When it comes to models for the physical world, an AI is not the whole architecture
To create useful AIoT systems, there has to be a process which moves information through physical devices and sensors, through connectivity, data pipelines, processing, AI, and applications, ending in decisions and action.
A good way to imagine the pipeline is
Devices → Connectivity → Data → Processing → AI → Decision → Action
The fun engineering part is not worrying about what kind of AI model to use, but designing the whole path around it.
1. Devices:Information Begins
Every AIoT system begins with information from the physical world.
Depending on what we're trying to accomplish, different devices and sensors might provide,
- temperature readings
- equipment telemetry
- location
- motion
- environmental readings
- operate events
It's easy to assume that more data is better, but when it comes to AIoT systems, noisy or useless data simply creates more work that doesn't improve the output.
Instead of worrying about how many devices we're going to use to gather information, a far more interesting question is what information do we need to acquire to understand a situation.
2. Connectivity: Moving Information
After devices have acquired information, it has to be delivered somewhere for processing.
This layer encompasses networks, gateways, wireless tech, and other means of delivery. We design this layer with things like latency, network availability, data size, device limitations, reliability, and the physical environment in mind.
For example, there's a big difference between a system which requires immediate action based on the information and something that passively stores it for later analysis.
There is no such thing as a universal architecture for this layer the needs of the application should define it.
3. Data Processing: The Input
Sensor data rarely goes directly from devices to an AI model.
It almost always has to be processed, normalized, filtered, validated, or aggregated before it's ready to use.
This is also the point where the debate over edge vs centralized processing comes into play.
Processing closer to the source can be beneficial for latency, bandwidth, or reliability, while some applications might make better use of central processing.
The needs of the system should decide this, not what's trendy.
4. AI and Analytics: Finding Patterns
After the data has been processed, we can begin using AI and analytics to find patterns, identify anomalies, discover relationships, and recognize changes.
Let's say we've created a system which monitors different kinds of industrial equipment.
An AI model analyzing this data might use historical and live telemetry to find equipment which acts abnormally compared to the majority.
However, the value of this information is severely limited because we can't put it in context.
A change in sensor readings can be the result of abnormal conditions or a regular occurrence depending on the situation.
Which is why we can't separate AI from the rest of the system.
*5. Applications: Where Information Goes
*
We need somewhere for the information to go after it's been processed by AI, and depending on the system, that might take many different forms.
It could be something as simple as a dashboard or an alert, or it might be an operational application, a workflow, or another software system. A deep learning model which makes great predictions is only as valuable as the people or systems which consume its output.
When it comes to architecture, we should always think about the complete path of information from model to consumer.
6. Decision and Action
The last two steps in the pipeline deal with making decisions and taking actions.
As an example, a typical workflow could look like:
detect → notify → investigate → decide → act
In some cases, we might want to automate this to detect → analyze → automate
However, that should always be a decision based on the requirements of the application.
There are times when a human has to review information before any form of automated action is possible.
Why This All Matters
It's easy to get caught up thinking about the AI layer because it can be the most interesting part of an AIoT system.
However, we have to remember that the output is only going to be as good as all the previous steps leading up to it.
Poorly chosen or implemented sensors will affect the outcome just as much as a problematic connectivity layer, data processing architecture, or AI model. Even the best model isn't valuable if it's disconnected from the systems and people which need to use its output.
This leads us to a slightly different way of looking at the architecture
Sense → Connect → Process → Analyze → Decide → Act
Models are an important part of this system, but they're not the whole picture.
Designing Around the Problem
When we begin working on a practical AIoT application, it's crucial to design around the problem.
Before choosing particular sensors, infrastructure, or AI models, we should ask ourselves,
What do we need to detect or understand?
What information will we need?
How reliable does that information have to be?
How much latency can we tolerate?
Should we process data at the edge or in the cloud?
Where can we use AI to improve the situation?
Who or what will consume the information?
What should we do after we've detected or analyzed something?
These questions will help us ensure that the architecture supports what we're trying to accomplish.
A Simple Mental Model
One way of looking at the pipeline is assigning different questions to different layers.
Devices: What is happening?
Connectivity: How does the information move?
Processing: How do we make the information usable?
AI: What patterns or changes can we identify?
Application: How should the result be presented?
Decision: What should happen next?
Action: What changes in the physical or operational environment?
Conclusion
AIoT is far more than simply adding an AI layer to top of IoT systems.
It's really an end to end architecture where physical signals become data, the data becomes information, AI discovers patterns, and those patterns can lead to decisions and actions.
The challenge for developers and engineers is far greater than a model for an AIoT system, it's an architecture which puts intelligence in the middle of the pipeline.
Devices → Connectivity → Data → Processing → AI → Decision → Action
The value of AIoT systems depends on how well these layers connect to support intelligent and informed operations and decisions.
Top comments (0)