| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 19.23 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A identificação precoce dos defeitos das rodas ferroviárias pode evitar danos graves quer ao nível do veículo quer da via reduzindo consideravelmente os custos de manutenção tanto para as administrações ferroviárias como para os operadores do material circulante. No presente trabalho é apresentada uma metodologia não supervisionada baseada em técnicas de inteligência artificial, que permite a identificação de multi-danos, por exemplo danos “Out of Roundness” e desequilíbrios de carga, com recurso a respostas dinâmicas induzidas na via pela passagem de um veículo ferroviário de mercadorias do tipo Laagrss. Nesta fase inicial, para testar a eficácia da metodologia, foram utilizados registos numéricos das passagens dos comboios que foram simulados com base em modelos numéricos do veículo, de irregularidades da via e das rodas e a interação roda-carril. Estas respostas são posteriormente sujeitas a um conjunto de etapas: i) análise outlier, ii) localização do dano, iii) identificação do tipo de dano e iv) classificação da severidade do dano. Na primeira fase, a deteção dos danos é realizada com base num limite de confiança usando o sinal completo. De seguida, é realizada a localização dos diversos danos presentes no veículo, através de uma segmentação do sinal, permitindo estabelecer janelas de corte de forma a isolar os vários danos, a fim de tornar o problema menos complexo. Numa terceira fase, já com os vários sinais cortados, as principais características são extraídas através da combinação da técnica da análise de componentes principais e de um autoencoder esparso, com a “Bottleneck layer” a desempenhar um papel crucial na obtenção dos indicadores. Estes indicadores após sofrerem uma fusão, com recurso à distância de Mahalanobis, são comparados em termos de extensão, aceleração e frequência, com o objetivo de distinguir qual o tipo de defeito presente no veículo, se poligonização, liso ou desequilíbrio de cargas. Por fim, com base numa metodologia desenvolvida nesta tese para tratar defeitos isolados, é possível distinguir as diferentes severidades do dano.
Early identification of railway wheel defects can prevent serious damage to both the vehicle and the track, considerably reducing maintenance costs for both railway administrations and rolling stock operators. To this end, this paper presents an unsupervised methodology based on artificial intelligence techniques, which allows the identification of multi-damage, for example “Out of Roundness” damage and unbalance loads, using dynamic responses induced on the track by the passage of a Laagrss-type rail freight vehicle. In this initial phase, to test the effectiveness of the methodology, numerical records of train passages were used and simulated using numerical models of the vehicle, track and wheel irregularities and wheelrail interaction. These responses are then subjected to a series of steps: i) outlier analysis, ii) damage localization, iii) damage type identification and iv) damage severity classification. In the first phase, damage detection is carried out based on a confidence limit using the complete signal. Next, the various damages present on the vehicle are located by segmenting the signal, allowing cutting windows to be established to isolate the various damages in order to make the problem less complex. In a third phase, with the various signals cut out, the main characteristics are extracted by combining the principal component analysis technique and a sparse autoencoder, with the "Bottleneck layer" playing a crucial role in obtaining the indicators. After being merged using the Mahalanobis distance, these indicators are compared in terms of strain, acceleration and frequency, with the aim of distinguishing the type of defect present in the vehicle, whether polygonization, flat or unbalance loads. Finally, based on a methodology developed in this thesis to deal with single damage, it is possible to distinguish the different severities of the damage.
Early identification of railway wheel defects can prevent serious damage to both the vehicle and the track, considerably reducing maintenance costs for both railway administrations and rolling stock operators. To this end, this paper presents an unsupervised methodology based on artificial intelligence techniques, which allows the identification of multi-damage, for example “Out of Roundness” damage and unbalance loads, using dynamic responses induced on the track by the passage of a Laagrss-type rail freight vehicle. In this initial phase, to test the effectiveness of the methodology, numerical records of train passages were used and simulated using numerical models of the vehicle, track and wheel irregularities and wheelrail interaction. These responses are then subjected to a series of steps: i) outlier analysis, ii) damage localization, iii) damage type identification and iv) damage severity classification. In the first phase, damage detection is carried out based on a confidence limit using the complete signal. Next, the various damages present on the vehicle are located by segmenting the signal, allowing cutting windows to be established to isolate the various damages in order to make the problem less complex. In a third phase, with the various signals cut out, the main characteristics are extracted by combining the principal component analysis technique and a sparse autoencoder, with the "Bottleneck layer" playing a crucial role in obtaining the indicators. After being merged using the Mahalanobis distance, these indicators are compared in terms of strain, acceleration and frequency, with the aim of distinguishing the type of defect present in the vehicle, whether polygonization, flat or unbalance loads. Finally, based on a methodology developed in this thesis to deal with single damage, it is possible to distinguish the different severities of the damage.
Description
Keywords
Danos nas rodas OOR Desequilíbrio de cargas Deteção e classificação de danos Comboios de mercadorias Sistemas de monitorização wayside Autoencoder esparso empilhado Inteligência artificial OOR wheel damage Unbalance loads Damage detection and classification Freight trains Wayside monitoring systems Stacked sparse autoencoder Artificial intelligence