The roller of belt conveyor is susceptible to rust, wear and other faults in long-term operation. The early fault acoustic signal is weak and easily submerged by strong noise in the industrial environment. The traditional diagnosis method has the problems of insufficient feature extraction and poor robustness. In this paper, a fault diagnosis method based on CNN-Swin Transformer dual-branch feature fusion network (CSF) is proposed. The time-frequency domain feature matrix is constructed by fusing variational mode decomposition and fast Fourier transform. Combined with the local feature extraction advantages of CNN and the global attention mechanism of Swin Transformer, the SE-CGA attention mechanism is designed to achieve deep feature extraction. Experiments show that the method achieves a test accuracy of 98.11 % on real industrial data sets, which is more than 9 % higher than that of a single CNN model. The recognition accuracy of 65.59 % is still maintained when superimposing-15 dB extreme noise, which verifies its diagnostic robustness and engineering application value in strong noise scenarios.