This sounds ambitious.
>> After annotating the dataset, I will train a machine learning model to detect the block diagrams.
This sounds like it should come first? This feels more approachable because, as Shaun said, there is a lot of code available in VIPM. There's also a ton on these (and NI's) fora that can easily be labeled as having a block diagram. Labeling a diagram as an "experimental setup" is more difficult; labeling it with the specific type of setup will be even harder and judging the quality of the programming (so you can "recommend how the investigation can continue from the similar setups") will be (nearly) impossible.
Good luck, though.