Typically speaking, artificial intelligence tends to focus on models, algorithms, and the sheer computational power of silicon circuits. However, as is the case with anything, the usefulness of AI is directly determined by the data that it can access.
This is where the Internet of Things becomes an interesting prospect
IoT devices can constantly gather information via sensors, machines, video, meters, and connected equipment from the physical world, which can then be processed and interpreted by AI to detect, predict, or automate patterns and behaviors.
These combined forces are sometimes referred to as AIoT, or Artificial Intelligence of Things.
How does the AI + IoT pipeline work?
A typical AIoT architecture can be thought of as a data pipeline:
Physical environment → IoT devices → Connectivity → Data processing → AI/ML model → Insight → Action
Let's take a factory setting as an example. Sensors attached to various pieces of machinery could be used to measure:
Temperature
Vibration
Pressure
Speed
Machine status
Power consumption
That data could then be fed into some sort of edge device or the cloud, where an AI could process historical and real-time data to detect patterns or anomalies. For instance, perhaps a machine's rate of vibration slowly climbs out of a tolerable range, which could then be detected by an anomaly detection model.
The important thing to note is that IoT provides the link to the physical environment, while AI provides a mechanism to extract value from the information.
Why IoT data is important for AI
Typically speaking, a dataset will be built from databases, documents, transactions, and user-generated content. IoT introduces yet another category of information that is continually being generated - the physical world.
This can provide valuable data points to an AI that previously did not exist, or allow for richer insights than just historical analysis.
In an industrial setting, this can be leveraged for use-cases such as:
Predictive maintenance
Production monitoring
Inventory analytics
Asset tracking
Energy reduction
Quality assurance
Supply-chain analytics
Environmental monitoring
However, the truth is that simply having more data does not guarantee success for an AI initiative.
The data-quality challenge
IoT devices may be able to gather massive quantities of data, but a lot of it may be noisy, missing values, have faulty timestamps, be duplicated, or otherwise unusable. Data scientists may need to spend considerable time and effort doing any number of ETL-type processes:
Data validation
Noise reduction
Handling of missing data
Normalization
Feature engineering
Real-time data processing
Data storage and management
Before actually feeding IoT data into an AI model.
This makes the architecture design just as important as the ML algorithm being used.
Edge AI vs. cloud AI
The question of where to process data is an important one as well. With cloud computing, IoT data can be sent to centralized servers, where more powerful models can be used for batch processing.
Meanwhile, edge AI processing keeps some of the analysis closer to the source. It can be useful in situations where latency is an issue, or where it would be too resource-intensive to send all of the data somewhere else.
A hybrid approach could also be taken, where certain transformations and simple analytics occur at the edge, but more involved processes like heavy lifting, machine learning, and model training are done elsewhere.
AIoT is more than connecting sensors to an AI model
One common misconception about AIoT is that it is simply about throwing sensor data at a machine-learning model. In practice, there is a long and involved pipeline that must be taken before any valuable insight can emerge. One must consider:
Sensors → connectivity → data ingestion → storage → processing → models → APIs/applications → operational workflows
The last step in this list is often the most important one, since the end result of an AI model is typically useless if it is never connected to an actual application. Someone must be notified, an inspection must be scheduled, or a process must be altered for an AIoT system to be truly useful.
This shift from simply gathering data to utilizing it for actionable insight is one of the more exciting prospects of industrial AIoT.
Where venture building fits in
The space of AIoT also allows for some interesting venture-building opportunities. By focusing on particular verticals and problems to solve, startups can begin to build out products, processes, and systems that utilize both AI and IoT effectively without having to develop something that is incredibly broad and generic.
Aperture Venture Studio is a venture studio that focuses on AI, IoT, and industrial applications.
Final thought
The future of artificial intelligence is not going to be defined by ever-bigger and more complex neural networks. While that is an exciting prospect to witness, the truth is that AI needs information about the physical world in order to be useful for most applications.
IoT provides the link to this information. Meanwhile, edge and cloud computing provide the horsepower to process this data. Finally, AI can be employed to detect patterns and behaviors that can then be acted upon.
The interesting engineering challenge is not "how do we build an AI model?" but rather "how can we build some sort of system that connects physical-world data to AI, and subsequently turns that into something useful?"
Top comments (0)