Elevator fault diagnosis incorporating multi-scale prototype diffusion enhancement

Abstract To address the challenges of limited fault samples, class imbalance, and similar time–frequency characteristics in elevator fault diagnosis, a small-sample fault diagnosis method integrating the multi-scale texture and prototype-constrained diffusion network (MTPDNet) data augmentation model with the EGA-ResNet diagnostic network is proposed. First, elevator car vibration signals are collected, and two-dimensional fault time–frequency maps are constructed using continuous wavelet transform. Second, MTPDNet is developed based on a class-conditional diffusion model, in which multi-scale time–frequency texture enhancement, prototype constraints, and a hybrid structural loss are introduced to improve the texture clarity, structural integrity, and class consistency of the generated samples. Finally, the generated samples are combined with real samples to train EGA-ResNet, where the efficient channel attention attention mechanism and generalized mean pooling are incorporated to enhance discriminative fault feature extraction. Experimental results show that the samples generated by MTPDNet achieve better visual quality and smaller distribution discrepancy from real samples, as evaluated by maximum mean discrepancy. EGA-ResNet achieves a diagnostic accuracy of 96.07%, outperforming the baseline ResNet. Under different class imbalance ratios, the proposed method consistently achieves superior diagnostic performance compared with real-sample-only training, loss-based imbalance handling, traditional data augmentation, and other generative augmentation methods. In addition, the model maintains diagnostic accuracies of 93.53%, 89.87%, and 84.80% under noise conditions of 20 dB, 10 dB, and 5 dB, respectively, demonstrating its robustness to noise interference. The proposed method effectively alleviates the adverse effects of limited fault samples and class imbalance and improves the accuracy and robustness of elevator fault diagnosis.