Using AI to predict accidents on construction sites

It's a familiar scenario in the aftermath of an accident: workers and supervisors recognise that there were early indicators of danger, which went unnoticed amidst the daily tasks. This realisation, often summarised as "the writing is on the wall" reveals a significant gap in safety management—the inability to spot and act on these early warning signs in a timely manner.

Construction sites are characterised by abundant, complex data that are challenging to interpret, making them ideal candidates for AI applications. However, a known obstacle until recently has been the non-digitised nature of such data within the construction industry. At Novade, we have been collaborating with Tiong Seng Contractors for several years to digitise their processes, thereby creating a rich repository of information. And in 2022, we started to explore the potential of using AI to predict accidents together. 

Our hard work has led to an exciting result: a tool that can help us anticipate and mitigate risks before they escalate. 

An AI capable of predicting 76% of accidents 

What does that actually mean? 
It means that our artificial intelligence (AI) model can predict, with over 70% accuracy, whether at least one reportable accident is likely to occur in the next month at a specific construction site. 

In data science terms: 

• Our model has a recall rate of 76%, meaning it can predict 76% of all months with accidents. It does not catch every incident, but it does successfully pinpoint the majority before they occur. The model tends to miss non-severe accidents – those that do not result in any lost man-days. 

• The precision rate of the model is 88%. This indicates that when our model predicts an accident, there is an 88% probability that the prediction will be accurate. This precision is crucial because predicting accidents with high frequency could lead to unnecessary alarms. It is essential to maintain a balance, only signalling high-risk situations to avoid the issue of "crying wolf." 

This balance between recall and precision illustrates our model's ability to be both precise and reliable in operational settings. 
Out of all months with accidents, the AI correctly predicts 76%. The rest are not detected by the AI. When the AI predicts an accident, it is correct 88% of the time. The rest are false alarms. 

Why is it important? 
These capabilities mark a significant advancement in construction safety, allowing for the proactive anticipation and prevention of many accidents. No prediction system is perfect, but our model serves as a reliable aid for safety managers to proactively address potential dangers. In other terms, the model helps with human oversight, to read the “writing on the walls” and alert teams early enough so that lives can be saved. This is a major shift from a traditionally reactive stance to a proactive safety ethos. 

A machine learning model to see through patterns in the field 

To demystify how our AI works, it is crucial to understand the processes behind its predictive capabilities. The AI operates on a machine learning model. Initially, the model is 'trained' on a dataset which encompasses years of detailed records from construction sites. This training phase allows the model to identify patterns and correlations that historically led to accidents.

Following the training, the model is 'tested' with a different set of data, including two years’ worth of site accident reports. This test is crucial; it evaluates the model's ability to predict accidents in the field based on its training. It functions as a sophisticated 'black box,' looking through patterns to find those predictive of accidents. These patterns are not always obvious; they are often a composite of many subtle indicators. 

A dataset made of more than 80 indicators 

To train the model, we used data collected from nine construction sites over a period of 22 months, including:

• Safety compliance processes: These are mandatory, daily procedures on site, such as permit to work (PTW) systems and toolbox meetings.

• Inspections & observations: This category includes regular checks of equipment and site conditions to ensure they meet safety standards, along with insights regarding safety behaviours and culture. 

• Reporting: This involves the systematic documentation of incidents and near-misses.

• Daily manhours and overtime: The number of worker hours spent on site each day, which can be an indicator of workload and potential fatigue. 

By consolidating all this information and configuring it into a dataset suitable for machine learning, we gained a comprehensive view of site safety. Ultimately, our dataset included over 80 indicators: 

• Some were straightforward data counts, such as the number of high-risk non-conformities. 

• Others tracked trends over time, such as whether accident rates have been increasing or decreasing at a specific site. 

• Some indicators focused on user behaviour, for example, the speed at which a user approves a permit to work or rectifies a non-conformity. 

This is the power of AI: the ability to analyse over 80 indicators and discern patterns and anomalies, to identify what matters and what creates the risk. 

