DEV Community

Nashtarin Nur
Nashtarin Nur

Posted on

The Edge-Native Control Loop: Building a 4-Engine Architecture for High-Throughput Operation

Modern industrial sites are generating unprecedented volumes of time-series data. With large-scale machinery, fast moving conveyors and autonomous materials handling systems, we are generating gigabytes of data every second. However, this data is often sent back to a central cloud system.

This leads to a problem known as decision latency.

If a high-speed packing line has issues or a multi-axis robot is overheating, the speed at which the data is sent to the cloud means you cannot react to the situation quickly enough. The analytics would not pick up on the issue until after the damage or issues have been caused.

Technical leaders are looking to shift intelligence from a faraway cloud into edge-Native Physical AI.

Four Engine Layers to Edge Native Systems

To get intelligence to work in real-time, there needs to be a continuous low-latency engine. This consists of four engine layers.

Engine Layer 1: Identification

To execute actions in the physical world, you need to know exactly where something is. In order for an operational metric to do something, the system must know what physical entity it belongs to and where it is. The Identification Engine takes data from active Ultra-Wideband networks, passive RFID, Bluetooth Low Energy and visual sensors to identify and locate everything in real-time. This will give a map of all the coordinates in real-time.

Engine Layer 2: Sensing

With identification and location sorted, the Sensing Engine takes multi-sensor data and processes it. This data includes sound, amplitude of vibrations, thermal and power usage. It will normalise this data. By doing this on the local hardware means the data does not suffer from cloud latency and will remove any background noise.

Engine Layer 3: Decision Making

The normalised data is sent into low-powered edge AI systems. This means the AI Decision Engine can take complex pattern changes and compare them to past data of the asset. It will calculate an instant risk score and show how likely something is to fail, affect product quality or cause issues.

Engine Layer 4: Physical Action

The final engine will take the decision and apply it to the physical world. The Physical AI Action Engine will interact with PLCs, asset management systems and materials handling systems. It will act within a set range and apply instant changes to the system. This can include changing the frequency of drive motors, changing the speed of automated guided vehicles or shutting the system down automatically to prevent catastrophic failure.

Safe Ranges for Autonomous Systems

When using autonomous systems on large scale physical assets there are risks involved. If the AI has full control of the asset and there is an issue with the models at the edge, the AI could cause issues or safety risks.

To counter this, Physical AI within an enterprise will have a set range of operations.

Low Risk Autonomous: Fully automated systems to complete low risk admin and operational tasks. For example, logging conditions, updating asset tags and queuing new parts.

Medium Risk Semi-Automated: Recommendations given from the algorithms. These would need to be confirmed by clicking once. Examples include changing conveyor belt speeds, changing furnace temperatures and scheduling maintenance.

High Risk Bounded: Hard coded safe systems are put in place. This will instantly shut the system down to keep it safe and will instantly notify the engineers in the facility.

By breaking down the intelligence into different engines, the enterprise can ensure they are not suffering from cloud latency. If you're a technical leader or co-founder wanting to look at edge AI and venture creation models, research can be found here at Aperture Venture Studio.

Top comments (0)