The reel bearing's operational state in a cigarette-making machine is crucial for both machine stability and product quality. To accurately assess its health, this paper introduces a multi-sensor monitoring system. It utilises various sensors, including vibration, acoustic emission, and temperature sensors, to collect real-time data on the reel bearing's five operational states: normal, inner ring fault, outer ring fault, rolling element fault, and compound fault. To tackle noise interference in multi-source signals and address heterogeneous feature distribution under complex conditions, wavelet decomposition is used. This technique extracts multi-scale time-frequency features to enhance the sensor data's multi-level features and eliminate redundancy. At the modelling level, this approach facilitates the development of a multi-channel feature modelling structure that integrates convolutional neural networks, long short-term memory networks, and transformers. This structure captures local patterns, global temporal dependencies, and long-range interactions, thereby improving the expression and discrimination of fault features in sensor data. Experimental results demonstrate that the multi-level feature fusion-based bearing fault diagnosis method achieves an accuracy of up to 99.73%, approximately 1.5 percentage points higher on average than other typical algorithms. This method shows significant potential for diagnostic accuracy and operational reliability in complex sensor network environments.
Ensuring the reliable operation of programmable logic controller control units is essential for industrial system safety. Traditional diagnostic approaches, which rely on manual inspection and delayed response, prove inadequate for the predictive maintenance requirements of smart manufacturing. This study tackles challenges in multi-sensor data fusion and model adaptability by developing an intelligent framework that combines digital twin technology with meta-learning. The approach integrates vibration and current signals through an attention-based fusion network, enabling rapid adaptation to new equipment with minimal data samples. Experimental results demonstrate exceptional performance: a Matthews correlation coefficient of 0.965 for fault classification and a root mean square error of 0.048 for remaining useful life prediction. These results significantly surpass current state-of-the-art methods, improving diagnostic accuracy by over 3.4% and prediction precision by more than 17% compared to the best existing baseline.
Wireless sensor networks are fundamental to the internet of things, yet are highly vulnerable to insider attacks that traditional cryptographic defences cannot mitigate. Existing trust evaluation models often rely on single-view learning or static fusion mechanisms, which lack adaptability and interpretability. To overcome the limitations of single-model approaches and static fusion mechanisms in wireless sensor network trust evaluation, this paper proposes a stacked ensemble deep learning framework. The framework coordinates three heterogeneous deep-learning base learners - a one-dimensional convolutional neural network, a long short-term memory network with attention, and a graph attention network - to extract behavioural, temporal, and topological trust evidence. Evaluated on a public wireless sensor network security dataset, the proposed model achieves a root mean square error of 0.083, a weighted F1-score of 0.941, and a normalised discounted cumulative gain at rank 10 of 0.953, demonstrating statistically significant improvements over recent state-of-the-art baselines.
