DEV Community

Nashtarin Nur
Nashtarin Nur

Posted on

The Post-Hype Edge Stack: Architecting a 4-Engine Pipeline for Deterministic Industrial Control

Engineering the Slope of Enlightenment: Deterministic Physical AI for Industrial Applications

As the first iteration of Generative AI experimentation begins to settle, industrial tech leaders are discovering that probabilistic software cannot control physical assets. Over the last 2 years, enterprises have invested heavily in building large cloud-based models and general chatbots which ultimately are not scalable due to high inference costs, lack of predictability and network lag.

To truly get to the post hype stage of industrial automation, we need to move away from large parameter models hosted in the cloud and instead ensure deterministic verification, efficiency at the edge and state management. To achieve this, we need to ensure AI is used as a tool in the background of operational technologies.

Four Engine Architecture for Post Hype Industrial Applications

To ensure predictable industrial automation we have to split our intelligence into different layers and ensure they can be executed at the edge. Our mature Physical AI systems will have 4 engines:

  1. Identification Engine

A common issue with early iterations of AI in industry was using ungrounded sensor data. Without knowing what asset you are looking at and its current location, you cannot determine if something is out of the ordinary or about to fail. Our Identification Engine takes data from Ultra Wideband (UWB) arrays, Passive RFID Gates, Bluetooth Low Energy (BLE) Beacons and Optical Tracking Nodes to give a constant state map of sub meter coordinates. This means every data packet can be mapped to a grounded asset.

  1. Sensing Engine

Sending raw sensor data to a centralised cloud is unsustainable and too slow. Our Sensing Engine runs on edge devices and takes multimodal data including Acoustic, Triaxial Vibration, Thermal and Current. This filters out local noise and normalises the data to ensure clean data is passed to the reasoning models.

  1. AI Decision Engine

In high consequence environments, you cannot have guess work. Our AI Decision Engine takes normalised data and compares it to behaviour baselines. This gives a risk score for the operation with exact probabilities for mechanical failure, bottlenecks and quality issues. This also shows clear audit trails.

  1. Action Engine

Without taking action, you will still have friction in your operations. Our Action Engine interacts with PLCs, supervisory systems and maintenance systems to take action. It only takes action within its confidence intervals and will take action based on the risk score. This includes Variable Frequency Drives, Autonomous Mobile Robots and Technician Orders.

Bounded Operation

A key point of our mature AI systems is that they are not fully autonomous. If you give full control over heavy machinery or high voltage systems to learning AI there is a high level of risk.

To remove any guess work our Physical AI systems in enterprises will have strict boundaries. This will ensure the automation is split into 3 categories:

Autonomous (Low Risk)

Fully automated tasks such as logging asset health, updating location inventory and adding standard consumables.

Semi-Autonomous (Medium Risk)

Tasks that will be done based on context but will require human verification. This will include changing line speed, thermal curing and changing batch order. This will have a 1 click verification from engineers.

Bounded (High Risk)

Hard coded safety systems that will not run any complex algorithms. This will shut down equipment safely if a serious anomaly is detected. This will alert control engineers on site.

To get to sustainable automation in industry you need to move past the wrapping up of software. If you are an engineering leader or founder looking into architecture designs for deterministic automation and deep tech ventures, you can see our ongoing research at Aperture Venture Studio.

Top comments (0)