Machine Learning Method Applied Online on a Waste Sorting Plant for Early Jam Detection
Résumé
In waste sorting plants, the operators endure conveyor belts jams because sorting waste is not easy due to its complex and unpredictable nature (shapes, weight, tangle, humidity, dirt …). Their job is also hard with patrolling up to 12 km per shift, and having to manually clear the jammed waste. In addition, preventing jams means more waste sorted, more materials to reuse and less to incinerate or to bury. To answer a current non-effective solution to detect jams on conveyor belts in waste sorting plants, a Machine Learning algorithm, the k-Nearest Neighbors is implemented online on site on different conveyor belts at the same time, to detect jams before the conveyor belts stop. This paper describes the different enhancements to a previous version of the k-Nearest Neighbors model by adding normalization, rejection, adaptation of the training set, on site implementation and results.