An IOT system isn't just a sensor connected to the cloud, the true challenge lies in how they communicate a reliable chain of information through to a use-case within a secure environment.
IoT systems gather information from sensors, machines, devices and the physical world. AI and machine learning can then analyze the information to spot irregularities, patterns, predictions and even make decisions.
A simple architectural design could go something like this.
Physical World → Sensors → Connectivity/Data → AI/ML → Decision → Action
Each step in the chain creates additional considerations for developer's looking to build an IoT system.
1. The Device Layer
We'll begin our exploration at the start of the pipeline - the device layer. Depending upon the application, this layer could feature:
- Temperature/Humidity Sensing
- RFID Scanners
- GPS Tracking Devices
- Cameras
- Machine Telemetry Sensors
- Environmental Sensors
- Industrial Controllers
- Smart Meters
These devices observe the physical world.
For instance, a piece of industrial machinery could be capturing information about it's vibration, temperature, speed and operating state.
It's also worth considering how valuable of a signal these devices provide.
Do we have missing, repeating, delayed or faulty information being received?
2. Connectivity and Data Transmission
We'll now move through the connectivity layer.
Depending on the application, an IoT system could utilize WiFi, Cellular, Ethernet, Bluetooth, LoRaWAN or various industrial communication protocols.
As with any IoT architecture, a lot of thought must go into connectivity, including: reliability, latency, bandwidth, device connectivity, security and intermittent connections.
Not all applications need to directly send events to the cloud. Some applications benefit greatly from processing closer to the source.
3. Edge Processing
A lot of solutions are now adopting a strategy best described by Edge Processing, where computations occur closer to the edge of an architecture in order to reduce latency or improve reliability.
For instance, A camera could potentially offload some processing to reduce the amount of information sent back to an on-premise security system.
A possible architectural design is as follows.
Device → Edge → Cloud/Data Platform → AI → Application
As discussed, the proper topology is usually determined by latency, computing power, security, connectivity and scalability requirements.
4. Data Ingestion and Storage
Now that we've transmitted the information, we will now turn our attention to data. It should come of no surprise that storage considerations become much more important when building an IoT application.
An IoT data platform will typically see a high volume of events arriving from devices in an application.
Some important considerations for developers should include: schemas, device identifiers, timestamps, validation, storage requirements, data retention, duplicates, and missing events.
A structured data schema would make data easier to consume for subsequent downstream processes.
Imagine we have an equipment event which conceptually features the following design.
device_id → timestamp → measurement → unit → location → status
Creating traceable and structured data makes the life of subsequent analytics and machine learning systems much easier to deal with.
5. Applying AI and Machine Learning
With a reliable source of information, an organization can begin to apply AI techniques to extract patterns in their data that can in turn make predictions or decisions.
Depending on the problem domain, an organization could benefit from utilizing anomaly detection, predictive maintenance, demand forecasting, computer vision, asset optimization, pattern recognition, classification or operational forecasting techniques.
For instance, imagine we have vibration measurements from a machine which we've used to train a machine learning model which will then recognize abnormal patterns.
The most important thing to remember is that AI is applied to information which in turn was captured from the physical world.
6. From Prediction to Decision
An AI model becomes more interesting when it's output can be integrated into a workflow which will allow it to contribute to a tangible decision.
For instance:
Sensor observes abnormal vibration pattern → Data Pipeline sees event → ML model identifies anomaly → application raises alert → maintenance crew investigates.
This represents a very common connection of an event in the physical world to an analytical workflow being executed in the digital world.
In highly automated plant-facilities, an authorized control system in an operational environment could also automatically execute a predefined action. This type of automation is useful but should also be carefully guarded by appropriate safeguards, permissions, validation and human-governance.
7. Reliability and Observability
An AI model could perform exceptionally well under certain conditions however begin to produce unreliable inferences in production due to issues with the data pipeline.
Some important considerations for an IoT developer should include: sensor calibration, quality of data, model drift, network failures, clock synchronization, duplicate or missing events, faulty devices, logging, model performance, and recovery strategies.
A good observability strategy will typically span all tiers within an application which makes observing just the application tier insufficient.
One effective monitoring strategy could potentially detect whether a problem originated from a device, network, ingestion layer, data store, model or the application itself.
8. Security Across the Stack
An enterprise will also have to consider security considerations across the entire architecture. An AI + IoT application represents a complex security surface which will also require operations and monitoring investments.
At various points within an architecture, security should be enforced between: devices, networks, APIs, cloud infrastructure, databases, machine learning models, and user access to applications.
Key security considerations should include: device and application authentication, encryption, secure communication channels, access control and credential rotation across APIs, data encryption at rest and in motion, and regular firmware and software updates.
A company cannot secure just the application if the underlying communication channels between a device and the cloud remain insecure.
9. Where could this apply?
The same architectural design could be utilized across industries with differing implementation considerations.
Manufacturing
Machine data can support predictive maintenance, operational analytics, anomaly detection and quality analysis.
Logistics
Vehicle and fleet information can support location tracking, route optimization and fleet management systems.
Agriculture
Environmental sensors can provide insights into field conditions which could be leveraged with analytics or AI to optimize input spending.
Construction
Connected equipment and sensors can provide insights into assets, environment and more.
Energy
Infrastructure monitoring can support anomaly, maintenance and safety predictions across oil rigs, windmills, power grids and more.
A common theme of these application domains is sense → connect → process → analyze → decide → act.
10. The Right Problem to Solve
One of the largest mistakes a team could make is choosing the wrong application for their technology.
A better approach would be something like this:
- Identify an operational problem
- Identify what information is required to solve a problem
- Determine what sensors or devices can capture the information
- Design the connectivity or data pipeline
- Decide what should be processed at the edge or in the cloud
- Determine what analytics or ML should be applied
- Connect model outputs to a useful application or workflow
- Monitor or observe how well your system is working after deployment.
By thinking in terms of a problem first, we avoid building yet-another-sensor-pipeline-without-reason.
The Big Picture
Why do we care about AI + IoT together? Essentially, we are connecting software intelligence with the physical world. We can now begin to use sensors to learn about the real-world. The data infrastructure now allows us to turn these observations into reliable information. AI can extract patterns from that information to predict future values or behaviors. Then, applications can integrate these insights to make informed decisions or authorize actions, when authorized.
This is why concepts similar to AIoT or Physical AI make so much sense for Industrial Technology.
For a glimpse into how these concepts are explored within an industrial context, Aperture Venture Studio makes an excellent reference.
The engineering challenge doesn't lie in connecting more things or deploying more AI models, it's about building secure and reliable observation pipelines which can convert physical-world data into a valuable asset for any organization.
Top comments (0)