Conventional thermal protection strategies for electromechanical compound transmission systems rely mainly on temperature thresholds and segmented power derating, resulting in reactive intervention and excessive power loss under prolonged overload and deteriorating cooling conditions. The proposed strategy treats thermal protection as a prediction-to-action problem: future temperatures of multiple critical components are estimated by an incrementally updated Neural ODE and incorporated, together with current thermal and operating states, into a PPO policy that outputs a continuous derating factor. This formulation enables one control decision to account for competing thermal risks while retaining vehicle power demand. The proposed framework is deployed on a physical controller and validated through controller hardware-in-the-loop experiments under multiple overheating scenarios. The temperature prediction RMSE remains below 0.43 °C. The temperature prediction model maintains good accuracy on previously learned conditions and reduces the average RMSE and MAPE by 16.71% and 9.91%, respectively, compared with LSTM. The closed-loop evaluation compares PPO with static stepwise derating, fixed-rate linear ramp derating, MPC, and SAC, together with a reference case without active derating. The experimental results show that, under eight representative thermal-risk scenarios, the PPO-based strategy can intervene proactively according to the system thermal state and continuously adjust the derating level, thereby effectively suppressing the temperature rise of the motor windings, bearings, and controller. Compared with static derating, the PPO-based strategy increases the NTSM by 0.0545–0.2094 and reduces the ITRI by 6.3–26.0%. The PDSR increases by 3.24–9.51 percentage points in five operating conditions, while decreasing by 2.82–2.88 percentage points in three high-speed-related conditions. Overall, the proposed strategy reduces the thermal risks of critical components and enlarges the thermal safety margin while balancing thermal safety and vehicle power demand.
High-speed traction motors in electric vehicles (EVs) subject their support bearings to severe mechanical and thermal loads. Bearing skidding and frictional heat generation can cause abnormal temperature rises, whereas uncertainty in convective heat transfer coefficients limits the accuracy of thermal simulations. This study investigates the skidding dynamics and temperature field of a motor deep-groove ball bearing (DGBB). A 35-degree-of-freedom dynamic model is established using the lumped-mass method to analyze rolling-element contact, skidding, and cage collisions at different speeds. A heat-generation and heat-transfer model is then developed based on bearing friction theory and moving-heat-source theory. The convective heat transfer coefficients are calibrated against measured temperatures using the non-dominated sorting genetic algorithm II (NSGA-II) within an integrated MATLAB–ANSYS framework, after which the temperature fields under constant- and variable-speed conditions are simulated. The results show that the rolling-element–raceway contact forces vary periodically, collisions between the rolling elements and cage are intermittent, and increasing speed intensifies the contact-force fluctuation and the ball–cage impacts and raises the skidding velocity, although the skidding ratio decreases. The component temperatures decrease in the order inner ring, rolling elements, outer ring and cage, and respond dynamically to changes in speed and load. The time-averaged cage speed agrees with the theoretical value to within 1% over 1000–10,000 rpm, and the calibrated thermal model reproduces the measured ring temperatures to within about 2 °C. The findings provide a useful reference for temperature prediction and thermal management of EV motor bearings.