IoT addressed an important problem: making physical objects observable.
They can give you information on temperature, vibrations, position, pressure, energy usage, equipment state, and thousands of other features.
But data collection is just the beginning.
The next and harder step is to make sense of this information.
This is where AIoT (Artificial Intelligence of Things) comes into play.
The idea of AIoT is to incorporate connected physical objects with data pipelines, ML, analytics, and applications to create systems that go beyond mere observation and make decisions about the physical world.
Let’s take a look at a generic AIoT architecture and discuss the different considerations and challenges.
An AIoT Stack
The typical industrial AIoT stack can be seen as a set of layers.
Physical Assets
↓
Sensors / Devices
↓
Edge Connectivity
↓
Data Ingestion
↓
Data Processing
↓
AI / ML Models
↓
Applications & Alerts
↓
Human or Automated Action
This is a good starting point for thinking about the types of engineering challenges you will encounter.
- Physical Assets
Industrial physical assets are:
Machines
Vehicles
Production equipment
Robots
Inventory
Energy infrastructure
Environment
And so on.
These are the objects you want to observe and make decisions about.
- Sensors and Devices
Sensors and devices bring physical characteristics of these machines to digital representations.
Depending on the industry, these sensors can measure:
Temperature
Vibrations
Pressure
Position
Motion
Voltage
Current
Humidity
Equipment state
And so on.
Sensors quality directly impacts the performance of the downstream ML model.
As always, garbage-in-garbage-out is in full effect.
Edge Considerations
The process of transforming analog values to digital reads is not the end of the pipeline.
Depending on the industry, the amount of data, and the latency budget, you might want to consider Edge computing.
For example, Edge nodes can pre-process the sensor data, filter out irrelevant information, detect relevant anomalies, and only report pertinent data to the central system.
Or the Edge node might take action on its own if the network connection to the central server is unavailable.
ML / Anomaly Detection for Predictive Maintenance
It’s tempting to jump to model creation.
However, as you build an AIoT production stack, you will realize that data pipeline often takes the largest share of the engineering budget.
The data pipeline has to reconcile data coming from various sources, vendors, and systems.
Time-series data has timestamps, units, sampling rates, and other information that needs to be normalized.
You also need the ability to ingest data from different vendors, systems, and devices.
Finally, the system should be able to process streaming data in real-time and store it in some time-series database for historical analysis and queries.
Once you have a working data pipeline, you can think about ML.
What ML Tasks Does the System Need?
Depending on your use case and operational goals, there are several different types of models you might need.
Anomaly Detection and Classification
Instead of hand-crafting rules, your model can detect and classify abnormal events.
Predictive Maintenance
You can train your model on historical data to predict future failures and breakdowns.
In an ideal world, a predictive maintenance system won’t give simplistic yes/no answers about whether equipment will fail.
A more realistic and effective approach is to give maintenance engineers early warnings and detect abnormal pattterns; let them decide whether to take any action.
Forecasting
AIoT systems often need to make predictions about demand, supply, inventory, and so on.
This is a good use for forecasting models that take time series data and predict next value in the sequence.
Computer Vision
Cameras can be another source of data in AIoT settings.
Vision models can be applied to do visual inspections of products, quality control, safety checks, and so on.
AIoT Is Not Just About ML
One of the common misconceptions about AIoT is that it is about ML models. However, in practice, such a system comprises multiple parts:
Sensors
+
Device Management
+
Connectivity
+
Streaming Data
+
Storage
+
Data Engineering
+
ML Models
+
APIs
+
Dashboards
+
Alerting
+
Human Workflows
ML models are only a part of the system.
In many ways, this is the engineering challenge of AIoT compared to, say, web development: industrial engineers don’t really care whether the model is fancy — they want the system to help them make decisions that improve their operations.
Human-in-the-loop Issues
You might be tempted to make your AIoT system fully automated, but true full autonomy is typically not possible.
Let’s say your AIoT system sees an unusual vibration pattern from one of your production machines.
Your ML model flags the event as suspicious, but what should be done? Some of the possibilities are that the machine is aging and nearing the end of its life, that the operating conditions changed, that there is a faulty sensor, or that the machine is due for servicing.
A human engineer might have more context and be able to reason about the situation better than an ML model.
This is another case where a hybrid or human-in-the-loop AI architecture can be more valuable.
Security
Any time you put industrial infrastructure online and connect it to an information system, you have to think about security implications. Some of the relevant security considerations for AIoT are:
Device authentication
Access control
Encryption and key management
Secure firmware
Network architecture
Model security
Overall data security
Incident response planning
And so on.
Depending on the architecture, an attacker could alter important physical objects, causing safety and operational issues.
It goes without saying that security must be designed from the start of the system development rather than being an afterthought.
AIoT in the Wider Economic Sense
The general purpose of building any industrial system is to turn it into something that creates value.
This is an economic issue at many levels and requires considering business, legal, and technical factors.
The AIoT technology field itself sits at a crossroads of several areas, notably AI, IoT, robotics, and operational systems.
Aperture Venture Studio (Aperture Venture Studio), for example, sees the convergence of AI, IoT, robotics, sensing, and operational technologies.
The main engineering challenge to build an AIoT system is not the connections between more devices but rather the feedback loop:
You can look at what is happening, understand it, predict future behavior, get insights or make decisions, and take some form of action that affects the objects in the system.
Summary
In sum, AIoT sits in the midst of several engineering specialties: IoT, embedded systems, Edge, cloud, data engineering, ML, networking, cybersecurity, analytics, and a few others.
A successful AIoT architecture will likely consider the problems from these points of view.
IoT is observation; AI is understanding and insight. Software connects everything, but domain knowledge rules the day.
From a software development perspective, this combination makes AIoT one of the most interesting engineering challenges of our time.
Top comments (0)