Prior-guided adaptive Kalman filtering for longitudinal speed estimation of four-wheel-drive electric vehicles under varying adhesion conditions

Longitudinal speed estimation becomes difficult when a vehicle runs through roads with changing adhesion, because the wheel-speed signals can be strongly affected by slip, wheel lock-up, and short road impacts. This problem is more obvious for four-wheel-drive electric vehicles, where the estimator usually has to work with production wheel-speed sensors and an onboard IMU. To improve the robustness of this low-cost sensor fusion problem, this paper presents a prior-guided adaptive Kalman filtering method for longitudinal speed estimation. The longitudinal acceleration measured by the IMU is first corrected by considering vehicle pitch, road slope, and lateral-motion coupling. A wheel-speed observation is then built from the consistency of the four-wheel speeds, so that obviously unreliable wheel-speed information can be weakened before it enters the filter. In addition, road-adhesion-related features are extracted online and used in a hidden Markov inference process. The obtained state probabilities are directly introduced into the prediction-stage process-noise covariance, allowing the filter to adjust the balance between acceleration-based prediction and wheel-speed correction before large wheel-speed residuals are accumulated. Real-vehicle tests were carried out on a four-wheel-drive electric SUV under three representative conditions, including high-to-low adhesion transition, low-to-high adhesion transition, and one-side road-step braking. To further examine case-to-case variability, nine additional high–low adhesion-transition runs were included, resulting in 12 real-vehicle validation cases in total. In these 12 cases, the proposed method achieved the lowest mean absolute error in all cases and the lowest maximum absolute error in 9 out of 12 cases. Its average mean absolute error and maximum absolute error were 0.1181 and 0.3480 m/s, respectively. Within the tested validation cases, these results indicate that the proposed method can improve transient robustness while keeping a simple recursive structure suitable for real-time rapid-prototyping implementation.