AI on the high seas
[NodeJS] [ReactNative] [Python] [Docker] [FastText]
Thanks to AI and excellent teams, our operational planners can now access a wealth of unused information. The result? Even better and safer operations offshore.
Nothing is left to chance when planning operations out to sea. Previously, our planners had to access several different systems to find all the relevant information - quite the time-consuming task.
But now they can hop into the Operational Planning Tool (OPT) dashboard to find all the information quickly. Additionally, a natural language processing (NLP) component is bringing previously hard to find information out in the open.
The AI tool digs into several data sources integrated in OPT, one of which are incident reports. Whenever an incident occurs on one of our assets, be it large or small, a case is created. This details the what, where and why of the event.
While the main purpose of these case reports is to prevent these incidents from happening again, we are now able to use them to provide our operational planners with a wealth of new knowledge.
“The business challenge has been to help our planners, discipline leaders and engineers improve risk management by providing a tool that links planned activities to lessons learned from the past," Jennifer Sampson, DCoE Knowledge AI team lead, says.
Want to stay updated on Loop?
Breaking new ground in the cloud
In order to understand how AI is working its magic we have to talk about software development first. Thanks to a team of 3-6 developers, UX and IT Business Analysts lead by Thomas Gudmestad and Lars Egil Obrestad, our natural language processing component has a huge data source to access and learn from in the cloud.
“OPT was the second project to go live on our Azure cloud platform, so we had to create our own path in many ways. We spent a lot of time working on data and getting all the different datasets to work together. Basically, we’ve all been learning while working."
Thomas Gudmestad
Two important aspects in their work was creating a scalable solution and an intuitive user interface - avoiding clutter from all the available data sources.
Like all software development teams they had an agile way of working, giving them a great deal of flexibility. The first MVP (minimum viable product) of OPT was based on Microsoft’s Power BI, but the later versions have been delivered as a web app.
“This meant we could start with a small proof of concept and improve from there,” Thomas explains.
Now that we’ve wrapped our heads around OPT, it’s time to head into the details. The AI component that’s helping our planners find relevant information to the task they’re planning is a machine learning network that runs in the background. Our Digital Center of Excellence (DCoE) Knowledge AI team are the ones that made it all happen.
“There are a lot of very experienced personnel who work on each asset, and they know just about all of the major incidents by heart. But they can’t possibly remember across all assets on the Norwegian continental shelf. Helping them do this is a key part of the work.,” Jennifer tells us.
But where is the learning component getting all this information from and how does it all work? Let’s head behind the code to take a closer look.
Natural language processing 101
1: It all begins with feeding the network data, both unstructured and structured.
2a: Pattern recognition looks for a specific pattern in the text, i.e a unique equipment ID tag.
2b: Entity extraction kicks in and looks for certain words, i.e “valve”, that could indicate equipment. Then it looks at the words surrounding “valve” to place it into context and determine if it’s a piece of equipment.
3: The results are fed into the knowledge graph.
The ABCs of NLP
While there are plenty of natural language processing tools (NLP) available for English, it took quite a while to find one for Norwegian. But eventually the team managed to find a component that could handle it letting us access this previously untapped well of knowledge, Eivind Sjaastad explains.
"We have tons of written reports that we're not capturing lessons learned from simply because it’s too time consuming to go through. If we can get all of this up and running on a large scale there are a lot of opportunities available."
Eivind Sjaastad
His teammate Claire Birnie tells us that since there were so few tools available for Norwegian they had to build one nearly from the ground up. They faced quite a few challenges, but trying to tackle those was really fun:
“We’re doing something for production, but at the same time we’re also in research and trying to find new interesting avenues."
Screenshots from the Learning Tab.
Highlighting help
When you’re teaching a computer to understand written language the context certain words are in is a key factor. So how did they do just that? By bringing in dedicated, skilled users from Oseberg South to do some digital highlighting.
“You need a certain level of knowledge to be able to tell what’s a task and what’s equipment in these reports. So we invited them to a couple of workshops where they basically spent hours simply highlighting what’s what in different reports,” Eivind says.
The team’s neural network wasn’t the only one that was struggling with understanding Norwegian. Claire, hailing from Scotland, found herself learning the language alongside the network she was training:
The computer wasn't the only one learning Norwegian - it had great company in Claire. (Video: Torstein Lund Eik)
Having taught the computer to read, one key component still remains to be explained: the mighty knowledge graph. It’s what makes the network able to connect the dots and bring up the results it thinks is relevant. Google uses a similar technology, and that’s what powers the field that can bring up images, quotes, movie appearances and a Wikipedia snippet when you do a search for i.e. ‘Albert Einstein’ in Google.
The allmighty knowledge graph
If our planners are replacing a component for a turbine on Oseberg South and input that into OPT, a knowledge graph like this one kicks in and does a search. It looks for relevant information found in work orders or incident reports across several plants and connects the dots.
Taking it a step further
When the learning tab brings up relevant reports or previous events to the planner, they can give the results a thumbs up or down, Jennifer explains.
“This lets them give us feedback on if the information was relevant or not. Then, we can feed that into the model and build off that to let it continually build upon itself.”
While this feature is still maturing and not showing as many results to rate as relevant or not as hoped, Ronny Larsen, the OPT development project leader, thinks it will become a great asset in the future.
"We're in it for the long haul and off to a great start. The skill sets and enthusiasm of the Knowledge AI team has been crucial in the success so far and will make sure we keep going forward."
Ronny Larsen
One of the team’s goals for the learning component was that it could easily be transferable to other projects or teams. They’re well on their way and are currently working on a similar tool for our onshore operations in the U.S.
It’ll be exciting to see what the future holds for NLP, whether it’s on land in the U.S. or out at sea here in Norway. Make sure you stay in the Loop by subscribing - that way all the software development stories will come flying into your inbox hassle-free.
People
Jennifer Sampson
Bjarte Johansen
Claire Birnie
Lars Egil Obrestad
Eivind Sjaastad
Ronny Larsen
With contributors from Tessella and Bouvet.