2026 International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2026)

This paper adapts the lightweight time-series foundation model SEMPO to three-class bearing fault diagnosis using the Paderborn University dataset. The method combines a low-rate SEMPO pathway with a full-rate multiscale event encoder and partition-restricted aggregation of three adjacent windows. The SEMPO pathway retains reversible normalization, FFT-based energy decomposition, multi-frequency views, patch embedding, encoder, domain mixture-of-experts and prefix modules, and decoder statistical pooling. The event pathway uses parallel convolutional kernels of 7, 31, and 127 samples to capture local impulses and oscillatory bursts that may be attenuated by downsampling. Evaluation uses class-stratified random window sampling at the N15_M07_F10 operating point, with 64%/16%/20% train/validation/test partitions and three independent seeds. The complete EventAggFocalMarginSupCon model achieves 98.16% accuracy and 98.16% Macro-F1, compared with 62.71% accuracy and 62.66% Macro-F1 for the SEMPO-only DualBase. Healthy, outer-race, and inner-race recalls reach 100.00%, 97.71%, and 96.77%, respectively. The protocol measures same-distribution window classification because windows from a physical bearing may occur in multiple partitions; it is not presented as unseen-bearing generalization.

bearing fault diagnosis, SEMPO, Paderborn dataset, stratified sampling, multiscale convolution, supervised contrastive learning