In this work, the development and validation of an AI- and sensor-based inline quality monitoring system for the analysis of particle size distributions (PSDs) of comminuted construction and demolition waste (CDW) material flows are described. In this, a custom-developed multitask CNN (CDW-MT-CNN) was developed using manually sieve analyzed particles. This model is able to rapidly and simultaneously predict the particle class and weight, essential for the determination of the PSD. The single particle data are then aggregated per raw image, usually consisting of around 1000 particles for full-scale experiments, to acquire a per-image PSD. The inline mounted RGB line scan sensor records high-resolution images in subsecond frequencies. With an inference time of around 54 ms for a single image, this model would be able to provide a PSD every minute in a full-scale plant. For the purpose of inline monitoring of CDW material flows in a comminution process, such intervals are sufficient according to experts and solve existing gaps regarding the upscaling of laboratory-developed systems. Together with the high predictive performance of the model, especially in terms of classification (82% accuracy), it is shown that this technology has potential for monitoring in full-scale plants, for instance by offering operators new insights to improve operation efficiency. Further research should focus on increasing the precision for weight prediction, for instance by increasing the labeled data set with a larger number of unique particles and on methods to verify the performance of the model on pilot or full-scale plants during live operation.
Göbbels et al. (Mon,) studied this question.