Connecting a sensor to the internet is fairly straightforward.
The harder problem consists in extracting valuable information from the data produced by thousands of connected devices that can aid in making a real world decision.
This is where AI and IoT come together: IoT provides the connection to the physical world via sensors, machines, and devices, whereas algorithms can help make sense of the data.
The joint solution is sometimes also called AIoT (Artificial Intelligence of Things).
An AIoT system can be thought of as a chain:
Sensors → Connectivity → Processing → AI/ML → Insights → Action
Each link has its own set of challenges in engineering.
- Sensors: Access to the Physical World
Sensors are where information about the real-world phenomena captured by IoT devices comes from.
Depending on the use case, this information can consist of temperature, vibrations, pressure, location, motion, machine status, or energy consumption.
The quality of this information is of utmost importance as it directly impacts other stages.
Do any measurements get missed? Are the timestamps correct? Is the sensor calibrated correctly? All of this affects downstream processing.
Before selecting the model that will turn sensor data into information, it is important to assess if the data is suitable for analysis.
- Connectivity: From Sensors to Processors
Once data is generated, it has to be transmitted somewhere: to a gateway, to another processor, or to an application. In practice, there is a suite of different connectivity options that can be used.
Depending on a use case, Wi-Fi, cellular networks, Bluetooth, UWB, MQTT, or industrial communication platforms can be used.
The choice of connectivity is important: the latency, range, bandwidth, reliability, power consumption, and security features of a chosen method have to match the constraints of a given AIoT system.
Connectivity is thus a design choice and an important aspect of system architecture.
- Processing: From Raw Sensor Readings to Information
Raw data from sensors is rarely ready for direct analysis or machine learning.
Typically, it has to go through a series of transformations:
Data collection
Validation
Cleaning
Normalization
Time-series extraction
Storage
Joining with other datasets
For example, readings about machine vibrations may sometimes be more informative when combined with readings about the machine’s operating conditions or maintenance history.
This is why data engineering plays such a large part in AIoT development.
- AI and Machine Learning
At some point, an AIoT system reaches the stage where patterns in information can be used to improve the outcomes of a given application or to make decisions.
For example, machine learning can help detect anomalies in equipment readings, classify images, recognize patterns in unstructured text, make suggestions, or plan resources.
In many cases, the type of model chosen is dictated by the application: sometimes, a deep learning model is not the best choice for a given task.
Edge and AIoT
Not all sensor data can or should be sent somewhere for processing: there are latency constraints and systems where constant connectivity is a weakness.
Edge computing helps distribute processing and decision-making closer to sources of data.
An AIoT system can be structured around an edge-cloud dichotomy:
Device → Edge Processors → Data Cloud → Machine Learning → Applications
For example, some preliminary data filtering can be performed at the edge, whereas complex pattern recognition can happen in the cloud.
How this balance is struck depends on network conditions, security requirements, and the characteristics of the application.
The Integration Challenge
Chances are, an AIoT system does not exist in a vacuum: there are legacy systems in the environment that cannot be replaced or are not flexible enough.
The heterogeneity of infrastructure and data formats can be an impediment to integration.
A factory environment, for example, can involve sensors, PLCs, SCADA systems, robots, MES, and enterprise applications.
Designing an AIoT system around such an environment is much more of a challenge than selecting the right machine learning model.
Another challenge is that a good machine learning model cannot make up for inadequate data integration.
It should also be born in mind that AIoT systems are fundamentally interconnected: an issue at any single point can bring the whole system to a standstill.
From Insights to Decisions
One of the common pitfalls of AIoT development is that the end-to-end nature of the system is not respected.
It is particularly important to think through what should happen with the insights generated by an AIoT system.
Let us imagine that a machine monitoring system has detected unusual vibrations that point to a possible equipment failure.
Here is how an end-to-end process can look like:
A sensor has picked up unusual vibrations.
The reading has been validated.
Historical data has been retrieved for context.
An AI model has inferred that an anomaly has occurred.
The system has evaluated the situation and determined that maintenance is required.
Maintenance personnel has received an alert.
The equipment has been taken off-line for inspection, which has now been recorded by the system.
Without such a feedback loop, the value of an AIoT system can be severely undermined.
An anomaly detection model might generate interesting insights, but without a corresponding process to act on these insights, it is a wasted opportunity.
Common AIoT Implementation Pitfalls
There are several pitfalls that are common for AIoT implementations:
• No well-defined use case
• Data quality issues
• Connectivity design issues
• Selection of an inadequate AI model
• Inadequate system integration
• No concrete action on insights
• Inadequate security considerations
In practice, these pitfalls are the result of the same cause: the failure to design an end-to-end system.
AIoT is a systems-level technology. Any corner can sabotage the launch or the performance of an AIoT product.
Start With the Problem
When designing an AIoT system, there is an attractive idea to start with the means: what sensors and IoT devices will we use, what cloud will we select, and what AI models will we employ.
In practice, it is much more useful to adopt a problem-driven approach.
Ask yourself these questions:
What is it that we want to learn?
What type of data will help us learn it?
What do we want to do with this information?
What action will this information enable?
These questions will shape the rest of your AIoT development and ensure that you have the right tools for the task.
Ultimately, it is not the right time to think about what AIoT can do for you: focus on what you can do for your business with AIoT.
Top comments (0)