Dennis Brandão

The integration of legacy industrial systems with modern intelligent technologies poses significant challenges for preserving and reusing engineering knowledge embedded in unstructured technical documentation. This limitation affects maintenance, diagnostics and decision‐making processes in industrial environments. This study investigates the use of Retrieval‐Augmented Generation (RAG) architectures for knowledge continuity in legacy industrial automation, using PROFIBUS documentation as a representative case of legacy industrial communication systems. Three configurations are comparatively evaluated: (i) a standalone large language model (LLM), (ii) a Single‐Agent RAG architecture and (iii) a Multi‐Agent RAG architecture with explicit supervisor‐worker orchestration. The evaluation combines quantitative lexical and semantic metrics (ROUGE, METEOR, MATTR and BERTScore) with qualitative assessment based on the LLM‐as‐a‐Judge paradigm, complemented by human expert judgement. Results indicate that the proposed Multi‐Agent RAG architecture with supervisor‐worker, which combines domain‐specialized retrieval with explicit multi‐agent orchestration, achieves higher contextual precision, factual accuracy and faithfulness than the evaluated baselines. These gains, however, come at the cost of increased computational complexity, highlighting a trade‐off between response quality and efficiency. Overall, the findings demonstrate that the proposed architecture provides a robust approach for reliable knowledge synthesis from legacy technical documentation, supporting maintenance activities and decision‐making in industrial environments.

The condition of abrasive grinding wheels after the dressing process can be visualized through acoustic maps, which enable the identification of patterns associated with the wheel topography and surface characteristics. Initially, acoustic maps were generated exclusively from acoustic emission (AE) signals. Due to the low sharpness of these maps, resulting from the inherent nature of acoustic emission signals, multimodal fusion techniques combining acoustic emission and vibration signals (AE + Vib) were developed, as well as approaches based on digital image processing (AE + Vib + DIP). Traditionally, the quantitative evaluation of these acoustic maps has been based on cross-correlation, in which regions of interest in the topological map are compared with predefined patterns. However, this approach does not consider analyses related to background noise in the images, which can affect the sharpness and contrast of the acoustic maps, making their visual interpretation more difficult. Therefore, this study effectively evaluates the background noise in acoustic maps obtained using different methods. The study investigates background homogeneity and the spatial definition of regions of interest through statistical metrics and digital image processing techniques. The results demonstrate that the integration of multiple information sources significantly reduces background noise, leading to substantial improvements in the overall quality of the acoustic maps.

The NIS2 Directive strengthens governance and time-bound incident reporting obligations for cybersecurity in critical sectors, while Industrial Ethernet and IT/OT convergence expose deterministic shop-floor networks to broader threat propagation. This paper introduces an engineering approach to the application of the NIS2 Directive for reducing the detection latency at the OT level without excessively increasing the costs. The proposed architecture is based on an innovative complementary device, the Early Warning System (EWS), tailored for OT monitoring and designed for integration with enterprise monitoring (Security Information and Event Management—SIEM) and operational response (SOC/MDR—Security Operations Center/Managed Detection and Response). The EWS adopts a silent device paradigm: because the honeypot has no production role, any interaction becomes a high-confidence signal, minimizing false positives. The EWS is freely available open-source and, to fit heterogeneous plants, it offers two deployment modes, both with zero impact on network latency and timing: Full Mode combines passive monitoring via SPAN/port mirroring with a low-interaction honeypot, while Light Mode provides a pure decoy for resource-constrained edge hardware. The implementation is containerized (Docker) and forwards events to an external SIEM. Two validation steps are presented: the first, on a real PROFINET robotic cell using the Full-mode, demonstrates the zero false positives in normal operations, the detection of reconnaissance scans, baseline deviations, and ARP spoofing attempts; while the second, on three emulated industrial control systems using the ICS-NAD dataset, shows the effectiveness in realistic industrial applications.

Over the last two decades, the automotive industry has spearheaded a shift toward data-centric manufacturing, where Real-Time Ethernet (RTE) networks defined in IEC61784-2 serve as critical components for ensuring deterministic communication at the Operation Technology level. Although AI-based systems offer significant potential for predictive maintenance and cybersecurity, their effectiveness is currently limited by a lack of structured datasets from real-world industrial environments. Most existing research relies on small-scale simulations or laboratory setups that fail to capture the scale and complexity of actual production. To address this gap, this paper introduces a novel methodology for repurposing network data collected throughout a plant’s lifecycle, specifically during the commissioning and validation phases of RTE networks according to IEC61918. An additional important contribution is the creation of the first multi-plant dataset for real RTE (PROFINET) traffic in the automotive sector, aggregating 300 GB of data from 54,000+ devices across nearly 700 production lines in 17 industrial sites. The work defines standardized methodologies and replicable processes for systematic data acquisition, validation, and labeling to ensure long-term usability for training AI models. Finally, four case studies (focused on performance, maintenance, security, and machine learning) show how this dataset can be used to enhance the reliability of modern smart manufacturing.

