Artificial intelligence offers great potential, for example in the quality control of manufacturing companies. However, training AI models is difficult and requires mathematical knowledge, as there are countless parameters that can be included in such an analysis. The barriers to entry are therefore high - small and medium-sized companies that do not have their own development department often shy away from AI applications. Expertise is also required during operation: if an AI algorithm has been trained and the product design or geometry of the component is then changed slightly, the algorithm initially recognizes this as an error. In this case, the AI must be retrained.

AI can also be operated without expert knowledge

The IDMT-ISAAC software from the Fraunhofer Institute for Digital Media Technology IDMT in Ilmenau helps users without expert AI knowledge to overcome these hurdles. IDMT-ISAAC stands for Industrial Sound Analysis for Automated Quality Control. "We want to enable SMEs to adapt and customize the AI algorithms themselves," says Judith Liebetrau, Group Manager Industrial Media Applications at Fraunhofer IDMT. "They can apply IDMT-ISAAC to their own audio data, retrain it and thus obtain fast and reliable results and decision-making aids for their quality assurance."

As experienced machine operators know, errors can often be recognized by the sound of the process. IDMT-ISAAC also relies on acoustics: the researchers have trained the system with recorded acoustic data from welding processes, the AI analyzes the typical process noises and draws conclusions about the quality of the respective weld seam from the audio data. At the heart of IDMT-ISAAC is a framework that allows users to change various parameters with just a few clicks, for example to teach the AI to change the geometry of the product. In summer 2021, the software should be adapted to live operation - meaning it could then immediately analyze real-time data from production and optimize quality assurance. In three to four years, it should also be able to actively intervene in production. However, the framework does not only offer new analysis options for welding. "We have integrated various methods into the modular system in order to be able to map other processes relatively quickly," explains Liebetrau. In the future, it should also be possible for companies to use their own software and access the Fraunhofer IDMT's AI via an interface on its server. However, regardless of whether the companies integrate the AI via the framework or access it via an interface: Data protection and data security are always observed, as the data is processed anonymously.

Understanding the AI's decisions

The software can be adapted to different user groups via various user profiles: AI beginners as well as AI experts. For example, it is very interesting for developers of AI algorithms to get a feel for how the AI makes its decisions and which sounds it uses to make them. "With the framework, we are therefore also moving towards explainable AI to make AI more comprehensible," says Liebetrau.