Steering feedback torque prediction using a driving-condition clustering-enhanced sequence-to-sequence network

Abstract

Accurate prediction of Steering Feedback Torque (SFT) is a key challenge in steer-by-wire (SbW) development, as it directly determines the realism and controllability of the steering experience. Traditional physics-based models rely on complex calibration and struggle to capture the nonlinear, time-varying dynamics inherent in real-world vehicle signals. To address this, a clustering-enhanced SFT prediction framework that integrates driving state clustering with a sequence-to-sequence (S2S) network is proposed. A multi-channel signal dataset, comprising steering, inertial, and wheel-speed measurements, is acquired from on-road experiments on an instrumented test vehicle across diverse driving scenarios. Driving features extracted from these signals are first clustered via unsupervised learning, and the resulting cluster membership probabilities are embedded as supplementary features into the multivariate time series input. This enhances the model's generalization across driving conditions and improves interpretability by linking model behavior to identifiable driving patterns. An S2S network is then trained to capture the temporal dependencies specific to the steering system. Hardware-in-the-loop experiments with driver-in-the-loop operation confirm that the proposed method outperforms conventional static regression models and non-clustered baselines in both common and extreme driving scenarios.