Semi-supervised weighted stacked autoencoder with spectral peak significance constraints for imbalanced fault diagnosis under low labeled rates

To overcome the severe performance degradation and majority-class bias of traditional fault diagnosis methods under practical scenarios of extreme label scarcity and inherent class imbalance, this paper proposes a semi-supervised weighted stacked autoencoder with spectral peak significance constraints (SSWAEF). A nearest-neighbor consistency voting strategy coupled with a reciprocal-class-size sampling mechanism is first developed to construct a high-quality, class-balanced pseudo-labeled dataset. At the feature learning stage, considering the physical nature that mechanical faults typically manifest as energy concentrations at specific frequencies, a physics-informed weighted loss is designed. By incorporating power spectral density peak significance, this constraint amplifies gradients in critical frequency bands, forcing the network to preferentially extract discriminative fault structures rather than fitting broadband noise. For the fine-tuning stage, a dual-weighted cross-entropy loss is constructed, which integrates class-balancing weights and instance-confidence weights to ensure robust learning from minority classes without being misled by low-quality pseudo-labels. Extensive experiments on the Paderborn University dataset, a laboratory dataset, and an industrial field dataset validate the superiority of the proposed method. Under the extreme scenario with an imbalance and labeled rate of 0.2/0.2, SSWAEF maintains high accuracies of 99.63, 92.96, and 94.44% across the three datasets, respectively, demonstrating its exceptional robustness and diagnostic performance.