서브메뉴
검색
A Physics Enhanced Residual Learning (PERL) Framework for Autonomous Vehicle
A Physics Enhanced Residual Learning (PERL) Framework for Autonomous Vehicle
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211153121
- ISBN
- 9798346813644
- DDC
- 385
- 저자명
- Long, Keke.
- 서명/저자
- A Physics Enhanced Residual Learning (PERL) Framework for Autonomous Vehicle
- 발행사항
- [Sl] : The University of Wisconsin - Madison, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 105 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
- 주기사항
- Advisor: Li, Xiaopeng Shaw;Chen, Sikai.
- 학위논문주기
- Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.
- 초록/해제
- 요약Autonomous vehicles (AV) driving in mixed traffic, comprising human-driven vehicles (HVs) and AVs, is notoriously challenging, underscored by complex driving environments and a multitude of uncertainties. Further, the AV system involves complex interactions cascading across the interconnected modules for perception, planning, and control. An error in the higher chain of these interconnected modules can propagate downstream, inducing instability in vehicle control. Such errors can also cascade through following vehicles and impact traffic dynamics at a broader level. To overcome these challenges, we propose the novel Physics-Enhanced Residual Learning (PERL) framework for AV operations in mixed traffic. PERL comprises two components: a physics-based model and residual learning. The physics model provides primary outputs, including prediction outputs or control outputs, and then its residuals are learned by a learning-based approach as corrections to the physics model to enhance the results.This dissertation comprises three major research thrusts: (1) Development of PERL frameworks for vehicle trajectory prediction (2) Validate the contribution of PERL-based prediction in AV control, and (3) Development of Physics-Enhanced Residual Policy Learning (PERPL) framework for vehicle control.The first part proposed the PERL framework and applied it to a vehicle trajectory prediction problem with real-world trajectory data of both HV and AV, using an adapted Newell car-following model as the physics model, and four kinds of neural networks (GRU, Convolution Long Short-Term Memory (CLSTM), VAE and Informer model) as the residual learning model. We compare this PERL model with pure physics models, NN models, and other physics-informed neural network (PINN) models. The result reveals that the PERL model yields the best prediction with limited training data and it has fast convergence during training. Moreover, the PERL model requires fewer parameters to achieve similar predictive performance compared to NN and PINN models. The second part proposes a PERL-based vehicle control method to mitigate traffic oscillation in the mixed traffic environment of connected and autonomous vehicles (CAVs) and HVs. This model includes the PERL prediction model and a controller. The PERL-based prediction model precisely predicts the behavior of the preceding vehicle, especially downstream speed fluctuations, to allow sufficient time for the driver to respond to these speed fluctuations. For the controller, we employ a Model Predictive Control (MPC) model that considers the dynamics of the CAV and its following vehicles, improving safety and comfort for the platoon formed including the CAV and following vehicles. The proposed model is validated through a Vehicle-in-the-loop (ViL) field test. Results validate the proposed method in damping traffic oscillation and enhancing the safety and fuel efficiency of the CAV and the following vehicles in mixed traffic with the presence of uncertain vehicle dynamics and actuator lag. The third part proposes the Physics-Enhanced Residual Policy Learning (PERPL) framework for vehicle control, leveraging the advantages of both physics-based models (data-efficient and interpretable) and RL methods (flexible to multiple objectives and fast computing). The physics component provides model interpretability and stability and the learning-based Residual Policy adjusts the physics-based policy to adapt to the changing environment, thereby refining the decisions of the physics model. This model is applied in decentralized control of a mixed traffic platoon of CAVs and HVs using a constant time gap (CTG) strategy, with actuator lag and communication delays. Experimental results demonstrate that this model has high extrapolation ability, achieving smaller headway errors and better oscillation dampening than the linear control model and reinforcement learning (RL) model in artificial extreme scenarios. At the macroscopic level, overall traffic oscillations are also reduced as the penetration rate of CAVs employing the PERPL-based controller increases.
- 일반주제명
- Transportation
- 일반주제명
- Computer engineering
- 일반주제명
- Automotive engineering
- 일반주제명
- Mechanical engineering
- 키워드
- Vehicle control
- 기타저자
- The University of Wisconsin - Madison Civil & Environmental Engr
- 기본자료저록
- Dissertations Abstracts International. 86-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017165080
■00520250211153121
■006m o d
■007cr#unu||||||||
■020 ▼a9798346813644
■035 ▼a(MiAaPQ)AAI31762628
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a385
■1001 ▼aLong, Keke.
