IoT (Internet of Things) systems are particularly good at collecting data.
Sensors can continuously report temperatures, vibrations, pressures, positions, and other physical world values. When the data starts to arrive, however, another question emerges:
How do we make sense of all these telemetry insights?
This is where Artificial Intelligence can expand the potential of IoT systems.
Combining the two is typically referred to as AIoT (Artificial Intelligence of Things), and the high-level architecture looks like this:
Physical Environment
↓
Sensors / Devices
↓
IoT Connectivity
↓
Data Processing
↓
AI / ML Models
↓
Insights / Decisions
↓
Alerts / Automation / Human Action
Interesting engineering problems arise between each layer. Let's look at a few examples.
IoT Gets Data, But What Do We Do With It?
Suppose you have an industrial machine producing thousands of sensor readings per hour. A typical implementation for an IoT application could be to store and visualize the readings in a dashboard.
Not a bad practice for monitoring, but a dashboard isn't likely to tell you why a value changed and if the change is worth investigating further. That's where an AI layer can help, though.
An AI layer can review historical and current telemetry for patterns, helping to identify if a sensor reading is abnormal.
One potential pipeline can look like this:
Normal operating pattern
↓
New sensor readings
↓
Feature extraction
↓
ML model
↓
Normal / anomalous behavior
The model can then take appropriate action, such as alerting or giving the operators more information to investigate the issue.
How it's implemented will largely depend on the equipment, data, model and operational requirements.
Predictive Maintenance
One obvious application for AIoT is predictive maintenance.
Rather than relying on a fixed schedule for maintaining the equipment, organizations can process the telemetry data of their equipment to find changes in its operating patterns.
For example, imagine vibration sensor data from a machine:
Timestamp | Vibration
10:00 | 2.1
10:01 | 2.2
10:02 | 2.1
10:03 | 2.8
A basic IoT implementation would likely be able to log and visualize the data, but not much else. An AI system, however, can use the telemetry in a larger context, such as looking at other similar patterns and determining if something should be flagged. The context can include additional telemetry, such as the equipment's temperature or other variables. It's unlikely that such a system would be able to predict exactly when the machine will fail (this is a common misconception about AI), but it can look for patterns that may be worth investigating further.
Anomaly Detection
Another common use case for AI + IoT systems is anomaly detection.
IoT solutions can generate an overwhelming amount of data for a human to investigate, so machine-learning techniques can help spot observations that are different in some way than others. For this specific application, techniques such as statistical detection, supervised, unsupervised or time-series learning, classification, neural networks and others may apply.
The important part is not to over-engineer your solution with complicated algorithms, but to pick the right model for your domain, data, latency and operational requirements.
Edge AI vs. Cloud AI
Where to run your inferences is another important architectural decision.
IoT data can be sent to the cloud for processing by AI models. Some of the advantages of doing this can be
centralized computing resources
easier model management
ability to process data at scale
convenient integration with cloud services
However, transmitting all that data may not always be feasible.
Another option is to use edge AI, where some processing can be done closer to the data source.
This is particularly useful when the application has strict requirements on latency, connectivity or data volume. A potential architecture could look something like this:
Sensor
↓
Edge Device
↓
Local AI Inference
↓
Immediate Response
+
Cloud
↓
Historical Analysis
↓
Model Training / Management
In practice, hybrid approaches are often viable.
The Data Engineering Problem
AIoT isn't only an AI problem, though. It's also a data engineering problem.
A model is only as good as the training and inference data that feeds it, and developers may need to deal with missing sensor readings, noisy data, varying sampling rates, calibration issues and more.
For example, if temperature is being sampled every second, but a pressure sensor is sampled every minute, how do you combine this data for your model? What about data versioning, network outages, feature engineering, and model drift? These are why, in practice, AIoT development involves collaboration between software developers, data engineers, ML engineers and domain specialists.
From AIoT to Physical AI
AIoT is also connected to the rising concept of Physical AI.
The idea is that AI can be integrated into systems where it interacts with the physical world rather than just relying on digital information.
A potential architecture can be thought of as:
Sense → Understand → Decide → Act
IoT devices can provide much of the "sense" and "connect" parts. AI can add "understand" and "decide". Robotics, machines, automation equipment and similar can provide "act". This creates an interesting engineering space between software, AI, IoT and physical systems.
Aperture Venture Studio is interested in this intersection through its work of exploring AIoT and Physical AI.
What Developers Should Think About
Creating an AI-enabled IoT system isn't a matter of connecting a machine-learning API to a sensor. A viable architecture must address considerations such as:
What data truly matters?
Where should data be processed?
How much latency can the application tolerate?
What happens when connectivity is lost?
How will sensor quality be monitored?
How will models be updated?
How will false positives be handled?
Where does human decision-making matter?
These questions can have much more impact on the final product than the choice of an AI model itself.
The Bigger Picture
AI adds value to IoT systems by providing an intelligence layer to the data coming from connected physical systems. IoT can answer "what is happening?", but AI can help investigate "is this happening often?" and a broader intelligent system can help answer "what should happen next?".
That progression from data collection → interpretation → decision-making support → action is why AIoT is interesting from an engineering point of view.
The goal isn't to use AI in every IoT application. The goal is to identify cases where intelligent analysis can make connected physical systems much more powerful.
Top comments (0)