The integration of Artificial Intelligence (AI) into manufacturing is transforming the industry by advancing predictive maintenance, quality control, and supply chain optimisation, while also driving the shift from Industry 4.0 towards a more human-centric and sustainable vision. This emerging paradigm, known as Industry 5.0, emphasises resilience, ethical innovation, and the symbiosis between humans and intelligent systems, with AI playing a central enabling role. However, challenges such as the “black box” nature of AI models, data biases, ethical concerns, and the lack of robust frameworks for trustworthiness hinder its widespread adoption. This paper provides a comprehensive survey of AI trustworthiness in the manufacturing industry, examining the evolution of industrial paradigms, identifying key barriers to AI adoption, and examining principles such as transparency, fairness, robustness, and accountability. It offers a detailed summary of existing toolkits and methodologies for explainability, bias mitigation, and robustness, which are essential for fostering trust in AI systems. Additionally, this paper examines challenges throughout the AI pipeline, from data collection to model deployment, and concludes with recommendations and research questions aimed at addressing these issues. By offering actionable insights, this study aims to guide researchers, practitioners, and policymakers in developing ethical and reliable AI systems that align with the principles of Industry 5.0, ensuring both technological advancement and societal value.

Artificial intelligence (AI) is increasingly adopted in manufacturing for tasks such as automated inspection, predictive maintenance, and condition monitoring. However, the opaque, black-box nature of many AI models remains a major barrier to industrial trust, acceptance, and regulatory compliance. This study investigates how explainable artificial intelligence (XAI) techniques can be used to systematically open and interpret the internal reasoning of AI systems commonly deployed in manufacturing, rather than to optimise or compare model performance. A unified explainability-centred framework is proposed and applied across three representative manufacturing use cases encompassing heterogeneous data modalities and learning paradigms: vision-based classification of casting defects, vision-based localisation of metal surface defects, and unsupervised acoustic anomaly detection for machine condition monitoring. Diverse models are intentionally employed as representative black-box decision-makers to evaluate whether XAI methods can provide consistent, physically meaningful explanations independent of model architecture, task formulation, or supervision strategy. A range of established XAI techniques, including Grad-CAM, Integrated Gradients, Saliency Maps, Occlusion Sensitivity, and SHAP, are applied to expose model attention, feature relevance, and decision drivers across visual and acoustic domains. The results demonstrate that XAI enables alignment between model behaviour and physically interpretable defect and fault mechanisms, supporting transparent, auditable, and human-interpretable decision-making. By positioning explainability as a core operational requirement rather than a post hoc visual aid, this work contributes a cross-modal framework for trustworthy AI in manufacturing, aligned with Industry 5.0 principles, human-in-the-loop oversight, and emerging expectations for transparent and accountable industrial AI systems.

Autonomous vehicles (AVs) require real-time, low-latency, and high-throughput communication to ensure safe and efficient operation. Fifth-generation (5G) and beyond communication networks, particularly New Radio (NR) operating at millimeter wave (mmWave) frequencies, offer promising solutions. However, higher frequencies introduce challenges like signal attenuation, blockage, and complexities of beam management. This paper provides a comprehensive survey of 5G and beyond communication networks, emphasizing their role in vehicle-to-everything (V2X) communication. It uniquely addresses the technical challenges of beam training and management, highlighting the application of artificial intelligence (AI) and machine learning (ML) techniques to optimize beam selection and reduce delays. Furthermore, the paper explores future directions, focusing on AI/ML-driven solutions to enhance beamforming and communication reliability in dynamic AV environments.

Artificial intelligence (AI) is increasingly used for tool-wear prediction in computer numerical control (CNC) machining, yet a high average accuracy can conceal systematic failure in the safety-critical high-wear region, where under-prediction makes a worn tool appear healthier than it is. This paper presents a trustworthy multimodal framework that audits, mitigates, and ultimately characterises this bias on the public MATWI dataset. A fusion model combining 512 image features with 84 force, acceleration, and acoustic features attains a strong test mean absolute error (MAE) of 30.9 µm, outperforming image-only (51.4 µm) and sensor-only (68.3 µm) baselines. A signed, stratified bias audit nevertheless reveals a pronounced directional error: the model predicts the healthy region to within 21.3 µm but under-predicts the worn region by approximately 85 µm (worn-region MAE 85.5 µm), with the error deepening monotonically as wear increases. A three-stage mitigation pipeline (data-level resampling, cost-sensitive and relevance-based training, and post-hoc recalibration) reduces the worn-region error by up to 18% but cannot eliminate it, and only at a measurable cost to the well-sampled region. The residual bias is shown to be both model-independent, persisting across six regressor families with a 69–127 µm residual gap, and optimisation-independent, surviving rigorous hyperparameter search and bias-aware loss functions behind an ≈ 85 µm error floor. These results establish the bias as data-intrinsic: it arises from the sparse sampling of the high-wear regime rather than from model or training choices, and the effective remedy lies in balanced, unbiased data collection rather than further algorithmic correction. The contribution is a reusable protocol for distinguishing model-reducible from data-intrinsic bias in safety-critical regression.

Artificial intelligence (AI) is increasingly used for tool-wear prediction in computer numerical control (CNC) machining, yet a high average accuracy can conceal systematic failure in the safety-critical high-wear region, where under-prediction makes a worn tool appear healthier than it is. This paper presents a trustworthy multimodal framework that audits, mitigates, and ultimately characterises this bias on the public MATWI dataset. A fusion model combining 512 image features with 84 force, acceleration, and acoustic features attains a strong test mean absolute error (MAE) of $30.9\mathsf{\mu }\mathrm{m}$, outperforming image-only ($51.4\mathsf{\mu }\mathrm{m}$) and sensor-only ($68.3\mathsf{\mu }\mathrm{m}$) baselines. A signed, stratified bias audit nevertheless reveals a pronounced directional error: the model predicts the healthy region to within $21.3\mathsf{\mu }\mathrm{m}$ but under-predicts the worn region by approximately $85\mathsf{\mu }\mathrm{m}$ (worn-region MAE $85.5\mathsf{\mu }\mathrm{m}$), with the error deepening monotonically as wear increases. A three-stage mitigation pipeline—data-level resampling, cost-sensitive and relevance-based training, and post-hoc recalibration—reduces the worn-region error by $18\%$ for the combined pipeline and by up to $22\%$ for the best single-stage configuration, but cannot eliminate it, and only at a measurable cost to the well-sampled region. The residual bias persists across six regressor families sharing the same fused representation, with a 69–$127\mathsf{\mu }\mathrm{m}$ residual gap, and survives rigorous hyperparameter search and bias-aware loss functions behind an $\approx 85\mathsf{\mu }\mathrm{m}$ error floor. At every maintenance threshold evaluated the model produces a $100\%$ false-safe rate.Under the feature representation and sampling of this dataset, the bias is therefore not addressable by model family, mitigation strategy, or optimisation, which identifies data adequacy rather than model design as the binding constraint. The conclusion is scoped accordingly: alternative feature representations, generative augmentation, and domain adaptation were not evaluated and remain open. The contribution is a reusable protocol for distinguishing model-reducible from data-limited bias in safety-critical regression.