A label-free adaptive signal reconstruction strategy within a supervised framework for rolling bearing fault diagnosis

Abstract

Rolling bearing fault diagnosis remains challenging because weak impulsive components are easily masked by mechanical interference, and the performance of variational mode decomposition is highly sensitive to parameter selection and modal screening. To address these issues, this study proposes a label-free adaptive signal reconstruction framework in which an improved crested porcupine optimizer is used to optimize VMD parameters without using fault label information. Label-free refers only to parameter optimization and IMF selection, not the classification stage. A composite IMF selection index combining envelope entropy and Pearson correlation is further designed to identify fault-sensitive modes for signal reconstruction. The reconstructed signals are transformed into concentrated time-frequency representations using synchrosqueezed wavelet transform, and a dual-branch ICNN-BiLSTM network is developed to learn multi-scale spatial textures and temporal dependencies. Experiments were conducted on the CWRU and Paderborn bearing datasets using physically isolated data partitioning, repeated trials, ablation studies, and comparative evaluations. The results demonstrated that the proposed method achieved average accuracies of 99.75% ± 0.22% and 99.88% ± 0.10% on the CWRU and Paderborn datasets, respectively. Precision, recall, macro-F1, per-class F1-score, and confusion matrix analyses further confirmed the class-wise diagnostic performance of the proposed framework. This framework alleviates classification confusion caused by damage patterns with highly similar physical morphologies. The results indicate that the proposed framework is effective for bearing fault diagnosis under the evaluated fixed-load benchmark conditions.