Bearing Fault Diagnosis with a KAN Transformer and a Multi-Granularity Spline Ensemble

This study investigates a multi-granularity KAN transformer for vibration-based bearing-fault diagnosis. Each of four encoder blocks contains three independently parameterized cubic-spline FFNs with grids of three, five, and seven, combined through trainable fusion weights. The 10-class network has 1,000,918 parameters. Raw sample splitting precedes overlapping-window extraction, and normalization uses training data only. On the 10-class CWRU protocol, the five-seed accuracy and macro-F1 are 99.41 ± 0.20% and 99.41 ± 0.20%, respectively, versus 97.95 ± 0.23% for a structurally matched transformer and 98.64 ± 0.16% for a parameter-matched transformer. On a recording-disjoint three-state Paderborn subset, the accuracy reaches 99.06 ± 0.14%. Three paired component comparisons show gains of 0.45, 0.41, and 0.59 percentage points, with a Holm-adjusted p≤0.0032. Per-class results, harmonic interventions, additive-noise sensitivity, edge-function decompositions, and inference profiling further characterize the design. The proposed model leads the evaluated controlled baselines; a comparison to all published benchmark configurations is outside the scope of these data.