Creating trustworthiness for industrial AI applications

Four dimensions for qualifying AI applications in production technology.

Rapidly growing volumes of data in industrial processes, advances in learning algorithms, and falling hardware costs have led manufacturing companies to discover the value of AI as well. AI and machine learning (ML) promise to improve the efficiency of production processes through data-driven modeling and advanced analytics. Although manufacturing companies recognize the importance and future potential of AI, they often fail to profitably integrate AI models into their processes and production systems.

The reasons for this lie in the challenging environment in which industrial AI applications are to be used: High complexity and low quality of the process data, uncertainty and intransparency of the AI models as well as highest quality and reliability requirements from the production area characterize the tasks associated with the implementation.

The fact that AI applications still have a certain degree of novelty means that companies still have little confidence in the use of AI and are reluctant to integrate it into their processes. In order to promote the spread of the technology, the "AI Certification" working group at the Fraunhofer IPT is investigating what requirements are placed on industrial AI applications and how they must be developed, rolled out and maintained.

To this end, the AI experts have designed a framework and procedure model for the development of trustworthy industrial AI applications that is specifically tailored to the challenges of production technology. The development framework follows a risk-based approach that considers the level of autonomy and the safety and economic criticality of the application. It provides rigorous specification and documentation guidelines and focuses on four dimensions identified as key requirements for qualifying AI applications in production technology:


Understandability and comprehensibility of how AI works and the decision-making process.


Extent to which the AI system behaves stably during training, validation, and operation under minor variations in the data.


Ability of the system to continuously respond to dynamics in the data and maintain functionality on a sustained basis.


Secured functioning of the AI application, especially in case of erroneous input data.

This taxonomy and a specially developed method library help to select suitable qualification methods and AI models that take into account all relevant user and regulatory requirements, the individual characteristics of the use case, and finally the properties of the data set. An iterative process model that combines the best of agile and certifiable software development ensures that the entire AI application – and not just the AI model itself – meets the high demands of industrial production. Involving users and process experts in the development of safe and transparent AI is of central importance to us. In this way, we ensure that AI applications are adapted to the needs of employees and that their AI competence and sovereignty in dealing with AI is strengthened.

Our goal is to unlock the maximum benefits of AI for the manufacturing industry and to increase technological maturity so that industrial AI applications become cost-effective to implement and use, rather than operating as isolated solutions.