Knowledge–data co-driven hierarchical traction control of distributed-drive electric vehicles without explicit online road-friction estimation

Traction control of distributed-drive electric vehicles (DDEVs) on cold-region, spatially non-uniform low-adhesion roads must coordinate slip suppression, yaw stability, energy consumption, and torque smoothness under uncertain tire–road interactions. Existing optimal-slip controllers rely on online road-adhesion estimation, whereas end-to-end learning controllers often lack an interpretable and actuator-feasible safety interface. This study develops a knowledge–data co-driven hierarchy that combines Long Short-Term Memory–Soft Actor-Critic (LSTM-SAC) with Super-Twisting Algorithm Sliding Mode Control (STA-SMC), denoted LSAC-STA-SMC, to achieve traction anti-slip control without explicit online road-friction estimation. The LSTM-SAC policy infers road-condition effects from historical vehicle-response signals and generates a four-dimensional structured reference-slip action, which is constrained by a safety projection before being tracked by the STA-SMC layer embedding wheel rotational dynamics and actuator limits. The method was evaluated in MATLAB/Simulink–CarSim co-simulation using controlled ablations, representative deep reinforcement learning baselines, an out-of-training checkerboard test, and N = 2 , 500 matched randomized trajectories with multi-source latency disturbances. Compared with an analytical optimal-slip baseline using ground-truth road adhesion, LSAC-STA-SMC reduced RMS Torque Rate by 54.6% and Slip Ratio > 15% Occurrence by 67.0%, with large distribution-level effects p < 0.001 , ε 2 ≥ 0.858 , while maintaining comparable energy consumption and traction performance. Vehicle experiments across six cold-region scenarios further demonstrated real-time implementation on a Powertrain Domain Control Unit at a 0.02-s control period. These results indicate that reference-slip-space learning with physics-based constrained execution offers a feasible route for multi-objective traction control within the evaluated non-uniform low-adhesion conditions.