DEV Community

Nashtarin Nur
Nashtarin Nur

Posted on

Dynamic Workload Routing at the Edge: Building a Lean 4-Engine Pipeline for Industrial IoT

As AI becomes a core part of enterprise expenditure, it’s clear that inference in the cloud is not scalable. This means that sending millions of live data streams to huge foundation models in the cloud will cause massive compute costs.

In high speed environments, paying per minute in API costs in the cloud to monitor a process that costs a fraction of a cent per cycle is not sustainable. In order to create sustainable automation, you need to deploy highly optimised models on the edge.

The 4-Engine Architecture

In order to create predictable unit economics in heavy industries, intelligence needs to be separated to ensure only meaningful changes are consuming any compute. A mature Physical AI architecture has 4 engines:

  1. Identification Engine

Sending inference on all sensors in your facility will cause wasted compute. The Identification Engine uses low power signals from UWB arrays, passive RFID gates, BLE beacons and optical tracking to create a sub-meter map of all assets in play. This ensures only real equipment will trigger any analytical compute.

  1. Sensing Engine

Sending raw high frequency sensor data will cause unnecessary costs. This engine runs on your edge microcontrollers and industrial gateways. It takes in multiple data types including 3-axis vibration, sound, temperature and current. This engine filters out background noise and creates a highly compressed feature vector.

  1. AI Decision Engine

Not all events require heavy neural network inference. This engine makes use of lightweight models to analyse feature vectors. These run on your edge hardware and compare the feature vector to physical metrics. Standard data will be scored in microseconds on site. Complex anomalies will be sent to more powerful edge servers.

  1. Action Engine

In order to get real value from your automation, you need to ensure the analytics are stopping downtime or waste. This engine connects to PLC’s, control systems and maintenance. It will use your risk score within defined confidence intervals to take action. This could include changing motor drive frequencies, adjusting pressure on pneumatic lines or creating prioritised tickets for technicians.

Defined Operating Limits

You should never sacrifice safety in your facility to save compute costs. Giving small algorithms the ability to change heavy machinery or high voltage systems without deterministic limits will cause issues if your sensors fail or the equipment is in an unknown state.

Enterprise Physical AI architectures will always have defined operating limits. These are split into 3 categories:

Autonomous - Low Risk

This will run automatically to complete simple tasks that will save you time. This includes recording asset usage history, updating inventory maps and creating standard consumables to replace.

Semi-Autonomous - Medium Risk

This will make changes to your operations based on context. They will stop and wait for an engineer to approve with one click. This could include changing conveyor speeds, adjusting curing temperatures or changing batch order.

Defined - High Risk

This will have hard coded safety features. If an extreme anomaly is detected it will override any algorithm and shut down the equipment safely. It will also alert the control engineers immediately.

Creating sustainable automation in your industry requires moving away from expensive prototype solutions and creating thin edge control loops. If you’re an engineering leader or technical founder looking to design your architecture for predictable automation and deep tech ventures then you can find out more about our research here at Aperture Venture Studio.

Top comments (0)