Translated by an automated translation plugin.
With the help of specific approaches, it is now possible for companies to design complex systems more intelligently: From sequential decision-making for robotic systems and the improvement of process parameters in production to process evaluation for quality control and error detection. While traditional methods reach their limits in the face of hundreds of interdependent variables, AI techniques such as reinforcement learning and Bayesian optimisation offer flexible, data-driven solutions.
The new SATW article in the SME magazine is aimed at decision-makers and experts who want to understand how AI can improve industrial processes in concrete terms and which prerequisites are crucial for success. The article was written by Johan Poccard and Dr Iason Kastanis from the Centre Suisse d'Électronique et de Microtechnique CSEM.
This article in the SME magazine is part of a series that has emerged as a follow-up activity to the publication Orientation AI: Challenges and Opportunities for Swiss SMEs. This SATW publication supports small and medium-sized enterprises in recognising the potential of artificial intelligence and planning concrete next steps.
The content was created in close collaboration with stakeholders from the SAIROP (Swiss AI Research Overview Platform) network. SAIROP promotes the exchange between science, business and society, makes Swiss AI expertise visible and provides orientation in the dynamic AI ecosystem.
Traditional approaches work well in structured environments. However, they reach their limits in dynamic systems with hundreds of interdependent variables. The number of parameters grows exponentially. AI methods offer more flexible, data-driven solutions here.
Ideal for optimising processes: Logistics systems plan hundreds of deliveries dynamically, optimise warehouse routes or reroute lorries based on real-time traffic data. In robotics, systems adapt their actions to changing environments.
An agent is trained in a simulated environment. Through rewards, it learns which actions lead to the desired result. The simulation makes it possible to run through thousands of scenarios safely and quickly before the system is used in real life.
The biggest challenge is the discrepancy between simulation and reality. What works in simulation can fail in practice due to vibrations or sensor noise. The solution: Realistic simulations through close collaboration between developers and process experts.
Bayesian optimisation finds the best parameters with just a few, targeted experiments. Physically informed neural networks combine data with physical laws and significantly reduce the need for large data sets.
Data collection is often expensive and time-consuming, especially for physical tests. In addition, processes change due to wear and tear or environmental changes. Models then often have to be retrained.
AI models recognise patterns in real-time data that indicate errors or failures - before they occur. Unlike rigid rule-based systems, modern processes also adapt to unexpected deviations.
User-friendly interfaces enable process experts to interact with models and visualise results without having to understand technical details. This speeds up the optimisation process considerably.
No. Every project requires careful modelling and validation. Companies should systematically document their experiences. This creates a library of methods that accelerates future implementations.
For example, CSEM optimised gear grinding processes with intelligent sampling strategies and multi-objective optimisation. Production speed and product quality were maximised at the same time.
Identify a specific use case: Is there a need to optimise processes, parameters or monitoring? Start with a manageable project, systematically document your experiences and gradually build up expertise. Collaboration with research partners such as the CSEM makes it easier to get started.
| Role | Title + Name |
|---|---|
| Text by | Iason Kastanis , Johan Poccard |
