| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 15.65 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A presente dissertação foca-se no desenvolvimento e aplicação de uma estratégia híbrida de técnicas de
Inteligência Artificial (IA) para detetar rodas com defeitos de circularidade no comboio ferroviário de
passageiros Alfa pendular. Este algoritmo é alimentado a partir das respostas dinâmicas adquiridas por
um sistema de monitorização wayside composto por um conjunto de acelerómetros instalados nos carris
da via-férrea.
O método foi desenvolvido com o auxílio de diversas simulações numéricas 3D relativas a passagens
ferroviárias, permitindo extrair a resposta dinâmica do sistema veículo-via. As simulações numéricas
abrangem duas possibilidades de circulação: condições normais (cenários de base) e condições anormais
(cenários de dano). Os dois tipos de cenários incluem condições de operação ferroviária variável,
nomeadamente, diferentes perfis de irregularidade e ruído no sistema de medição. Relativamente aos
cenários de dano, foram considerados dois tipos de defeitos geométricos possíveis de serem modelados,
um referente à presença de lisos na superfície de rolamento e outro caracterizado por um desgaste
ondulatório periódico ao longo do perímetro da roda, denominado de poligonização.
A metodologia desenvolvida inclui um processo de treino com o auxílio de uma rede neuronal artificial,
designada por Autoencoder Esparso, de forma a serem extraídos indicadores de dano. Os dados de
entrada para este Autoencoder compreendem as respostas dinâmicas da via adquiridas numericamente.
A sensibilidade do indicador de dano é incrementada através da aplicação da distância Mahalanobis.
Posteriormente, é aplicada uma análise de outliers, para detetar os danos previamente simulados, e uma
técnica de clusters para classificação do dano em duas fases: i) identificação do tipo de dano; ii)
identificação da severidade de cada tipo de dano.
Por fim, a metodologia desenvolvida é sujeita a uma validação com base num conjunto de respostas
dinâmicas adquiridas experimentalmente.
This dissertation is focused on developing and applying a hybrid strategy of Artificial Intelligence (AI) techniques to detect out-of-roundness (OOR) wheels on the Alfa pendular train. This algorithm is supplied with dynamic response data obtained through a wayside monitoring system comprising a set of accelerometers installed on the railway tracks. The method was developed with several 3D numerical simulations of railway passages, allowing the dynamic response of the vehicle-track system to be extracted. The numerical simulations cover two circulation possibilities: i) normal conditions (baseline scenarios) and ii) abnormal conditions (damage scenarios). Both scenarios are characterized by a set of traffic variabilities, as well as the presence of irregularity profiles and noise. Regarding the damage scenarios, two types of geometric defects were considered, one referring to the presence of defect on the wheel tread and a defect around the complete wheel circumference called polygonization. The core of the methodology developed includes a training process with the aid of an artificial neural network, called the Sparse Autoencoder, to extract the damage index. The input data for this Autoencoder comprises the dynamic track responses acquired numerically. The damage index obtained is increased by applying the Mahalanobis distance. Consequently, an outlier analysis is applied to detect previously simulated damage and a cluster technique is used to classify damage in two phases: i) identifying the type of damage (e.g. wheel flat or polygonal wheel) and ii) identifying the severity of each type of damage. Finally, the methodology developed is subject to validation based on a sample of dynamic responses acquired experimentally.
This dissertation is focused on developing and applying a hybrid strategy of Artificial Intelligence (AI) techniques to detect out-of-roundness (OOR) wheels on the Alfa pendular train. This algorithm is supplied with dynamic response data obtained through a wayside monitoring system comprising a set of accelerometers installed on the railway tracks. The method was developed with several 3D numerical simulations of railway passages, allowing the dynamic response of the vehicle-track system to be extracted. The numerical simulations cover two circulation possibilities: i) normal conditions (baseline scenarios) and ii) abnormal conditions (damage scenarios). Both scenarios are characterized by a set of traffic variabilities, as well as the presence of irregularity profiles and noise. Regarding the damage scenarios, two types of geometric defects were considered, one referring to the presence of defect on the wheel tread and a defect around the complete wheel circumference called polygonization. The core of the methodology developed includes a training process with the aid of an artificial neural network, called the Sparse Autoencoder, to extract the damage index. The input data for this Autoencoder comprises the dynamic track responses acquired numerically. The damage index obtained is increased by applying the Mahalanobis distance. Consequently, an outlier analysis is applied to detect previously simulated damage and a cluster technique is used to classify damage in two phases: i) identifying the type of damage (e.g. wheel flat or polygonal wheel) and ii) identifying the severity of each type of damage. Finally, the methodology developed is subject to validation based on a sample of dynamic responses acquired experimentally.
Description
Keywords
Wayside condition monitoring Out of roundness wheel damage passenger train Alfa pendular Sparse Autoencoder damage index outliers? analysis cluster analysis