How NLP is overcoming the document bottleneck in digital threads

Enterprises are increasingly adopting new digital technologies to help streamline and automate the design and manufacturing of physical products. Digital twins help organize and share much of this technical data. Digital threads connect changes to this data across a product or processes lifecycle. 

However, some of the most crucial manufacturing data is managed as PDF documents and handwritten notes, also known as “unstructured data.” At the Digital Twin Summit, executives from XSB, an industrial artificial intelligence (AI) company, explained how natural language processing (NLP) techniques are bridging the gap between text documents, digital twins and digital threads. 

“We are using an ensemble of artificial intelligence technologies, to basically read the document to extract the information and to infer additional information,” said Rupert Hopkins, CEO of XSB. 

For example, a document may reference a material specification only used in aluminum casting workflows. The extracted data from these documents could also help plan out lead time, new work instructions, or supply chain requirements.

An old-fashioned process

Andrew Bank, strategy and business development manager at XSB, said that manufacturers are increasingly digitizing product information, accounting data and customer data. But other critical engineering data ends up in PDF files, including product specs, drawings and work instructions. These documents include critical data like tables, graphs, equations and references to other documents. 

“A static document is a terrible container for data,” said Banks. 

There are often dozens of explicit and implicit references between documents — which creates a bottleneck every time an engineer needs to figure out how to respond to a new customer order or supply chain constraint. A single line specifying a new requirement might mean hopping through ten documents to figure out what will be required to tune machine settings, procure new raw materials and adjust work processes. 

Engineers today typically copy and paste or re-key data from one document or system to another. A minor adjustment on the factory floor or the supply chain can propagate across all the interrelated requirements. 

“If there’s no dynamic live link between those derivatives downstream and the authoritative source data change, management and impact assessment become almost impossible and very difficult at the least,” Bank said. 

Creating a data graph

Graph databases are an ingredient of digital twins, since they provide a way to help connect context across different data sources and for other users. XSB has developed tools that use AI and semantic ontologies to transform a static collection of data into digital models represented using a graph structure. Banks claimed that some manufacturing-intensive organizations can save up to 65% when they move from static documents to these digital models. 

The data is stored in the Swiss Knowledge Graph, which attributes meaning and context to each piece of data. This allows teams to pull  this data into the product lifecycle management (PLM), manufacturing execution systems (MES) and Microsoft Office. 

PLM records often refer to industry standards stored in other systems or attached as a file attachment. XSB has developed plugins for PLM tools like PTC Windchill, and Siemens Teamcenter that bring metadata from the document into the PLM interface. 

XSB claims one customer is using the tool to automate a process for approving configuration changes from within SharePoint. It says others are using it to ensure work instructions are automatically updated in response to new customer requests. Additionally, XSB says another component testing company uses the tool to pull test requirements out of technical data packages to ensure labs test the correction materials. 

Connecting the thread

Pulling requirements out of complex engineering documents can be tedious. Engineering requirements are often expressed using “shall” statements. A statement such as, “The part shall be made of brass, with a finish of desert sand and a size of AA,” actually specifies three different requirements that all need to be analyzed and accounted for separately. 

Hopkins said most humans only demonstrate 75% accuracy at pulling all the requirements out of a complex document. He claims that XSB achieves 80-90% accuracy on loosely trained systems and 95-100% accuracy when combined with statistical process controls. 

Digital models help inform a digital thread about the relationships between the regulations, materials, processes and specifications associated with a product. This can help to plan more efficiently around changes to work instructions or material regulations. 

“If all of a sudden cadmium is being pulled from the supply chain, we need to alert the twin to that kind of change and those changes today are mostly document-driven,” Hopkins said. This mirrors how companies like John Snow Labs use NLP to digitize medical records in healthcare. More importantly, it illustrates how NLP technology could play a valuable role in creating digital twins and digital threads in other industries such as construction, power, telecommunications and logistics.