The present thesis discusses about the possibility of performing defect identification on board of a high-speed train wheelset by means of axle-box acceleration measurements. Two defects that commonly affect this mechanical system are specifically addressed: the Wheel-Flat on the circumferential wheel profile and the transversal crack in the axle resistant cross section. Both the defects are firstly studied by means of a literature review and a preliminary numerical analysis performed through a Multi-Body model. The core of the thesis then moves to the experimental campaign performed at Lucchini BU300 test bench in Lovere (Italy), where a full-scale wheelset of an ETR1000 vehicle has been tested at different speeds from 0 to 300 km/h, both in presence and in absence of the two defects. Based on the database acquired during the experimental campaign a Machine Learning procedure is proposed in order to identify the presence of the defects. Two different classification algorithms are shown: the k-Nearest Neighbor and the Linear Regression. Satisfying results are obtained for the Wheel-Flat identification algorithm that confirm a good performance of the features selected as results of the numerical simulations, allowing to conclude that a Wheel-Flat greater than 30 mm length can be easily detected by means of Axle-Box Acceleration measurements onboard. For the crack identification algorithm, the classical features coming from literature review and previous works proved to be inadequate to correctly identify the presence of the defect. Thus, an alternative feature selection procedure is proposed. The results obtained highlight the possibility to identify the presence of a crack interesting an area equal or greater than the 9.6% of the total resistant cross section, in absence of wear on the wheel profile. In presence of unknown wear of the wheel profile the algorithm performances drop, requiring future works to address the problem of identification in case of co-existence of crack in the axle and wear on the wheel profile.
La presente tesi tratta della possibilità di effettuare l’identificazione di difetti a bordo di un assile di un treno ad alta velocità mediante misure di accelerazione in boccola. In particolare, vengono affrontati due difetti che comunemente interessano questo sistema meccanico: il Wheel-Flat sul profilo della ruota e la cricca trasversale nella sezione resistente dell'assile. Entrambi i difetti sono studiati in primo luogo attraverso una revisione della letteratura e successivamente per mezzo di un'analisi numerica preliminare eseguita mediante un modello Multi-Body. Il cuore della tesi riguarda la campagna sperimentale eseguita presso il banco prova Lucchini BU300 di Lovere (Italia), dove una sala montata in scala reale di un veicolo ETR1000 è stata testata a diverse velocità tra 0 a 300 km/h, sia in presenza che in assenza dei due difetti. Sulla base del database acquisito durante la campagna sperimentale, viene proposta una procedura di Machine Learning per identificare la presenza dei difetti. Vengono presentati due diversi algoritmi di classificazione: il k-Nearest Neighbor e la Regressione Lineare. Per l'algoritmo di identificazione del Wheel-Flat si ottengono risultati soddisfacenti che confermano una buona performance delle caratteristiche selezionate a seguito dei risultati delle simulazioni numeriche, consentendo di concludere che un Wheel-Flat di estensione superiore a 30 mm può essere facilmente individuato mediante le misure di accelerazione in boccola. Per l'algoritmo di identificazione delle cricche, le caratteristiche classiche ricavate dalla revisione della letteratura e dai lavori precedenti si sono rivelate inadeguate a identificare correttamente la presenza del difetto. È stata quindi proposta una procedura alternativa di selezione delle caratteristiche. I risultati ottenuti evidenziano la possibilità di identificare la presenza di una cricca di estensione pari o superiore al 9,6% della sezione resistente totale, in assenza di usura sul profilo della ruota. In presenza di usura del profilo della ruota, le prestazioni dell'algoritmo calano, richiedendo lavori futuri per affrontare il problema dell'identificazione in caso di coesistenza di cricca nell'assale e usura sul profilo della ruota.
Smart wheelset: high-speed wheelset defect identification by means of axle-box acceleration measurements
Cii, Stefano
2021/2022
Abstract
The present thesis discusses about the possibility of performing defect identification on board of a high-speed train wheelset by means of axle-box acceleration measurements. Two defects that commonly affect this mechanical system are specifically addressed: the Wheel-Flat on the circumferential wheel profile and the transversal crack in the axle resistant cross section. Both the defects are firstly studied by means of a literature review and a preliminary numerical analysis performed through a Multi-Body model. The core of the thesis then moves to the experimental campaign performed at Lucchini BU300 test bench in Lovere (Italy), where a full-scale wheelset of an ETR1000 vehicle has been tested at different speeds from 0 to 300 km/h, both in presence and in absence of the two defects. Based on the database acquired during the experimental campaign a Machine Learning procedure is proposed in order to identify the presence of the defects. Two different classification algorithms are shown: the k-Nearest Neighbor and the Linear Regression. Satisfying results are obtained for the Wheel-Flat identification algorithm that confirm a good performance of the features selected as results of the numerical simulations, allowing to conclude that a Wheel-Flat greater than 30 mm length can be easily detected by means of Axle-Box Acceleration measurements onboard. For the crack identification algorithm, the classical features coming from literature review and previous works proved to be inadequate to correctly identify the presence of the defect. Thus, an alternative feature selection procedure is proposed. The results obtained highlight the possibility to identify the presence of a crack interesting an area equal or greater than the 9.6% of the total resistant cross section, in absence of wear on the wheel profile. In presence of unknown wear of the wheel profile the algorithm performances drop, requiring future works to address the problem of identification in case of co-existence of crack in the axle and wear on the wheel profile.| File | Dimensione | Formato | |
|---|---|---|---|
|
Smart Wheelset_ High-Speed Wheelset Defect Identification by Means of Axle-Box Acceleration Measurements.pdf
non accessibile
Dimensione
28.25 MB
Formato
Adobe PDF
|
28.25 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/10589/191932