CaMT-iTire: Context-Aware Multi-Task Learning for Intelligent Tire Framework

Intelligent tire systems (ITS) enable direct sensing of tire and road interaction using sensors mounted inside the tire, providing information that is difficult to obtain from conventional vehicle-mounted sensors. However, acceleration signals measured from tire-mounted inertial measurement unit (IMU) sensors are influenced by multiple physically coupled factors, including vertical load, wheel alignment, and varying operating conditions. This coupling limits the robustness of conventional single-task estimators and multi-task learning approaches in real-world driving environments. In this paper, we propose CaMT-iTire, a context-aware multi-task learning framework designed to jointly estimate wheel load and wheel alignment parameters (toe and camber angles) from a single tire accelerometer sensor mounted on inner liner of a tire. The proposed framework combines a learnable Sinc-based filterbank with a condition-aware convex band-mixing layer to perform adaptive frequency-selective feature extraction from raw acceleration signals. In addition, a Progressive Layered Extraction (PLE) backbone with context-aware gating is introduced to mitigate inter-task interference by dynamically routing features according to operating conditions. Extensive experiments conducted on both controlled drum-test and real world on-road driving data demonstrate that the proposed CaMT-iTire framework consistently outperforms conventional single-task and standard multi-task learning baselines. The results show improved estimation accuracy and robustness, particularly across the operating conditions covered in our datasets, highlighting the effectiveness of the proposed context-aware multi-task learning approach for intelligent tire.