ESPKARN: a multichannel framework for bearing fault diagnosis in the presence of imbalanced samples

The representation of minority-class samples is enhanced by a multichannel fusion framework, where discriminative features are extracted from multidimensional signals to mitigate class imbalance in bearing fault diagnosis. However, the application of multichannel feature fusion in real-world engineering systems is hindered by two factors: temporal-spatial misalignment between heterogeneous sensor data streams and feature space deterioration due to overlapping channel information of minority-class samples. To overcome these challenges, an entropy-guided spatial pyramid Kolmogorov–Arnold residual network (ESPKARN) is proposed. Initially, the application of spatial pyramid pooling (SPP) is innovatively broadened to capture the physical spatial relationships within multichannel data, resolving the problem of spatial semantics loss in time-series data. Subsequently, a novel residual module, augmented with the Kolmogorov–Arnold Network, is introduced to refine feature fusion further. The fused features are then subjected to residual learning and nonlinear mapping mechanisms to optimize the overall feature representation. Finally, a uniquely multichannel framework is proposed, complexity analysis using multiscale sample entropy, multiscale spatial fusion via SPP, and nonlinear optimization within Kolmogorov–Arnold enhanced Residual. ESPKARN is extensively evaluated using two sets of multichannel bearing data featuring varying imbalance ratios. In a severe class-imbalance scenario where healthy data samples are 35 times more numerous than each fault class, the proposed method achieves accuracy rates of 99.80% and 88.62%, respectively, outperforming other fault diagnosis approaches.