Discovering Safety is appealing to industry to get involved in developing a machine-learning tool that will use natural language processing to gain insights from data on Loss of Containment (LoC) incidents. In spite of decades of advances in safety, LoC events continue to lead to avoidable death, injury and economic damage. The ability to learn from LoC is an urgent need if we are to drive down the numbers of people killed or injured at work in process industries such as gas, petroleum or nuclear.
Catastrophic accidents continue to occur across the world, such as the dockside explosion in Beirut. In August 2020, 218 people were killed when a cargo of ammonium nitrate being stored in a hangar exploded.
As Matt Clay, the data engineer in charge of the project, explains: “In spite of advances in occupational safety, these accidents continue to happen and the mountain of data they generate is growing exponentially. We need partners in the process industries to help us develop a tool that will give them insight into the events that lead to LoC incidents. We have the technology to do this, but the knowledge, insight and expertise of industry practitioners will be the key to unlocking its potential.”
Discovering Safety has recently launched a survey that will help gather the information we need to develop a flexible tool based on industry insight. Machine-learning can help digest vast quantities of data, but developing a reliable tool that industry can use with no compromises to accuracy, anonymity or the reputation of their sectors will require human input as well as the latest technology.
“We’re asking for help from right across the process industries, from oil and gas to pharmaceuticals and also the nuclear sector, to help us develop a tool that can have a transformative effect on safety,” says Matt.
The survey will be followed up with detailed interviews with industry partners.
Industry representatives will recall the significant LoC event resulting in the fire at the Buncefield fuel storage facility. Although no-one was killed, the fire led to an estimated £1 billion damage to the UK economy. Loss of containment events continue to occur globally and in the UK and have the potential to cause huge loss of life from a single incident.
One way of gaining insight into why catastrophic accidents occur is to look at smaller, related precursor events that happened before. These could include smaller chemical escapes due to overfilling of a tank, or instances where operating pressures are exceeded but mitigation barriers are effective in preventing a release.
The Discovering Safety tool will be applied not only to data that is gathered when an incident has occurred, but to relevant stages on the journey towards them – such as audit reports that highlight areas of weakness and suggest remedial action.
“In getting to the causes of LoC incidents we’ll be going a step back,” says Matt, “to an earlier, small petrol leak, a tank that was overfilled or where the organisation found, in the course of maintenance, a flaw in a trip which was still working.”
By identifying these early-stage warning signs, and looking at early failures in prevention, we’ll be giving industry the tools to put preventative measures – more proactive maintenance regimes, for example – in place.
Another challenge is that the process industries at the heart of LoC incidents are highly sensitive organisations and the data they produce is highly complex. Matt explains:
“Discovering Safety is already successfully using machine-learning on occupational safety data from the construction industry,” says Matt. “But accidents in the construction environment often fall into common categories such as falls from height. LoC incidents are harder to categorise and more wide-ranging, stemming from very diverse industries from gas processing to pharmaceuticals.”
This additional complexity means that any tool developed to learn from LoC incidents needs to combine human expertise with the processing power of artificial intelligence and its ability to digest vast quantities of free text. Discovering Safety is asking industry to get involved and help us to learn about the issues at stake and develop a tool that addresses them.
The datasets which the project is seeking to use are challenging to extract meaningful insights from, as they stem from a wide variety of industries with different terminologies and ways of working.
“It’s proving very challenging to get insights from a chaotic dataset with no codification, different writing styles and terminologies, and a much larger vocabulary than, for example, the construction industry,” Matt explains. “Training a machine to extract the data from this uncodified mass of data is a huge challenge, there is lots of chaff amongst the wheat – some documents provide lots of insight, and others none.”
Matt’s team has developed a way of pre-screening the data using keyword searches (on a myriad of sector-specific terms such as ‘overpressure’ and ‘valve stuck’) and a system of aggregated scoring. “This allows us to winnow down the mass of documents – eliminating those which contain the terms but nothing else – such as lists of chemicals.”
This is essential as the process of Natural Language Processing (NLP) used in text-mining tools is very processor-hungry, taking seconds to strip one block of text down into parts of speech.
The project will be helped as other Discovering Safety projects develop – in particular the automatic data anonymisation tool being created with Ohalo. This will help overcome one of the main obstacles to sharing data on LoC incidents – the need to maintain commercial and public confidence, which is often required for the industries’ licence to operate.
“We completely understand the reticence of some industries and hope that, by taking part in our secure and confidential survey, they will help us develop a tool that saves lives and economic interests while protecting industries from reputational damage,” says Matt.
By working together to develop a tool that blends the best of industry and sector expertise with the data-digesting power of machine-learning, Discovering Safety will provide a tool that can save lives and businesses. With funding from Lloyd’s Register Foundation, the learning we generate will be shared across the world. Other potential benefits include more information-sharing between diverse sectors of the process industry, in particular the nuclear industry and onshore processing, where the risk profile is diverse and complex.
Related Content