| Title |
Frequency Selective Feature Extraction from In-tire Acceleration Signals for Tire State Estimation |
| Authors |
허수웅(Suwoong Heo) ; 문혜원(Hyewon Moon) ; 이권우(Kwonwoo Lee) ; 정승현(Seunghyun Chung) ; 유성한(Sunghan You) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.6.81 |
| Keywords |
Learnable Sinc filterbank; Tire state estimation; Tire acceleration signals; Intelligent tire system; Deep learning |
| Abstract |
Tire acceleration signals are non-stationary and are jointly affected by speed, pressure, load, and alignment conditions. Therefore, fixed frequency bands or generic convolution feature extraction are limited in capturing condition-dependent spectral variations and target-specific frequency selectivity. This paper proposes a frequency-selective feature extraction framework that combines a learnable Sinc filterbank with condition informed band mixing. The proposed method first decomposes tri-axial in-tire acceleration signals into interpretable multi-band frequency representations, then adaptively reweights those bands according to operating conditions, and finally estimates a cycle-level state value through a TCN-based temporal encoder. The same framework is trained independently for wheel load, toe angle, and camber angle estimation. Experiments on drum-test data collected from a commercial TBR tire show consistent performance improvements over the baseline across all estimation targets. Ablation results further verify the contributions of the learnable Sinc filterbank and condition informed band mixing, and a representative case analysis shows that the proposed model learns target-relevant frequency representations in an interpretable manner. |