Artificial intelligence solutions applied in industrial settings often share a common challenge: achieving the connection between intelligent software and the physical world.
The data generated by manufacturing plants, warehouses, construction sites and other industrial facilities through various sensors, RFID tags, connected devices and existing enterprise systems need to be processed and analyzed in order to create value in the form of operational insights, automation and informed decision making.
Let us discuss why this is a systems-level problem, how this impacts the creation of an industrial AIoT architecture, and what engineers should think about before jumping to model development.
- The Industrial AIoT Architecture
In its most basic form, any working solution will involve multiple layers:
Physical: The machinery, objects, assets, sensors and connected devices existing in the physical world;
Data collection: RFID, IoT sensors, connected equipment and existing data sources to gather information about the state of the physical world;
Integration: The ability to join disparate data sources with operational technology systems;
Intelligence: The application of analytics and AI to extract actionable insights;
Application: The ability to make information accessible, often in the form of an interface or set of rules.
The specifics will vary depending on the needs of the industrial setting, network infrastructure and other factors.
A crucial insight to keep in mind is that none of these components can be designed in isolation. Software intelligence cannot make up for insufficient data collection, poor integration with existing systems or lack of application-specific customization.
- Prioritize Good Data Over Neural Networks
The problems often start before the model selection phase.
In most cases, industrial facilities do not have perfect data: events may be mislabeled, timestamps incorrect, relevant objects not present etc.
This issue compounds when systems that generate data use different identifiers for the same object or do not expose necessary parameters.
A solid implementation requires careful consideration of data validation, normalization and reconciliation, including label consistency, value integrity, temporal correctness and observability of data flow.
Another important aspect is validation: ensuring that the data actually reflects the state of the physical world.
All these elements contribute to the systems-level foundation necessary for auditing, debugging and model validation.
Without these elements in place, even the best neural network will not be sufficient for creating an accurate AIoT solution.
- Systems First: Device Abstraction, Interfaces and Interoperability
Most industrial facilities are not fully digitalized from the ground up.
There is often a mix of legacy equipment, proprietary communication channels, various vendors and other factors to consider.
The correct approach is to analyze the existing infrastructure and identify points of integration.
Some projects benefit from having a thin layer of abstraction between low-level device communication and application logic.
In any case, critical infrastructure should be evaluated based on power and connectivity availability, access to documentation, security requirements, existing data formats and other relevant factors.
The solution must be able to operate in potentially harsh conditions.
This may require additional processing at the edge or other measures to ensure reliability, availability and integrity of information.
- Prove Validity of the System, Not Just the Model
The proof of concept for an industrial AIoT project should not only involve the creation of a working prototype.
It is important to test the system as a whole and ensure that the end result fulfills the intended purpose.
A working but fragile system that generates correct results under ideal conditions is not significantly better than a demonstrator.
When applicable, the model should be tested with real-world data.
Using the example of predictive maintenance, this means that a prototype should be evaluated against a real-world asset or assets.
Same applies to other aspects of the application: logistics tracking, inventory management, analytics and reporting functions.
If the system is intended to operate in a network partitioned environment, its ability to operate in such conditions should be validated.
- Building Industrial Tech Companies: Thoughts on Systems-Level Engineering
These principles apply to the work of any company involved in the development of industrial technology.
A working product is only as good as the quality of systems engineering that went into it.
Any solution targeting the physical world implies the necessity of systems thinking that takes into account the realities of the industrial environment, data collection and integration needs, as well as application-specific logic.
I hope that by writing this article I managed to convey the importance of this philosophy to anyone involved in the development of industrial AIoT products or applications.
As the systems architect at Aperture Venture Studio, I am often required to advise our machine learning engineers on the intricacies of industrial applications and data flow.
Discussion: When designing industrial AIoT platforms, which layer tends to create the greatest engineering challenge: device connectivity, data integration, analytics, or deployment in real operational environments?
Top comments (0)