自行車智能動力輔助系統控制之研究;Study of an Intelligent Power-Assist Control for Human–Electric Bikes

NCU Institutional Repository > 工學院 > 機械工程研究所 > 博碩士論文 > Item 987654321/98577

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98577

題名:	自行車智能動力輔助系統控制之研究;Study of an Intelligent Power-Assist Control for Human–Electric Bikes
作者:	邱威穎;Ciou, Wei-Ying
貢獻者:	機械工程學系
關鍵詞:	電動輔助自行車;強化學習;智能動力輔助系統;能量管理;Electric-Assist Bicycle;Reinforcement Learning;Intelligent Power Assist System;Energy Management
日期:	2025-05-20
上傳時間:	2025-10-17 12:56:58 (UTC+8)
出版者:	國立中央大學
摘要:	本文針對外掛式電動輔助自行車開發一套基於強化學習的智能動力輔助系統。該系統旨在提供符合駕駛者需求且可動態調整的輔助扭矩，以期提升續航里程、優化騎士的騎乘體驗，並滿足其個人化需求。相較之下，現行市售電動輔助自行車多採用傳統控制策略，難以根據即時騎乘狀況與駕駛者意圖動態調整輔助扭矩，常導致能源效率不彰。為克服此限制並發展節能型智能輔助系統，本研究首先建構了完整的自行車動態模型；進而，運用多項式逼近法建立擴展狀態估測模型，並結合自適應卡爾曼濾波器(Adaptive Kalman Filter, AKF)同步估算驅動力矩及車速，作為智能輔助決策與控制策略的關鍵依據。在此基礎上，本文採用強化學習演算法，並提出兩種基於驅動力矩控制策略模型(RL-T_f)及駕駛者扭矩控制策略模型(RL-T_user)的智能動力輔助控制策略，以模擬方式發展能適應駕駛者需求、騎乘環境及路況變化的智能動力輔助策略與能量管理機制，以輸出最佳化的輔助扭矩。並以模擬方式將RL-T_f與RL-T_user分別與傳統的固定比例控制策略(Proportion-assisted PowerControl, PAPC)及常見的輔助自行車速度控制(Pedelec Velocity Control, PVEC)進行比較，而PAPC與PVEC兩種控制方法應用於本研究測試上兩者表現相近，故本研究以PAPC作為主要對照基準。RL-T_f、RL-T_user兩者分別相比於PAPC，每公里能量消耗改善率分別為65.21%與36.36%，顯示RL-T_f、RL-T_user皆具備顯著的能量效率提升。為了驗證智能輔助策略，透過建置自行車實驗系統進行驗證，從實驗結果可知，在能量效率方面，RL-T_f能量效率均優於PAPC與PVEC，RL-T_f之能量效率相較於PAPC可提升58.41%。而RL-T_f較RL-T_user，於每100公尺可有效降低316.61J的電池能量損耗，並可同時減少駕駛者12.27%的踩踏力。針對即時踩踏輸入的反應能力，RL-T_user能提供立即性且較大的輔助扭矩，並減少駕駛者25.29%的踩踏力。根據電池續航里程評估，在相同電池配置(6.8 Ah/36 V)條件下，RL-T_f預估可行駛里程為295.61公里，而RL-T_user僅為143.41公里，雖兩者皆能有效提升電動輔助自行車的能量利用效率與續航里程，但RL-T_f更能展現顯著續航優勢。從模擬與實驗的輔助扭矩誤差分析可知，RL-T_f之相對誤差為5.33%，而RL-T_user之相對誤差為6.88%，整體誤差不大且趨勢一致，顯示模擬與實驗的輔助扭矩相近，驗證模型之準確性與可行性。在此模擬基礎上，已進一步評估並驗證智能動力輔助系統可依據實際情況進行線上學習，適應不同操作情境，並動態調整輔助扭矩，從而有效減少能源浪費，提升騎乘舒適度。綜合而言，本文所提出智能輔助策略皆能有效提升電動輔助自行車的運作效能與續航表現，並可符合實際應用需求。 ;This study presents the development of an intelligent power assist system (IPAS) for add-on electric-assist bicycles, leveraging reinforcement learning (RL). The system is engineered to deliver dynamically adjustable assistive torque tailored to driver demands, aiming to enhance cruising range, optimize the driving experience, and accommodate personalized requirements. In contrast, conventional control strategies predominantly employed in commercially available electric-assist bicycles often exhibit limitations in dynamically modulating assistive torque in response to real-time driving conditions and driver intent, frequently leading to suboptimal energy efficiency. This research firstly established a bicycle dynamic model to address this constraint and engineer an energy-efficient intelligent assist system. Subsequently, an extended state estimation model was formulated utilizing a polynomial approximation method, integrated with an Adaptive Kalman Filter (AKF), to estimate driving torque and vehicle speed concurrently. These estimations serve as critical inputs for the intelligent assist decision-making and control architecture. Building upon this foundation, RL algorithms proposed two distinct intelligent power assist control strategies: one predicated on a driving torque control strategy model (RL-Tf) and the other on a driver torque control strategy model (RL-Tuser). These strategies, along with associated energy management mechanisms, were developed through simulation to adapt to driver demands, diverse driving environments, and varying road conditions, thereby optimizing assistive torque output.Comparative simulations were conducted, benchmarking RL-Tf and RL-Tuser against traditional Proportion-assisted Power Control (PAPC) and the commonly used Pedelec Velocity Control (PVEC). Given that PAPC and PVEC demonstrated comparable performance in preliminary evaluations for this study, PAPC was selected as the primary reference. Relative to PAPC, RL-Tfand RL-Tuser achieved improvements in energy consumption per kilometer of 65.21% and 36.36%, respectively, signifying substantial enhancements in energy efficiency. For experimental validation, an experimental system for bicycles was constructed. Experimental results corroborated that, regarding energy efficiency, RL-Tf surpassed both PAPC and PVEC, registering a 58.41% improvement in energy efficiency compared to PAPC. Furthermore, RL-Tfrelative to RL-Tuser reduced battery energy expenditure by 316.61 J per 100 meterswhile concurrently diminishing driver pedaling effort by 12.27%. Concerning responsiveness to instantaneous pedaling input, RL-Tuser furnished more immediate and substantial assistive torque, culminating in a 25.29% reduction in driver pedaling effort.Battery endurance evaluations, conducted under identical battery configurations (6.8 Ah/36 V), projected a cruising range of 295.61 km for RL-Tf, in contrast to 143.41 km for RL-Tuser. Although both strategies proficiently augmented the energy utilization efficiency and cruising range of the electric-assist bicycle, RL-Tfmanifested a markedly superior endurance advantage.Analysis of assistive torque error between simulation and experimental data revealed relative errors of 5.33% for RL-Tfand 6.88% for RL-Tuser. These modest error margins and congruent trends indicate a robust correlation between simulated and experimentally measured assistive torques, thereby substantiating the accuracy and feasibility of the proposed models. Building upon this simulation framework, it was further ascertained and validated that the IPAS possesses the capability for online learning contingent upon real-world conditions, adapting to heterogeneous operational scenarios and dynamically recalibrating assistive torque. This adaptability effectively curtails energy wastage and augments driving comfort. In summary, the intelligent assist strategies proposed in this paper effectively enhance the operational performance and endurance of electric-assist bicycles, thereby meeting practical application requirements.
顯示於類別:	[機械工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	13	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....