We are in the midst of a major technological inflection point: knowledge graphs have developed into a foundational component of the modern Industry 4.0 technology stack. I have this perspective because in my role at Cambridge Semantics I constantly engage with business and technology leaders and discuss data strategies, initiatives, and future-forward planning. I’m not alone in this viewpoint; the software industry sees this market opportunity too, and most major technology vendors have positioned offerings.
Knowledge graphs are not a “new” technology. For years, proponents of knowledge graphs have touted their data integration capabilities. Today, knowledge graphs are increasing automation and reducing decision cycles, drastically optimizing data usage and efficiently driving down costs. This is because knowledge graphs semantically harmonize access to diverse data sources, resulting in machine readable and human understandable data. Now, the suite of demands and challenges in Industry 4.0 environments have pushed beyond traditional data landscape capabilities and propelled market interest toward knowledge graphs.
Before diving into the current explosion of interest, let me quickly establish the key requirements for the knowledge graph stack.
- The ability to construct an ontology or semantic model which describes the entities and relationships existing in the system.
- The ability to integrate data from diverse sources into an RDF-based knowledge graph in alignment with the ontology.
- The ability to execute performant queries on that data. This requirement is often overlooked, but Architectures which don’t include a distributed database with an OLAP (online analytical processing) design are inherently limited.
Set those aside for a minute, and let’s discuss some key factors driving the growth of knowledge graph technologies.
Manufacturing Systems and Processes Have Complex Relationships
Digital twins are inherently intricate, combining data from several sources as they reflect an entity’s state across its lifecycle. The state of the entity and its relationships cannot always be easily modeled using a traditional schema. For example, a bill of materials for a given design has a hierarchical relationship between entities and sub-assemblies within a particular entity which present a challenge for a relational database. A knowledge graph is a natural way to express these relationships and create validation rules to ensure compliance.
Another complex use case is the enterprise supply chain which can be modeled in a network (graph) form. Supply chains are akin to Complex Adaptive Systems which feature ongoing emergent and divergent relationships. The data architecture supporting the robust supply chain must be adaptable and gracefully cope with change as a constant. Knowledge Graphs — especially those built on W3C OWL/RDF — are ideal for such systems.
Semantic Reasoning Augments Traditional AI
Knowledge graphs have long been understood to be particularly useful in analyzing specific types of systems. Graph algorithms provide a natural means of calculating shortest paths, system throughput, community detection, node importance, and node similarity (graph embeddings).
The Graph Search Neural Network [is] a way of efficiently incorporating large knowledge graphs into a vision classification pipeline. We show in a number of experiments that our method outperforms standard neural network baselines for multi-label classification.
– Dr. Kenneth Marino, PhD Thesis
What’s become increasingly clear is that these techniques can also augment traditional machine learning. Researchers and academics have realized that robotic vision algorithms can benefit from logic available solely through semantic reasoning. One early adopter, Amazon, is betting that this technology can eventually lead to a fleet of autonomous robots. Having a semantic understanding of the fulfillment center is optimizing the movement of robots as they navigate the busy floor and in the future may help with robotic grasping and other tasks.
Knowledge Graphs Enable Q & A Engagement
Artificial Intelligence is certainly having a moment. ChatGPT has captured the public’s attention, garnering headline articles in the New York Times and other newspapers. Personally, I’m more interested in the capabilities of the GPT-3 model (third-generation Generative Pre-trained Transformer). Over the past few months, I’ve experimented with GPT-3’s capabilities via their public API and I was stunned. I provided as input a semantically described ontology; from this GPT-3 was able to generate SPARQL queries from natural language questions. I tested with non trivial questions that require multiple-hop relationships amongst connected entities as well as aggregations. Because the ontology is represented with semantic terms, GPT-3 was able to translate synonyms or derivatives of the language I used onto the semantic model. An example query (input) and the automatically generated output is below.
Input: What are the components that occur more than twice in the bill of materials for the tubes with attached sensors that have a temperature reading over 140?
Researchers from all over the world are trying the same thing. The implications of this are manifold. The next logical step in my view is the introduction of a simple chatbot or user interface where managers or technicians can ask any question of a manufacturing system and receive real-time data in response.
A Vision For the Future
At this point, many companies are still cobbling together their knowledge graph strategy, but there are some firms in the vanguard.
Bosch is one company that has gone all-in on knowledge graph and semantic technology. Not surprisingly, Bosch is one of Cambridge Semantics largest clients. Anzo, our knowledge graph product, is the only full stack knowledge graph platform with a massively parallel processing (MPP) architecture that allows users to create knowledge graphs by overlaying and combining data from any data source - structured or unstructured. Each data source is activated in the RDF graph engine as an in-memory data layer, on top of which additional layers can be added that logically connect, extend, and transform the knowledge graph.
Without exception, these three use cases can benefit significantly from this architecture:
Complex Manufacturing Systems
Creating an ontological model of a complex system isn’t a straightforward exercise, and mapping data onto that model can be even trickier. The Anzo Graph Data Interface (GDI) can do a parallel load directly from enterprise data sources to RDF, and then data layers can perform in-memory transformations. This means that regardless of the shape, complexity, or dirtiness of source data, it can be loaded sight unseen, then restructured and cleansed using powerful transformation queries. This stands in stark contrast to the approach of label property graph (LPG) databases which will require pre-defining the schema and creating pipelines that extract and transform source content in a specific format.
Semantic reasoning itself is a computationally expensive exercise, and generally requires several iterations to perfect. Anzo’s data layers streamline the exercise of testing different permutations of modeling and connecting the knowledge graph, which can accelerate the generating of one of the most important inputs to AI. Furthermore, Anzo offers a set of out of the box algorithms and a framework where data scientists can implement user defined extensions (UDXs) that leverage the MPP performance of the underlying graph engine.
The nature of questions enterprise leaders generally ask are often complex, as in computationally and logistically expensive. In the example query I posted above, all sensor values for all the tubes in the system need to be scanned, and then the common components determined. The execution of that query benefits significantly from Anzo’s distributed OLAP architecture. Unlike data virtualization approaches that require managing network load and caching, Anzo creates a single in-memory database layer. Application end users can receive punctual answers and downstream operational systems aren’t impacted by the stress of complex questions.
If you have thoughts on additional use cases or would just like to have an interesting discussion on the possibilities of knowledge graph please message me at email@example.com or visit our suite of resources.
“Digital Manufacturing Platforms in Industry 4.0” https://www.assemblymag.com/articles/97286-digital-manufacturing-platforms-in-industry-40
“Towards Knowledgeable AI: Agents that See, Speak, Act, and Know” https://www.ml.cmu.edu/research/phd-dissertation-pdfs/kdmarino_phd_machinelearning_2021.pdf
“Amazon Testing Robots to Transport Oversized Items in Fulfillment Centers” https://www.therobotreport.com/amazon-testing-robots-to-transport-oversized-items-in-fulfillment-centers/
“SGPT: A Generative Approach for SPARQL Query Generation From Natural Language Questions” https://jens-lehmann.org/files/2022/ieee_access_sgpt.pdf
“Bosch Semantic Stack” https://www.bosch-connected-industry.com/de/en/portfolio/bosch-semantic-stack