The safety of a vehicle is a major concern for car manufacturers, given that over one million of fatalities are recorded yearly, due to the loss of vehicle stability. Despite the worrying data, a deeper understanding of the phenomena behind the loss of control is still lacking. The complexity of the problem skyrockets due to the presence of the driver, whose models are still unable to catch his behaviour in a trustworthy way. Generally, stability is assessed locally: a system is considered stable if, starting from an equilibrium position, the system comes back to the same equilibrium position, after a small disturbance. However, it does not address what "small" means in a practical application. It is easy to recognize, even for novice drivers, that the vehicle could become unstable after a sufficiently big perturbation, e.g. strong gusts of wind or evasive maneuvers, even if the car is running on a locally stable equilibrium. The stability criterion explained in the thesis can be used together with Milliken Moment Method to improve the safety of both autonomous and normal vehicles. The motion of the system vehicle and driver is heavily influenced by saddle-type limit cycles. Using Floquet theory, a real-time assessment of the stability both in a straight line and in a curved path can be done. In order to obtain realistic saddle-type limit cycles, a physics-based neural network is trained using a 14 DoFs model. The information that can be extracted by a digital twin are by far more reliable than a single-track vehicle model. The limit cycles shrink when the speed increases, until a Hopf bifurcation is found at relatively high speed, where limit cycles collapse into a point. The numerical results obtained in this work are then checked by doing experimental tests using a driving simulator.
La sicurezza di un veicolo è un obiettivo fondamentale per i costruttori automobilistici, dato che oltre un milione di incidenti mortali sono registrati annualmente, a causa della perdita di stabilità del veicolo. Malgrado i dati preoccupanti, non si ha ancora a disposizione una conoscenza profonda del problema. La complessità del problema aumenta esponenzialmente a causa della presenza del pilota, i cui modelli non riescono ancora a predire in maniera affidabile il suo comportamento. Generalmente, la stabilità è studiata localmente: dato un piccolo disturbo, il sistema è stabile se torna allo stesso equilibrio di partenza. Tuttavia, non è specificato chiaramente cosa "piccolo disturbo" significhi in un'applicazione pratica. Anche guidatori inesperti possono riconoscere che disturbi sufficientemente forti, come raffiche di vento o manovre evasive, possono destabilizzare un sistema localmente stabile. Il criterio di stabilità spiegato in questa tesi può essere utilizzato, insieme al Milliken Moment Method, per migliorare la sicurezza di veicoli normali e a guida autonoma. Il moto del sistema veicolo/guidatore è influenzato da cicli limite sella. Utilizzando la teoria di Floquet può essere svolta una valutazione in tempo reale sulla stabilità globale, sia in rettilineo che in curva. Per ottenere dei cicli limite sella attendibili viene utilizzata una rete neurale, che riproduce il comportamento di un modello a 14 GdL, la cui affidabilità è migliore rispetto a un semplice modello a bicicletta. La dimensione dei cicli limite sella si riduce al crescere della velocità, fino al rintracciamento di una biforcazione di Hopf, ove il ciclo collassa in un punto. I risultati numerici ottenuti in questa tesi sono supportati da numerosi esperimenti svolti su un simulatore di guida.
Real-time assessment of global stability in vehicle/driver systems via a physics-based neural network
Fontana, Matteo
2023/2024
Abstract
The safety of a vehicle is a major concern for car manufacturers, given that over one million of fatalities are recorded yearly, due to the loss of vehicle stability. Despite the worrying data, a deeper understanding of the phenomena behind the loss of control is still lacking. The complexity of the problem skyrockets due to the presence of the driver, whose models are still unable to catch his behaviour in a trustworthy way. Generally, stability is assessed locally: a system is considered stable if, starting from an equilibrium position, the system comes back to the same equilibrium position, after a small disturbance. However, it does not address what "small" means in a practical application. It is easy to recognize, even for novice drivers, that the vehicle could become unstable after a sufficiently big perturbation, e.g. strong gusts of wind or evasive maneuvers, even if the car is running on a locally stable equilibrium. The stability criterion explained in the thesis can be used together with Milliken Moment Method to improve the safety of both autonomous and normal vehicles. The motion of the system vehicle and driver is heavily influenced by saddle-type limit cycles. Using Floquet theory, a real-time assessment of the stability both in a straight line and in a curved path can be done. In order to obtain realistic saddle-type limit cycles, a physics-based neural network is trained using a 14 DoFs model. The information that can be extracted by a digital twin are by far more reliable than a single-track vehicle model. The limit cycles shrink when the speed increases, until a Hopf bifurcation is found at relatively high speed, where limit cycles collapse into a point. The numerical results obtained in this work are then checked by doing experimental tests using a driving simulator.File | Dimensione | Formato | |
---|---|---|---|
2024_10_Fontana_Tesi.pdf
solo utenti autorizzati a partire dal 18/09/2027
Descrizione: Testo della tesi
Dimensione
28.75 MB
Formato
Adobe PDF
|
28.75 MB | Adobe PDF | Visualizza/Apri |
2024_10_Fontana_ExecutiveSummary.pdf
solo utenti autorizzati a partire dal 18/09/2027
Descrizione: Executive summary
Dimensione
3.35 MB
Formato
Adobe PDF
|
3.35 MB | 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/10589/228173