A fire destroyed the SGB2 data center of OVHcloud in Strasbourg, France, last year. The fire was so severe that Octave Klaba, founder and CEO of French colocation and cloud provider, tweeted about how important it is for customers to activate their Disaster Recovery Plans. While firefighters responded quickly to the scene, they were unable to save the building, although they managed to save adjacent data center facilities.

At a conference organized last year by the European Society for Automatic Alarm Systems (EUSAS), incidents like this were a hot topic. The question was: can AI further improve the fire safety of data centers?

During the conference, Guillermo Rein, Professor of Fire Science at Imperial College London, presented a fire protection system combining building sensors, computer modeling, and artificial intelligence. Known as the Fire Navigator, it is designed to forecast the movement of fire in a large building, providing crucial information about the development of flames and smoke at an early stage. It combines fire safety with building information modeling (BIM) by leveraging the data already produced by sensors such as smoke and heat sensors. A fast and simple wireless system collects all sensor data, and by applying algorithms, it can determine the location, time, spread of flame, and smoke rate of an ignition.

French security and environmental consultant Ibrahim Daoudi issued a warning, however. In his research, he has investigated a number of vulnerabilities associated with artificial intelligence used in security and safety products. According to him, there are basically three categories of vulnerabilities. In the first category, attackers attempt to deceive the model by generating fake data. The second category involves physical attacks. The vulnerability is actually based on opponents, but it is now applied to real objects. The third type of vulnerability comes from traditional attacks on information systems that can lead to the contamination of the model itself or the training data.

In contrast, if used correctly, smoke can be detected in video sequences using temporal information. Andreas Wellhausen, of Bosch Sicherheitssysteme in Germany, presented his work on video smoke detection using temporal approaches based on deep learning. He presented two methods. First, a combination of convolutional neural networks (CNN) and long-short-term memory networks (LSTM), and second, inflated 3D architecture (i3D), which consists of 3D convolutions. These are two state-of-the-art approaches to extracting spatial and temporal information from video sequences. Smoke can be detected and localized within such sequences using cell-wise classification, a new approach. In addition, he demonstrated the advantage of temporal approaches over CNN methods, which are commonly used to solve detection problems in Computer Vision.

While the rapid adoption of AI creates exciting new opportunities, it also poses an important question: do current fire safety regulations apply to AI? Orgalim's Tadas Tuménas, who represents 770.000 tech companies in Europe, discussed how and if this technology should be regulated. His view is that a new EU legislative framework is needed, which should be built around a widely accepted definition of artificial intelligence.