Many existing Internet of Things (IoT) systems were focused on the problem of data retrieval — getting information from a sensor to a dashboard.
A sensor would measure a given value (say, temperature). A gateway would retrieve that information and send it via a network stack to a data platform, then to an application, and ultimately to a dashboard for viewing and analysis.
This was an effective way of solving a particular problem: making previously inaccessible information visible and accessible.
But what comes next for developers who want to build the next generation of connected systems?
In particular, how can they productize an AI model to make it valuable within an IoT architecture?
Traditional IoT Pipeline
A basic IoT pipeline can be visualized as something like the following:
Device
Sensor
Gateway
Network
Data Platform
Dashboard
The architecture is great for gathering and displaying information.
Suppose, for instance, that you want to monitor a set of manufacturing machinery.
You might have a pipeline that ingests temperature, pressure, and vibration readings along with data about the machine’s state.
After that, you can view the information on a dashboard.
But does the dashboard indicate anything about the normality of the data or suggest any particular course of action?
Adding an Intelligence Layer
Let’s add an intelligence layer:
- Physical Asset
- Sensors
- Connectivity
- Data Platform
- AI / Analytics
- Contextual Insight
- Application
- Human or System Action
Essentially, we’ve added two new steps to the process: contextual insight and a recommendation for human or system action.
This transformation appears to be simple — just add an AI model.
But in practice, it can be more involved than it appears.
Specifically, an AI model must analyze a signal or set of signals and transform it into some sort of recommendation, prediction, or explanation.
Intelligence Is About Context
For instance, consider a scenario in which a manufacturing machine’s vibration levels have increased.
A basic IoT system could notify an operator that something is happening.
An intelligent system could ask more questions: Is the machine running at a higher speed than usual?
Has the production line changed?
Has the vibration increased gradually over several days, or has it changed suddenly?
Has the machine been maintained recently?
Have similar patterns been seen in the past, and if so, what were their results?
The AI model might not be able to answer every question, but the system could integrate and analyze multiple signals.
These signals would be more valuable than a single sensor on its own.
No Need For Total Autonomy
Some people think that adding AI to an IoT system inherently requires that the machine make decisions on its own, but this isn’t the case.
Here’s an example of a potential pipeline:
- Sensor Event
- Data Validation
- Context Enrichment
- AI Analysis
- Prediction / Recommendation
- Human Review
- Operational Action
The example shows that human review is still a part of the process — and for some use cases, this is the correct approach.
AI doesn’t have to replace every aspect of a person’s job; it only has to make their jobs easier.
Edge Intelligence
Not all models have to run on the cloud.
For some industrial applications, latency is critical, or the connection can be spotty at times.
In these instances, some of the intelligence layer can be moved to an edge device:
- Machine
- Edge Device
- Local Processing
- Immediate Decision
- Cloud
- Historical Analysis
This approach creates a hybrid system, one in which some processing takes place locally and some processing takes place on the cloud.
There’s no need to make every model a distributed, or even an edge, model.
The appropriate location for it depends on the use case.
The Real Engineering Challenge
The real challenge of AIoT development often has little to do with the model itself: developers have to consider the pipeline, data quality, device identity and health, time sync, connectivity, versioning, data drift, edge vs. cloud, security, observability, and human review.
A model’s accuracy doesn’t necessarily dictate the result of the system — the engineering around it does.
A New Long Tail
This evolution leads to a new long tail:
- Connect
- Collect
- Detect
- Understand
- Predict
- Recommend
- Act
IoT solved the first two.
AI can solve the rest, but it depends on the use case.
A recommendation still has to fit within the context of a broader application or even business process.
It isn’t enough to simply have an autonomous, self-driving machine — the right tool for the job has to be selected and put into the correct workflow.
The Future of Connected Systems Was Never About the Sensors
This isn’t to say that the sensors themselves aren’t important; it does mean that developers shouldn’t stop at the model.
The true opportunities in AIoT development lie in the space between.
Aperture Venture Studio is one such company, focusing on the intersection of AI and IoT to develop solutions for physical-world and industrial applications.
Top comments (0)