IoT systems excell at collecting data
Sensors measure things like temperature, vibration, pressure, movement and operational state; RFID and BLE track and identify; GPS provides position data about vehicles and mobile assets.
But getting data is not the same as leveraging it.
Within industrial contexts, the problem is getting this information to support operational analytics or processes that can help people make decisions
And that's where AIoT comes in (the combination of Artificial Intelligence and the Internet of Things)
Getting Started: A Simple AIoT Architecture
If you begin to think of industrial AIoT applications, it's helpful to remember, at a high level, that you're trying to transform reality into information people can use.
Using a simple example, an AIoT system takes the form:
Physical Assets → Sensors & Identification Tech → Connectivity & Data Collection → AI/ML Layer → Insights → Operational Action
Where you choose to focus within this stack will determine your success.
1. Physical assets
You can think of this layer as the actual "environment" within your industrial IoT application.
Assets could be:
Manufacturing equipment
Vehicles
Tools
Inventory
Industrial infrastructure
Environmental systems
And so on; you'll need to know what you need to track in order to address an operational issue.
Physical assets are things you operate within a business process. If you're trying to find out why a machine tool fails unexpectedly (for example), you're seeking information about a physical condition, process or event.
2. Sensors and identification technologies
In order to evaluate physical conditions or events, you'll need to think about what technologies you need to apply.
RFID could help you identify relevant assets and maintain inventory information
BLE may give you positioning data indoors
GPS or GNSS are useful for things that move around outside
Or other IoT sensors could read temperature, humidity, vibration, pressure, or machine status.
Pick the right tool for your environment: indoor positioning is dramatically different than tracking a vehicle across great distances, for instance.
3. Connectivity and data collection
Unless you're collecting and organizing information generated by sensors, you can't hope to get value from it.
Gateways, wireless networks, APIs, and other tools can help you collect, organize and distribute information.
At this stage in your project, you'll need to think about the importance of your data format, timing, identifiers for devices, handling connectivity outages, and how you'll deal with interoperability.
As you may guess, the more messy or poorly handled your data is at this stage, the harder it will be throughout the rest of your project.
But Where Does AI Come In?
Your enterprise or operational analytics (using AI or machine learning algorithms) can help you understand patterns in the data you've collected.
Let's say you have a machine that generates temperature and vibration data constantly.
As a first pass, you might use an operational data visualization to see when these measurements are above a certain threshold.
More analytically sophisticated systems, however, will try to model the change in conditions over time.
Such analytics may be used for use cases around anomaly detection or predictive maintenance, equipment and asset monitoring, analyzing utilization, or looking at production analytics or forecasts; these represent ways of using machine learning and data science to model equipment, process, or operational conditions and detect patterns or irregularities.
The key point is that these kinds of analytics are not substituting for IoT; they're building on top of it.
From Prediction to Real-World Action
Getting a prediction from an AI model is one thing; determining what to do about it is another.
Using the previous example, we might have an anomaly detection model that spots unusual activity from a machine.
We can use predictive maintenances tools such as time-series databases to spot patterns and forecast likely failure points, but someone has to take action to prevent that equipment from failing.
A more complicated example might see sensor data going through a pipe → data collection and preprocessing → AI/ML model picks up an insight → alert to humans → they examine it and decide a specific maintenance action should be taken.
For less important use cases, you might instead have your system automatically trigger software tools; for physical work, it may make more sense to require a person to explicitly trigger an action before it occurs.
This is why your workflow is so important in AIoT: your ability to get your people, analytics, and processes working together effectively.
Is Good Enough Data Better Than Complicated ML?
While it can be seductive to start an AIoT project by picking a machine learning model to solve your problem, the truth is that more interesting projects often benefit massively from investing in the underlying data quality and structure.
Pragmatic questions you should think through when starting such a project include:
Are your sensor readings correct or noisy?
Are your timestamps consistent?
Can you uniquely identify every asset and record?
Will you have reasonable coverage for your use case?
Does your system properly communicate with yours?
Do you have enough information to make the model you want?
How will you evaluate your model on actual data?
Complicated ML models built atop inconsistent data may yield interesting results, but it's probably not a great use of everyone's time.
Getting Started with an AIoT Project: Going Small
You do not need to look at every asset in your enterprise to generate value with AIoT.
Instead, focus on a specific problem that you'd like to solve in the near future--say, a single class of equipment--and work on building a data foundation and analytical model around that.
You should define metrics that estimate the value such insights could deliver: how much downtime could you avoid, how much faster can you perform a maintenance check, how much equipment utilization information could you gain, or how often false-positive alerts would occur?
This lets you establish a feedback loop between your physical assets and data collection process, analytics and modeling, and actual people in an operational setting.
This will give you valuable insights for your next steps.
AIoT: The Big Picture
AIoT is about connecting from the physical world, up through data management and analytics, to support real-world decision-making.
Connect → Collect → Analyze → Understand → Act
This kind of architecture could be applied across industries and in many different settings: manufacturing, logistics, transport, construction, power utilities, and so on.
To learn more about AI and IoT for industrial and connected systems, I'd encourage you to look at Aperture's website to begin learning more about connected technologies and applications.
The most successful AIoT projects begin with a focused operational problem, a solid data foundation, validation of their analytical value, and connecting those insights and value to appropriate, impactful actions within an enterprise workflow.
Technology matters, but connecting people and data is key.
Top comments (0)