Milling is one, if not the most important, manufacturing process based on machining in today’s industry. Since this is a continuous process, mechanical tool wear becomes an active concern that affects the final product quality. Tool condition monitoring (TCM) is critical in milling to enhance product quality and prevent tool failures, while tool wear estimation is the first step to achieving a complete system. Several intelligent models have been proposed in the literature to estimate mechanical tool wear, but the use of different combinations of recent algorithms in a hybrid way has not been entirely explored. This study introduces a hybrid tool wear estimation model for milling, combining an Adaptive Neuro-Fuzzy Inference System (ANFIS) with Micro Evolutionary Particle Swarm Optimization (MEPSO) on both antecedent and consequent parameters. MEPSO is a recent modification of the classic particle swarm optimization that has not yet been applied in tool wear estimation or in a hybrid setting with ANFIS. The proposed hybrid ANFIS-MEPSO model addresses tool wear prediction challenges by optimizing ANFIS parameters through MEPSO while employing a feature selection method in the ANFIS-MEPSO inputs. A dataset containing experimental milling data, including vibration, acoustic emission, and electrical current signals, was used to train and test the model with 8 features extracted and a feature selection algorithm to estimate tool wear, achieving promising results with MSE, MAE, and R² metrics of 0.0045, 0.0578, and 0.9058 on testing data. A sensitivity analysis was conducted to better understand the changes in performance by modifying the optimization hyperparameters. It was revealed that stable and optimal performance occurred with σ and γ between 0.6–0.8 and at least 20 particles, while higher values (0.9) led to inconsistent results. This research highlights ANFIS-MEPSO’s viability for tool wear monitoring in milling, suggesting potential for broader application and further refinement.

Acoustic maps are valuable tools for visualizing mechanical conditions in industrial processes, but their quality is often compromised by noise, interference, and low signal sensitivity. This work presents an enhanced methodology that combines the fusion of acoustic emission (AE) and vibration signals with digital image processing techniques to improve the visual quality of the generated maps. Pixel formation was refined using bidimensional rms, improving smoothing. Subsequently, a distribution adjustment block generated a composite signal with balanced contributions from AE and vibration, increasing image interpretability while preserving the relevance of each sensing modality. Experiments were conducted on an aluminum oxide grinding wheel employed in tangential surface grinding, which was dressed with a single-point diamond tool at different depths, with reference patterns in the shapes of plus (+) and tee (T) machined onto its surface. The resulting acoustic maps achieved a correlation index of 0.830 with the reference images, representing an average improvement of 22.87% compared to AE-only approaches. The proposed methodology proved particularly effective under intermediate and deep dressing conditions, offering a robust, noninvasive, and efficient solution for grinding wheel monitoring, with potential compatibility with digital manufacturing systems.

Communication networks are widely used in smart manufacturing systems due to their flexibility, low cost, mobility, and easy deployment, providing an efficient way to establish communication between devices in the plant. In many cases, the communication between devices may suffer from instabilities, leading to delays in communication and, consequently, the observation of events in a different order from their occurrence in the system. The use of unreliable communication, in general, leads to a degradation of data-driven systems, which can make them useless. In this paper, we address the problem of fault diagnosis of Networked Discrete Event Systems (NDES) resilient to observation delays in the communication network. We consider that each event is communicated together with its sequence number, and we show how the sequence number can be used to improve the fault diagnosis capability in the presence of observation delays, without the need to reorder the sequence of events that has occurred in the system. We also propose the implementation of a delay-resilient diagnoser that diagnoses the fault event based on the observed sequence of events together with their sequence numbers.

This paper presents a methodology for analyzing industrial PROFINET networks. It focuses on integrating automated topology reconstruction based on collected data and data preprocessing to estimate cumulative distribution functions and probability density functions for failures. Results establish a foundational framework for predictive maintenance through probabilistic analysis. The approach enables temporal error pattern visualization and supports statistical reliability threshold estimation. The work contributes to predictive maintenance in Industry 4.0 environments.

The integration of virtual Programmable Logic Controllers (vPLCs) into industrial automation systems introduces the potential for enhanced maintainability and scalability through container-based automation. Unlike traditional hardware-based PLCs, vPLCs operate within edge computing environments, leveraging lightweight virtualization to provide flexibility and support modern microservices architectures. However, the open question is: can vPLCs meet the stringent real-time performance requirements of industrial control applications, particularly in communication with sensors and actuators? The paper objective is to fill this gap. Differently from other works in literature, the performance of the real-time data exchange between virtual PLCs and sensors/actuators is evaluated. Inparticular, the paper presents and describes a methodology designed for comparing real PLC and virtual PLC in real-time industrial automation scenarios. The methodology includes the definition of specific performance metrics, the design of a standardized experimental setup to characterize both device real-time performance and uncertainty sources, and the development of analytical models to support simulations and digital twin applications. The proposed method of comparison is demonstrated in a reference use case including real-time Ethernet connectivity; results lead to: a) important conclusions about methodology effectiveness; and b) the analytical model of the considered use case. Indetails, the analysis indicates that vPLCs exhibit approximately 50% higher jitter, suggesting a minimum recommended PROFINET cycle time of 2 ms for optimal performance. The findings contribute to the broader understanding of vPLC capabilities in industrial automation, offering practical insights for industries aiming to transition to modern, containerized control systems without compromising real-time communication performance.