>Business >Safety comes first Artificial Intelligence for autonomous information centre cooling and industrial control

Safety comes first Artificial Intelligence for autonomous information centre cooling and industrial control

Several of society’s most dire issues have grown gradually more complicated, so the search for answers can appear overwhelming. DeepMind and Google, the focal point of this blog post by AICoreSpot, they hold the belief that if they can leverage artificial intelligence as a utility to find out new knowledge, solutions will be simpler to reach. 

5 years ago, in 2016, DeepMind teamed up and produced AI-driven recommendation systems to enhance the energy efficiency of Google’s already very optimized data centres. The line of thought was not very complex: even minor enhancements would furnish considerable energy savings and minimize CO2 emissions to assist in the war against climate change. 

Presently, this system is being taken to the next level: rather than human-driven recommendations, the AI framework is directly controlling data centre cooling, while staying under the specialist supervision of our data centre operators. This forerunner, first-of-its-kind cloud-based control system is now securely providing energy savings in several Google Data Centres. 

How it function 

Each five minutes, the cloud-driven artificial intelligence takes a snapshot of the data centre cooling system from several sensors and inputs it into the deep neural networks, which forecast how different combinations of possible actions will impact future energy consumption. The AI framework then recognizes which actions and behaviours will reduce the energy consumption while fulfilling a robust grouping of safety constraints. These actions are transmitted back to the data centre, where the actions receive verification by the local control system and then have implementation. 

The concept had its evolution from feedback from the data centre operators who had been leveraging the AI recommendation system. They informed us that even though the system had instructed them some new best practices – like spreading the cooling load across more, instead of less, equipment-implementing the recommendations needed too much operator effort and supervision. Naturally, they wished to now if we could accomplish similar energy savings with no manual implementation. 

Thankfully, the answer was yes. 

Developed for reliability and safety 

Google’s data centres consist of thousands upon thousands of servers that drive famous services which includes Google Search, Gmail, and Youtube. Making sure that they run dependably and effectively is a mission-critical matter. The AI agents have been developed with the underlying control infrastructure from scratch with reliability and safety as chief concerns, and leverage eight differing mechanisms to make sure the system will behave as expected at all times. 

One simple strategy that has had implementation is estimation of uncertainty. For each possible action – and there are billions possible – or artificial intelligence agent quantifies its confidence that this is a good behaviour. Actions with reduced confidence are eradicated from consideration. 

Another strategy is two-layered verification. Optimal behaviour computed by the Artificial Intelligence are put up against an internal grouping of safety limitations defined by the data centre operators. After the instructions are transmitted from the cloud to the actual data centre, the local control system authenticates the instructions against its own grouping of constraints. This redundant check makes sure that the system stays within range of local constraints and operators maintain complete control of the operational boundaries. 

Most critically, the data centre operators are constantly in control and can opt to exit AI control mode at any point. In these situations, the control system will transfer seamlessly from AI control  to the on-site rules and heuristics that define the automation industry currently. 

While the original recommendation system had operators vetting and involved in the implementation of actions, the new AI control system goes into direct implementation of the actions. It was purposefully constrained (the system’s optimization boundaries to a narrower operating regime to prioritize safety ad dependability, implying there is a risk/reward trade off with regards to energy reductions. 

Regardless of being in place for only a matter of a few months, the system is already providing consistent energy savings of approximately 30% on average, with further predicted enhancements. That’s due to the fact that these systems improve over the course of time with increased amount of information. Our optimisation boundaries will also be expanded as the technology experiences maturity, for even bigger reductions. 

Our direct artificial intelligence control system is identifying yet more unique ways to handle cooling that have shocked even the data centre operators. Dan Fuenffiner, one of Google’s data centre operators who has functioned extensively along with the system, commented: “It was amazing to see the AI learn to reap the benefits of winter scenario, and generate colder than normal weather, which minimizes the energy needed for cooling within the data centre. Rules don’t improve over time, but artificial intelligence does. 

It is exciting times that the direct AI control system is functioning safely and reliably, whereas consistently providing energy savings. But, data centres are just the start. Over the longer-term, these possibility to apply this tech in other industrial situations, and assist in managing climate change on an even large scale. 

Add Comment