Scalable online estimation with performance guarantees: Application to traffic network monitoring

Sun, Ye

Permalink

https://hdl.handle.net/2142/100892 Copy

Description

Title

Scalable online estimation with performance guarantees: Application to traffic network monitoring

Author(s)

Sun, Ye

Issue Date

2018-01-18

Director of Research (if dissertation) or Advisor (if thesis)

Work, Daniel B.

Doctoral Committee Chair(s)

Work, Daniel B.

Committee Member(s)

• Dullerud, Geir E.
• Ouyang, Yanfeng
• Sowers, Richard B.

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Traffic networks
• Traffic estimation
• Observability
• Kalman filter
• Switches
• Estimation error
• Computational modeling
• Consensus filter
• Sensor scheduling

Abstract

This work is motivated by the need for scalable online estimation with provable performance in cyber-physical systems, especially in traffic monitoring applications. While model-based traffic estimation has achieved great success via experimental deployments, there are substantial open questions on the theoretical understanding of the performance of these estimators. This gap is largely due to the nonlinearity of the underlying traffic flow and the nonobservability of the estimation problems. The main contribution of this dissertation is to explicitly address performance guarantees of filtering algorithms on traffic networks, with specific focus on systems that are unobservable, or even switch among observable and unobservable scenarios. We first consider one-dimensional road sections to establish the main proof techniques before considering more general road networks. To tackle the non-linearity issue in traffic models, the Lighthill-Whitham-Richards partial differential equation (LWR PDE) is transformed to a discrete-time switched linear form, i.e., the switching mode model (SMM). We provide a rigorous analysis on the performance of the Kalman filter (KF) on the SMM. Although the error dynamics of the KF is very likely to diverge under general unobservable systems, we show that in the context of traffic estimation, a uniform upper bound for the mean error exists when the system is unobservable. This is done by exploring the interactions between the physical properties of traffic flows, the stability conditions in the discretization scheme, and the information update in the filter. We also derive error bounds for the KF when the system switches among the observable and unobservable modes of the SMM. The above analysis is then extended to traffic networks with junctions. To support the analysis, we develop a switched linear model describing traffic dynamics on a freeway section with a junction inside. The model, namely the switching mode model with junctions (SMM-J), combines the discretized LWR PDE with a junction solver. Based on the SMM-J, the error bounds of the KF are extended to freeway networks. This dissertation also studies two essential problems related to the scalability issue in the estimation of general cyber-physical systems: (i) state space scalability, where the enormous state dimension causes computational burden on estimators, and (ii) data scalability, where massive data transmission incurs considerable energy, bandwith, or monetary costs. First, we design a distributed local Kalman consensus filter (DLKCF) for large-scale estimation, where the entire state is partitioned into local sections, and the computation task is distributed to local agents. In addition, a consensus term is designed to promote agreement on the estimates of neighboring agents. We also derive the error bounds of the DLKCF used for traffic estimation. Next, we study sensor scheduling schemes designed to select the most informative data to transmit to the estimator, thus reducing data transmission while preserving estimation accuracy. In this context, we propose a filtering algorithm that extracts the implicit information in the scheduling policy and update both the state estimate and the error covariance when data transmission is not triggered, which achieves better estimation accuracy compared to existing algorithms that only update the error covariance in the absence of data transmission.

Graduation Semester

2018-05

Type of Resource

text

Permalink

http://hdl.handle.net/2142/100892

Copyright and License Information

Copyright 2018 Ye Sun

Owning Collections