Développement et optimisation des stratégies de réponse à l’obsolescence via des modèles prédictifs

Abstract

The popularity of electronic devices has sparked research to implement components that can achieve better performance and scalability. During the last decades, electronic technology has developed very rapidly, which made the life cycle of the components very short compared to the product life cycle. Support and maintenance of systems containing certain obsolete components is often difficult and costly. Due to this problem, companies face significant challenges when they use systems with a long life cycle. To overcome this problem, forecasting appears to be one of the most efficient solutions. It plays a key role in proactive and strategic management levels. Through obsolescence forecasting, companies can ensure support of parts in service. This thesis proposes methods for obsolescence forecasting using different types of forecasting, obsolescence risk forecasting and life cycle forecasting. The first part of the thesis focuses on obsolescence risk forecasting. It introduces machine learning applied for obsolescence of electronic components and presents a comparison of different algorithms that are capable to forecast obsolescence risk of a large sample of data. Random forest algorithm has shown a high degree of accuracy for obsolescence risk forecasting. This research focuses also on developing an approach based on two meta-heuristic methods which are: genetic algorithm (GA) and particle swarm optimization (PSO) which search for optimal parameters and features selection to construct random forest (RF). This approach was proven to optimize the parameter of RF as well as identify the right features extraction to improve the accuracy of RF. The experimental results obtained show that GA-RF outperform the other algorithms. The second phase of risk modeling aims to combine unsupervised learning with supervised learning in order to create groups of risk assessment. The second part of this thesis focuses on the obsolescence life cycle forecasting. It presents a review of gaps and limits in the actual approaches and proposes a method that can better forecast the life cycle. The proposed method introduces a stochastic approach to estimate the obsolescence life cycle using Markov chain and homogeneous compound Poisson process. This approach uses multiple states of the life cycle curve based on the change in demand rate and introduces hidden Markov theory to estimate the model parameters. We demonstrate the proposed approach and highlights its benefits and limitations.

Date	11 Oct 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Yvan Beauregard (Supervisor) & Thien-My Dao (Co-supervisor)