Machine Learning and IoT as Enablers of Intelligent Industrial Transformation

1. Introduction

The rapid expansion of cyber-physical systems, sensor networks, and data-driven automation platforms has profoundly reconfigured the operational logic of contemporary industrial environments. Within the paradigm of the Fourth Industrial Revolution—commonly designated as Industry 4.0—the convergence of machine learning (ML) and the Internet of Things (IoT) has emerged as a defining technological axis, reshaping processes that range from real-time monitoring and predictive maintenance to supply chain optimisation and cybersecurity [1,2,3]. This Special Issue was conceived at this intersection, with the purpose of consolidating peer-reviewed contributions that advance both the theoretical foundations and the practical deployment of ML- and IoT-based systems in industrial and related contexts.

The thematic scope of the call was deliberately broad, encompassing anomaly and intrusion detection in wireless sensor and SCADA networks, demand forecasting in supply chains, digital twin frameworks, augmented reality in industrial processes, and intelligent agricultural systems underpinned by Artificial Intelligence of Things (AIoT) architectures. This breadth reflects the interdisciplinary nature of Industry 4.0 research and the recognition that no single methodology or application domain can fully capture its complexity [4,5]. Following a rigorous double-blind review process, eight contributions were accepted—comprising original research articles and critical review papers—whose collective scope is summarised below.

Security and intrusion detection: Two contributions address cybersecurity in IoT-based networks from complementary methodological perspectives. The first [6] presents a machine learning framework optimised for wireless sensor network (WSN) environments, combining Adaptive Synthetic Sampling (ADASYN) for class-imbalance correction with a dual-stage feature selection strategy that integrates Feature Importance Selection (FIS) and Recursive Feature Elimination (RFE), followed by an XGBoost classifier trained under five-fold cross-validation. The reported overall accuracy of 99.87% confirms that explicit imbalance treatment and dimensionality reduction are critical determinants of IDS performance in resource-constrained deployments [7]. The second contribution [8] benchmarks XGBoost, LightGBM, and Random Forest under a Top-K feature selection strategy on the CICIoMT2024 dataset, finding that Random Forest achieves the best balance between predictive performance (91.7% accuracy; F1-score 93%) and computational efficiency, with a 35% reduction in processing time when restricted to the Top-10 feature subset. Together, these papers establish ensemble learning—particularly gradient-boosting variants—as a productive methodological direction for real-time IoT security applications.

Anomaly detection in industrial control systems: A third contribution [9] conducts a rigorous empirical comparison of four unsupervised anomaly detection algorithms—Autoencoder (AE), LSTM-Autoencoder (LSTM-AE), One-Class SVM (OCSVM), and Isolation Forest (IF)—applied to SCADA telemetry from an urban wind turbine characterised by an inverted class imbalance in which operational anomalies constitute 75.7% of records. The AE trained exclusively on the rare normal state achieves the highest overall performance (AUC 0.9667), while the IF, despite a strong discriminative AUC of 0.8616, fails entirely to classify the normal class (${\mathrm{Recall}}_{\mathrm{normal}}=0.00$) when the optimal F1-score threshold is applied. This result constitutes an empirically grounded demonstration of the methodological superiority of reconstruction-based approaches when the statistically dominant class corresponds to the anomalous state, and provides quantitative guidelines for threshold calibration in prognostics and health management (PHM) systems operating under non-standard distributional assumptions.

Data-driven decision support in supply chain management: The fourth contribution [10] proposes a novel Decision Support System for demand forecasting that models the prediction task as a multivariate time series problem on a causal dependency graph. A Graph Convolutional Network (GCN) captures both the temporal dynamics of each variable—including the Consumer Sentiment Index, Consumer Price Index, Personal Income, and Unemployment Rate—and the structural interdependencies encoded in the graph topology. Comparative evaluation against conventional ML baselines demonstrates that the GCN approach more accurately captures abrupt transitions in demand behaviour, where competing methods tend toward excessive smoothing. This contribution reframes demand forecasting as a graph-structured inference problem, opening productive directions for incorporating macroeconomic knowledge into operational forecasting models [11].

Operational technologies for process enhancement: Two papers address the deployment of digital twins and augmented reality (AR) in industrial process optimisation. The digital twin contribution [12] presents a comprehensive strategic framework for implementing Digital Twin (DT) technology in manufacturing environments, integrating IoT, AI, and ML to construct accurate digital replicas of production systems and enable real-time simulation, predictive maintenance, and process optimisation. A detailed case study demonstrates empirical reductions in downtime and improvements in maintenance scheduling, responding to a recognised gap between conceptual DT frameworks and empirically validated deployment guidance [13,14]. The scoping review on augmented reality [15]—conducted following PRISMA 2020 guidelines with Cochrane bias assessment—synthesises 38 articles published between 2019 and 2024. Key findings confirm that AR consistently reduces error rates and task execution times, improves training outcomes, and optimises maintenance and assembly workflows. Unity 3D is identified as the dominant graphics engine; persistent barriers include high initial costs and hardware limitations.

Systematic reviews: smart manufacturing and intelligent agriculture: The final two contributions provide field-level overviews of broader application domains. The critical review on smart manufacturing [16] surveys IoT systems and ML algorithms across security, predictive maintenance, process control, and additive manufacturing, offering a comparative analysis of architectural configurations, sensing solutions, data strategies, and communication protocols. The review on AIoT in agriculture [17]—following PRISMA 2020 guidelines and covering publications from 2018 to 2025—examines the transition from IoT-based monitoring to AIoT-driven intelligent agriculture, with particular attention to deployment challenges in low-income countries where infrastructural deficits and economic constraints constrain adoption. Applications reviewed include smart irrigation, pest detection, yield prediction, and livestock management. A tailored AIoT architecture for resource-constrained environments is proposed, integrating edge intelligence with governance and capacity-building frameworks.

2. Cross-Cutting Observations

Reading the eight contributions collectively, several cross-cutting themes merit explicit acknowledgment. The problem of data imbalance—whether standard or inverted—emerges as a persistent methodological challenge across intrusion detection, anomaly analysis, and operational classification tasks; resampling strategies and reconstruction-based unsupervised approaches represent complementary mitigation paths depending on distributional characteristics. A notable convergence toward ensemble and graph-based learning architectures is also evident: XGBoost, LightGBM, and Random Forest dominate the supervised security contributions, while GCNs are deployed for relational forecasting, consistent with the broader maturation of methods that exploit ensemble diversity and relational inductive biases. At the system level, the review articles collectively signal a shift from isolated algorithmic proposals toward integrated architectures that address the full data lifecycle—from acquisition and preprocessing through inference and decision support—reflecting the increasing operational maturity of Industry 4.0 deployments. Finally, the inclusion of a contribution specifically addressing AIoT in low-income agricultural contexts broadens the discourse beyond the typically high-income-country framing of Industry 4.0 scholarship, underscoring the potential of these technologies for sustainable and equitable development.

3. Conclusions

The eight contributions assembled in this Special Issue constitute a substantive addition to the rapidly evolving body of knowledge at the intersection of ML, IoT, and Industry 4.0. Looking forward, the deployment of ML at the network edge requires continued innovation in model compression and federated learning. Interpretability remains a critical barrier in safety-critical environments where regulatory requirements demand explainable inference. Furthermore, the extension of intelligent systems into resource-constrained and low-income contexts constitutes both a scientific challenge and a social imperative that the research community should address with equity-conscious design principles.

We extend our sincere gratitude to all authors for their rigorous and timely submissions, the reviewers whose evaluations were instrumental in ensuring scientific quality, and the editorial team for their professional coordination throughout the process.