IoT systems can observe a massive amount of information from the physical world: sensors capture temperatures, systems share activity details, vehicles share location data, and connected devices continuously produce events
But getting that data is only part of the story.
The more interesting challenge is in using that intelligence to create value
This is where AIoT comes in. Instead of simply observing and reporting, an AIoT system can take that data and create intelligence that informs decisions and potentially takes action
California's tech ecosystem includes companies working in AI, IoT, robotics, and autonomous systems - all fields related to tying together software capabilities and the physical world.
What is AIoT
We can depict a simple IoT architecture like this:
Devices → Connectivity → Data → Application
An AIoT approach adds intelligence in the mix:
Devices → Connectivity → Data Processing → AI/ML → Decision → Action
Let's say we have an industrial machine that has some sensors attached. Those could be measuring vibrations, temperatures, pressures, statuses, or similar variables. A simple IoT application would stream or store this data for later analysis
An AIoT approach could analyze this data (current and past) to spot important patterns or support decisions.
The value proposition is in what this entire architecture can do as a whole, not any single element.
A practical architecture
A good way to think about an AIoT architecture is as this:
Physical World → Sensors → Connectivity → Data Processing → AI/ML → Application → Decision → Action
Each stage handles an element of the overall task.
1. Physical World
This can represent anything that we need to observe.
Depending on the application, this could include machines that are running, vehicles that are being driven, products being transported, construction activities or worksites, or other elements.
2. Sensors and Devices
Most of these will be sensors that capture data about the physical world. Based on the application, it can involve temperature, location, motion, vibrations, pressures, status checks, or other variables.
The overall quality of this layer will impact later stages: bad sensor placement, lost or missing signals, or other issues that reduce data quality will limit what we can do later in the architecture.
3. Connectivity
This layer handles transporting the data from the sensors to someplace where it can be processed. Depending on the use case, this can involve Wi-Fi, cellular connections, Bluetooth, LoRaWAN, or other approaches.
The choice of connectivity approach will impact performance, range, power usage, reliability, and other considerations.
4. Data Processing
This layer involves any number of data processing steps after the data leaves the sensors. We can conduct data filtering steps to focus on important variables, outlier rejection to improve quality, normalization, data aggregation and enrichment, or a variety of other transformations
This stage becomes especially important if the application produces high volumes of data.
5. AI and ML
At this point, we have valuable data that can feed into AI and ML applications. We can train an ML model or use existing anomaly detection, classification, forecasting, or regression algorithms to generate important insights.
But note that we're only dealing with what we have: the AI/ML model is only as good as the previous stages, and mistakes there will impact results later.
Edge vs Cloud
An important consideration in these systems is the choice between edge processing and cloud processing.
While cloud computing provides immense processing capabilities, we'll sometimes have a reason to leave that data closer to where it was created.
One common use case involves latency: if a physical application requires low latency processing, getting data to the cloud and back again may be too slow.
We may need to consider an architecture that involves:
Device → Edge Processing → Cloud/Data Platform → AI/Analytics → Application
Depending on use cases we can have some processing at the edge, and some in the cloud.
Applications of AIoT
These approaches can cut across a variety of industries and areas:
Manufacturing
Connected machines can produce valuable insights about their operation and performance; AIoT can bring in machine learning capabilities to analyze this.
Anomaly detection is one example here, or using this data for predictive purposes.
Logistics
IoT-connected devices can provide information about the status of vehicles, assets, and shipments. Combined with this analytics layer, we can draw insights about the state of our logistics operations.
Construction
Construction sites can have vehicles, materials, and workers. These represent potential uses for both IoT and AIoT approaches for connected logistics and analytics capabilities.
A sensor layer allows the site managers to see what's going on; AIoT adds value by transforming that into actionable intelligence for logistics and productivity insights.
Agriculture
Using sensors in agriculture, AIoT can enable valuable applications by analyzing weather, temperatures, and other conditions to help farms and agricultural operations make better decisions
Autonomous Systems
Autonomous robots represent another AIoT application. This can involve both perception and planning capabilities: sensing the world around, making decisions about actions, and executing them in the physical world.
These represent just a few potential areas, but there's much further to go both due to this architecture being relatively new and because a growing number of fields are exploring these.
What Does California Do?
There are some companies in California working on these AIoT applications. Samsara provides solutions in the connected operations space, and Nuro focuses on the autonomous systems side. But more interesting, perhaps, are the companies working to explore this space, tying together AI and IoT.
(https://apertureventurestudio.com/) is an example of one such company; looking at the portfolio shows projects around AIoT and the emerging space of Physical AI, covering the spectrum from AI to IoT and connected devices.
In short, while these approaches cut across a range of industries and fields, there are still some companies in California interested in these concepts.
AIoT Is Bigger Than You Think
It can be very easy to file AIoT and think "Oh, I just need to attach the AI part on top of the IoT architecture." When we actually need to consider the architecture, it's considerably more complex.
Developers and architects looking to build AIoT applications should consider these points:
What physical data do we want to collect?
What sensors would we need to collect this data?
How often do we want to collect this data?
Can we trust that the data collection will be reliable?
Do we want to process this at the edge or in the cloud?
How would devices communicate in the event of disconnected periods?
Where would this AI/ML model live?
How would its output get ingested into applications or decisions?
What happens if the sensor gets damaged?
How would we support this architecture at scale?
These and many other questions are often as important as the AI model itself. Most people aren't really aware that building a good AIoT application involves as much engineering skill as it does machine learning expertise.
From Connected Devices to Physical Intelligence
Overall, one way to think about the AIoT architecture is:
Sense → Connect → Process → Analyze → Decide → Act
IoT deals mostly with the sensing and connecting layer; AI can help with processing and analyzing the data. Applications then tie this into decisions and actions. The value proposition is that AIoT applications can impact the physical world, using these principles.
We won't belabor the point: that's why it's interesting!
Final Thoughts
AI + IoT is increasingly shaping up as a space that is becoming more relevant in our world. IoT provides the ability for data collection and sensing; AI enables analytics capabilities that can inform decisions. Edge and cloud platforms provide powerful computing capabilities for any processing or analysis we need to undertake.
For developers, it's worth considering the broader picture here. AIoT applications aren't as simple as "throw an AI model on this IoT data" they involve much more complexity. This is what makes the whole thing interesting: turning this concept into practical applications that are relevant, valuable, and sustainable.
What matters is creating a reliable value chain that goes:
Physical World → Data → Intelligence → Decisions → Action
That is what transforms connected devices from being interesting data sources into something that can really shape our world.
Top comments (0)