In recent decades, a major concern in the automotive industry has been how to improve and enforce safety. The development of systems that enhance safety has expanded into various areas, covering everything from everyday road use to high-speed racing on race tracks. To work effectively, these safety systems need accurate knowledge of the vehicle’s dynamics at every moment in time. Traditionally, physics-based methods have been used to meet this need. These methods, like the dynamic bicycle model, consistently capture the vehicles behavior, but often rely on detailed, vehicle-specific parameters. They also require careful adjustments depending on the conditions they face. These requirements can reduce flexibility and accuracy, especially when conditions change. In this thesis, we explore a data-driven alternative that uses neural networks to model both longitudinal and lateral vehicle dynamics. This approach aims to reduce the need for detailed physical parameters and to lessen dependence on specific road conditions. In particular, this work introduces and evaluates two neural network architectures, a recurrent and a feedforward network, trained on a datasets representing various road grip conditions and driver performance levels. The training uses a simulated dataset generated with a dynamic bicycle model, ensuring the understanding of known physical constraints and real vehicle behavior. Following this, high-fidelity data from CarRealTime simulation software further enhances the network’s robustness across dynamic driving scenarios. Through comparisons with physics-based models, this study assesses the potential of neural networks to improve dynamics modeling and adaptability. While the recurrent network has proven ineffective, a well-trained feedforward network can provide accurate estimation of important vehicle dynamics parameters, such as accelerations, velocities and yaw rate, particularly in high-speed or complex handling maneuvers, where it has the potential to surpass traditional models.
Negli ultimi decenni, l’industria automobilistica si è concentrata sul miglioramento della sicurezza, sviluppando sistemi che la rafforzano in vari ambiti, dall’uso quotidiano su strada fino alle gare in pista. Per funzionare al meglio, questi sistemi richiedono una conoscenza accurata della dinamica del veicolo. Tradizionalmente, per assolvere a questo scopo sono stati utilizzati metodi basati sulla fisica che descrivono il comportamento dei veicoli. Questi metodi, come il modello dinamico bicicletta, offrono una stima consistente del dinamica, ma si basano su parametri specifici per ogni veicolo. Inoltre, necessitano un’affinamento basato sulle condizioni che andranno ad affrontare. Questi requisiti possono ridurre la flessibilità di utilizzo e l’accuratezza, sopratutto quando le condizioni di utilizzo cambiano. In questa tesi viene esplorata un’alternativa, basata su reti neurali, per la modellazione della dinamica longitudinale e laterale del veicolo. Questo approccio verte a limitare la necessità di una dettagliata parametrizzazione dei veicoli e a ridurre la dipendenza da condizioni specifiche della strada. In particolare, vengono introdotte e valutate due reti neurali, una rete ricorrente e una rete feed-forward, addestrate con l’utilizzo di diversi dataset comprensivi di varie condizioni di attrito del fondo stradale e livelli di prestazione di guida. L’allenamento è svolto dapprima con un insieme di dati generati con l’utilizzo del modello bicicletta, fornendo una base di conoscenza del comportamento del veicolo e dei suoi limiti. Successivamente, viene ulteriormente affinato grazie all’utilizzo di dati ad alta fedeltà generati con CarRealTime, per ottimizzare la robustezza alle diverse condizioni. Grazie ad una comparazione con il modello fisico tradizionale, questa ricerca valuta il potenziale delle reti naurali per migliorare la stima dello stato del veicolo e la loro flessibilità. Mentre la rete ricorrente si è dimostrata inadatta, la rete feed-forward è stata in grado di fornire stime molto accurate di parametri quali le accelerazioni, le velocità e il tasso di imbardata, in particolar modo in manovre ad alta velocità.
Neural networks based modeling of longitudinal and lateral vehicle dynamics
STUCCHI, WILLIAM
2023/2024
Abstract
In recent decades, a major concern in the automotive industry has been how to improve and enforce safety. The development of systems that enhance safety has expanded into various areas, covering everything from everyday road use to high-speed racing on race tracks. To work effectively, these safety systems need accurate knowledge of the vehicle’s dynamics at every moment in time. Traditionally, physics-based methods have been used to meet this need. These methods, like the dynamic bicycle model, consistently capture the vehicles behavior, but often rely on detailed, vehicle-specific parameters. They also require careful adjustments depending on the conditions they face. These requirements can reduce flexibility and accuracy, especially when conditions change. In this thesis, we explore a data-driven alternative that uses neural networks to model both longitudinal and lateral vehicle dynamics. This approach aims to reduce the need for detailed physical parameters and to lessen dependence on specific road conditions. In particular, this work introduces and evaluates two neural network architectures, a recurrent and a feedforward network, trained on a datasets representing various road grip conditions and driver performance levels. The training uses a simulated dataset generated with a dynamic bicycle model, ensuring the understanding of known physical constraints and real vehicle behavior. Following this, high-fidelity data from CarRealTime simulation software further enhances the network’s robustness across dynamic driving scenarios. Through comparisons with physics-based models, this study assesses the potential of neural networks to improve dynamics modeling and adaptability. While the recurrent network has proven ineffective, a well-trained feedforward network can provide accurate estimation of important vehicle dynamics parameters, such as accelerations, velocities and yaw rate, particularly in high-speed or complex handling maneuvers, where it has the potential to surpass traditional models.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Stucchi_Tesi.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
10.74 MB
Formato
Adobe PDF
|
10.74 MB | Adobe PDF | Visualizza/Apri |
2024_12_Stucchi_Executive Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
663.82 kB
Formato
Adobe PDF
|
663.82 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/229998