■24512▼aA Physics Enhanced Residual Learning (PERL) Framework for Autonomous Vehicle
■260 ▼a[Sl]▼bThe University of Wisconsin - Madison▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a105 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-06, Section: B.
■500 ▼aAdvisor: Li, Xiaopeng Shaw;Chen, Sikai.
■5021 ▼aThesis (Ph.D.)--The University of Wisconsin - Madison, 2024.
■520 ▼aAutonomous vehicles (AV) driving in mixed traffic, comprising human-driven vehicles (HVs) and AVs, is notoriously challenging, underscored by complex driving environments and a multitude of uncertainties. Further, the AV system involves complex interactions cascading across the interconnected modules for perception, planning, and control. An error in the higher chain of these interconnected modules can propagate downstream, inducing instability in vehicle control. Such errors can also cascade through following vehicles and impact traffic dynamics at a broader level. To overcome these challenges, we propose the novel Physics-Enhanced Residual Learning (PERL) framework for AV operations in mixed traffic. PERL comprises two components: a physics-based model and residual learning. The physics model provides primary outputs, including prediction outputs or control outputs, and then its residuals are learned by a learning-based approach as corrections to the physics model to enhance the results.This dissertation comprises three major research thrusts: (1) Development of PERL frameworks for vehicle trajectory prediction (2) Validate the contribution of PERL-based prediction in AV control, and (3) Development of Physics-Enhanced Residual Policy Learning (PERPL) framework for vehicle control.The first part proposed the PERL framework and applied it to a vehicle trajectory prediction problem with real-world trajectory data of both HV and AV, using an adapted Newell car-following model as the physics model, and four kinds of neural networks (GRU, Convolution Long Short-Term Memory (CLSTM), VAE and Informer model) as the residual learning model. We compare this PERL model with pure physics models, NN models, and other physics-informed neural network (PINN) models. The result reveals that the PERL model yields the best prediction with limited training data and it has fast convergence during training. Moreover, the PERL model requires fewer parameters to achieve similar predictive performance compared to NN and PINN models. The second part proposes a PERL-based vehicle control method to mitigate traffic oscillation in the mixed traffic environment of connected and autonomous vehicles (CAVs) and HVs. This model includes the PERL prediction model and a controller. The PERL-based prediction model precisely predicts the behavior of the preceding vehicle, especially downstream speed fluctuations, to allow sufficient time for the driver to respond to these speed fluctuations. For the controller, we employ a Model Predictive Control (MPC) model that considers the dynamics of the CAV and its following vehicles, improving safety and comfort for the platoon formed including the CAV and following vehicles. The proposed model is validated through a Vehicle-in-the-loop (ViL) field test. Results validate the proposed method in damping traffic oscillation and enhancing the safety and fuel efficiency of the CAV and the following vehicles in mixed traffic with the presence of uncertain vehicle dynamics and actuator lag. The third part proposes the Physics-Enhanced Residual Policy Learning (PERPL) framework for vehicle control, leveraging the advantages of both physics-based models (data-efficient and interpretable) and RL methods (flexible to multiple objectives and fast computing). The physics component provides model interpretability and stability and the learning-based Residual Policy adjusts the physics-based policy to adapt to the changing environment, thereby refining the decisions of the physics model. This model is applied in decentralized control of a mixed traffic platoon of CAVs and HVs using a constant time gap (CTG) strategy, with actuator lag and communication delays. Experimental results demonstrate that this model has high extrapolation ability, achieving smaller headway errors and better oscillation dampening than the linear control model and reinforcement learning (RL) model in artificial extreme scenarios. At the macroscopic level, overall traffic oscillations are also reduced as the penetration rate of CAVs employing the PERPL-based controller increases.
■590 ▼aSchool code: 0262.
■650 4▼aTransportation
■650 4▼aComputer engineering
■650 4▼aAutomotive engineering
■650 4▼aMechanical engineering
■653 ▼aAutomated vehicles
■653 ▼aConnected vehicles
■653 ▼aDeep reinforcement learning
■653 ▼aPhysics-informed neural network
■653 ▼aVehicle control
■653 ▼aVehicle trajectory prediction
■653 ▼aVehicle trajectory planning
■690 ▼a0709
■690 ▼a0464
■690 ▼a0548
■690 ▼a0540
■71020▼aThe University of Wisconsin - Madison▼bCivil & Environmental Engr.
■7730 ▼tDissertations Abstracts International▼g86-06B.
■790 ▼a0262
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17165080▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


