Abstract
In bearing fault diagnosis for rotating machinery, multi-sensor data-such as acoustic and vibration signals-are increasingly leveraged to enhance diagnostic performance. However, existing methods often rely on complex network architectures and incur high computational costs, limiting their applicability in real-time industrial environments. To address these challenges, this study proposes a lightweight and efficient multi-sensor ensemble framework that achieves high diagnostic accuracy while minimizing computational overhead. The proposed method transforms vibration and acoustic signals into spectrograms, which are independently processed by modality-specific lightweight convolutional neural networks (CNNs). The softmax outputs from each classifier are integrated using an AdaBoost-based ensemble strategy that emphasizes high-confidence predictions and adapts to sensor-specific misclassification patterns. Experimental results on benchmark datasets-UORED-VAFCLS, KAIST, and an in-house bearing dataset-demonstrate an average classification accuracy exceeding 99.90%, with notable robustness against false positives and missed detections. Furthermore, the framework significantly reduces resource consumption in terms of FLOPs, inference latency, and model size compared to existing state-of-the-art multi-sensor fusion approaches. Overall, this work presents a practical and deployable solution for real-time bearing fault diagnosis, balancing classification performance with computational efficiency without resorting to complex feature fusion mechanisms.