Uncertainty-Aware Adaptive Multi-Modal Fusion for Robust Bearing Fault Diagnosis Using Transformer Networks in Noisy Environments

This paper presents an Uncertainty-Aware Adaptive Multi-Modal Fusion (UAMF) framework for robust bearing fault diagnosis leveraging synchronized vibration and motor current signals. The proposed method converts multi-channel time-series data into normalized grayscale images, which are processed through a shared convolutional neural network (CNN) to extract local features. Subsequently, modality-specific Transformer encoders are employed to model temporal dependencies within each signal source. To address input uncertainty and sensor noise, Monte Carlo Dropout is applied during inference to estimate epistemic uncertainty for each modality. A key contribution of UAMF is its dynamic fusion strategy, which adaptively assigns higher weights to modalities with lower uncertainty, thereby reducing reliance on noisy or unreliable inputs. The resulting fused features are passed through a fully connected classification layer for fault prediction. Experimental validation on a benchmark rotating machinery dataset demonstrates that UAMF achieves superior diagnostic performance—reaching up to 99% accuracy—and exhibits enhanced robustness under various noisy conditions compared to conventional CNN- and Transformer-based baselines, including FAN-BD. The proposed framework improves diagnostic reliability in practical, sensor-degraded industrial environments.