کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Cross-Attention on Multitaper-Reassigned Time-Frequency Maps for Mental Workload Classification

نویسندگان :

Sedighe Mirbolouk¹ Hashem Kalbkhani²

1- دانشگاه ارومیه 2- دانشگاه صنعتی ارومیه

کلمات کلیدی :

Mental workload،multitaper spectrogram،time–frequency representation،cross-attention،reassignment،EEG

چکیده :

Mental workload (MWL) decoding from electroencephalography (EEG) is a core capability for neuroergonomics and adaptive human–machine systems. Yet it doesn't remain easy due to strong nonstationarity and substantial inter-subject variability. This paper proposes an end-to-end framework, cross-attention on multitaper–reassigned time–frequency maps (CA-MTR-TF), that couples a variance-stabilized time–frequency (TF) front-end with an interaction-aware attention backbone. For each EEG channel, we construct a three-stream TF tensor comprising the multitaper power map, a reassigned TF map for sharper localization, and a stability-derived confidence map computed from multitaper variance. The network then applies a redesigned TF cross-attention module that explicitly models time↔frequency dependencies and uses confidence to downweight unreliable TF regions. Experiments on two public MWL benchmarks (STEW and EEGMAT) report consistent improvements over feature-based and attention-light variants, achieving 97.4 ± 0.4% accuracy on STEW and 98.8 ± 0.3% on MAT. These results indicate that combining multitaper stability, reassignment sharpening, and confidence-guided TF cross-attention provides a robust representation for subject-wise MWL decoding.