Downer Group specialises in urban services to create and sustain the modern built environment.

When it comes to managing risk, there’s a lot to be said for the vision of sci-fi movies like the investigation wall in the 2002 film Minority Report, the bridge computer in Star Trek and Iron Man’s J.A.R.V.I.S. system in the Marvel universe. Systems that collect and analyse evidence at a massive scale to predict potential futures before they occur (without the pre-cognitive humans in a pool of goo, of course), and bridge computers that can estimate the likelihood of success across millions of potential futures in seconds, highlight the possibilities of predictive technology and real-time data analysis.

The art of the possible and the science of current technological advances are inspiring a leading integrated service provider in Australia and New Zealand to explore how artificial intelligence (AI) could help keep its frontline workers safe.

Downer Group specialises in urban services to create and sustain the modern built environment. This work ranges from designing, building, and sustaining assets, infrastructure, and facilities for a range of sectors, including transport, energy and telecommunications, to providing essential services in hospitals, aged care facilities and educational institutions. The company employs around 44,000 people across more Australia and New Zealand and operates according to the deeply held value of Zero Harm, with safety the first of its four strategic pillars.

Downer has embedded a critical risk management program and strong risk identification, management, and reporting processes across its operations. At the core of this program are the company’s critical controls – the things that make a significant difference in protecting people, the environment, and the community, and that must work when needed. Put simply, Downer wants to make sure it focuses on the right things and that:

• they know what their major risks are and what causes them
• they know how to manage major risks
• they do what is needed to manage these risks
• they ensure what they do to manage these risks is working.

In 2021, Downer Group partnered with Microsoft and IT services specialist Crayon to start building an AI-powered health, safety, and environment management system. Together, they reviewed Downer’s roadmap for embedding data-driven decision-making into frontline and strategic Zero Harm management processes and systems, with the goal of creating a solution that helps prevent problems before they happen.

“To achieve our goal, we need to know what to do to make sure we go home from work safe and healthy every day, and protect our environment and communities, and we need to do it in practice,” says Dr Matthew Hancock, General Manager of Zero Harm Risk at Downer and business lead for the project.

It’s simple to say, but complex to achieve.

Critical risk management programs rely on detailed analysis of how to control potential causal pathways for each type of risk that could credibly lead to a fatality. This is often done using a technique called bowtie analysis. Teams across each business unit at Downer Group had developed more than 300 detailed bowtie analysis covering approximately 70 different themes of risk, outlining more than 15,000 controls in the process.

While this analysis helped each business unit focus on their critical controls in certain scenarios, it also posed a number of challenges for Downer. In order to reduce duplication and improve consistency, all of this information needed to be analysed and standardised without losing sight of the specific requirements and language of each part of the business.

On top of this, the company needed to make sure that corporate standards, frontline procedures, regulations, training, lessons learnt, corrective actions and monitoring activities were all aligned, and that updates could be efficiently propagated throughout the system as things changed and lessons were learnt. It also needed to consider the broader business requirements and rapidly evolving regulatory environment when designing these Zero Harm processes – managing health, safety and the environment is not a one-size-fits-all kind of problem. To stay safe and profitable, you need the right people to do the right work at the right time.

To ensure that all these processes work on a day-to-day basis, Dr Hancock says they must be useful and user-friendly for managers and frontline workers.

“Our vision is that the work we do will, over time, feed into frontline decision support systems that help our teams decide on the best way to do their work when they need to do it and give them up-to-date best practice and consolidated lessons learnt on how to do it well,” he explains.

“We learn an incredible amount every year, and smarter systems will help us systematise these lessons and get them to the people who need them when they need them.

“We also want it to be really easy to keep track of what is and isn’t happening, and to check if what is being done is working. If we can learn which things work well in an easy and unobtrusive way and share this efficiently with others doing the same work in real time, we could see amazing benefits in Zero Harm performance and productivity.”

Applying machine learning to risk management

Downer Group has worked with Crayon on a Microsoft-based AI-infused platform to connect information more effectively. In the project’s first phase, Downer and Crayon worked together to discover how they could use machine learning algorithms to reveal and interpret controls from the company’s documentation. The first iteration of the platform was developed in just five weeks, which helped prove the approach and build relationships.

The proof of value focused on enhancing existing work by Downer to integrate the cause-and-effect bowtie analyses and associated controls and drew on compiled documents detailing how risk is managed at the company – from group-level standards and industry codes of practice to local-level safe work method statements.

Dr Bruno Beltran, a data scientist with Crayon who has been working with Dr Hancock, says: “There are two things which Downer tracks very carefully, and those are all of the risks that their workers are exposed to, and the controls that are supposed to be mitigating those risks, or even ideally, helping to eliminate the exposure to those risks of the workers at the front line.

“Whenever a user is logging information about the day’s work or whenever we’re looking retrospectively at incident reports, it could be incredibly useful to have a natural language processing model in Azure Machine Learning (ML) that can take those input documents, whether that be real-time logs or historical documents, and automatically identify from the very detailed lists of risks and controls that Downer has built up over the years, using all of their subject matter expertise, which ones are available or are present in these documents.

We might see that a worker will be using scaffolding to put up special wall material on a given day. That would, of course, expose them to certain risks, including working at height and different kinds of equipment, and one thing that we could do is have the natural language processing model extract what kinds of controls their manager is going to be putting in place like having proper barriers, not having people walking underneath and so forth.”