Once we started training and testing our models, it emerged that only a dozen of these indicators had significant impact on the predictions. Tiong Seng Contractors found that the critical indicators identified by the model aligned with their on-site team's intuition. What's particularly valuable about the model is its ability to prioritise and order these indicators, determining which ones are the most influential. This revelation also highlights that advancements in safety can begin even without the application of AI; they start with the digitisation of site processes and the systematic monitoring of data. 

AI is only part of a safety digitalisation journey 

The accuracy of an AI model is directly tied to the quality and quantity of its training data, something notoriously difficult to achieve on a construction site. Acquiring this data is the first and most challenging step in the digital journey, particularly in the construction industry where many companies have yet to digitalise their field processes. 

Digitalising processes in the field involves extensive work: 

• Identifying and defining key processes 
• Digitalising and optimising them 
• Training users 
• Increasing adoption 

Like any change management, this approach is time-intensive. For Tiong Seng Contractors, it took years of partnership with Novade to create the dataset that would later be instrumental in training the AI. It also required effort and consistency, especially in monitoring site adoption. In such cases, it's crucial to select a tool that can help track both the volume and quality of data, ensuring your digitalisation project stays on course, until you gain enough confidence in your data to train an AI tool. 

Some might worry about the cost and efforts required by such an initiative, but there are many benefits that will emerge early on: 

• Digitalising site processes often leads to increased efficiency, with a return on investment that can be quite substantial. 

• The advantages of a data-driven approach can be realised well before accumulating enough data for AI use. These include improved on-site communication, streamlined and immediate reporting, the ability to manage projects and performance using shared metrics, and employing descriptive analytics to identify straightforward patterns. You can also benchmark teams, contractors, and projects, and pinpoint common root causes. 

Bottom line: Once you measure it, you can improve it! Initiating the collection of safety data is a positive step towards improving safety on site. 

Why Tiong Seng Contractors embarked on this journey 

As part of Tiong Seng’s efforts to push its digitisation to better its EHS performance at its worksites, the company began a strong collaboration with Novade in exploring how its past data could be transformed from mere retrospective information into trends which would provide actionable intelligence used by the site management to better their safety standards. 

“Having access to this kind of analysis is incredibly cost-effective. It allows us to take a more objective look at our projects, and Novade has made the whole process much easier. We're now able to see things from a different perspective, which is helping us to make better decisions. It's like having a new pair of glasses that helps us see our projects more clearly.”
  - Kevin Seet, Head of Corporate Division, Tiong Seng Contractors 

AI helps human decision but does not replace human judgment 

The advancement of AI is as promising as it is critical, necessitating responsible use and discernment. Our model, despite predicting 70% of accidents and identifying high-risk sites, is not foolproof — it does not account for the remaining 30%. A low-risk prediction by the AI doesn't guarantee the absence of accidents; risk, by nature, holds the possibility of the unforeseen. 

This model is intended as a strategic tool for prioritisation and alerts, not as an all-seeing solution. Caution should be exercised towards any claims of absolute foresight. 

Yet, when used appropriately — as a tool — we stand on the cusp of a new era. An era where AI not only enhances safety but also contributes to saving lives, heralding a significant leap forward in the construction industry. 

“We were sitting on a gold mine of data, but we needed help consolidating it in a systematic and meaningful way to make sense of the data! With the help of Novade, the data collection process is seamless and now, with the data from all our projects over the years, we can use it to proactively identify and prevent risks. This is giving us a wealth of insights. It's a game-changer for safety in our industry.”
  - Kevin Seet, Head of Corporate Division, Tiong Seng Contractors

About Hélène Menthon 
Hélène, a seasoned construction and technology professional with over 7 years of experience spanning Europe and Asia, currently serves as the Senior Manager for Data Analytics at Novade. She is a passionate advocate for leveraging data-driven insights to revolutionise the construction industry, focusing on enhancing safety, quality, and overall performance.