(32aq) Gen AI Application in Event Categorization

Over seventy years ago, the idea of “a machine that thinks” was already taking shape. In 1950, Alan Turing, famous for having cracked the German ENIGMA code during the Second World War and often referred to as the “father of computer science”, implemented the “Turing Test” in which a human tried to distinguish whether the answers to a question were given by a computer text or by another human being.

A few years later, the term “Artificial Intelligence” emerged, coined by John McCarthy, and has been gaining more and more ground within companies, since several benefits have been identified, such as: optimizing customer service, convenience and scalability, increasing the automation process, reducing errors, risks and operating costs and improving decision-making.

In this sense, Vale - a Brazilian multinational mining company and one of the largest mining companies in the world, with a staff of approximately 185,000 people, including its own professionals and permanent contractors - deals with various fronts and initiatives focused on data and has a specific corporate area whose main objective is to capture and transform Health, Safety and Environment (HSE) information into strategic assets through a data-oriented culture.

To improve its results, the company has been using Artificial Intelligence (AI) to add even more value to its data processes and is already beginning to see significant impacts that make it possible to generate relevant insights in an intelligent and effective way, which will help make work safer.

One of the first initiatives implemented with the use of Artificial Intelligence in Vale's HSE area, aligned with the objective of having more reliable, traceable and quality data, improving the targeting of the analyses conducted by the Data Intelligence area, was related to the recording of events that generated injuries with first aid, material losses or unsafe conditions. The recording of these events has been motivated by the leadership as they are considered fundamental for the proactive management of operations and people’s safety, contributing to the prevention of fatalities.

However, due to the high volume of events reported (around 1,000 per week), it has become unfeasible to carry out the manual sanitation process - an estimated average of 6 hours per day - to guarantee the quality and reliability of the information in accordance with the concept defined by the company. Therefore, in order to guarantee increased accuracy and assertiveness in the database of events of this nature, Generative Artificial Intelligence (GenAI) technology was implemented with the aim of confirming the classification of these events.

To apply the GenAI model, Meta Llama2-70B-Chat was chosen - a language model pre-trained on 2 billion tokens with quality equivalent to OpenAI's ChatGPT. To ensure the best accuracy of the model, a few steps were implemented: a volume of almost 100,000 records was worked on and processed so that it could be used for machine learning; based on the first results obtained by the model, the normative focal point worked on the testing and validation process so that the model could be adjusted to provide more assertive responses.

Further tests were carried out - seven in total - until a good accuracy result was achieved, guaranteeing 96% correct classification of the events recorded. Through the use of this technology, combined with the management of the issue via an online visual solution, it has been possible to:

• Increase the quality and reliability of the category of these events, improving the targeting of analysis and decision-making, moving the company towards a Data-Driven culture;
• Significantly reduce the total number of hours spent by the normative focal point;
• Gain agility in the sanitation process through innovation and technology.

The great expectations for the study are to extend the application of this technology to all HSE events to improve the quality of the events recorded, further increasing the model's performance and offering operational areas a robust tool that will help them learn the concept through automatic categorization.