Deep Learning

A new look into the depths of technical drawings: A deep learning framework reads CAD drawings more accurately than ever before, recognizing geometrical dimensioning and tolerancing, dimensions, and every other detail. What used to be tedious manual labor is now carried out by an AI that understands the special features of every line and label. This promising technology not only increases accuracy but also speeds up the processing of drawings considerably. The system thus opens up new avenues for precision in production.

The complex dynamics of block warehouses pose major challenges to the manual stocktaking process. Frequent relocation of pallets, crates or pallet cages without fixed storage locations leads to a time-consuming and error-prone inventory process, wherein goods often have to be searched for and damages due to improper storage can occur. The use of (semi-)autonomous drones offers a promising solution to enable automated stocktaking, especially if these are appropriately equipped for optical goods detection.

Deep learning-based error detection methods outperform traditional methods because of the continuously increasing complexity of technical systems and inherent flexibility and scalability of Deep Learning techniques. This article introduces the KrakenBox – an autonomous Deep Learning-based error detector for industrial Cyber-Physical Systems (CPS). It exploits a lightweight, Long Short-Term Memory (LSTM) network capable of online error detection that can be deployed on an embedded platform such as NVIDIA Jetson AGX Xavier or even Google Coral Edge TPU. This article describes the architecture of the KrakenBox and demonstrates its application with two case studies.

Manufacturing companies must work economically while delivering quality - in some industries with a zero-defect tolerance. Quality control often is carried out manually and with a time delay, thus errors can only be corrected at a late stage. Using artificial intelligence (AI), visual quality control can be automated, carried out in real time and integrated into the production process - making it more accurate, efficient and cost-effective. A case example shows the advantages of tackling AI issues in interdisciplinary teams with partners.

Due to competitive pressure pharmaceutical companies are also driven to increase the efficiency of their processes. In this paper an approach for the predictive detection of malfunctions of filling systems for powdery pharmaceutical products using machine learning is presented. The focus is on the prediction of filling deviations with recurrent neural networks, with the objective to detect a drift in the process stability to intervene accordingly.

Intensified networking and digitalization of systems affect an increasing number of sectors. At the same time a great variety of different concepts, ideas, expectations as well as fears have emerged around Industry 4.0. A look into the newspapers is enough to understand that the profound connection of critical structures does also hold profound dangers. For the future it is crucial to consider a way of using the new mass of data and information to protect these structures. Evaluating big data and transforming it into smart data with support of Artificial Intelligence will be a significant security factor in the future.