Cross-domain intelligent fault diagnosis with superlet spectrograms and a parameter-efficient CNN-Transformer hybrid

Smart fault diagnosis is a crucial component of modern industrial systems, offering early detection of machine conditions to prevent costly breakdowns and safety issues. While state-of-the-art deep learning models are capable of near-perfect classification on controlled laboratory data, they often perform poorly in practice because of noisy environments, class imbalance and distribution shifts across operating conditions. This paper addresses these challenges by re-formulating industrial acoustic monitoring as a supervised image classification problem with superlet time-frequency representations, which capture both the transient dynamics and phase-envelope. An efficient CNN-Transformer hybrid model is proposed by placing a lightweight Transformer encoder on an ImageNet-pretrained ResNet-50 backbone with spatial attention gating to learn both local and global spectro-temporal features. The proposed architecture is lightweight, with 25.6M parameters and 4.1G FLOPs, supporting efficient resource usage. An asymmetric focal contrastive learning strategy is also proposed to improve anomaly discrimination under extreme imbalance. Experimental results on the MIMII dataset demonstrate that the proposed framework achieves 94.82% accuracy, 91.08% balanced accuracy, 91.72% F1-score, and 0.9758 ROC-AUC under a strict group-aware evaluation protocol, while requiring only 25.6M parameters, 4.1G FLOPs, and 4.5 ms inference time per sample. In cross-domain evaluation on the CWRU dataset, the model further achieves a zero-shot AUC of 0.8886 and a fine-tuned AUC of 1.0000 within three epochs. These results validate that the proposed framework provides an effective, robust, and generalizable solution for industrial fault diagnosis.