Thomas Wolf’s research while affiliated with Zolitron Technology GmbH and other places

Optimization of waste collection is an important component in achieving sustainable solid waste management for modern cities. A basic prerequisite for route optimization is, naturally, information about the current fill levels of the container. This paper proposes a novel method for reliable monitoring of fill levels in glass containers. A data-based, hybrid approach was taken, by using vibration data recorded from an accelerometer mounted on each container. Firstly, a deep learning model is used to estimate the mean and variance of the current fill level. Two types of Bayesian neural networks and one ensemble-based model have been compared for this purpose. The final fill-level estimate is computed by a discriminative Kalman filter, improving estimation quality and reducing variance by fusing the instantaneous estimates with a simple, linear statistical model of the filling process. The proposed method was tested on two large real-world datasets with over 200 containers and 300,000 recorded samples. Because of the dataset imbalance, the models were optimized using label distribution smoothing. The final results demonstrate improvements across all fill levels, with the most significant enhancement observed in the estimation of the especially important and difficult case of higher fill levels, where the estimation error was reduced by up to 60%.

Monitoring the fill levels of glass containers is important for smart cities, to simultaneously save energy and traffic by preventing unneeded pick-up routes, and to support the circular economy by ensuring that containers are always available for new recycling glass. Here, we present a novel and highly energy-efficient method for reliable monitoring of glass container fill levels. This was achieved by framing the problem as a classification problem of the container fill state, and by using a dataset consisting of over 100,000 accelerometer recordings from 106 different containers for training hybrid models that combine the best aspects of deep learning and probabilistic inference. We propose the use of hybrid models, via optimal sequential decision making based on a probabilistic output of the deep neural network. With this approach, the overall accuracy increases by more than 10% while preventing sudden changes in state prediction. Finally, we have optimized the network efficiency. For this purpose, we investigated four techniques of explainable artificial intelligence methods for time series to investigate which feature are important for classification. The final results show that this allows for training a classification model of roughly comparable performance by using only 5% of the input features, which leads to an additional improvement of 97 % in terms of energy consumption of the smart sensor.