ABB and Salzburg University File Patent on AI-Powered Energy Optimization for Industrial Drives

The Problem: Models Don't Capture What Actually Happens

Industrial automation has a dirty secret. Control engineers spend countless hours building mathematical models of drive systems, tuning them to predict energy consumption with increasing precision. The problem is that no model captures everything. There are losses in bearings, heat dissipation in cables, friction in seals that vary with temperature and wear. Some of these can be measured, but modeling them accurately across all operating conditions remains impractical.

The result is that even well-tuned systems waste energy. A robot performing thousands of pick-and-place cycles per day might be consuming 15-20% more power than necessary because its motion profiles were optimized against an incomplete model.

Conventional approaches hit a wall here. You can add more sensors, more complex models, more parameters to tune. But you're still trying to analytically describe something that's inherently messy and context-dependent.

The Solution: Let the Machine Figure It Out

The Josef Ressel Center for Intelligent and Secure Industrial Automation (JRZ ISIA) at Salzburg University of Applied Sciences, working with ABB's Machine Automation Division (B&R), took a different approach. Instead of trying to model the losses, they decided to bypass the models entirely.

The team deployed reinforcement learning agents directly on physical drive systems. The agent interacts with the machine, running motion sequences and measuring actual energy consumption. It learns, through trial and error, which motion profiles minimize energy use under real operating conditions. No complete system model required.

This isn't a new idea in robotics research, but there's a catch. Standard RL approaches are notoriously data-hungry and slow. Training a neural network to control a complex multi-axis system can take days or weeks, which makes it impractical for industrial deployment where downtime costs money and systems frequently change configuration.

The Salzburg team's contribution is a new mathematical formulation of the learning strategy that drastically reduces both the data requirements and training time. They haven't published exact figures yet, but the implication is that what previously took days now takes hours. The approach is specifically designed for cyber-physical systems where the RL agent must safely interact with real hardware.

The Outcome: Patent Filed, Research Moving Forward

The team has filed a joint patent application covering this energy-optimized motion control methodology. The scope includes industrial robots, machine tools, and automated production lines, specifically targeting the dynamic motion sequences like positioning, acceleration, deceleration, and cyclic movements that dominate energy consumption in these systems.

The research foundation goes back to 2020, originating in the EU Interreg project KI-Net. Since 2022, the work has continued at JRZ ISIA with B&R and other industry partners including COPA-DATA. This patent represents the transition from research to practical application.

Whether this approach becomes standard practice depends on whether the reduced training time proves scalable across different machine types and whether the energy savings hold up in long-term industrial deployment. But for engineers wrestling with the gap between model-based predictions and actual energy consumption, this represents a fundamentally different path forward.

The patent application is pending. No timeline for commercial implementation has been announced.

M4S TAKE

My take: AI claims need scrutiny. The useful implementations reduce cycle time or defect rates in measurable ways. Vague promises about 'optimization' without specific metrics are usually marketing.

Simon McLoughlin