Refine
Document Type
- Article (2)
- Master's Thesis (2)
Has Fulltext
- yes (4)
Keywords
- Neuronales Netz (4) (remove)
As a customer, it can be frustrating to face an empty shelf in a store. The market does not always realize that a product has been out of stock for a while, as the item is still listed as in stock in the inventory management system. To address this issue, a camera should be used to check for Out-of-Stock (OOS) situations.
This master thesis evaluates different model configurations of Artificial Neural Networks (ANNs) to determine which one best detects OOS situations in the market using images. To create a dataset, 2,712 photos were taken in six stores. The photos clearly show whether there is a gap on the shelf or if the product is in stock. Based on the pre-trained VGG16 model from Keras, two fully connected layers were implemented, with 36 different ANNs differing in the optimization method and activation function pairings. In total, 216 models were generated in this thesis to investigate the effects of three different optimization methods combined with twelve different activation function pairings. An almost balanced ratio of OOS and in-stock data was used to generate these models.
The evaluation of the generated OOS models shows that the FTRL optimization method achieved the least favorable results and is therefore not suitable for this application. Model configurations using the Adam or SGD optimization methods achieve much better results. Of the top six model configurations, five use the Adam optimization method and one uses SGD. They all achieved an accuracy of at least 93% and were able to predict the Recall for the OOS class with at least 91%.
As the data ratio between OOS and in-stock data did not correspond to reality in the previously generated models, the in-stock images were augmented. Including the augmented images, new OOS models were generated for the top six model configurations. The results of these OOS models show no convergences. This suggests that more epochs in the training phase lead to better results. However, the results of the OOS model using the Adam optimization method and the Sigmoid and ReLU activation functions stand out positively. It achieved the best result with an accuracy of 97.91% and a Recall of the OOS class of 87.82%.
Overall, several OOS models have the potential to increase both market sales and customer satisfaction. In a future study, the OOS models should be installed in the market to evaluate their performance under real conditions. The resulting insights can be used for continuous optimization of the model.
During spaceflight, humans experience a variety of physiological changes due to deviations from familiar earth conditions. Specifically, the lack of gravity is responsible for many effects observed in returning astronauts. These impairments can include structural as well as functional changes of the brain and a decline in cognitive performance. However, the underlying physiological mechanisms remain elusive. Alterations in neuronal activity play a central role in mental disorders and altered neuronal transmission may also lead to diminished human performance in space. Thus, understanding the influence of altered gravity at the cellular and network level is of high importance. Previous electrophysiological experiments using patch clamp techniques and calcium indicators have shown that neuronal activity is influenced by altered gravity. By using multi-electrode array (MEA) technology, we advanced the electrophysiological investigation covering single-cell to network level responses during exposure to decreased (micro-) or increased (hyper-) gravity conditions. We continuously recorded in real-time the spontaneous activity of human induced pluripotent stem cell (hiPSC)-derived neural networks in vitro. The MEA device was integrated into a custom-built environmental chamber to expose the system with neuronal cultures to up to 6 g of hypergravity on the Short-Arm Human Centrifuge at the DLR Cologne, Germany. The flexibility of the experimental hardware set-up facilitated additional MEA electrophysiology experiments under 4.7 s of high-quality microgravity (10–6 to 10–5 g) in the Bremen drop tower, Germany. Hypergravity led to significant changes in activity. During the microgravity phase, the mean action potential frequency across the neural networks was significantly enhanced, whereas different subgroups of neurons showed distinct behaviors, such as increased or decreased firing activity. Our data clearly demonstrate that gravity as an environmental stimulus triggers changes in neuronal activity. Neuronal networks especially reacted to acute changes in mechanical loading (hypergravity) or de-loading (microgravity). The current study clearly shows the gravity-dependent response of neuronal networks endorsing the importance of further investigations of neuronal activity and its adaptive responses to micro- and hypergravity. Our approach provided the basis for the identification of responsible mechanisms and the development of countermeasures with potential implications on manned space missions.
Table Tennis Tutor: Forehand Strokes Classification Based on Multimodal Data and Neural Networks
(2021)
Beginner table-tennis players require constant real-time feedback while learning the fundamental techniques. However, due to various constraints such as the mentor’s inability to be around all the time, expensive sensors and equipment for sports training, beginners are unable to get the immediate real-time feedback they need during training. Sensors have been widely used to train beginners and novices for various skills development, including psychomotor skills. Sensors enable the collection of multimodal data which can be utilised with machine learning to classify training mistakes, give feedback, and further improve the learning outcomes. In this paper, we introduce the Table Tennis Tutor (T3), a multi-sensor system consisting of a smartphone device with its built-in sensors for collecting motion data and a Microsoft Kinect for tracking body position. We focused on the forehand stroke mistake detection. We collected a dataset recording an experienced table tennis player performing 260 short forehand strokes (correct) and mimicking 250 long forehand strokes (mistake). We analysed and annotated the multimodal data for training a recurrent neural network that classifies correct and incorrect strokes. To investigate the accuracy level of the aforementioned sensors, three combinations were validated in this study: smartphone sensors only, the Kinect only, and both devices combined. The results of the study show that smartphone sensors alone perform sub-par than the Kinect, but similar with better precision together with the Kinect. To further strengthen T3’s potential for training, an expert interview session was held virtually with a table tennis coach to investigate the coach’s perception of having a real-time feedback system to assist beginners during training sessions. The outcome of the interview shows positive expectations and provided more inputs that can be beneficial for the future implementations of the T3.
Untersuchung des Potentials Neuronaler Netze für Regelungsprozesse am Beispiel eines Betonverteilers
(2018)
Ziel der Arbeit ist die Entwicklung eines Ansatzes zur Nutzung Neuronaler Netze für die Steuerung eines Betonverteilers. Nach der Erläuterung der Grundlagen Neurona-ler Netze wird ein Netz entworfen und ein Test anhand realistischer Testdaten durch-geführt. Nach einer ausführlichen Beschreibung des Umfelds des Anwendungsfalls (Betonverteiler), sowie der Grundlagen Neuronaler Netze, ihrer Lernverfahren und ihrer Einsatzmöglichkeiten in der Regelungstechnik erfolgt der Entwurf eines Ansatzes zur Konstruktion eines Netzes für den Betonverteiler als Anwendungsfall sowie die Beschreibung der Implementation von Simulationsmodell und neuronalem Netz.