By encoding the business logic to map the territory between the tasks, risks and controls that should be in place, Downer and Crayon hope to create a tool that ensures frontline workers have the right information at their fingertips to do their job safely.

The blueprint for the platform calls for Downer’s risk and controls data, stored in an Azure Data Lake, to be analysed using a machine learning model developed in Azure ML. An application programming interface (API) can then make that output accessible to user-facing interfaces as required.

At present, the Downer team has integrated the APIs into a Visio-based data mapping tool and several VBA-enhanced Excel spreadsheets used for prototyping user interfaces and workflows. This includes functionality like splitting documents into sentences, classifying each sentence, and then efficiently reviewing and improving classifications using semantic ‘next best’ suggestions and search functionality, and annotating documents sentence by sentence with hyperlinks to a control library. Simple tools such as these make training supervised learning models easy for subject matter experts, and immediately benefit knowledge workers who are preparing documentation by allowing them to access best practice relevant to their current work with the click of a button from their Word document.

The company is also looking to adopt more sustainable and user-friendly tools for frontline and office-based workers. Dr Hancock says deploying Microsoft Power Apps integrated with SharePoint would be a natural fit for Downer, given its team members regularly use SharePoint and Microsoft Teams.

“The ability to easily connect the strong back-end and developer tools with teams through SharePoint and Power Apps has also got us pretty excited,” he says.

Five-week sprint proves process

The first pass of Downer’s AI-infused health and safety management platform was developed in just five weeks using an agile methodology. Crayon also created machine learning operations that would train the model on new data as it arrived, and the platform is already classifying controls into a pre-defined framework with 80 per cent accuracy.

Downer’s internal machine learning team is now using this framework and approach to design its own models. Dr Hancock says that transfer of technology and know-how was critical.

This approach also allows Downer to focus on the specific needs and fast-track the outcomes of the communities of practice that it has established around risks.

“We have communities of practice – groups of subject matter experts – across all the big types of risks. For example, working at heights, working with electricity, driving vehicles, fires and explosions, and hazardous chemicals,” Dr Hancock says.

“These groups of people are articulating in detail the types of risks we have to manage across the company within their subject matter area, verifying all the different types of controls that could be put in place and working through which control strategies are most appropriate for each type of risk.

“We are then going to start connecting all our documentation into this framework, starting with our frontline safe work method statements.”

Reaping the benefits and looking to the future

As more types of information and more systems are connected to Downer’s knowledge framework, Dr Hancock expects the safety and productivity benefits to grow exponentially.

“There is so much value-adding work we are currently unable to do because of the effort required to do the basics,” he says.

“If we can significantly reduce the effort required in some areas, then we can put more effort into doing intelligent and thoughtful work which, in turn, improves business performance and drives down injuries and fatalities.”

One of the projects Downer is considering is integrating its machine learning platform with the company’s incident management system.

When an investigator is entering information into the system to report an incident, the machine learning platform will analyse what they are writing about and generate prompts and create structures to increase the efficiency and accuracy of the investigation. The incident reports will then be fed back into the Azure Data Lake, where they can be analysed for additional safety insights, creating a virtuous cycle of information collection and procedure optimisation.

Another example is to augment regulatory compliance activities. “We need to keep up to date on regulatory compliance, so we carry out regulatory mapping and track regulatory developments,” Dr Hancock explains. “There’s a whole bunch of work done to map regulations for what we do and make sure we understand and respond to the changes.

“If we can connect all our controls with relevant regulation and tie this into our frontline procedures with machine learning, that’s going to cut down the work required to make sure we are compliant to a fraction of the time it takes today.”

Dr Hancock is also keen to explore how AI could be used to create chatbots that increase the productivity of frontline workers and help ensure they go home safely at the end of each day.

“We’re looking at connecting the work we do with risks and controls so that if you say, ‘I’m digging a trench to lay a pipe’, it’s going to say, ‘Here are the different risks that you are likely to face and these are the critical controls that will help you manage those risks’,” he explains.

Looking forward, Dr Hancock is excited about advances like Microsoft’s Megatron-Turing Natural Language Generation model and how they could help the company rapidly integrate lessons learnt into practice at scale.

“In the medium term, I would love to see something like a super-smart Cortana for health, safety and environmental management, maybe with a bit of machine-generated, Wikipedia-style best practice thrown in there as well” he says.

“Our longer-term vision is to integrate next-level predictive capacity to further tailor and strengthen the decision support advice until you get to the point where you have a Minority Report / Star Trek computer kind of thing, where we’re systematically identifying where harm is becoming increasingly likely, and then taking appropriate and timely action to stop it before it happens.”

Dr Hancock acknowledges that there is a long journey ahead to achieve this vision, but he is confident that it is a journey worth taking.

“This isn’t something anyone can do on their own,” he says.

Rhonda Robati, Senior Vice President for the Asia Pacific at Crayon, says: “This is a great example of collaboration between Crayon, Microsoft and utilisation of our global resources in delivering expertise and tangible results to our customer, Downer.”

Dr Hancock says the company wants to get to the point where it can model complex systems and unknowns, and predict potential futures based on what it knows and what it sees happening in the world.

“If we can efficiently consider complex sets of events with material consequences at scale – circumstances we haven’t seen before – then we can start to identify when the holes in the Swiss cheese are lining up for a big event in the real world, and act to prevent disaster,” he says.

“And if we can manage complex risks with confidence, then we can take on the more challenging and transformational work, and ultimately improve the lives and wellbeing of more people. “

Tags: AIMicrosoftTechnology