IoT systems enable the collection and analysis of signals originating from machines, vehicles, sensors, equipment, and other physical assets
The initial engineering challenge focuses on collecting that data and getting it into a usable form. But from here, the real engineering work begins, namely applying intelligence and context to generate decisions, recommendations, or other actionable outcomes.
The area of AIoT, or Artificial Intelligence + Internet of Things, applies AI and machine learning to generate insights and decisions from signals originating in the physical world. This combination covers a range of capabilities from simple identification to complex inference and decision-support systems.
The AIoT Stack
A simplified architectural view of AIoT might look like this:
Physical Assets
↓
Sensors & Identification
↓
Connectivity
↓
Data Pipeline
↓
AI / ML
↓
Insights
↓
Decision
↓
Action
Every layer contributes to the ultimate outcome by generating, refining, or applying intelligence. Physical assets expose their values through sensing or identification mechanisms, which may rely on different technologies as appropriate. Connectivity enables information exchange, which can take many forms depending on requirements and capabilities. Pipelines allow the data to be organized for analysis. AI methods analyze the information, and operational applications expose actionable outcomes.
Note that the value of the system depends on every layer. Not on the sophistication of the methods at any one layer.
Context Matters
Let us suppose we have a machine that generates vibrations as part of its operation. The vibration signal, by itself, might not be very informative.
A model analyzing these measurements might need to combine it with additional data, such as the identity of the machine, its operating state, production schedule, maintenance history, temperature, load, or other measurements, to generate useful insights.
Or take a similar example of location data. A vehicle’s position on its own is often not as valuable as combining it with other information, such as a route, delivery schedule, or the vehicle’s state.
Another way to think about it is that data collection is only valuable when it is contextualized, i.e., when it is related to the things that matter for the specific application. Context is one of the keys to building an AIoT system because it enables models to generate useful outcomes.
Identification and Sensing Are Different
In practice, AIoT systems combine multiple technologies at different layers of the stack. For example, RFID tags can be used to identify objects, GPS/GNSS to determine their position, and BLE/UWB for proximity, and IoT sensors to measure their state or the state of the environment.
Each type of information provides different context and can be combined in different ways. An RFID tag might expose the identity of an object, while a sensor might expose its state or the state of the environment. A positioning system might expose the coordinates, BLE might expose proximity, and a video camera can expose visual content. All of that information can be combined in novel ways to generate insights.
AIoT systems therefore often feature hybrid architectures utilizing multiple modalities. These might include computer vision, natural language processing, and conventional machine learning models analyzing telemetry data.
Where Does AI Come Into Play?
The role of AI in AIoT is multifaceted.
Depending on the application, different models and techniques can be used. For example, anomaly detection models can be utilized to expose irregularities in the sensed data, or forecasting models can predict future values. Classification models can recognize objects or patterns, regression models can estimate numeric values, or neural networks can analyze images, audio, or other complex signals. These techniques can be combined, used at different stages of the pipeline, or augmented with operational data to generate decisions.
For example, an AI model analyzing sensor data might be used to expose anomalies, which would then be reviewed by an operator
Sensor Signal
↓
Data Validation
↓
AI Analysis
↓
Anomaly / Prediction
↓
Human or System Review
↓
Operational Response
Again, the decision-making or recommendation is not made by the AI model itself, but it informs the operators or supports other systems in making those decisions.
Note that there are additional opportunities for the application of AI, namely when information from multiple sources needs to be combined. This can be done at any point in the information pipeline, depending on the application.
The Engineering Challenges
Designing AIoT applications introduces a number of challenges that are not always present in conventional software systems:
Data quality
Modern machine learning methods often require high-quality data, which can be available only after extensive preparation.
Interoperability
Industrial systems are often heterogeneous, featuring different vendors and technologies. Building pipelines and data models that interoperate with existing infrastructure is often a significant engineering challenge.
Latencies
Depending on the application, AI signals may need to be produced in real time or within a specific latency window.
Security
The application surface for attacks increases significantly with the inclusion of connected physical objects. Security, both of the objects themselves and of the data and signals, becomes a vital consideration.
Humans-in-the-loop
Even when AI signals are available, the decisions may need to be made by humans, who may want reassurance or explanation before taking action. This can complicate the design of the application flow.
From IoT to Physical AI
AIoT can also be seen as the bridge towards a more general concept of Physical AI. AIoT focuses on the intersection of IoT and AI. But AIoT is really only a part of a larger space of AI systems interacting with the physical world. This can take the form of observing, understanding, deciding, and acting.
It is a spectrum, and the boundary between AIoT and Physical AI is often blurred. However, the distinction can help us understand the differences between the various applications. AIoT focuses mostly on observing and analyzing. Physical AI enables capabilities such as observing, understanding, deciding, and acting.
When it comes to industrial applications, the question of interest is how can we design such systems so that they are accurate enough, appropriate for the domain and its actors, and secure. The long-term promise of AIoT is that it enables such systems at scale.
For a more detailed perspective on how AIoT and Physical AI relate to one another, see the blog post at Aperture Venture Studio.
Build AIoT from Operational Needs
Finally, note that AIoT applications are often most valuable when they are designed from an operational perspective. Rather than the question being "Where can we apply AI?", it is often a question of "What operational challenge do we need to address?"
What physical processes are difficult to observe?
What information is unavailable, fragmented, or difficult to combine?
What decisions must be made, based on that information, and how timely do they need to be?
What actions are needed, once we have that information, and where do humans need to be involved in that process?
These are all questions that help us bring the technology and its capabilities back to operational considerations, informing the architecture and implementation. The overall goal of AIoT applications is to generate insights and decisions based on information that originated in the physical world.
Top comments (0)