While AI and IoT are often treated as separate fields, there are real world applications where both areas are required to build a working solution.
IoT provides the connections and sensors to the physical world, while AI analyzes the data and builds models that can recognize patterns, detect anomalies or make decisions.
Both combined are known as AIoT (Artificial Intelligence of Things).
But building an actual system involves much more than attaching a sensor to an AI model.
Define a Problem Before Thinking About Technology Stack
A frequent mistake is to design the technology stack before identifying what problem the AIoT system is supposed to solve.
Ask yourself the following questions (among others):
- What physical process can benefit from better information?
- What information is currently missing?
- How often should measurements be taken?
- What decision does the information help to automate or improve?
- Does the application require real-time processing?
- What happens if the connection is interrupted?
For example, an industrial environment may benefit from more information about the state of equipment, inventory, the environment or the movement of assets and people.
Based on this, the rest of the system can be designed.
A Simple Architecture of an AIoT System
A simple view of an AIoT system can be represented by the following elements:
Physical Environment → Sensors → Connectivity → Data Platform → AI/Analytics → Decision or Action
Each layer plays a specific role.
Sensors are used to capture physical values such as temperature, location, movement, pressure or others.
Connectivity transports this information using existing connectivity solutions (WiFi, cellular, Bluetooth, Low Power Wide Area Network, etc).
The data platform is responsible for storing this information so it can be queried, processed or analyzed.
Finally, the AI part makes it possible to recognize patterns, detect anomalies, classify events and support a decision or take an action.
Data Quality Comes First
An AI model can only be as good as the data that is used to train and validate it.
In practice, this means that an AI model cannot fix fundamentally flawed data.
In the real world, data quality issues can come from many sources:
- Missing measurements
- Duplicated entries
- Wrong timestamps
- Sensor malfunctions
- Connectivity issues
- Outliers
- Device tampering
- Environmental factors
- A single telemetry entry can look like this:
- timestamp
- device_id
- location
- parameter
- value
- unit
- quality_status
This is only an example in practice, any number of additional metadata fields can be present.
It becomes even more important to capture this information when dealing with tens of thousands of data points across multiple devices and parameters.
Before training an AI model, it can be helpful to define validation rules for impossible values, missing fields, timestamp freshness and other issues.
Processing at the Edge or in the Cloud?
Depending on the use-case, it may make more sense to process information at the edge (closer to the source) or in the cloud.
There are valid reasons for implementing gateways that process information at the Edge before sending updated statistics or alerts further.
Such a configuration can reduce bandwidth usage, address latency issues, improve privacy by keeping raw data on the device or handle large volumes of information better.
Meanwhile, the cloud offers processing power, storage, additional dashboards, better model training possibilities or management of multiple devices.
In practice, many AIoT systems use a mix of both approaches:
Device → Edge/Gateway → Cloud → Analytics
The choice depends on the specific requirements of the project.
Where Can AI Add Value?
AI is useful when there are enough data points and an actionable insight that can improve operations.
For example, an AI model can:
Recognize patterns of abnormal equipment behavior
Help to predict maintenance events
Optimize inventory flows
Detect environmental changes
Improve asset tracking
Help to ensure worker safety
When designing an AIoT system, it is important to understand the purpose of the model.
This helps to avoid the temptation to use an AI model just because it is available.
A good question to ask is: What decision will be improved by recognizing patterns in this data?
Prototyping an AIoT System Can Begin Small
An AIoT proof of concept does not have to use a large set of devices or a complex model.
A prototype can use:
- 1 sensor or data source
- a microcontroller or gateway
- connectivity
- a small database
- basic data validation
- some analytics
- a dashboard or visualization
Once a working prototype has been built, the system can be expanded by adding more devices, using a more realistic model or connecting additional data sources.
At the same time, such a prototype can help evaluate if the system is actually useful.
Other Considerations: Reliability and Security
When building industrial AIoT systems, it is important to remember that the tools exist in the real world.
This means that such systems must be able to handle device failures, sensor issues or network outages.
At the same time, such systems should be designed with security in mind.
Some important considerations include:
- Device authentication
- Secure connectivity
- Authorization and access control
- Updates and patch management
- Data integrity
- Logging and monitoring
- Failover strategies
- Connectivity backup
- Security should be designed into a system from the ground up.
The Challenge of AIoT Engineering: Bringing Everything Together
AIoT systems bring together multiple areas of engineering:
Hardware, connectivity, software development, data engineering, AI, security and operations.
The complexity of such systems is what makes them interesting, challenging and rewarding to build.
Good practice shows that the best AIoT systems address a specific need.
They start with a well-defined problem, gather the required information and apply the right amount of intelligence to get tangible results.
AIoT is not about putting AI on top of just any IoT system. It is about creating a reliable link between the physical and digital worlds.
What engineering challenge do you think is the hardest in the context of AIoT: data quality, connectivity, model accuracy or scale?
Top comments (0)