Hannes Vietz

Abstract

Indoor positioning systems are an enabling technology for many current developments in the manufacturing field like digital twins and robot fleet management. Utilizing 5G for positioning promises high accuracy, reliability, and costefficiency due to shared hardware usage for communication and positioning. Which positioning technique suits 5G-bases positioning best for manufacturing is still an open research question. This paper presents a deep learning approach for 5Gbased positioning. The first results of our research work in progress obtained at the research factory ARENA 2036 indicate a positioning accuracy in the centimeter range.

Abstract

Trotz hoher Lösungspotentiale des maschinellen Lernens für gängige Probleme der Automatisierungstechnik, finden sich in der Praxis wenig Anwendungsbeispiele. Um der Ursache hierfür auf den Grund zu gehen, zeigen die Autoren anhand von vier beispielhaften Anwendungsfällen die Hürden für konventionelles maschinelles Lernen auf und benennen Lösungsansätze mittels industriellem Transfer-Lernen. Zum großflächigen Einsatz derartiger Ansätze, fehlt es an Voraussetzungen, deren Schaffung jedoch anders als beim konventionellen maschinellen Lernen grundsätzlich möglich ist. Der Artikel schließt mit einer Betrachtung dieser Voraussetzungen und macht Vorschläge, wie sie zu erfüllen sind.

Abstract

Despite the high solution potential of machine learning for common problems in automation technology, there are only few examples of its application in real-world manufacturing practice. In order to find the reason for this phenomenon, the authors identify the hurdles for conventional machine learning using four exemplary use cases namely self-learning robots, wear prediction, visual object detection, and predictive quality in manufacturing. While these use-cases differ in principle, the problems engineers face when using conventional machine learning approaches to solve them are related, such as the lack of manifold training data or high dynamics of industrial processes. The authors showcase that utilizing deep transfer learning and continual learning approaches in the industrial context – subsumed under the term industrial transfer learning – can overcome these hurdles. Even for industrial transfer learning, there is a deficiency regarding preconditions for the large-scale deployment of such approaches, but unlike in conventional machine learning, it is principally possible to establish those. The article concludes with a discussion of these prerequisites and makes suggestions as to how they could be fulfilled.

Abstract

Vision applications are becoming increasingly important for product quality surveillance in manufacturing. Training consistently well-performing visual detection algorithms based on convolutional neural networks is very challenging. Typically, there is too much training data for engineers to keep track of possible gaps in it. But even small cases of missing training data e.g. certain viewing angles can lead to trained CNNs that are unable to detect objects, that seem obvious to engineers i.e. cognition gaps. This paper presents how synthetic training data can be created in a targeted manner to close cognitive gaps of a CNN for specific use-cases. The proposed methodology uses 3D rendering to create new image data by variating scene parameters. The created data is used to reveal a cognition gap of a CNN. We show that by using this created synthetic data to train the CNN the cognition gap can be successfully closed. This is evaluated with the well-known AlexNet CNN used as a visual bicycle detector. The bicycle example is used as a stand-in for a geometrically interesting, but simple product, that is manufactured in large and growing amounts.

Abstract

In partially automated manufacturing, humans work together with mobile robots. Trajectory prediction, i.e. predicting future positions of human workers, improves collaboration and coexistence between humans and robots on the shop floor. In this paper, we discuss the interrelated research questions of how human motion trajectories can be predicted and how mobile robots such as Autonomous Mobile Robots and Automated Guided Vehicles can take such predictions into account in their pathfinding and navigation. On the robot side, advanced D* pathfinding algorithms allow robots to take dynamic obstacles into account. For trajectory prediction, the position of human workers is determined by an Ultra-Wideband-based Real-Time Locating System. A trajectory prediction framework is introduced to support the implementation and use of pattern- and planning-based trajectory prediction algorithms. The evaluation is based on scenarios from the addressed problem area of manufacturing.

Abstract

Developing consistently well performing visual recognition applications based on convolutional neural networks, e.g. for autonomous driving, is very challenging. One of the obstacles during the development is the opaqueness of their cognitive behaviour. A considerable amount of literature has been published which describes irrational behaviour of trained CNNs showcasing gaps in their cognition. In this paper, a methodology is presented that creates worst-case images using image augmentation techniques. If the CNN's cognitive performance on such images is weak while the augmentation techniques are supposedly harmless, a potential gap in the cognition has been found. The presented worst-case image generator is using adversarial search approaches to efficiently identify the most challenging (worst) image. This is evaluated with the well-known AlexNet CNN using images depicting a typical driving scenario.

Abstract

Data-science-driven development projects are increasingly gaining the attention of small and medium sized enterprises. Since SME are often lacking the necessary competencies in data science, cooperation with other companies or universities is required. The efficient handling of data is one of the main challenges in joint cross-enterprise development projects. Actual cost driver is the development of data by labeling and classifying the data by domain experts, which is very time-consuming and labor-intensive with large amounts of data. Furthermore, clearance processes also have a high potential to cause delays before data can be shared with project partners. Moreover, before the actual work can begin, it is often necessary to clean up and repair incomplete or noisy data. The concept of Data Administration Shell presented in this paper addresses the challenge of structured information sharing and information management in joint cross-enterprise engineering. The Data Administration Shell links data sets to information regarding data origin and already performed analyses including their results and program scripts. Adding relations and documentation facilitates the reuse of data sets for subsequent projects. For this purpose, the Data Administration Shell adapts the concepts serving the information sharing in the research field of manufacturing and Digital Twin. The evaluation of the Data Administration Shell was based on time-series measurement data from a production process optimization scenario. Here, the Data Administration Shell manages the data sets of time series data and facilitates the joint cross-enterprise engineering of data-driven solutions.

Abstract

Recurrent neural networks and exceedingly Long short-term memory (LSTM) have been investigated intensively in recent years due to their ability to model and predict nonlinear time-variant system dynamics. The present paper delivers a comprehensive overviewof existing LSTM cell derivatives and network architectures for time series prediction. A categorization in LSTM with optimized cell state representations and LSTM with interacting cell states is proposed. The investigated approaches are evaluated againstdefined requirements being relevant for an accurate time series prediction.These include short-term and long-term memory behavior, the ability for multimodal and multi-step ahead predictions and the according error propagation.Sequence-to-sequencenetworks with partially conditioning outperform the other approaches, such as bidirectional or associative networks, and are best suited to fulfill therequirements.

Abstract

Fault prediction based upon deep learning algorithms has great potential in industrial automation: By automatically adapting to different usage contexts, it would greatly expand the usefulness of current predictive maintenance solutions. However, restrictions regarding the centralized accumulation of data necessary for such automatic adaption call for a distributed approach to training these algorithms. Therefore, in this paper, a continual learning based algorithm for fault prediction is presented, allowing for distributed, cooperative learning by elastic weight consolidation. This algorithm is then evaluated on a large NASA turbofan engine dataset and shows promising results regarding the performant training on decentral sub-datasets for industrial automation scenarios.