Using Natural Language Processing to improve safety management at Downer Group

Key business challenge

Safety is one of Downer’s four pillars, and Zero Harm – managing health, safety, environmental, and community risks - is embedded in Downer's culture and is fundamental to the company's future success. 

Continuing to understand and manage risks and responding to and learning from incidents plays a fundamental role in Downer’s Zero Harm goal.

Deriving insight from the data collected about health, safety, and environmental activities and incidents is essential to achieving improved performance. Downer has an embedded critical risk management program that focuses on:

• Understanding the risks that matter most and what causes them
• Knowing how best to manage critical risks and focusing on the work that matters most
• Making sure the work to manage critical risks is done
• Making sure what is done to manage critical risks is working
  
   

 

Engagement objectives

Downer Group’s vision is to develop an easy-to-use zero harm management platform that makes it quick and simple for the front line to access the information they need to work safely. Artificial intelligence will be a key part of how they do this.

“We want to make sure it's really straightforward for our people on the ground to know what they need to do when they need to do it, and how to do it well. If we can help people verify that the right work has been done and then check if what they're doing is working using highly intelligent systems it could really change the game,” said Dr. Hancock.

The Downer team’s vision is to:

• Significantly enhance their capability to categorize, analyze and report on their data to enable better prioritization of corrective and preventative action
• Enhance front-line planning and decision-making using data analytics and AI
• Ultimately enable real-time targeted corrective action generation to prevent incidents before they happen

The team at Downer is building a harmonizing framework to map, interpret and integrate thousands of control requirements across their management system – from Group Standards and Procedures to front line Safe Work Method Statements/work instructions.

This work will enable Downer to improve information to the front line and leverage large volumes of information in their data lake to generate insights on coverage, overlaps, and gaps in their documents, systems, and risk analyses.

Thinking about where this could go in the future, Dr. Bruno Beltran, the data scientist at SoftwareOne (formerly Crayon) who led the project with Downer, offered an example. “We might see that a worker will be using scaffolding to put up some kind of special wall material on a given day.

That would, of course, expose them to certain risks, including working at height and with different kinds of equipment. NLP and correlation analysis can assist with identifying the types of controls that are commonly applied in these situations, like having proper barriers, not having people walking underneath, and so forth.”

SoftwareOne helped develop a solution to assist Downer Group’s safety team with this work.

Our solution

Downer asked our team to help them apply natural language processing (NLP), a branch of AI that uses software to comprehend human text or speech, to tagging Downer Group’s safety reports and documents with the safety controls that were mentioned in each one. A machine learning model garners insights from the data lake (Azure) and an application programming interface (API) makes the insights accessible to the users.

“There are two things which Downer tracks very carefully: the risks that their workers are exposed to and the controls that are supposed to be mitigating those risks. Whenever a user is entering information about the day's work, or whenever we're looking retrospectively at incident reports, one task that could be incredibly useful is to have a natural language processing model in Azure Machine Learning take those documents and automatically identify risks and controls,” says Dr. Beltran, the data scientist at Crayon who worked on this project with the Downer Group.

NLP is making streamlining and optimization of the management system practical and manageable for the Downer Group zero harm team. They have set up Communities of Practice to bring together subject matter experts for specific types of safety risk who are using tools built on top of the NLP foundations our team delivered to systematically improve standards, procedures, front-line work practices, and training.

The main initial challenge for building such a tool was the generation of a logical structure and grouping of controls that enabled a wide range of subject matter experts on diverse topics to collaborate in a single shared space. By leveraging state-of-the-art transformer-based natural language models, we enabled Downer to significantly improve their ability to organize unstructured content purely based on “semantic similarity.”

Through the Azure Machine Learning Studio platform, we built a supervised machine learning model leveraging semantic NLP technology that could assist the Downer team in analysis of their documentation, providing prioritized lists of suggestions on how content linked to their control library and allowing them to correct any inaccurate linkages or add new library entries where required. Initial experience showed that the accuracy of suggestions increased rapidly as training progressed, requiring less and less human intervention over time.

This approach allows the business to integrate management system documentation with increasing speed into a harmonized framework leveraging the power of AI.

The results

A model that produced up to 80% accuracy when identifying controls in management system documentation–achieved in only five weeks.

Within our initial five-week sprint, the Downer and SoftwareOne teams worked together to deploy the model’s first iteration. Even within that short period, the iteration has already proven itself capable of identifying semantically similar concepts with minimal overlap in actual words used, allowing the teams to harmonize content even where different terminology was used in each part of the business.

This has saved significant staff time. The model, which currently has an impressive accuracy of around 80% with minimal training, will only improve as more people use it.

The solution is enabling a range of tools being developed in-house by Downer that will begin to transform the way health, safety, and environment are managed across the Group. In the future, Downer Group’s employees may be able to simply talk to an app to receive accurate and relevant best practices for managing risk associated with the task they are about to do, all thanks to an artificial intelligence-infused safety management platform.

A reflection of the project result’s impact is the universally positive feedback we received from both the customer and Microsoft.

“Recognizing that in five weeks, not only building the model but also getting it to the point where we can use it ourselves and create new models with it, is impressive. For me, it's not just fishing for us and giving us a fish, it's actually teaching us to fish in five weeks, which is just phenomenal,” Hancock mused. Ms. Flitcroft reflects that: “The NPL tool has significantly improved our ability to get across large volumes of information, the time saved through this tool cannot be overstated, it’s been game changing”.