Industrial data needs new targeted ontologies. They aren’t for humans
A pair of engineers inspect a series of charts and graphs with ontologies optimized for human comprehension (Source: Getty Images/Cultura RF).
In times of old, industrial data was optimized for human understanding. Sure, much of it was used, stored, accessed, processed, graphed or even consumed digitally. But, from an ontology perspective, humans were the primary audience.
As digital tools like machine learning (ML) and artificial intelligence (AI) increase in popularity, it becomes important to optimize industrial data for both humans and computers.
Consider the data siloed between product lifecycle management (PLM) and enterprise resource planning (ERP) software. While human intuition is well adjusted to understand the cross-system overlap and relationships between the data stored within these tools, AI trained using the information on these silos will initially see them as unrelated data. The assumed relationships made by the AI based on this training can therefore become a significant source of hallucinations.
As a result, a new imperative is emerging where engineers must organize a lot of their industry data into ontologies optimized for AI and human comprehension. Or as William Hahn, director of Solutions Consulting at Siemens Digital Industry Software, said it, “Industrial data is becoming more machine-native, requiring ontologies that encode relationships across systems—while still supporting human oversight.”
What are ontologies and why are they important?
The National Institutes of Health (NIH) explains that “ontologies are essential to science because they identify and clarify the entities and concepts that people want to talk about and study, and they identify the key relationships among those concepts.”[1] In other words, ontologies are the ways we categorize and conceptualize information to best communicate their connections, properties and relations.
To demonstrate an ontology, the NIH cites the Diagnostic and Statistical Manual of Mental Disorders (DSM) as a tool to describe a patient’s mental disorders amongst people of different professions. When doctors select a code from the DSM to describe a patient’s condition it, “enables payors to know the purpose of some clinical encounter, … pharmacists to know the purpose of some prescription for medication, [and] social workers to negotiate follow-up care with other health care facilities.”[1]
How ontologies relate to industrial engineering
Many ontologies exist in the industrial world. But a great example includes the drafting of detailed drawings and the visual language associated within these documents. Consider blueprint, isometric and assembly drawings. They all have their own set of rules, shorthand and best practices that help individuals, who have never met or spoken, communicate how to build and manufacture objects.
These drawings are also a great example of how ontologies evolve, as much of the logic behind these documents live on in modern CAD software. The primary benefit of this change is that CAD models are salient to humans and machines. Thus, they work well with other industrial technologies like computer-aided engineering (CAE) simulations, computer numerical control (CNC) and more.
As these digital tools became increasingly popular, it started the industry trend to adopt ontologies optimized for both computers and people. But in the age of AI, this trend becomes a necessity.
The ontological future of the industrial sector
Ontologies optimized for the human brain tend to be based on human senses. Thus, they are typically visual, tactile or auditory in nature (imagine drawings, models and written language). For example, consider a spreadsheet and a graph. Numbers organized in rows and columns can be helpful to humans when the data is compact. The data, however, typically needs to be graphed to easily relay to humans the relationships and trends within large swaths of data.
Hahn describes that ontologies for computers are more like a fabric. He said, “Imagine more of a web of data where each data point is a node, [and] there’s connections … between each node. You can get this … complex web of data that crosses across disciplines and across systems. We can start creating all this interconnectedness regardless of data model … it doesn’t matter what your data model in PLM is, or your data model in ERP. You kind of extract that out into an ontology and say ‘here’s my web of data, all the data objects and the interconnected relationships in one system and the other system. [We bridge how] those data nodes connect to the data nodes in this other system. And we start building out a fabric, a web of data that’s completely interconnected. [This enables AI systems] to truly understand not just the relationships within systems, but the relationships across systems.”
As more data is used, stored, processed, graphed and consumed by data centers, ML algorithms, automated machinery and other staples of the AI sphere ontologies will become more machine centric. They, however, will not move fully away from humans—as governance, compliance and validation still require oversight.
Why industrial AI needs its own ontologies
Consider again the AI algorithm trained on ERP and PLM data without the ontologies it needs to understand how the data within those systems overlap and relate. Hahn says, “it has to guess, based on contextual clues, what those relationships are. And every time we know AI guesses, it just opens up the door for hallucination.”
Since many engineers know about the overlaps and relationships between ERP and PLM data, they can teach them to AI algorithms via ontologies. Thus, what was once seemingly unrelated data—from the AI’s perspective—now has a clear relationship. Hahn noted that this will not eliminate the full risks of AI hallucinations, but it will certainly address one significant source of them.
To learn more about ontologies and their importance to industrial AI, listen to the podcast: Ontologies: How to make sense of the data explosions in process industries.
[1] National Academies of Sciences, Engineering, and Medicine; Division of Behavioral and Social Sciences and Education; Board on Behavioral, Cognitive, and Sensory Sciences; Committee on Accelerating Behavioral Science through Ontology Development and Use; Beatty AS, Kaplan RM, editors. Ontologies in the Behavioral Sciences: Accelerating Research and the Spread of Knowledge. Washington (DC): National Academies Press (US); 2022 May 17. 4, How Ontologies Facilitate Science. Available from: https://www.ncbi.nlm.nih.gov/books/NBK584333
