Data Analysis Using AI Can Prevent Accidents in the Chemical Industry

“If this works, it’s a real game-changer, and we can prevent many issues, both in terms of safety and environmental impact,” 

says Gui Hoedemakers, Health, Safety, and Environment (HSE) Manager at AnQore. 

“Society increasingly demands environmentally responsible behavior. As the chemical industry, we must respond appropriately. We are working hard on this, but it’s not something you achieve overnight.”

Fewer Accidents

The chemical industry has experienced numerous accidents over the decades, from the Bhopal disaster in India (1984) to the Cindu explosion in Uithoorn (1992). Even Chemelot, a chemical complex in Limburg, saw a toxic nitric acid cloud escape in 2019—fortunately, without damage or casualties. Due to greater focus on safety (HSE), the number of accidents, including workplace incidents and environmental emissions, has significantly decreased this century.

Last year, the 60 factories at Chemelot recorded only 21 minor incidents. The complex aims to become the safest, most competitive, and most sustainable in Western Europe by 2025. To achieve this, Chemelot is working with Brightsite, a knowledge center involving partners such as TNO, on various projects. One of these projects focuses on significantly improving the safety of chemical processes through computers with machine-learning algorithms.

Preventing Incidents

Such computers can analyze data step by step, learning to recognize patterns and trends and eventually making predictions to prevent incidents. A predictive model is currently being tested at AnQore, formerly part of DSM and Europe’s largest supplier of raw materials like acrylonitrile.

“We often experience incidents, often at unexpected moments, that we want to prevent from recurring,” says Hoedemakers. “For example, leaks or emissions that harm the environment. Often, we can’t prevent them because we lack signals to respond to early enough.”

Hidden Signals

Chemical companies have enormous amounts of data in their systems, containing hidden signals, Hoedemakers explains. “But these signals are currently not being noticed. The information often exceeds human cognitive capacity and becomes unmanageable for operators, who are already dealing with complex systems. So, I asked Brightsite: if humans can’t process this, can machines aggregate such data and become intelligent enough to make predictions and display them on a screen? This way, timely intervention or a time-out can occur.

"This applies to safety but also to environmental concerns, like leaks of highly hazardous substances such as hydrogen cyanide and acrylonitrile. These substances must be handled with extreme care, and we are committed to preventing such incidents.”

First, Learning the Safety Language

Many incidents are the result of a combination of factors, such as delayed maintenance combined with pushing production capacity to its limits, or the introduction of new technologies with unknown risks. With increasing automation and the emergence of autonomous systems and robots, facilities are becoming increasingly opaque to humans.

For the safety pilot, AnQore provided all its historical factory data to train the computer. The first task was to teach it the vocabulary of safety jargon.

Computer Issues Warnings Two Days in Advance

Once the computer understood safety terms, it scanned and analyzed all historical texts and records—from meeting notes to maintenance reports—focusing on periods before incidents occurred. “Surprisingly, the computer detected weak signals and could present them clearly,” says Hoedemakers. “It would alert us: ‘Be careful, something is about to go wrong.’ The computer can look up to two days ahead, identifying weak signals. This means it could alert us two days before a potential incident, allowing us to act.

“With data analysis, 50% of past incidents could have been prevented.”

Analyzing Numbers as Well

The goal is for the computer to eventually issue live warnings by connecting it to all company systems. However, there are still hurdles to overcome. Hoedemakers wants to take it further by having the computer analyze not only words but also numbers.

“More data will make the system more effective. We want to go live only once we can make cross-references between numbers and words. I expect we’ll achieve this step in two years.”

Available to All

This early warning model could then be deployed throughout the chemical industry to prevent incidents. “We’re collaborating with partners interested in this, both within and outside Chemelot,” says Hoedemakers. “Safety should benefit everyone, so this system should be accessible to all.”

This article originally appeared on Change Inc. as part of a campaign in collaboration with Green Chemistry, New Economy, and journalist André Oerlemans. Read the full series on this site or at Change Inc.