Renewable energy sources are a growing branch of industry. One such source is wind farms, which have significantly increased their number over recent years. Alongside the increased number of turbines, maintenance problems are growing. There is a need for newer and less intrusive predictive maintenance methods. About 40% of all turbine failures are due to bearing failure. This paper presents a modified neural direct classifier method using raw accelerometer measurements as input. This proprietary platform allows for better damage prediction results than convolutional networks in vibration spectrum image analysis. It operates in real time and without signal processing methods converting the signal to a time–frequency spectrogram. Image processing methods can extract features from a set of preset features and based on their importance. The proposed method is not based on feature extraction from image data but on automatically finding a set of features from raw tabular data. This fact significantly reduces the computational cost of detection and improves the failure detection accuracy compared to the classical methods. The model achieved a precision of 99.32% on the validation set, and 96.3% during bench testing. These results were an improvement over the method that classifies time–frequency spectrograms of 97.76% for the validation set and 90.8% for the real-world tests, respectively.

In this study, a predictive maintenance (PdM) system focused on feature selection for the detection and classification of simulated defects in wind turbine blades has been developed. Traditional PdM systems often rely on numerous, broadly chosen diagnostic indicators derived from vibration data, yet many of these features offer little added value and may even degrade model performance. General feature selection methods might not be suitable for PdM solutions, as information regarding observed faults is often misinterpreted or lost. To address these issues, a structured feature selection method based on correlation analysis supplemented with comprehensive visual evaluation was proposed. Unlike generic dimensionality reduction techniques, this approach preserves critical domain-specific information and avoids misinterpretation of fault indicators. By applying the proposed method, it was possible to successfully filter out redundant features, enabling simpler machine learning (ML) models to match or even surpass the performance of more complex deep learning (DL) architectures. The best results were achieved by a deep neural network trained on the full dataset, with accuracy, precision, recall, and F1 score of 97.30%, 97.23%, 97.23%, and 97.23%, respectively, while the top-performing ML model (a voting classifier trained on the reduced dataset) attained scores of 97.13%, 96.99%, 96.95%, and 96.94%. The proposed method for reducing condition indicators successfully decreased their number by approximately 3.27 times, simultaneously significantly reducing computational time of prediction, reaching up to 50% reduction for complex models. In doing so, we lowered computational demands and improved classification efficiency without compromising accuracy for ML models. Although feature reduction did not similarly benefit the metrics for DL models, these findings highlight that well-chosen, domain-relevant condition indicators can streamline data input and deliver interpretable, cost-effective PdM solutions suitable for industrial applications.

Optimal control of multivariable systems is a complex dynamic process that minimizes the cost function to obtain the optimal control strategy. Unfortunately, for nonlinear systems, it is not possible to use the traditional linear quadratic regulator (LQR), which would be optimal over the entire range of parameter variation. The problem of nonlinear multivariable systems and their optimal control is very momentous. The solution presented in this paper is based on the application of Reinforcement Learning (RL) networks in controlling a five-degree-of-freedom overhead crane system. Additionally, unlike the classical approach, the algorithm is adapted to directly analyze tabular data of inputs and outputs of the controlled model instead of analyzing its state as feedback (model-free). Implementing the new control structure for the multivariable system improved control quality compared to the classical LQR controller with linearization at the operating point. In addition to quality, the resource indicators, which in the LQR controller are represented by the matrix R, have been significantly improved. The architecture of the neural control system is presented, ensuring that over the entire range of nonlinearity, the quality of control is preserved while reducing the cost of its resource intensity. Obtaining optimal control with reduced resources for its implementation induces a wide range of applications of such neural control in engineering systems. The effectiveness of the proposed control system has been demonstrated in simulation studies. The simulation results present the system’s excellent control performance and adaptability over the entire range of object nonlinearity. The neural algorithm resulted in significantly shorter adjustment time and better control quality with significantly less system resource consumption and increased system dynamics.

This paper presents methods for modelling and optimising the industrial limestone grinding process carried out using a real limestone plant. Two key process evaluation indicators were developed: specific electric energy consumption and an extended indicator that also includes gas usage. Using process data collected from the SCADA system and results from industrial factorial experiments, regression artificial neural network models were developed, with controllable process parameters used as inputs. In the next phase, black-box optimisation was performed using Bayesian and genetic algorithms to identify optimal mill operating settings. The results demonstrate significant improvements in energy efficiency, with energy savings reaching up to 48% in selected cases. The proposed methodology can be effectively applied to enhance energy performance in similar industrial grinding processes.

Vibration-based bearing fault diagnosis is a widely studied predictive maintenance problem, but reported results are often difficult to compare. Performance depends not only on the model itself, but also on the evaluation protocol, the train--test split, and the type of domain shift considered. In particular, leakage-prone window-level splitting and loosely defined source--target settings can lead to overly optimistic conclusions that do not reflect real transfer performance across changing operating conditions, acquisition regimes, or bearing identities. To address this issue, this paper introduces a leakage-safe and reproducible benchmark for cross-domain bearing fault diagnosis on the Case Western Reserve University and Paderborn University datasets. The benchmark defines six fixed source--target scenarios, enforces recording-level train--test separation, and evaluates both machine-learning and deep-learning baselines under a common protocol. Final reporting is based on a consistent evaluation setup, with repeated-seed follow-up used where necessary to support reliable conclusions for deep-learning models. The results show that scenario difficulty is highly heterogeneous. Some transfer settings are effectively saturated, while others remain substantially more challenging. Deep-learning models often achieve stronger performance, but their conclusions can be sensitive to initialization and require repeated-seed validation. Overall, the benchmark provides a reproducible basis for scenario-level evaluation and more reliable comparison of cross-domain bearing diagnosis methods. The code for this study is publicly available at https://github.com/1Sensor/pdm-bench.

In predictive maintenance (PdM) applications, sensor signals are often affected by various forms of noise that can degrade model accuracy and reliability. While many studies use Gaussian noise to simulate environmental disturbances, real-world sensor data is typically impacted by a broader range of noise types, including impulsive, frequency-dependent, and drift-based distortions. This paper presents a modular noise injection framework designed to augment PdM datasets with diverse synthetic and measured noise signals. The framework simulates realistic sensor and system-level disturbances and is evaluated on vibration data from an online shaft monitoring test stand. We assess the impact of noise-based augmentation on both limited and comprehensive datasets using six machine learning models. Results demonstrate that properly designed noise injection can substantially enhance model robustness and performance, particularly in data-constrained settings. Notably, we observed that models trained on raw time-domain data, while offering greater potential for learning rich signal characteristics, also exhibit higher sensitivity to aggressive augmentation. Our findings highlight the importance of carefully tuning noise augmentation strategies to match the measurement context and model architecture.