical framework structure), and design synthesis (integrating outputs into a final framework). The resulting framework consists of four phases: Contextual Assessment, Skills Gap Analysis, Skills Delivery Mechanism, and Stakeholder Engagement. Validation was conducted through expert review and a pilot study, leading to refinements that improved usability, strategic alignment, and practical applicability. The framework offers a scalable approach for SME managers, training providers, and policymakers seeking to accelerate 4IR readiness and workforce development in South Africa’s manufacturing sector. Limitations related to the single pilot validation context are acknowledged, and directions for future research are proposed.