مشخصات پژوهش

Kolmogorov-Arnold Networks (KANs) represent a paradigm shift in neural architectures by employing learnable univariate functions on edges, inspired by the Kolmogorov-Arnold representation theorem. In this study, Convolutional KANs (CKANs), including B-spline-based and ReLU-spline-based variants, are benchmarked against traditional Convolutional Neural Networks (CNNs) for bearing fault detection using the Case Western Reserve University (CWRU) dataset. Ten fault types are classified from drive-end vibration signals, and performance is evaluated on raw data as well as under noisy conditions with signal-to-noise ratios (SNR) ranging from -6 dB to +6 dB. Accuracy, precision, recall, and F1-score are employed as key evaluation metrics. The experimental results indicate that CKANs fail to outperform CNNs, exhibiting higher sensitivity to noise, greater susceptibility to overfitting without regularization, and increased computational demands in terms of parameters and FLOPs. Although ReLU-KANs alleviate some efficiency issues, they remain slower. These findings highlight the limitations of CKANs in classification tasks and suggest that tailored architectures are required to fully exploit their interpretability advantages.