Structuring the Edge in the factory: how to deploy reliable AI as close as possible to the machines?