Automatic Calibration of Freeway Models with Model-Based Fault Detection
- Author(s): Dervisoglu, Gunes
- Advisor(s): Horowitz, Roberto
- et al.
This dissertation presents system identification, fault detection and fault handling methodologies for automatically building calibrated models of freeway traffic flow. Using these methodologies, data driven algorithms were developed as part of a larger scheme of a suite of software tools designed to provide traffic engineers with a simulation platform where various traffic planning strategies can be analyzed. The algorithms that are presented work with loop detector data that are gathered from California freeways.
The system identification deploys a constrained linear regression analysis that estimates the so-called fundamental diagram relationship between flow and density at the location of a given sensor. A triangular fundamental diagram is assumed that establishes a bi-modally linear relationship between flow and density, the two modes being free flow and congestion. An approximate quantile regression method is used for the estimation of the congested regime due to this mode's high susceptibility to various external factors.
The fault detection algorithm has been developed to facilitate the automatic model building procedure. The macroscopic cell transmission model, which is the model assumed in this study, requires consistent observations along the modeled freeway section for an accurate calibration to be possible. When detectors are down or missing, the model has to be modified to a less accurate representation to conform with a configuration where a sensor is assigned to each cell of the model. In addition, on most California freeways the ramp flows in and out of the mainline are not observed. Since the estimation of these unknown inputs to the system also hinge on healthy mainline data, the identification of faulty mainline sensors becomes crucial to the automatic model building process. The model-based fault detection algorithm presented herein analyzes the parity between simulated and measured state, along with estimated unknown input profiles. Subsequently, it makes use of a look-up table logic and a threshold scheme to flag erroneous detectors along the freeway mainline.
Finally, the fault handling algorithm that accompanies the fault detection aims to revert the model to its original configuration after the aforementioned modifications are made to the model due to missing or bad sensors. Using a relaxed model-constrained linear optimization, this algorithm seeks to fill in the gaps in the observations along the freeway that are a result of poor detection. This method provides a reconstruction of the unobserved state that conforms with the rest of the measurements and does not produce a state estimate in a control theoretical sense.