AI Technologies
Sensor Fusion & Physical AI
Combining information from different physical measurement sources – camera, radar, LiDAR, and other sensors – as the basis for systems that act in the real world.
AI Technologies
Hybrid AI & Decision Intelligence
Machine learning combined with mathematical modelling, rules, and technical system models – for statements that show how certain they are.
Methods and techniques
- Physical models
- State estimation from incomplete, noisy measurements
- Statistical evaluation and uncertainty modelling
- Tracking multiple hypotheses in parallel until the data supports a decision
- Plausibility checking
- Explainability and traceability
Application examples
- Common operational picture of a supply network: detecting a leak early, localising it, and distinguishing it from a sensor fault
- Forecasts with an uncertainty band and a rationale in natural language
What matters
The method is not limited to one field of application: the mathematics stays the same, and the model of the field of application is created together with the user. It is strategically relevant above all for safety-relevant, industrial, and complex technical tasks.
The mathematics behind it is the same we have been using to fuse sensor data since 2005. What is new is the AI layer, which proposes hypotheses and explains the situation in words.
Example: a water supply network
A supply network consists of nodes and pipes. Measurements are taken at only a few points. When a leak occurs, the pressure drops first at one node, then at its neighbours. The task: detect the leak early, localise it, and distinguish it from a mere sensor fault.
- Computation core. A physical model of the network; different network shapes are created through configuration.
- State estimation. The complete network state is estimated from incomplete, noisy measurements, with an uncertainty band.
- Monitoring of the measurement channels. Faults such as offset, drift, outliers, or a frozen value are detected continuously.
- Parallel hypotheses. In case of ambiguity, multiple interpretations are tracked in parallel and rated by probability.
- AI layer. Language models propose explanations and translate operator knowledge, for example from maintenance reports, into prior knowledge.
- Situation report. The situation is summarised in natural language, including the limits of its informative value.
Why multiple hypotheses are necessary
If a measurement is assigned to the wrong node, a classic filter fails: the estimate jumps back and forth, and no amount of retuning helps, because the assumption about the assignment is wrong. If the system tracks the possible assignments in parallel, it finds the right one. The same applies to the question of whether an anomaly is an event in the network or a fault of the sensor; both produce similar measurement patterns.
Further topics
Industries
Public Sector and Common Operational Pictures
Combine heterogeneous data and spatial information into common operational pictures and support decisions in a traceable way.
Industries
Manufacturing and Mechanical Engineering
Preserve experience-based knowledge, manage variants, and use image and sensor data to assess condition and quality.
Would you like to make AI usable in your company?
Whether a first AI adoption or a complex development task – we discuss the requirements and suitable next steps with you.