In engineering applications, the bearing faults diagnosis is essential for maintaining reliability and extending the lifespan of rotating machinery, thereby preventing unexpected industrial production downtime. Prompt fault diagnosis using vibration signals is vital to ensure seamless operation of industrial system avert catastrophic breakdowns, reduce maintenance costs, and ensure continuous productivity. As industries evolve and machines operate under diverse conditions, traditional fault detection methods often fall short. In spite of significant research in recent years, there remains a pressing need for improve existing methods of fault diagnosis. To fill this research gap, this research work aims to propose an efficient and robust system for diagnosing bearing faults, using deep and Shallow features. Through the evaluated experiments, our proposed model Multi-Block Histograms of Local Phase Quantization (MBH-LPQ) showed excellent performance in classification accuracy, and the audio-trained VGGish model showed the best performance in all tasks. Contributions of this work include: Combine the proposed Shallow descriptor, derived from a novel hand-crafted discriminative features MBH-LPQ, with deep features obtained from VGGish pre-trained of Convolutional Neural Network (CNN) using audio spectrograms, by merging at the score level using Weighted Sum (WS). This approach is designed to take advantage of the complementary strengths of both feature models, thus enhancing overall bearing fault diagnostic performance. Furthermore, experiments conducted to verify the approach’s performance is assessed based on fault classification accuracy demonstrated a significant accuracy rate on two different noisy datasets, with an accuracy rate of 98.95% and 100% being reached on the CWRU and PU datasets benchmark, respectively.

Early fault detection in photovoltaic grid-connected systems is crucial for optimizing energy output and ensuring system reliability. However, challenges such as noise interference, limited computational resources, and the need for high diagnostic accuracy make fault detection complex and resource-intensive. This study introduces a novel fault diagnosis approach aimed at enhancing diagnostic accuracy and efficiency, even in noisy environments, while being efficient enough for real-time deployment on resource-limited devices. Raw one-dimensional time series data are transformed into 2D pixel-wise images to meet the requirements of a lightweight MobileNet architecture, specifically chosen for its ability to efficiently extract discriminative features in this context. Alongside, an ensemble stacked classifier incorporating Support Vector Machines, Random Forest, and Decision Trees is employed, enabling the use of MobilNet features for improving the classification. This combination provides insights into the relative strengths of each model type, showcasing the benefits of contrasting traditional machine learning with more complex deep learning methods for fault detection in PV systems. The model demonstrates high test accuracy on both noiseless (97.36%) and noisy datasets (91.67%), with precision, recall, and F1 scores above 91% across all conditions, while maintaining low processing times. This approach not only advances fault diagnosis in grid-connected PV systems but also offers a practical framework for selecting optimal models tailored to specific application requirements.

Occupancy detection plays a pivotal role in optimizing energy consumption, enhancing occupant comfort, and enabling intelligent decision-making in smart building environments. While traditional centralized learning methods often compromise user privacy and scalability, federated learning (FL) offers a promising privacy-preserving alternative. In this study, we propose Federated Cascade Learning with Long Short-Term Memory and Attention (FCL-LSTMA), a novel decentralized architecture that synergistically combines FL and cascade learning to enable accurate and efficient occupancy detection across distributed environments. The FCL-LSTMA framework operates through a two-stage training process: initially, Block B1 leverages 1D convolution, batch normalization, and a parallel LSTM layer for foundational feature learning. Once stabilized, the block is frozen and extended to Block B2, which incorporates attention mechanisms and an additional LSTM layer to refine temporal feature extraction. This progressive training ensures improved stability, model depth, and generalization without compromising data privacy. Experimental results on four benchmark datasets demonstrate the superiority of the proposed framework. FCL-LSTMA achieves outstanding accuracy for binary occupancy detection: 99.99% on the Living Room dataset, 98.95% on the UCI Occupancy Detection dataset, 99.95% on both the UCI Room Occupancy Estimation and IEQ datasets. For multi-class tasks, the model achieves 99.98% and 99.95% on the Living Room and UCI Room Occupancy Estimation datasets, respectively. Furthermore, it demonstrates computational efficiency with inference times ranging from 0.2028 to 0.562 milliseconds and memory usage between 0.17 MB and 0.49 MB. These results confirm the robustness, scalability, and privacy-preserving capabilities of FCL-LSTMA, making it a compelling solution for real-time occupancy detection in smart building systems.

Recent advancements in technology, particularly in artificial intelligence and privacy-preserving tools, have facilitated the development of more sophisticated and secure approaches to addressing this challenge. This paper introduces FedWell, a novel privacy-preserving Federated occupant Stress monitoring approach for smart building environments. Our system integrates physiological data from smart yoga pillows (SaYoPillow) and wearable environmental sensors to train a lightweight, edge-deployable Artificial Neural Network (ANN) model. The FedWell framework ensures data privacy while enabling collaborative learning across distributed clients using Federated Averaging (FedAvg) aggregation. We conducted experiments using a comprehensive dataset of physiological parameters for stress level detection. The results demonstrate the global model’s exceptional performance, achieving 99.95% accuracy in stress level recognition, with precision of 99.95%, recall of 99.93%, and an F1-score of 99.94%. The global model exhibits a minimal loss of 0.0019% and a low communication cost of 0.08 Mb, highlighting its efficiency for real-time applications.