Many modern industrial environments have automated monitoring systems using rule based logic. Within high-pressure fluid lines, stamping machines or high-speed sorting conveyor belts, legacy IoT solutions use static thresholds. These systems will alert you if the temperature or vibrations go above/below a certain value.
These are easy to set up however cause real problems.
If the temperature increases due to high load batch processing, then an alert will trigger in the same way as a critical failure would. This means no contextual awareness of what is happening and causes lots of alert fatigue as there will always be noise.
Instead industrial systems are using contextual intelligence through Physical AI.
The 4 Engine Layers
To change from using static alerts to deterministic contextual physical actions we need 4 engine layers to process multi-variable data.
Engine 1 – Accurate Location Identification
Before any contextual intelligence can happen we need to know where an issue has occurred. Our Identification Engine pulls information from active UWB, passive RFID tags, BLE beacons and optical sensors to create an accurate real-time map of where assets are.
Engine 2 – Multi-Variate Edge Sensing
Our Sensing Engine will pull high frequency data from the assets. This could be sounds, vibration, temperature, fluid movement, or current draw. We can normalise this data on the edge to remove local noise.
Engine 3 – Contextual Intelligence
Normalised data is fed into our edge based machine learning models. Instead of just seeing if variables are above/below a value, our AI Decision Engine looks at how multi-variate patterns change against the behaviour we have learnt. From this we calculate a risk score for the current operation. This tells us how likely our equipment is to fail, or cause a defect/issue based on the current situation.
Engine 4 – Physical AI Action
Finally we can use our intelligence to trigger physical actions. Our Physical AI Action Engine will take the high confidence decisions that our algorithms have made, and use them to trigger physical changes. This will use our pre-defined values to make changes to PLCs, create automated work orders in their asset management systems or change the duty cycle of the equipment.
Governance & Risk Envelopes
There are concerns around using autonomous physical actions. If we allow our software intelligence to make changes to heavy machinery or high voltage equipment without any oversight there is a risk. If our models at the edge encounter something they haven’t seen before.
To ensure complete stability, Physical AI solutions in enterprises will use risk envelops.
Low Risk (Fully Automated): Fully automated tasks that cause little impact. This could be creating logs, updating location tags or ordering new consumables.
Medium Risk (Partially Automated): Recommendations for changes to the operation. These will need to be confirmed by the user with one click. This could be small changes to conveyor belt speeds or preventative maintenance.
High Risk (Enveloped): Pre-defined safety measures. These are hard coded to ensure we are making safe quick decisions. In the case of a major anomaly our equipment will be turned off and engineers notified instantly.
By breaking down our intelligence into modular layers we can prevent alert fatigue. If you’re an engineering leader or technical co-founder wanting to see examples of modern Physical AI solutions and venture building models then look at the research from Aperture Venture Studio.
Top comments (0)