This thesis places in the context of a joint project by Politecnico di Milano and Scuderia Ferrari, that study the aerodynamic behavior of Formula One car to improve its global performances. Formula One, indeed, is the kind of race that pushes more than all the others the technological bounds to an extreme level. For this reason, constructors invest a large part of their budget in research, to keep up with the evolution of the competitor's cars. Indeed, one of the most critical task is to develop effective methodologies for the estimation of aerodynamic forces on the vehicle, which are useful to both provide a direction for the car development and check the performance of the car during race. This thesis tackles the problem of estimating in a precise and robust way the aerodynamic load of a Formula One cars. In the specific, the existing methods for this problem had to face a tradeoff between interpretability of the method and precision of the estimates. We used data-driven techniques coming from the Machine Learning field that exploits the data gathered during the wind tunnel tests, and use the measurements from a small set of pressure sensors to reconstruct the pressure profile of the entire car and, after that, uses a model based on the physic of the car to provide an estimate of the aerodynamic load. Moreover, we proposed a methods to aggregate the data coming from a set of wind tunnel test, to better generalize the aerodynamic load estimation over newly seen aerodynamic configurations of the car. The combined use of these two techniques allows to have methods whose results are easy to evaluate by Ferrari engineers and, at the same time, to compute a precise estimation of the car loads which is robust w.r.t. changes in the car aerodynamic configuration. These properties have been confirmed by a wide experimental campaign on real-world data, coming from wind tunnel tests, conducted comparing the developed techniques with a set of baselines from the ML field and the method currently used by Ferrari engineers.
Questa tesi si pone nel contesto di un progetto di ricerca tra Politecnico di Milano e Scuderia Ferrari. Ferrari si occupa da tempo di studiare il comportamento aerodinamico delle vetture di Formula Uno al fine di aumentare le performance globali della vettura. La Formula Uno, infatti, è la classe di competizioni automobilistiche che spinge maggiormente i limiti tecnologici ad un livello estremo. Per questa ragione, i costruttori investono molto in ricerca per rimanere ai più alti livelli di performance. La stima dei carichi aerodinamici a cui un veicolo è sottoposto si è dimostrata fondamentale per sviluppare componenti più efficaci in ogni situazione e fornire una corretta analisi delle performance delle vetture durante le competizioni. Questa tesi affronta il problema di come stimare in modo preciso e robusto i carichi aerodinamici delle monoposto di Formula Uno. Nello specifico, i metodi che svolgono questo compito devono avere un buon bilanciamento tra l'interpretabilità dei risultati e la precisione nelle stime. Sono state utilizzate tecniche data-driven, provenienti dall'ambito del Machine Learning, per sfruttare al meglio i dati raccolti in galleria del vento e, a partire da un piccolo sottoinsieme di sensori, ricostruire il profilo di pressione della vettura. Questo procedimento permette, come risultato finale, di stimare il carico aerodinamico complessivo della vettura attraverso un modello fisico. Oltre a questo, sono stati proposti dei metodi per aggregare dati provenienti da differenti test di galleria del vento, il che permette di generalizzare meglio la stima dei carichi aerodinamici di nuove configurazioni. L'uso combinato di queste tecniche consente di avere metodi i cui risultati sono facilmente valutabili da parte degli ingegneri Ferrari e, allo stesso tempo, permette una stima robusta e resistente alle variazioni nella configurazione aerodinamica. Queste caratteristiche sono state ampiamente sperimentate su dati reali di galleria del vento e i metodi sono stati confrontati con altri basati sul Machine Learning oltre che con il metodo correntemente usato dagli ingegneri Ferrari.
Improving aerodynamic load estimation algorithms for F1 racing cars
MUSSI, MARCO
2018/2019
Abstract
This thesis places in the context of a joint project by Politecnico di Milano and Scuderia Ferrari, that study the aerodynamic behavior of Formula One car to improve its global performances. Formula One, indeed, is the kind of race that pushes more than all the others the technological bounds to an extreme level. For this reason, constructors invest a large part of their budget in research, to keep up with the evolution of the competitor's cars. Indeed, one of the most critical task is to develop effective methodologies for the estimation of aerodynamic forces on the vehicle, which are useful to both provide a direction for the car development and check the performance of the car during race. This thesis tackles the problem of estimating in a precise and robust way the aerodynamic load of a Formula One cars. In the specific, the existing methods for this problem had to face a tradeoff between interpretability of the method and precision of the estimates. We used data-driven techniques coming from the Machine Learning field that exploits the data gathered during the wind tunnel tests, and use the measurements from a small set of pressure sensors to reconstruct the pressure profile of the entire car and, after that, uses a model based on the physic of the car to provide an estimate of the aerodynamic load. Moreover, we proposed a methods to aggregate the data coming from a set of wind tunnel test, to better generalize the aerodynamic load estimation over newly seen aerodynamic configurations of the car. The combined use of these two techniques allows to have methods whose results are easy to evaluate by Ferrari engineers and, at the same time, to compute a precise estimation of the car loads which is robust w.r.t. changes in the car aerodynamic configuration. These properties have been confirmed by a wide experimental campaign on real-world data, coming from wind tunnel tests, conducted comparing the developed techniques with a set of baselines from the ML field and the method currently used by Ferrari engineers.File | Dimensione | Formato | |
---|---|---|---|
thesis_2dic.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
1.02 MB
Formato
Adobe PDF
|
1.02 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/10589/152241