Artificial Intelligence for Waste Characterisation and Real-Time Mass Balances

Abstract

Urban mining is defined as the process of reclaiming raw materials from spent products, buildings and waste and is often put forward as one of the solutions to fulfill our ever increasing demand for materials in highly urbanized areas. The strong temporal, seasonal and regional variations of this waste and its intrinsic heterogeneity are a source of high insecurity and risk for the waste processing industry and puts high pressure on the applied processing technologies. Processing plants should therefore continuously adapt to changing input variations to warrant optimal material valorisation. However, due to the lack of suitable (continuous and fast) characterisation methods this is often not possible. As a consequence, the input variability translates directly to the output streams. The variable quality of secondary materials strongly decreases market interest in these materials and hampers the transition to a circular economy. Quality assessment is traditionally performed by superficial visual inspection or manual separation of too small and possibly non-representative samples, and is therefore often not reliable. In addition the task is tedious, time-intensive, subjective and rather unpleasant. To meet this need for a rapid, continuous, automatic, objective and reliable characterisation technology, a device, combining different sensor types, was built.

The technology will allow to optimize existing and to develop new recycling processes, and assess secondary raw material quality, based on accurate, representative and objective data.

Current sensor techniques in waste characterization mainly focus on surface properties, e.g. near-infrared, colour, hyperspectral or X-ray fluorescence. However waste material is often dirty and the surface properties are not representative for the bulk of the material. To overcome this limitation, a technology that sees “through” the material was adopted: X-ray Transmission (XRT). By measuring at two energy levels, called Dual Energy (DE-XRT), it is possible to determine material properties such as the average atom number and density. To accurately interpret the information gathered by DE-XRT, extra information such as the 3D shape and volume of the object is employed. This is measured by 3D laser triangulation (3DLT). 3DLT is a well-known technology in the industry that can measure the geometry of object at high resolution (sub-mm) using a laser and a camera. The combination of these technologies allows to fully characterise a waste stream on the level of individual particles with respect to volume, mass, shape and composition. Using this information, accurate mass balances can be measured. In addition, the material and shape measurement is complemented by an RGB detector, bringing in additional information which can be used to better differentiate the materials using image processing and machine learning algorithms.

The development of methods to extract the relevant information from the sensor data is the topic of ongoing research. The technology will allow to optimise existing and to develop new recycling processes, and assess secondary raw material quality, based on accurate, representative and objective data. During the conference, a real demo case will be presented based on mixed construction and demolition waste typically collected by SUEZ in Belgium.

Keywords: waste, characterisation, multi-sensor, sorting, machine learning, artificial intelligence, urban mining