The classical Supervised Learning approach, a field of Machine Learning, is a data-driven approach that aims at building a model able to represent the hidden relationship existing among the data collected from a real process. It has been shown, both in scientific works and practical scenarios, that this approach provided effective solutions in most of the cases. Unfortunately, the accuracy of the developed models depends on the availability of a set of data which are able to properly describes the problem. In practical situations, this might not be the case. A significant example is provided by the models designed for the prediction of the tyre deflation in truck tyres. Indeed, we do not have any data regarding the deflation process of newly installed tyres to use for model estimation. This is generally called the Cold Start problem. This work, being part of the Cyber Fleet project, aims at studying the relationships between tyre's features and how the similarity among tyres can be exploited to improve the prediction in the cases in which the Cold Start problem arises. The final goal is to develop a new model, complementary to the existing ones, able to obtain a prediction even for newly installed tyres, as a combination of the information already gathered on previously seen tyres. We tested this method on two real-world datasets coming from sensors installed on fleets of trucks, namely Crescini and Rivetti. The results showed that the proposed Cold Start approach is able to provide a meaningful prediction for the tyre deflation predictive models, even when no data are available about a specific tyre. Unfortunately, the accuracy of the developed models depends on the availability of a set of data which are able to properly describes the problem. In practical situations this might not be the case. A significant example is provided by the models designed for the prediction of the tyre deflation in truck tyres. Indeed, we do not have any data regarding the deflation process of newly installed tyres to use for model estimation. This is generally called the \emph{Cold Start} problem. This work, being part of the Cyber Fleet project, aims at studying the relationships between tyre's features and how the similarity among tyres can be exploited to improve the prediction in the cases in which the Cold Start problem arises. The final goal is to develop a new model, complementary to the existing ones, able to obtain the model parameters as a combination of the information already gathered on previously seen tyres. We tested this method on two real-world datasets coming from sensors installed on fleets of trucks, namely Crescini and Rivetti. The results showed that the proposed Cold Start approach is able to provide a meaningful prediction for the tyre deflation predictive models, even when no data are available about a specific tyre.
Utilizzando approcci data-driven e grazie a tecniche di Supervised Learning, un sottocampo del Machine Learning, è possibile ottenere modelli che ben rappresentano processi reali, in grado di comprendere le relazioni che ne stanno alla base. È stato mostrato, sia nella letteratura scientifico, sia in applicazioni reali, che tali tecniche sono in grado di risolvere efficacemente problemi relativi a diversi contesti applicativi. Tuttavia, l'accuratezza di tali modelli dipende dalla disponibilità di un insieme di dati realmente rappresentativo del problema. In situazioni reali tali dati potrebbero non essere sufficienti per modellizzare appropriatamente il problem analizzato. Un esempio significativo è rappresentato dai modelli sviluppati per predire l'andamento della pressione all'interno degli pneumatici montati su camion. Infatti, quando un nuovo pneumatico viene montato per la prima volta, non si hanno dati disponibili per poter stimare tale modello. Questo problema viene definito col nome di Cold Start. Questo elaborato, essendo parte del progetto Cyber Fleet, ha come fine lo studio delle relazioni esistenti tra caratteristiche fisiche degli pneumatici, e come le similarità tra differenti pneumatici possano essere utilizzate per migliorare le predizioni in caso di Cold Start. L'obbiettivo finale è lo sviluppo di un nuovo modello predittivo, complementare a quelli già esistenti, in grado di ottenere delle predizioni anche per pneumatici installati da poco tempo, utilizzando le informazioni provenienti da pneumatici utilizzati nel passato. Il metodo è stato testato su due casi reali, utilizzando i dati provenienti dai sensori montati su due flotte di camion, Crescini e Rivetti. I risultati ottenuti mostrano che, il metodo proposto per risolvere il problema Cold Start, è in grado di effettuare predizioni significative sull'andamento della pressione all'interno di uno pneumatico, anche quando i relativi dati non sono disponibili.
Analysis of the cold start problem in predictive algorithms for cyber tyres
SALMOIRAGHI, ANDREA
2016/2017
Abstract
The classical Supervised Learning approach, a field of Machine Learning, is a data-driven approach that aims at building a model able to represent the hidden relationship existing among the data collected from a real process. It has been shown, both in scientific works and practical scenarios, that this approach provided effective solutions in most of the cases. Unfortunately, the accuracy of the developed models depends on the availability of a set of data which are able to properly describes the problem. In practical situations, this might not be the case. A significant example is provided by the models designed for the prediction of the tyre deflation in truck tyres. Indeed, we do not have any data regarding the deflation process of newly installed tyres to use for model estimation. This is generally called the Cold Start problem. This work, being part of the Cyber Fleet project, aims at studying the relationships between tyre's features and how the similarity among tyres can be exploited to improve the prediction in the cases in which the Cold Start problem arises. The final goal is to develop a new model, complementary to the existing ones, able to obtain a prediction even for newly installed tyres, as a combination of the information already gathered on previously seen tyres. We tested this method on two real-world datasets coming from sensors installed on fleets of trucks, namely Crescini and Rivetti. The results showed that the proposed Cold Start approach is able to provide a meaningful prediction for the tyre deflation predictive models, even when no data are available about a specific tyre. Unfortunately, the accuracy of the developed models depends on the availability of a set of data which are able to properly describes the problem. In practical situations this might not be the case. A significant example is provided by the models designed for the prediction of the tyre deflation in truck tyres. Indeed, we do not have any data regarding the deflation process of newly installed tyres to use for model estimation. This is generally called the \emph{Cold Start} problem. This work, being part of the Cyber Fleet project, aims at studying the relationships between tyre's features and how the similarity among tyres can be exploited to improve the prediction in the cases in which the Cold Start problem arises. The final goal is to develop a new model, complementary to the existing ones, able to obtain the model parameters as a combination of the information already gathered on previously seen tyres. We tested this method on two real-world datasets coming from sensors installed on fleets of trucks, namely Crescini and Rivetti. The results showed that the proposed Cold Start approach is able to provide a meaningful prediction for the tyre deflation predictive models, even when no data are available about a specific tyre.| File | Dimensione | Formato | |
|---|---|---|---|
|
2017_07_Salmoiraghi.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
3.14 MB
Formato
Adobe PDF
|
3.14 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/10589/135123