Mature Industrial Tech should be invisible. When an electrical grid, water treatment plant or a high speed assembly line is running at peak health, engineers will not interact with any new software; instead relying on deterministic infrastructure.
Most early stage AI implementations are the opposite. Using stateless cloud based models and brute force Retrieval Augmented Generation (RAG), they treat every query as one-off. As there is no state management of physical assets, environment changes and previous maintenance actions, stateless models will hallucinate, increase compute requirements and fail when deployed within existing OT systems.
Industrial automation must be moved away from stateless cloud models to modular state management and edge native Physical AI.
Four Engines for Stateful Edge Native Execution
Physical AI must be invisible by having an accurate real-time view of the physical world. To ensure this, advanced systems will use 4 engines to provide modular state management.
- Identification Engine
Stateless models will fail within a factory. If you know there was a spike in vibration but do not know which machine it happened on, where it is in relation to other assets and what batch it is running, it is of no use. The Identification Engine uses data from Ultra Wideband (UWB) arrays, passive RFID gates, Bluetooth Low Energy (BLE) beacons and optical trackers to provide a persistent state map of your physical world. Every telemetry packet will be linked to a confirmed physical asset with sub-meter accuracy.
- Sensing Engine
Sending high frequency raw telemetry data to the cloud will cause high bandwidth usage. The Sensing Engine processes data at the edge using local edge nodes. It receives multi-sensor data including tri-axis vibration, acoustic emissions, thermal changes and electrical draw. It filters out noise and provides clean state vectors.
- AI Decision Engine
Instead of analysing large amounts of unstructured text, the AI Decision Engine will compare the structured state vectors from the edge with physical baselines and degradation curves. This will provide a quantified risk score using maths to provide exact probabilities of wear, bottlenecks and quality issues, alongside providing auditable logic paths for plant engineers.
- Action Engine
This provides intelligence within your existing control loops. The Action Engine will interface directly with PLCs, control systems and maintenance systems. It will only act within your defined confidence intervals. It will make changes to the physical world including changing the frequency of drive motors, changing the path of autonomous mobile robots.
Defined Operating Parameters
To ensure your AI infrastructure is invisible, you must define operating parameters. Leaving your AI to make changes to high voltage equipment and high speed assembly lines without deterministic parameters will create unacceptable risks.
To ensure determinism, your Physical AI within your enterprise will have defined operating parameters. These will be split into 3 levels.
Autonomous Low Risk
This will complete tasks automatically. Examples include updating your spatial inventory, providing asset histories and ordering consumables.
Semi Autonomous Medium Risk
This will change processes based on context. It will stop and await approval from an engineer to make changes. Examples include changing conveyor belt speeds, thermal curing and resequencing batches.
Defined High Risk
This will have hard coded safety parameters. If an anomaly is detected, this will disregard the AI and shut down equipment to a safe state. This will alert the control engineers instantly.
To get your industrial AI into production, you must move away from fragile stateless cloud models and use modular edge native control loops. If you are an engineering leader or founder looking to build deterministic automation and deep tech ventures, you can see our current research at Aperture Venture Studio.
Top comments (0)