Multi-Channel Feature Fusion for Gearbox Fault Diagnosis based on CNN-Transformer

It is common to deploy multiple vibration sensors at different locations on machines for comprehensive data acquisition and more accurate fault diagnosis. This paper presents a novel multi-channel fusion model for end-to-end fault diagnosis. The proposed architecture employs a hybrid approach, utilizing a one-dimensional Convolutional Neural Network (1D-CNN) for local feature extraction and a Transformer network for capturing global features. These extracted features are then adaptively weighted and fused across channels. Experimental evaluation using the XJTU Gearbox Datasets demonstrates that our model achieves superior diagnostic accuracy compared to single-channel approaches, with improvements of 4.9% and 2.2% over the two single-channel baselines on the mixed dataset.