Background
Production lines continuously generate sensor data, yet conventional monitoring systems are expensive, require machine-specific calibration and do not provide real-time metrics. Trends such as mass customisation, volatile supply chains and rising energy costs demand new levels of process transparency. Existing AI approaches (deep learning) require large amounts of training data and, as black-box systems, are often not certifiable. What is needed is explainable, scalable real-time intelligence without interrupting production.
Technical Description
The patented method measures electrical power consumption via a non-invasive sensor and streams the data in real time to a computing unit (edge or cloud). An optimised streaming algorithm (based on Dynamic Time Warping) reduces complexity from N² to O(N) per time step, ideal for Edge AI / TinyML. Calibration is performed in just three cycles (few-shot), without manual labelling. A two-stage logic detects deviations even whilst the process is running. The key metrics (duration, energy, similarity) are physically interpretable, certifiable and can be directly integrated into SCADA, MES and ERP systems, or used as a data source for AI agents and digital twins.
Possible Applications
The technology addresses key challenges in Industry 4.0 and Industry 5.0 and can be applied across a range of sectors:
• Predictive maintenance: Real-time detection of wear and anomalies based on performance profiles – without the need for specialised sensors
• Brownfield digitalisation: Retrofitting existing machines with a current sensor – without requiring PLC intervention
• AI agents and digital twins: Explainable metrics as structured input for autonomous agents and digital twins
• Energy management: Process-specific consumption tracking for ISO 50001 and CO₂ accounting per workpiece
Procedural flow (© TUB/S.Emec)
