Abstract
With the increase in waste flows, industrial sorting has become a major issue. The main challenge is to minimise sorting errors in order to avoid a significant degradation in the quality of the final recycled product leading to poor recyclability of second life materials. In this thesis, we have data provided by an optical technology in an industrial context (Pellenc ST company). This technology is equipped with a medium infra-red camera capable of providing information regardless of the colour of the plastics (previous cameras were not suitable for dark and black plastics). However, even with the most recent acquisition technologies based on spectral imaging, plastic recognition remains a difficult task for certain types of plastic.Indeed, the presence, on the one hand, of uncertainty about the nature of the material due to measurement variabilities such as the ageing of plastics, surface hygroscopy, the addition of anti-UV opacifiers (carbon blacks), etc. associated with the presence, on the other hand, of imprecision due to incomplete information, i.e. the absence of certain characteristics, hamper the use of classic classification algorithms.Beyond the recognition stage, the project focuses on the sorting phase: to which container should the fragment be sent given the imprecise and uncertain information on the nature of this plastic fragment on the one hand, and the objectives and constraints of the sorting process on the other hand.In addition, impurity thresholds are set for each container.In this thesis we propose a sorting approach adapted to imperfect data such as those obtained in the industrial context of the Pellenc ST company. The proposed approach consists of two phases: a classification phase and a sorting phase. For the classification phase, we proposed a conservative classification algorithm in the framework of belief functions based on : 1) an automatic partial re-labelling of "difficult" learning examples, 2) the construction of a mass function as information on the type of plastic and 3) a decision rule where the decision maker could introduce his constraints of prudence and precision. Concerning the sorting phase, we posed an optimisation problem in the framework of belief functions to choose the appropriate container for a given fragment. This problem consists in minimizing the quantity of rejected fragments (lost for recycling), maximizing the gain, i.e., purity of expensive plastics, while respecting constraints on the proportions of impurities. The experiments we have conducted on plastics sorting show the gain in quality of the recycled product that our conservative approach provides compared to deterministic or probabilistic approaches.