International Scientific Journals
of Scientific Technical Union of Mechanical Engineering "Industry 4.0"

The integration of engineering support tools is essential for the efficient modeling, simulation, and optimization of complex technical systems. This paper presents a dynamic model of a micro-combined heat and power (mCHP) system, developed to validate the feasibility of integrating various computational environments. The approach leverages modular architectures, enabling seamless data exchange between distinct software platforms, thus supporting both detailed thermodynamic analysis and real-time performance optimization. The flexibility of this approach allows for the inclusion of diverse analytical frameworks, including neural network-based optimization, data-driven control strategies, and alternative programming languages, without being limited to a single computational tool. This adaptability makes the proposed architecture particularly suitable for evolving engineering applications, where rapid prototyping and iterativ e refinement are critical. The study highlights the potential of such integrated environments to enhance the design and operational efficiency of energy systems, providing a scalable foundation for future expansions.

This study presents a correlation-based approach for both detecting sensor malfunctions and identifying redundant sensors in a multi-sensor condition monitoring system. Sensor malfunctions were detected using a threshold-based method that flagged correlation drops, with persistence criteria applied to eliminate false positives. While no persistent malfunctions were observed during the study, the developed algorithm remains suitable for real-time deployment. Correlation analysis also revealed that the 3axis_Y signal exhibited the highest average correlation with others, indicating redundancy. Five machine learning models were trained and evaluated with the Leave-
One-Run-Out strategy to guarantee generalization across acquisition sessions. The findings demonstrated that correlation-driven sensor selection and anomaly detection are effective tools for optimizing predictive maintenance systems, improving model generalization, and simplifying sensor networks without sacrificing reliability.

This study evaluates the performance of newly designed deep-learning model—bidirectional long short-term memory network (Bi-LSTM), with baseline to a conventional multilayer perceptron (MLP)—for classification faults of rolling-element bearings from raw vibration signals. The models are benchmarked against a previously optimised one-dimensional convolutional neural network (1-D CNN), originally obtained via Bayesian hyperparameter search. A carefully selected dataset of 3600 one-second segments was captured under varying speed conditions and dynamically enhanced with Gaussian noise during processing. On the test set, the Bi-LSTM achieves 100 % accuracy, the 1-D CNN 97.9 %, and the MLP 53.3 %. Training dynamics, confusion patterns, and model complexity were thoroughly analysed, highlighting the trade-offs between accuracy, latency and deployment cost in edge-computing scenarios.

This paper proposes a novel architecture for pdm-tools, a dedicated library designed to streamline the development process of predictive maintenance (PdM) algorithms within the Industry 4.0 paradigm. pdm-tools tackles the time-consuming nature of PdM algorithm development by encompassing five key components within a unified framework. pdm-tools tackles the time challenge by offering a unified workflow: generate synthetic vibration data (gearboxes, bearings, shafts) for rapid prototyping, eliminating initial real-world data collection. Data preparation and feature extraction ensure readiness through scaling, filtering, and extracting diverse condition indicators, empowering robust algorithms. Finally, user-defined model development and evaluation allow training, optimization, and selection of optimal PdM algorithms (machine learning/deep learning). By integrating these functionalities within a single architecture, pdm-tools empowers rapid development and evaluation of PdM algorithms, ultimately leading to a more efficient and effective implementation of predictive maintenance strategies within the Industry 4.0 landscape.

The development of Industry 4.0 and Predictive Maintenance Systems allows for effective utilization of equipment by incorporating ML methods for identifying tool condition. However, including large number of Condition Indicators for machinery monitoring increases computational complexity, hence the response of the system elongates. Therefore, it is important to check the utility of indicators and reduce them. In this paper, we investigate different feature selection methods: AIC (Akaike Information Criterion), BIC (Bayesian Information Criterion), Random Forest, Lasso Regression (L1 Regularization) for NASA Gearbox Fault Detection Dataset, PHM 2009. We processed the raw data and calculated CI from time domain, frequency domain and envelope. An SVM Classifier model was trained on full collection of indicator and reduced, then performances were compared. The obtained results highlight the advantage of feature selection, proving that effective PdM systems can be based on diminished number of Health Indicators.

The Industry 4.0 revolution has significantly influenced the control of non-linear Multiple Input Multiple Output (MIMO) systems, particularly through the application of artificial neural networks (ANNs). This paper explores current trends in the control of non-linear MIMO objects, emphasizing the role of ANNs in enhancing performance and efficiency. Key developments, methodologies, and case studies are reviewed to illustrate the impact of ANNs on non-linear MIMO control

Controlling Multi-Input Multi-Output (MIMO) systems, such as portal conveyors, poses significant challenges due to their inherent complexity and variability. Traditional control methods often fall short in handling the dynamic and nonlinear nature of these systems. This paper presents a novel reinforcement learning (RL) approach, leveraging the twin-delayed deep deterministic policy gradient (TD3) algorithm, to develop a follow-up controller that is robust to changes in system parameters. Our simulation experiments demonstrate the effectiveness of this method.

Industry 4.0 and the rise of predictive maintenance have led to increased interest in developing efficient and accurate methods for detecting tool wear in manufacturing processes. In this work, we investigate the use of machine learning techniques, specifically Long Short- Term Memory (LSTM) networks, for the detection of tool wear using NASA Milling Data Set. We first preprocess the raw milling data and extract relevant features using signal processing techniques. Then, we train and evaluate the performance of the LSTM network for detecting tool wear in the milling process. Our results demonstrate the effectiveness of the LSTM network for detecting tool wear, achieving high accuracy and recall scores. Additionally, we compare our results to those obtained using traditional statistical process control methods and highlight the advantages of using machine learning techniques for tool wear detection in the context of Industry 4.0 and predictive maintenance. Our work provides a practical example of how machine learning techniques can be applied to manufacturing processes to improve efficiency and reduce downtime.