DEV Community

Eman Tanveer
Eman Tanveer

Posted on

Building AIoT Systems: Connecting AI Models to the Physical World

AI applications are typically discussed in terms of models, API, data pipelines, or user interfaces.

When is an AI system required to understand something that is happening in the physical world?

A machine in a factory changes it's vibration pattern, an asset moves to an unexpected location, a camera sees an anomalous occurance, or a warehouse's flow of inventory changes drastically.

In these examples, the AI model is only one component to the system.

The true engineering challenge involves tying sensors, edge hardware, data infrastructure, AI model, and physical processes together into a reliable production system.

This is the basis for AIoT.

What Is AIoT?

AIoT combines AI with The Internet of Things.

A (very) simplistic architecture for AIoT would look like:


Physical Environment

↓

Sensors / Cameras / RFID / Devices

↓

Edge Processing & Connectivity

↓

Data Ingestion

↓

Storage / Stream Processing

↓

AI / ML Models

↓

Decision & Automation Layer

↓

Human or Machine Action

Enter fullscreen mode Exit fullscreen mode

Each layer introduces different engineering requirements.

The interesting aspect of this stack is that the AI Model itself is not always the most challenging component.

1. Data Starts at the Edge

Industrial environments can often be large producers of data.

Depending on the domain, this could take the form of:

Temperature

Vibration

Location

Motion

Video

Machine telemetry

Equipment status

Environmental conditions

And other forms.

The act of sending every raw signal to a cloud data warehouse may not always be optimal depending on the application.

Using edge processing lowers bandwidth requirements, reduces latency, and can allow for localized decision making.

For developers, this adds another layer to the system design challenge - deciding what computation should be done on the device, the edge gateway, or somewhere else.

2. Connecting Heterogeneous Systems

The reality of most industrial environments is that they are not designed around a unified set of technologies.

A production environment may involve legacy equipment, IoT hardware, proprietary systems, databases, industrial protocols, and many other technologies.

This presents an integration challenge that forms a large part of AIoT engineering.

It's not enough to simply gather data - data from different systems must be made:

Consistent

Time aware

Identifiable

Contextual

Accessible

Without that foundation, an AI model can offer limited value.

3. AI Needs Context

Let's say a sensor reports an abnormal temperature value.

What does this value mean?

Where was this value measured?

What was the machine doing?

What is a "normal" reading for this machine?

Has this value been seen before?

Are other sensors reporting similar values?

Has maintenance been performed?

This is why AIoT systems need more than typical ML inference - they need contextual data models.

These kinds of information can help an AI system make decisions about physical entities.

4. Real-Time Data Changes the Architecture

Most traditional AI applications are concerned with historical data.

AIoT systems, on the other hand, often need to reason about a continuous stream of data.

This introduces additional considerations such as:

Latency: How quickly does the system need to react?

Reliability: What happens if connection is lost?

Ordering: Are the events arriving in order?

Storage: What information should be stored?

Scalability: How many events will need to be processed?

A predictive-maintenance system and an analytics dashboard both might use similar data, but can have very different architectural requirements.

5. From Prediction to Action

There's an important difference between being able to predict something, and taking an action to change the outcome.

An AI model may be able to predict that a machine has higher chances of failure.

To actually make use of this prediction, another system layer is needed to:

  1. Report the alert

  2. Create a recommended maintenance plan

  3. Send a message to an operator

  4. Schedule a work order

  5. Take other actions

Depending on the application, there may be an appropriate degree of automation.

Where safety is a concern or high value is placed on manual decisions, humans can be an important part of an AIoT system.

6. Physical AI Introduces New Testing Problems

Testing applications that operate in the physical world are significantly more complex than testing a generic API.

Developers must take into account scenarios such as:

Sensor failure

Missing data

Incorrect data

Network interruption

Physical conditions

Model uncertainty

Device failure

Conflicting sensors

A system that works properly in development can fail dramatically in production - especially in the face of unaccounted edge cases.

This is why simulation and system observability are so important.

7. Digital Twins Can Provide Another Layer

Digital twins can give additional value to AIoT systems.

By providing a virtual representation of a physical entity, twins can provide a layer of abstraction.

Twins can be used in a variety of ways such as:

Connecting the current state of a machine to a software representation

Using sensor data to update the digital representation of the machine

Applying predicted states and behaviors to a machine model

Maintaining historical information about a machine

AI can then operate not on raw sensor data, but instead a richer representation of the physical system.

Where AIoT Is Being Applied

The architecture outlined above can be applied to a variety of applications:

Industrial asset tracking

Predictive maintenance

Inventory optimization

Workforce safety

Computer vision

Robotics

Industrial automation

Environmental monitoring

Operational intelligence

Companies building businesses around these areas include Aperture Venture Studio, who are building applications for physical-world and industrial AIoT use cases.

A Practical Development Approach

When building an AIoT application, a good sequence looks like:

  1. Define the physical problem

Do not start with the AI Model. First, determine the physical problem to be solved.

  1. Identify the required signals

Determine what information is needed to solve the problem.

  1. Design the data pipeline

Decide how devices, edge systems, databases, and AI services will communicate.

  1. Establish a baseline

Measure the performance before implementing automated systems.

  1. Introduce AI

Use AI to improve upon the current methods.

  1. Add automation carefully

Move from insights to recommendations and, where applicable, actions.

  1. Monitor the complete system

Monitor not just model accuracy, but also device health, data quality, latency, system reliability, and results.

The Bigger Engineering Challenge

AIoT is fundamentally a systems-engineering challenge.

While the model is important, the sensors feeding it, the network carrying it, the infrastructure processing it, the software providing context, and the workflow taking actions are all of critical importance.

As AI makes its way into factories, warehouses, logistics systems, robotics, and other physical environments, developers will need to think beyond the model.

The question developers must ask is:

How can we create AI systems that can operate and understand the physical world?

This question requires AI, IoT, distributed systems, edge computing, data engineering, automation, and domain expertise to work together.

Top comments (0)