AI and Sustainability - Session III
Detection of wind turbine motion between satellite bands with convolutional neural networks
Due to the design of the MutliSpectral Instrument on-board Sentinel-2 Satellite, each spectral band observes the ground surface at different times. Recent studies have examined this temporal band offset, mostly to predict aircraft or ship velocities but also for the movement of waves. However, these methods come with the premise of somewhat consistent backgrounds and clear horizontal movements. This paper presents an approach for detecting motion between blue, green and red satellite bands in changing environments to evaluate wind turbine motion using a convolutional neural network (CNN). With this technique, an automatic recognition of turbine activity has been developed, avoiding the barely visible vertical motion in the labelling stage by focusing on the shadow spin. This has been done by creating open source data that is used for training and validation. By focusing on a binary classification approach between spinning wind turbines and non-turbine images, it has been found that a classification is possible and accurate with this method. In addition, limitations and peculiarities of the data and the band offset are described, including an analysis of the occlusion sensitivity. This detection can be useful for precise remote sensing of activity at a given location and is therefore not only of interest to the wind energy industry, which currently only works with proprietary data for energy efficiency or other activity based turbine improvements, but also for environmental monitoring and protection.
Proof of concept for a new battery sorting method based on deep learning image classification
Battery recycling requires efficient sorting based on chemical composition. Traditional methods like X-Ray or Electromagnetic Sensors lack automation, with X-Ray sorting 26 batteries and electromagnetic sorting only 6 batteries per second. We propose using deep learning image classification to detect battery manufacturer and product series. Our prototype includes a conveyor belt, webcam, ring light, and Nvidia Jetson AGX Orin. With a dataset of 9 battery series, we achieved over 99% validation accuracy using a pretrained MobileNetV2 model. The model can classify 50 images per second with limited hardware. This approach offers potential for automated sorting, significantly improving recycling throughput and efficiency. Further research should expand the dataset and explore applicability to other battery types, optimizing the model and hardware configuration.
Graph-Based Time Token Recognition using Graph Neural Networks and Stanza Library
This publication aims to explore the capabilities of the Stanza NLP library on one hand and Graph Neural Networks on the other hand by combining them in a time token classification task. After providing information on the German ”WikiWarsDE” dataset that is used for training, the evaluation dataset consisting of environmental documents as well as the transformations applied to both datasets to enhance the model, the setup of the Neural Network is presented. Afterwards, some early test results are evaluated before possible enhancements to the model are suggested.
Development of a smart farming dashboard based of 5G mobile Data
This work in progress paper is written as a short description mainly of the backend of project 5G, which is in the field of smart farming. The project focuses on using different technologies and machines for weed management. This work in progress paper highlights the need for efficient weed management. It discusses the problems which are associated with weed management and it raises questions that need to be addressed in this domain. Moreover, the topic of using weed management 5G, UAV (unmanned aerial vehicle) and field robotics in agricultural and farming services is an important topic at present. Besides, the work in progress paper shows possible technical concepts and processes which can be implemented into smart farming to increase its efficiency. This paper discusses special methods, which can be used in weed management by using AI (artificial intelligence). In addition to the project description, the paper includes an evaluation of the current state of the research and an outlook of potential future research.