Partial Label Learning for Bearing Fault Diagnosis via Residual Shrinkage and ETF-Guided Prototype Learning

Accuratebearing fault diagnosis is challenged by noise-contaminated vibration signals and ambiguous fault annotations. To address these challenges, this paper proposes RS-ETFPLL, a residual shrinkage and equiangular tight frame (ETF)-guided partial label learning (PLL) framework. Vibration signals are transformed into time–frequency representations using the continuous wavelet transform, and an adaptive residual shrinkage encoder is employed for noise-robust feature extraction. Fixed simplex ETF prototypes provide a stable and uniformly separated class geometry. Both the classification and prototype branches are trained with candidate-set supervision, while pseudo-soft labels obtained by normalizing classification predictions within candidate sets guide prototype learning to facilitate label disambiguation. Experiments on the CWRU, MFPT, and PU datasets demonstrate that RS-ETFPLL outperforms six representative PLL baselines across the evaluated ambiguity levels. In the main comparison at an incorrect-label inclusion probability of q=0.5, it achieves mean accuracies of 99.59%, 98.99%, and 98.01%, surpassing the best baseline in each setting by 2.03, 4.46, and 2.16 percentage points, respectively. At the same ambiguity level and a signal-to-noise ratio of −2 dB, the corresponding accuracies are 81.72%, 82.08%, and 73.78%. These results demonstrate the effectiveness of RS-ETFPLL for bearing fault diagnosis under the evaluated noise and candidate-label ambiguity conditions.