Nonlinear Modeling, Sparse Estimation, and Input Design for Vehicle Dynamics

QC 20251124

Abstract

With the increasing diversity of road environments and the growing intelligence of transportation systems, autonomous vehicles are exposed to more complex, variable, and nonlinear working conditions. To ensure driving safety, autonomous driving systems require higher demands on the modeling accuracy of vehicle dynamics and the reliability of state estimation. However, key physical quantities such as wheel center torque and vehicle mass are difficult to measure directly due to the high cost and deployment complexity of high-precision sensors. In practice, accurate wheel center torque estimation is important for durability assessment and helps reduce the risk of premature component failures. Likewise, in environments where direct measurements are impractical, such as mining operations, vehicle mass is difficult to obtain despite its strong influence on control performance and transport efficiency. Consequently, developing accurate estimation methods for such hard-to-measure quantities becomes crucial. To address these challenges of autonomous vehicles under complex and variable conditions, this dissertation makes three core contributions.

First, addressing the strong nonlinearities commonly found in part-time four-wheel-drive vehicles, particularly the dead-zone phenomenon prevalent in wheel center torque signals, this dissertation proposes a soft sensor architecture integrating a multiple-input single-output finite impulse response (MISO-FIR) estimator with a logistic regression classifier. The classifier utilizes low-frequency input data to identify system dead-zone states, after which torque estimation is performed using readily available sensor signals under non-dead-zone conditions. Real vehicle data experiments demonstrate that the proposed method significantly improves torque estimation accuracy and computational efficiency compared to what is achieved with standard linear and nonlinear models, e.g., neural networks, effectively addressing the challenges of modeling a highly nonlinear phenomenon.

Second, to address issues of parameter redundancy, excessive model quantity, and the lack of a unified sparse modeling criterion in multi-condition vehicle modeling, this dissertation proposes a parameter merging and input selection method based on l1-norm regularization. By integrating MISO-FIR modeling with l1-norm and a model difference l2-norm, the method effectively reduces model redundancy, eliminates irrelevant inputs, and enhances parameter estimation efficiency. Theoretical upper bounds are derived for the hyperparameters of both the l1-norm regularization and the model difference l2-norm. Validation experiments using quarter-vehicle suspension tests and real-world vehicle data confirm that the proposed method effectively reduces parameter redundancy and promotes sparsity, while maintaining high estimation accuracy across multiple conditions.

Finally, shifting from passive modeling to active input design, this dissertation proposes an application-oriented input design (AOID) framework for vehicle mass estimation. Targeting improved estimation accuracy, the method optimizes acceleration and velocity trajectories within operational constraints, thereby enhancing system identifiability and parameter excitation. The dissertation systematically analyzes three common input design objectives, namely maximum estimation accuracy, minimum experimental duration, and minimum driving distance, and derives structurally optimal input strategies considering physical constraints. Finally, real-world experiments conducted on a heavy-duty truck with different payloads verify the designed acceleration and velocity trajectories are feasible and yield estimation performance consistent with theoretical expectations.

The above three methods jointly form an integrated framework for system identification under complex vehicle conditions: 1) a soft sensor for accurate torque estimation with strong nonlinear effects; 2) a regularization-based method for parameter estimation, model merging and sparsity promotion with theoretical regularization bounds; and 3) an AOID framework for mass estimation under different optimization objectives and real-world tests. Together, the framework improves the flexibility of modeling strategies and supports practical implementation, contributing to more reliable estimation and control in representative intelligent vehicle scenarios.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-373227