Multimodal mechanical equipment fault diagnosis using dual-channel convolutional neural network

Abstract Mechanical fault diagnosis models often suffer from incomplete feature characterization due to single-modality inputs and generalization under complex noise and operating-condition variations. This paper proposes a lightweight multi-modal network, TSAF-Net. First, a random-overlapping segmentation strategy is employed to perturb sampling positions and improve information coverage, while generating synchronized dual-modal inputs: 1D time-domain segments and their corresponding CWT-based time–frequency grayscale maps, enabling complementary representations of transient impacts and time–frequency textures. Second, TSAF-Net adopts a parallel two-branch architecture: the 1D branch uses a CNN to screen hierarchical local temporal patterns and compress the sequence, upon which a Transformer models long-range dependencies for joint local–global representation learning; the 2D branch leverages depthwise separable convolutions for efficiency and integrates CBAM to emphasize fault-sensitive regions in both channel and spatial dimensions. The fused features are finally used for fault classification. Experiments on the CWRU bearing and gearbox datasets demonstrate that TSAF-Net outperforms representative baselines, and ablation studies confirm the effectiveness of the dual-modality design and fusion modules in reducing misclassification. Under robustness evaluations, TSAF-Net achieves 91.9% accuracy at SNR = −3 dB, and maintains over 90% accuracy under multiple non-Gaussian noises and unseen load conditions, indicating strong robustness, generalization, and computational efficiency for industrial diagnosis.