Proposal for a new approach to forecast spare parts demand
Data
2021-08-03
Tipo
Dissertação de mestrado
Título da Revista
ISSN da Revista
Título de Volume
Resumo
A previsão de peças de reposição para apoiar as operações de manutenção é essencial para garantir a disponibilidade dos mais diversos serviços. Este documento tem como objetivo investigar quais técnicas têm sido propostas para prever a demanda por peças de reposição, bem como em que medida essas técnicas foram validadas em ambientes industriais. O documento também visa desenvolver uma nova metodologia de previsão de peças de reposição por meio da seleção ou combinação de técnicas de previsão. Primeiramente, conduzimos uma Revisão Sistemática da Literatura sobre o tema de Previsão de Demanda de Peças de Reposição. As principais conclusões da revisão são de que falta um estudo comparativo quanto ao desempenho entre as técnicas analisadas e que muitos estudos ainda lutam com a necessidade de grandes volumes de dados e grandes conjuntos de variáveis. O presente estudo baseia-se nessas lacunas e implementa um algoritmo de meta-aprendizagem. Uma base de dados real de demanda de peças de reposição do setor aeroespacial é usada como entrada para um Classificador Random Forest para selecionar entre os modelos Random Walk, Theta, ARIMA, Exponential Smoothing e Neural Network para prever a demanda e calcular as combinações de previsão. Os novos métodos são então comparados com métodos convencionais e com um software de uso comprovado pela indústria. A metodologia de previsão de meta-aprendizado proposta mostra-se uma opção válida e competitiva para prever a demanda de peças de reposição e apresenta bons resultados quando comparada aos métodos convencionais de previsão e também quando comparada a uma solução comprovada usada pela indústria que possui seu próprio algoritmo de seleção de métodos de previsão. As previsões de combinação linear ponderada mostraram bons resultados gerais e conseguiram melhorar a precisão média das previsões. Finalmente, uma aplicação web simples foi desenvolvida como uma prova de conceito para uso industrial.
Forecasting spare parts to support maintenance operations is essential to guarantee the availability of a wide variety of services. This document aims to investigate what techniques have been proposed to predict the demand for spare parts, as well as the extent to which these techniques have been validated in industrial environments. The document also aims to develop a new forecasting methodology for spare parts through the selection or combination of forecasting techniques. First, we conducted a Systematic Literature Review on the theme of Spare Parts Demand Forecasting. The main conclusions are that there is a lack of a comparative study with respect to the performance between the techniques analyzed and that many studies still struggle with the need for large volumes of data and large sets of variables. The present study builds on these gaps and implements a meta-learning algorithm. A real data-base of spare parts demand in the Aerospace sector is used as input for a Random Forest Classifier to select between the Random Walk, Theta, ARIMA, Exponential Smoothing and Neural Network models to forecast demand and to calculate forecast combinations. The new methods are then compared with conventional methods and with an industry-proven software. The proposed meta-learning forecasting methodology shows a valid and competitive option to forecast spare parts demand and shows good results when compared to conventional forecast methods and also when compared to a proven industry-used solution which has its own forecast method selection algorithm. The weighted linear combination forecasts have show good overall results, and managed to improve the average forecast accuracy. Finally, a simple web application was developed as a proof-of-concept for industrial use.
Forecasting spare parts to support maintenance operations is essential to guarantee the availability of a wide variety of services. This document aims to investigate what techniques have been proposed to predict the demand for spare parts, as well as the extent to which these techniques have been validated in industrial environments. The document also aims to develop a new forecasting methodology for spare parts through the selection or combination of forecasting techniques. First, we conducted a Systematic Literature Review on the theme of Spare Parts Demand Forecasting. The main conclusions are that there is a lack of a comparative study with respect to the performance between the techniques analyzed and that many studies still struggle with the need for large volumes of data and large sets of variables. The present study builds on these gaps and implements a meta-learning algorithm. A real data-base of spare parts demand in the Aerospace sector is used as input for a Random Forest Classifier to select between the Random Walk, Theta, ARIMA, Exponential Smoothing and Neural Network models to forecast demand and to calculate forecast combinations. The new methods are then compared with conventional methods and with an industry-proven software. The proposed meta-learning forecasting methodology shows a valid and competitive option to forecast spare parts demand and shows good results when compared to conventional forecast methods and also when compared to a proven industry-used solution which has its own forecast method selection algorithm. The weighted linear combination forecasts have show good overall results, and managed to improve the average forecast accuracy. Finally, a simple web application was developed as a proof-of-concept for industrial use.