A contribution to machine learning applications in logistics and maintenance problems

• Ano de defesa: 2021
• Autor(a) principal: GONZÁLEZ, Hanser Steven Jiménez
• Orientador(a): CAVALCANTE, Cristiano Alexandre Virgínio
• Tipo de documento: Tese
• Tipo de acesso: Acesso embargado
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco
• Programa de Pós-Graduação: Programa de Pos Graduacao em Engenharia de Producao
• País: Brasil
• Palavras-chave em Português: Engenharia de Produção; Aprendizagem profunda; Racionamento de estoque; Dropshipping; Aprendizagem profunda por reforço; Sistemas de múltiplos componentes; Manutenção imperfeita
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/43731

As the time goes by, organizations acknowledge more and more the role of business support functions for the achievement of competitiveness and a sustainable performance. Considering that, it is important to propose novel mathematical models that enable the improvement of these functions. In the recent years, ML-based models have gained popularity in areas such as robotics, natural language processing, manufacturing, logistic and maintenance management. They have proven to be efficient in these complex domains in which the relation between some variables is sometimes unknown or in which the problem dimensionality and the solution space are high. Accordingly, in this thesis, we propose a maintenance and a logistic model based upon Machine Learning technics (ML) that have the capacity of dealing with the complexity of the problems approached when some real-life characteristics are taken into account. The first proposed model is based upon Deep Learning and aims to classify e- commerce orders in dropshipping systems as soon as they are placed on the internet. The model fulfils the gap in the literature in which models force e-taler to cumulate batches of orders before engaging in any order classification and inventory rationing. The second model is a Condition-based maintenance policy for multi-component systems based upon Deep Reinforcement Learning and Goal Programming. The model fulfills a gap in the literature in which real industrial system factors such as multiple degradation states, imperfect maintenance and multiple conflicting criteria are not considered. In order to validate the efficacy of each model, numerical experiments and sensitivity analyses were conducted using simulation. Results showed that the proposed models enable the improvement of key indicator performances such as order fulfilment rate, total e- tailer’s profit, maintenance cost rate and average system’s reliability, in different scenarios.