Chaotic dynamic feature enhancement and adaptive integrated learning framework for bearing fault diagnosis under unbalanced small sample conditions

To address the challenges of insufficient training samples and class imbalance in bearing fault diagnosis, a novel diagnosis framework termed ChaoFEX-Ensemble is proposed. This method integrates time-frequency domain weighted preprocessing with nonlinear dynamic feature extraction using chaotic neurons, followed by adaptive classification via ensemble learning. Firstly, one-dimensional time-domain vibration signals are transformed into the frequency domain via fast Fourier transform (FFT) and subsequently fused to form enhanced time-frequency representations. These signals are then input into a chaotic neural network, wherein chaotic GLS (Generalized Logistic Sequence) neurons are stimulated to generate ChaoFEX features, thereby mitigating the issue of limited sample sizes. To handle class imbalance, the EasyEnsemble classifier is employed, which constructs multiple balanced subsets and trains a series of weak learners for robust classification. Finally, bearing fault samples are evaluated using k-fold cross-validation to determine the optimal hyperparameters. The proposed method is validated on the PU dataset, a laboratory bearing test platform, and an industrial dataset. The macro F1-scores achieved are 0.98 and 0.97 on the PU and experimental platform datasets, respectively, demonstrating the effectiveness and generalization capability of the proposed approach.