基於強化學習於小型電動全地形車輛驅動控制與能量管理之研究;Study on Traction Control and Energy Management of Small Pure Electric All-Terrain Vehicles Based on Reinforcement Learning

NCUIR > Engineering College > Graduate Institute of Mechanical Engineering > Electronic Thesis & Dissertation > Item 987654321/99430

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/99430

Title:	基於強化學習於小型電動全地形車輛驅動控制與能量管理之研究;Study on Traction Control and Energy Management of Small Pure Electric All-Terrain Vehicles Based on Reinforcement Learning
Authors:	羅裕譯;Lo, Yu-Yi
Contributors:	機械工程學系
Keywords:	電動全地形車輛;強化學習;驅動力控制;能量管理;Electric all-terrain vehicle;Reinforcement learning;Traction control;Energy management
Date:	2025-10-02
Issue Date:	2026-03-06 18:58:33 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究針對小型純電動全地形車(All-Terrain Vehicle, ATV)提出一套智慧型驅動力控制演算法，旨在改善駕駛穩定性並提升續航表現。為發展與驗證此控制演算法，本研究首先建立了結合ATV縱向動態模型與駕駛者行為模型的模擬平台。在進行效能評估時，本研究設計了一種基準控制模式，名為SPORT模式，以模仿一般未經能量優化的駕駛行為，將油門輸入與馬達扭矩進行直接線性對應。此基準模式將作為對照組，用以量化本研究所提出之智慧控制演算法在節能與穩定性方面的具體成效。在模擬系統中，首先進行了系統參數識別與狀態估測，以獲得如車重及車速等關鍵控制參數。其中，車重採用遞迴最小平方法(Recursive Least Squares Estimation, RLSE)，在車輛滑行過程中進行即時辨識。模擬結果顯示，在輪速訊號之訊號雜訊比(Signal-to-Noise Ratio, SNR)分別為40 dB與30 dB的條件下，車重估測誤差可控制在5%與12%以內，證實了所提方法在雜訊環境下的穩健性。由於本研究之小型電動ATV僅配置輪速感測器與加速規，因此利用卡爾曼濾波器同時估測驅動力與車速，並解決加速規存在的零偏誤差。模擬結果顯示，當加速規訊號存在均值為2 m/s²的變動零偏誤差時，此方法能將偏移誤差修正至-0.084 m/s²。修正後的加速度訊號與輪速訊號進行訊號融合後，所估測的車速誤差可限制在 ±0.2 m/s 的範圍內，並可有效估測驅動力。在驅動控制方面，本研究採用最大可傳遞扭矩估測(Maximum Transmissible Torque Estimation, MTTE)方法以動態限制馬達最大輸出扭矩，避免車輛行駛於低摩擦路面上可能發生的打滑狀況。將本研究改良之最大可傳遞扭矩估測技術與傳統文獻所提方法的模擬結果進行比較，可知本研究透過整合卡爾曼濾波器之改良MTTE方法(KF-MTTE)，較傳統使用低通濾波器之MTTE方法(LPF-MTTE)能更準確地達到驅動力控制，並有效將滑移率抑制於合理範圍內。且在低摩擦係數路面進行全油門驅動測試時，本研究之改良KF-MTTE方法所對應的滑移率可有效控制於0.2以內，已較SPORT模式所對應的輪胎最大滑移率0.9顯著改善，驗證本研究提出之KF-MTTE方法可顯著提升車輛於惡劣路面下的驅動穩定性。本研究針對評估了兩種無模型強化學習(Reinforcement Learning, RL)演算法，即Q-learning與SARSA，以實現最佳化能量管理策略，並動態決定最佳的馬達輸出扭矩。在10公里的模擬測試中，傳統SPORT模式的能量消耗為2946 kJ。而應用Q-learning與SARSA所得的最佳能量管理策略，其能量消耗分別為2543 kJ與2479 kJ，相較於SPORT模式分別降低了13.7%與15.8%。此結果驗證了強化學習演算法能有效調整扭矩輸出，從而提升能源使用效率。此外，雖然SARSA在特定模擬測試中展現較佳的節能效果，但由於其on-policy特性，探索能力相對受限。相較之下，Q-learning具備更強的探索性與跨情境適應能力，因此本研究進一步選擇Q-learning作為線上強化學習之方法。驗證結果顯示，線上學習能在新任務模式下持續修正策略，在500公尺行駛任務下消耗135 kJ，相較於離線策略消耗142 kJ，降低約4.9%的能量消耗。由於離線訓練已涵蓋多樣駕駛模式，形成了有效基礎策略，使降幅有限，但線上學習仍展現出即時調整與提升效能的優勢。綜合而言，本研究在模擬測試階段所開發的智慧控制策略，成功整合了即時估測、驅動力控制與強化學習技術。此策略展現出顯著的節能效益與穩健性，為未來智慧型越野電動車控制技術的發展提供重要的參考依據。 ;This study proposes an intelligent drive control algorithm for a small pure electric all-terrain vehicle (ATV), aiming to improve driving stability and extend driving range. To develop and validate the proposed algorithm, a simulation platform combining the ATV longitudinal dynamics model and the driver behavior model was first established. A simulation platform combining the ATV longitudinal dynamics and driver behavior models was established for development and validation. For performance evaluation, a baseline SPORT mode—directly mapping throttle input to motor torque—was designed as a conventional non-optimized reference. In the simulation, system parameter identification and state estimation were first performed to obtain key control parameters such as vehicle mass and speed. Vehicle mass was identified in real time during free-rolling tests using the Recursive Least Squares Estimation (RLSE) method. Results showed that under wheel speed signals with SNRs of 40 dB and 30 dB, mass estimation errors remained within 5% and 12%, demonstrating robustness against noise. Since the ATV is equipped only with wheel speed sensors and an accelerometer, a Kalman filter was applied to simultaneously estimate driving force and vehicle speed while correcting the accelerometer’s zero-bias error. When the accelerometer signal contained a variable bias with a mean of 2 m/s², the method reduced the error to -0.084 m/s². After fusing the corrected acceleration with the wheel speed signal, the estimated vehicle speed error was constrained within ±0.2 m/s, while the driving force was accurately estimated. For drive control, the proposed KF-MTTE method effectively limited motor torque, suppressing slip ratio within 0.2 under full-throttle low-friction tests, compared with 0.9 in SPORT mode, thereby enhancing stability. For drive control, the proposed KF-MTTE method effectively limited motor torque, suppressing slip ratio within 0.2 under full-throttle low-friction tests, compared with 0.9 in SPORT mode, thereby enhancing stability. This study further evaluated two model-free reinforcement learning algorithms, Q-learning and SARSA, for optimizing energy management and torque control. In a 10 km simulation, the SPORT baseline consumed 2946 kJ, while Q-learning and SARSA reduced consumption by 13.7% and 15.8%, respectively, confirming the effectiveness of RL-based torque adjustment. Although SARSA achieved slightly better savings, its on-policy nature limits exploration. Q-learning, with stronger adaptability, was thus adopted for online learning. Validation results demonstrated real-time adaptability, with online learning consuming 135 kJ compared to 142 kJ under the offline strategy in a 500 m task, achieving an additional 4.9% reduction in energy consumption. In summary, the proposed intelligent control strategy integrates real-time estimation, drive control, and reinforcement learning, achieving notable improvements in both energy efficiency and operational robustness, and offering insights for the future development of intelligent control technologies in electric all-terrain vehicles.
Appears in Collections:	[Graduate Institute of Mechanical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	60	View/Open

社群 sharing

Loading...