A Hybrid Data-Driven Soft Sensor Framework for Torque Estimation

Efficient torque estimation plays an important role for the real-time durability analysis of vehicle components. It is desired to replace expensive torque sensors by applying soft sensor methods to accurately estimate the torque signals. However, due to the occurrence of the dead-zone phenomenon in torque signals on part-time four-wheel-drive (PT-4WD) vehicles when accelerating or braking (BK), normal linear identification methods, like multiple-input single-output finite impulse response (MISO-FIR), may not be directly applicable for torque estimation. A novel data-driven soft sensor method including an estimator and a classifier is proposed in this article. First, a logistic regression (LR) based classifier uses the low sampling-frequency input data to detect the time periods when the dead-zone phenomenon happens. Then, an estimator, using MISO-FIR, is applied to estimate the target output based on the other known sensor signals over time periods detected by the classifier. As a result, a complicated nonlinear system identification problem has been solved with the proposed bounded-input bounded-output and explainable method. The proposed hybrid method is validated on multiple experiments with historical datasets where the dead-zone phenomenon occurs. Extensive experiments demonstrate that the proposed method outperforms multiple baseline methods in the comparison study, achieving a smaller normalized mean squared error (NMSE) and a larger goodness of fit (FIT) with a considerably lower computational cost.